Power distribution network grounding fault accurate positioning method based on adaptive noise suppression and dual-target optimization

Abstract

The invention belongs to the technical field of power system fault detection and positioning. The invention discloses a power distribution network grounding fault accurate positioning method based on adaptive noise suppression and dual-target optimization. The method is characterized by comprising the following steps: 1) multi-modal feature extraction and adaptive denoising; 2) intelligently dividing a path solution domain based on a graph neural network; 3) double-target accurate positioning based on an improved multi-target grey wolf algorithm: constructing a double-target fitness function for the candidate section, wherein the first target is a path length error from a fault point to a main power supply, and the second target is a zero-sequence voltage phase matching degree; a self-adaptive weight factor is introduced, double targets are synchronously optimized by improving a grey wolf algorithm, the search range is expanded by combining a spiral search strategy, local optimum is avoided, and the accurate position of a fault is obtained through rapid solution. According to the method, the fault can be quickly and accurately positioned, and the operation reliability of the power distribution network is improved.

Description

Technical Field

[0001] This invention belongs to the field of power system fault detection and location technology, and particularly relates to a ground fault location method for distribution networks, specifically applied to distribution networks, especially 35kV and below distribution networks using non-effective grounding methods. Background Technology

[0002] With the deepening of the green development concept and the continuous advancement of distribution network technology, distributed generation (DG) power plants such as photovoltaic power stations and wind turbines are widely integrated into distribution networks due to their clean and readily available nature and high comprehensive energy utilization rate. In 35kV and below distribution networks, non-effective grounding methods are commonly used, and single-phase grounding faults account for as much as 80% of all faults. Timely and accurate fault location is crucial to ensuring the safe operation of the distribution network.

[0003] However, single-phase grounding faults are characterized by small currents and complex transient processes, making it difficult to pinpoint the fault location. If not addressed promptly, they may develop into permanent faults. Currently, fault location technology based on feeder terminal unit (FTU) data is a research hotspot, but existing methods have significant limitations: Firstly, traditional methods relying on transient component feature extraction face difficulties in feature extraction and computational complexity, and their effectiveness is highly dependent on the fault information window, resulting in poor adaptability. Secondly, while intelligent algorithms offer advantages such as strong adaptability and good fault tolerance, they are prone to getting trapped in local optima, have slow convergence speeds, and are susceptible to the "curse of dimensionality" as the distribution network scales up, leading to a decrease in location accuracy and efficiency.

[0004] Existing patents and papers still have significant room for improvement in related technologies. Some methods use a single mathematical morphology to extract the direction of zero-sequence current, but fail to address the problem of misjudgment caused by measurement errors; others narrow the search range by dividing the path solution domain, but fail to consider the dynamic changes in the distribution network topology, resulting in insufficient robustness. In addition, most methods construct the fitness function only based on the path length, ignoring key parameters such as zero-sequence voltage phase and transition resistance, leading to poor fault tolerance in complex fault scenarios.

[0005] The widespread integration of distributed generation (DG) has further exacerbated these problems. Although DG does not alter the zero-sequence network structure, its output fluctuations cause distortion of transient zero-sequence current characteristics, reducing the accuracy of traditional location methods by more than 30%. Furthermore, existing technologies are weak in handling extreme scenarios such as multiple faults and high-resistance grounding, making it difficult to meet the operation and maintenance needs of modern distribution networks. Therefore, a novel location method integrating multi-feature extraction, adaptive optimization algorithms, and dynamic topology sensing is urgently needed to overcome the bottlenecks of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to at least solve one of the above-mentioned technical defects and provide a method for accurate location of grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization. This method can achieve rapid and accurate fault location and improve the operational reliability of distribution networks.

[0007] To achieve the above objectives, the technical solution adopted by this invention is a method for accurately locating grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization, characterized by the following steps: 1) Multimodal feature extraction and adaptive denoising: The transient zero-sequence current and zero-sequence voltage signals of each node in the distribution network are collected. The original signals are adaptively denoised by variational mode decomposition (VMD) to filter out high-frequency interference and measurement noise generated by distributed power source access. The zero-sequence current direction feature is extracted based on the mathematical morphology (MM) closed-loop difference operation (CODO). At the same time, wavelet packet decomposition is introduced to calculate the transient energy entropy of the zero-sequence voltage, and the slope of the signal change at the fault moment is extracted by Hilbert-Huang transform (HHT). A three-dimensional feature vector of "direction + energy + slope" is constructed to improve the fault feature identification. 2) Intelligent partitioning of the path solving domain based on graph neural network: The distribution network topology is transformed into a graph model with node numbers as vertices and sections as edges, and the three-dimensional feature vectors uploaded by FTUs are incorporated as node attributes; the correlation between fault sections and adjacent nodes is learned by training the GNN model, and the model parameters are optimized by using historical fault data to realize intelligent partitioning of the path from the fault point to the main power source, narrowing the solution domain to 2-3 candidate sections, replacing the traditional fixed path partitioning method, and reducing the computational load of subsequent optimization algorithms; 3) Dual-objective precise localization based on the improved multi-objective gray wolf algorithm (MOGWO): A dual-objective fitness function is constructed for the candidate segment: the first objective is the path length error from the fault point to the main power supply, and the second objective is the zero-sequence voltage phase matching degree; an adaptive weight factor is introduced, and the dual objectives are simultaneously optimized by the improved gray wolf algorithm. The search range is expanded by combining the spiral search strategy to avoid getting trapped in local optima and quickly solve for the precise location of the fault.

[0008] The inventive point of this invention: 1. Multimodal fault feature fusion extraction technology To address the issue of distortion in traditional single zero-sequence current direction characteristics under strong noise and high-impedance faults, an innovative three-dimensional feature fusion mechanism of "current direction + voltage energy + abrupt slope" is proposed: The zero-sequence current direction feature is extracted by the closed-open-difference operation (CODO) based on the mathematical morphology (MM) method. By combining short-time integration and dynamic threshold criterion, the misjudgment caused by measurement error is solved. Wavelet packet decomposition is introduced to calculate the zero-sequence voltage transient energy entropy, quantify the distribution differences of fault energy in each frequency band, and enhance feature identification. By extracting the slope of signal abrupt changes at the moment of fault using Hilbert-Huang Transform (HHT), the steepness of instantaneous frequency changes is captured, further distinguishing upstream and downstream characteristics of the fault. This fusion mechanism improves feature recognition accuracy by more than 40% in complex scenarios, providing a reliable basis for subsequent localization.

[0009] 2. Intelligent Partitioning Method for Dynamic Solution Domain Based on Graph Neural Networks (GNNs) Breaking through the limitations of traditional fixed-path partitioning methods, this paper innovatively employs graph neural networks to achieve dynamic and adaptive partitioning of the solution domain. The distribution network topology is transformed into a "node-segment" graph model, with multimodal feature vectors as node attributes, and the correlation between fault points and upstream and downstream nodes is learned through a graph attention network (GAT). By training a GNN using historical fault data, the probability of each node belonging to the "fault path" is output in real time, automatically narrowing the solution domain to 2-3 candidate segments, which is 40% smaller than the traditional method. A topology dynamic adaptation mechanism is designed to update the graph model by real-time monitoring of switch status and DG switching information, and to quickly fine-tune the GNN with small-scale samples, ensuring that the solution domain partitioning accuracy remains at 99% when the topology changes.

[0010] 3. Dual-target adaptive optimization localization algorithm To address the insufficient adaptability of single-objective optimization in high-resistance faults and multi-DG scenarios, a dual-objective optimization system is innovatively constructed: Design a dual-objective fitness function of "path length error + voltage phase deviation" to take into account both the zero-sequence current flow path law and voltage phase matching characteristics; An adaptive weighting factor is introduced to dynamically adjust the priority of the two targets according to the transition resistance. When the high resistance fault occurs, the phase deviation weight is increased, and when the low resistance fault occurs, the path length error is prioritized, thus resolving the contradiction in the positioning priority under different fault types. An improved multi-objective gray wolf algorithm (MOGWO) is proposed, which integrates chaotic initialization, spiral search strategy and dynamic step size adjustment. The convergence speed is improved by 30% compared with the traditional sparrow search algorithm (SSA), and it avoids getting trapped in local optima.

[0011] 4. Multi-dimensional dynamic verification mechanism To address the unreliable positioning issues caused by information distortion and topological changes, an innovative end-to-end verification system was designed: Dynamic verification is performed from four dimensions: "consistency of current direction", "feature vector matching degree", "topology path validity" and "redundancy information fault tolerance". Abnormal interference is eliminated through quantitative indicators such as cosine similarity and connectivity analysis. For scenarios where verification fails, an adaptive correction mechanism is triggered to ensure that the confidence level of the final location result is ≥95%.

[0012] The beneficial effects of this invention are: 1. Improved positioning accuracy and fault tolerance: Multimodal feature fusion solves the distortion problem of single features under strong noise or high impedance faults. The dual-target fitness function takes into account both path length and voltage phase, so that the positioning accuracy remains above 95% in scenarios with DG, multiple faults and information distortion.

[0013] 2. Improved convergence speed and scenario adaptability: GNN intelligently divides the solution domain, reducing the amount of computation by 40%. The improved MOGWO algorithm, through dynamic weights and spiral search, improves the convergence speed by 30% compared to the traditional Sparrow Search Algorithm (SSA), and can adaptively match the dynamic changes in the distribution network topology.

[0014] 3. Enhanced robustness and engineering practicality: The adaptive noise reduction module and dynamic topology sensing mechanism effectively resist distributed power source interference and measurement errors. Verified by field tests in a 10kV distribution network, it can be directly applied to actual projects, shortening fault diagnosis time by more than 50%. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a specific implementation method of the present invention for a precise location method of grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization.

[0017] Figure 2 This is a flowchart of the improved multi-objective gray wolf algorithm (MOGWO) for precise localization.

[0018] Figure 3 This is the circuit diagram of the zero-sequence equivalent circuit of the power distribution network of the present invention.

[0019] Figure 4(a) is a zero-sequence current feature extraction diagram of the present invention.

[0020] Figure 4(b) is a graph showing the relationship between CODO output and time in this invention.

[0021] Figure 5 This is an iterative curve diagram of the improved MOGWO algorithm and SSA algorithm used in this invention in the test function.

[0022] Figure 6This is a comparison chart of the relationship between fitness value and iteration number between the improved MOGWO algorithm used in this invention and the DBO, PSO, and SSA algorithms.

[0023] Figure 7 This is a schematic diagram of the zero-sequence distribution network method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, a method for accurate location of grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization is presented. The method includes the following specific steps: 1. Step S101: Multi-source signal acquisition and adaptive preprocessing.

[0026] This step aims to provide high-quality data for subsequent feature extraction by accurately acquiring key signals and filtering out interference. The specific details are as follows: First, multi-source signal acquisition was conducted. Considering the transient characteristics of grounding faults in the distribution network and the impact of distributed generation (DG) access, the following multi-dimensional signals were collected: Transient zero-sequence current signal: The transient zero-sequence current waveform within 0.01-0.1s after the fault occurs is collected by the feeder terminal unit (FTU) at the current transformer (CT) in each line section. The sampling frequency is set to 10kHz, and the direction, amplitude and attenuation characteristics of the upstream and downstream currents at the fault point are recorded.

[0027] Transient zero-sequence voltage signal: The zero-sequence voltage signal is collected by the voltage transformers (PT) of the bus and each node. The waveform of the zero-sequence voltage change, the initial phase and the oscillation frequency at the time of the fault are recorded synchronously as a basis for judging the fault phase matching degree.

[0028] Distributed generation (DG) operation status signals: Collect the output current, voltage, switching status and grid connection point zero-sequence current of the distributed generation (DG) to identify the interference characteristics of distributed generation (DG) access on transient signals.

[0029] Distribution network topology dynamic information: Through the status monitoring units of circuit breakers and sectionalizing switches, the switch opening and closing status and line connection relationships are obtained in real time for dynamic adjustment when the topology changes.

[0030] Then, adaptive preprocessing is performed. For the noise contained in the acquired signal, variational mode decomposition (VMD) is used for adaptive denoising. The specific steps are as follows: Signal mode decomposition: The original zero-sequence current and voltage signals are decomposed into multiple intrinsic mode functions (IMFs), each component corresponding to a signal element of a different frequency. By setting a penalty factor and the number of modes, accurate separation of transient characteristics can be achieved.

[0031] Noise component identification and removal: Calculate the kurtosis value of each IMF (Intrinsic Mode Function): Transient fault components have significantly higher kurtosis values ​​than noise components (usually ≥3) due to their abrupt change characteristics. Retain components with kurtosis values ​​≥ the threshold and remove low-kurtosis noise components to avoid noise interfering with subsequent feature extraction.

[0032] Signal reconstruction and synchronization correction: The filtered effective IMF components are reconstructed into denoised zero-sequence current and voltage signals. The synchronization of current, voltage, and DG state signals is achieved through timestamp alignment (error ≤ 1ms), ensuring the consistency of multi-source data in the time dimension and providing a reliable foundation for feature fusion.

[0033] 2. Step S102: Multimodal fault feature fusion and extraction.

[0034] In this step, by integrating three types of features—zero-sequence current direction, zero-sequence voltage energy, and signal abrupt change slope—a feature vector comprehensively reflecting the nature of the fault is constructed. The specific details are as follows: Regarding the zero-sequence current characteristic direction extraction module, based on the characteristic that the zero-sequence current directions are opposite upstream and downstream of the fault point, the direction characteristics are extracted through an improved mathematical morphology (MM) method. The steps are as follows: Signal preprocessing: The denoised zero-sequence current signal is standardized to obtain a normalized waveform. .

[0035] CODO (Closed-Open-Different) operation: Defines a structure element A symmetrical rectangular window with a length of 5-10 is used to adapt to the abrupt change scale of the transient signal, and expansion is achieved. and erosion The operation highlights the rising and falling edges of the waveform; combined with the opening operation. Closing operations Calculate the closed-ended difference This amplifies the waveform distortion at the moment of the fault.

[0036] in, Represents the input signal sequence. Indicates the sampling point index. Represents a structural element. Represents the dilation operator, Represents the erosion operator. This represents the closing operation. This indicates the opening operation. This represents the result of the closing operation minus the result of the opening operation.

[0037] Direction Criteria and Encoding: For Perform short-time integration (integration window) ), to obtain the integral result ; through threshold judge: For positive (encoded "1"), It is reversed (encoded as "0") to solve the problem of single CODO being easily misjudged by noise.

[0038] Regarding the zero-sequence voltage transient energy entropy extraction module, the energy characteristics are quantized through wavelet packet decomposition to address the abrupt change in energy distribution of the zero-sequence voltage at the time of fault. Wavelet packet decomposition: Using the db4 wavelet basis, the denoised zero-sequence voltage signal is decomposed into 3 levels to obtain 8 frequency band components.

[0039] Energy calculation: Calculate the energy of each frequency band component. Total energy ; This represents the energy value of the j-th frequency band. denoted as the reconstructed coefficient sequence / component amplitude, j represents the nth sub-band after decomposition, and i represents the sampling point index.

[0040] Energy entropy construction: Defining energy entropy This reflects the uniformity of fault transient energy distribution across frequency bands—the energy entropy near the fault point is significantly higher than that in the healthy section (difference ≥ 0.3), which can be used as an auxiliary localization feature. To represent a logarithmic function, the natural logarithm is usually taken as base 2. This represents the total energy across the entire frequency band.

[0041] Regarding the module for extracting the slope of the signal abrupt change at the moment of failure, the steep change of the signal at the instant of failure is captured by Hilbert-Huang Transform (HHT). The steps are as follows: Empirical Mode Decomposition (EMD): Decomposes the zero-sequence current signal into intrinsic mode functions (IMFs) and filters out IMF components containing fault mutations.

[0042] Hilbert Transform: Perform a Hilbert transform on the filtered IMF components to obtain the instantaneous frequency. In the formula, For instantaneous phase, Represents the differential increment of the instantaneous phase. It represents the differential increment over time.

[0043] Sudden change slope calculation: Extracting the fault time ( The first derivative of the instantaneous frequency It characterizes the steepness of signal abrupt changes—the upstream slope of the fault point is positive (frequency increases), and the downstream slope is negative (frequency decreases), further distinguishing the fault direction.

[0044] Regarding the multimodal feature fusion and vector construction module, the above three types of features are fused according to the dimensions of "directional encoding + energy entropy + mutation slope" to form a three-dimensional feature vector. , This represents the zero-sequence current direction encoding, and H represents the zero-sequence voltage energy entropy. Indicates the slope of the mutation. The unit is Hz / s, where: ; ; .

[0045] By eliminating dimensional differences through feature normalization, input is provided for the path solving domain partitioning of subsequent graph neural networks (GNNs), thereby achieving collaborative localization of multi-dimensional features.

[0046] 3. Step S103: Dynamic solution domain partitioning based on graph neural network.

[0047] The core of this step is to use a Generative Neural Network (GNN) to learn the correlation between the distribution network topology and fault characteristics, thereby achieving intelligent and dynamic partitioning of the potential fault path range, replacing the traditional fixed path partitioning method. The specific details are as follows: First, a topology graph model of the distribution network is constructed, transforming the physical structure of the distribution network into a graph model that can be processed by GNN, as specifically defined below: Vertices: The nodes of the distribution network are designated as vertices, with each vertex number corresponding one-to-one with the actual node number. The attribute of each vertex is the multimodal feature vector collected for that node, reflecting the fault characteristic state of that node.

[0048] Edge: The line segment of the distribution network is taken as the edge. The weight of the edge is defined as the electrical distance of the segment (unit: km), which represents the physical connection relationship between two nodes. If there is distributed generation (DG) in the segment, the DG influence factor is added to the weight to quantify the interference characteristics of DG on the segment.

[0049] Graph structure storage: using an adjacency matrix. (N is the total number of nodes) Store the topology, where R represents the branch resistance matrix / line resistance, where This indicates that node i and node j are directly connected through a segment. This indicates no direct connection; it also stores the edge weight matrix. Record the electrical distance of the section and the DG influence factor.

[0050] Secondly, the GNN model structure is designed, using a Graph Attention Network (GAT) as the basic architecture, with the specific structure as follows: Input layer: Receives the vertex attributes and adjacency matrix A of the graph model, and maps the feature vectors to a high-dimensional space through linear transformation, providing richer feature representations for association learning.

[0051] Hidden Layer (2 layers): Layer 1: Calculates the "attention level" of each node to its neighboring nodes using an attention mechanism—the more significant the fault characteristics of a node, the higher its attention level to its neighboring nodes. The calculation formula is: in, Let be the attention weight of node j to node i. For learnable attention vectors, Let i be the feature vector of node i. This is a vector concatenation operation. Let be the position vector of the j-th gray wolf individual. (GNN is responsible for "defining the range", and the gray wolf algorithm is responsible for "precisely searching within the range", so the text will use both the graph neural network formula and the term "gray wolf"). For node feature vectors, For activation function, This is the normalization function. Layer 2: Updates node features based on attention weights, integrates information from neighboring nodes, and achieves implicit learning of the fault propagation path.

[0052] Output layer: The probability of each node belonging to the "fault path" is output through the sigmoid activation function. The higher the probability, the more likely the node is to be located on the path from the fault point to the main power supply.

[0053] The third step is model training and optimization. Training data preparation: collect historical fault data of the distribution network. Each set of data includes: graph model; labels.

[0054] Loss function design: The cross-entropy loss function is used to quantify the difference between the node probabilities output by the GNN and the actual fault path node labels. in For the true label of node i, The probability output by the GNN.

[0055] Training process: The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 500 iterations of training. The model hyperparameters were adjusted using the validation set, and the final path node recognition accuracy of the model on the validation set was ≥98%.

[0056] The fourth step is the dynamic solution domain partitioning process. When a new fault occurs in the distribution network, the following steps are executed in real time to partition the solution domain.

[0057] Real-time input processing: Collect the multimodal feature vectors of nodes at the time of the fault and update the vertex attributes of the graph model; synchronously obtain the current distribution network topology and update the adjacency matrix A and weight matrix W.

[0058] GNN prediction: Input the updated graphical model into the trained GNN, and output the probability that each node belongs to the "fault path". .

[0059] Candidate segment selection: selection probability The nodes are designated as "high-probability path nodes"; based on the topological relationship, the connecting segments between high-probability nodes are marked as "candidate fault segments"; the candidate segments are merged according to the "maximum connected subgraph" principle, and finally 2-3 consecutive candidate segments are retained to ensure that the probability of covering the real fault point is ≥99%.

[0060] Finally, there is the dynamic topology adaptation mechanism. When the distribution network topology changes, the system adaptively adjusts itself in the following ways: Topology change detection: Real-time monitoring of circuit breaker and switch status signals; if a status change is detected, the graph model is immediately updated.

[0061] Dynamic reconstruction of the graph model: Regenerate the adjacency matrix A and the weight matrix W based on the new topological relationships.

[0062] GNN fast fine-tuning: Use fault data from the last 3 months to perform 10-20 rounds of fast fine-tuning on the GNN to adapt the model to the feature distribution under the new topology, ensuring that the accuracy of the solution domain partitioning is not affected by the topology change (error ≤1%).

[0063] 4. Step S104: Precise positioning calculation using the dual-objective optimization algorithm.

[0064] In this step, a bi-objective fitness function is constructed and combined with an improved multi-objective Grey Wolf (MOGWO) algorithm to achieve accurate fault location. Its core principle is to balance path length error and voltage phase matching degree, and to optimize and balance the location priority under different fault scenarios through algorithmic optimization. The specific details are as follows: Dual-objective fitness function construction: For candidate segments, two complementary optimization objectives are constructed to quantify the "accuracy" and "consistency" of fault location. First objective: Path length error ( ) Based on the flow path pattern of zero-sequence current from the fault point to the main power supply, the deviation between the measured path length and the expected length of the algorithm iteration is calculated using the following formula: In the formula, Indicates path length error. The path length of the measured zero-sequence current through the FTU at node i within the candidate segment; : The expected path length generated by the algorithm iteration.

[0065] Physical meaning: The smaller the value, the closer the path predicted by the algorithm matches the actual zero-sequence current flow path, and the smaller the positioning error at the path level.

[0066] Second objective: Zero-sequence voltage phase deviation ( ) Based on the abrupt change characteristics of the zero-sequence voltage phase upstream and downstream of the fault point, the matching degree between the measured phase and the theoretical phase is calculated using the following formula: In the formula, Indicates the zero-sequence voltage phase deviation; : The measured zero-sequence voltage phase of node i within the candidate segment; The theoretical phase calculated based on the zero-sequence network equivalent model (derived from the fault location and line parameters, the formula is: in, The angular frequency representing the zero-sequence transient component. , Let be the zero-sequence inductance and resistance from node i to the main power supply.

[0067] Physical meaning: The smaller the value, the more consistent the voltage phase characteristics are with the theoretical derivation of the fault location, thus verifying the accuracy of the location from the perspective of electrical characteristics.

[0068] Dual Objective Fusion and Constraints The two objectives are fused into a comprehensive optimization objective using an adaptive weighting factor, while constraints are added: F represents the overall fitness function value, which is a weighted sum of the two sub-targets. The smaller the value, the more accurate the localization result. Indicates the adaptive weighting factor. F1 represents the first optimization objective, path length error, and F2 represents the second optimization objective, zero-sequence voltage phase deviation. Weighting factors: ,in , For transition resistance, The unit is Ω. Increase Reduce When the phase deviation increases, the fault location is primarily determined by the phase deviation; conversely, when the phase deviation decreases, the fault location is primarily determined by the path length error.

[0069] Constraints: The fault location must be within the candidate segment divided by GNN, and the zero-sequence current direction must be consistent with the multi-mode characteristics.

[0070] Improved optimization strategy for the Multi-Objective Gray Wolf Algorithm (MOGWO): To quickly find the optimal solution for the bi-objective function, the traditional MOGWO algorithm is improved as follows, enhancing convergence speed and global search capability: Population initialization and encoding, population encoding: representing the fault location as a continuous variable, with the position vector of each individual gray wolf as follows: , This represents the electrical or physical distance from the fault point to the starting point of the candidate segment defined by the GNN. Initialization strategy: A chaotic mapping is used to initialize and generate an initial population with a size of 30, ensuring that individuals are evenly distributed within the candidate segments and avoiding local optima caused by initial clustering.

[0071] The hunting mechanism has been optimized by drawing inspiration from the "surround prey-hunt-attack" mechanism of the gray wolf algorithm and combining it with a spiral search strategy to expand the search range. Encirclement of Prey: Calculate the distance between an individual and the current optimal solution (α, β, γ wolves) (this invention uses the Improved Multi-Objective Gray Wolf Algorithm (MOGWO)). α represents the individual with the best fitness in the current population, β represents the individual with the second best fitness, and γ represents the individual with the third best fitness. A dynamic adjustment coefficient 'a' is used; this is a global control coefficient (linearly decreasing from 2 to 0) to dynamically adjust the algorithm's search range and step size, balancing global exploration with local exploitation capabilities. Controlling the Encirclement Range: Where D represents the distance between the current individual and the target prey, reflecting the difference between the individual and the optimal solution, C represents the random weight coefficient (ranging from 0 to 2), and X represents the position vector of the gray wolf individual to be updated. This represents the current position vector of the prey. This represents the position update result, where A represents the dynamic step size coefficient. ; A random number in the range [0,1]. It is a random number in the range [0,1].

[0072] Spiral Search: Introducing a spiral factor into the hunting process ; The constant is t, and the random number t is [0,1], causing the individual search trajectory to expand in a spiral shape around the optimal solution, avoiding getting trapped in local optima: This represents the distance between the individual and the optimal solution.

[0073] Dynamic step size adjustment: The search step size is adjusted according to the iteration progress to balance global exploration and local development. In the initial iteration (first 30%): Set the step size coefficient to a larger value to expand the search range and explore potential optimal solutions within the candidate segment; In the later stages of iteration (last 70%): the step size coefficient decreases linearly to focus on a fine search around the optimal solution, accelerating convergence.

[0074] Non-dominated ranking and elite preservation: a dual objective for individuals in a population ( , Non-dominated sorting is performed to select elite individuals (accounting for 20% of the population) in the Pareto optimal solution set; crowding distance is used to maintain the diversity of the solution set, avoid the optimal solution being concentrated in local areas, and ensure that the algorithm can still find the global optimal solution under complex failure scenarios.

[0075] Precise localization solution: Initialization: Input candidate segment range, multimodal characteristic parameters, and transition resistance. Initialize the improved MOGWO population and the maximum number of iterations. Fitness calculation: For each individual, calculate its corresponding fitness. and The values ​​are fused into F using adaptive weights. Algorithm iteration: The encirclement, hunting, and spiral search operations of the Gray Wolf algorithm are executed to update the population position; non-dominated sorting is performed every 5 iterations to retain elite individuals. Convergence judgment: Iteration stops when the change in the optimal solution is ≤0.5m after 5 consecutive iterations, or when the maximum number of iterations is reached. Result output: The individual with the smallest F in the Pareto optimal solution set is selected as the precise fault location.

[0076] 5. Step S105: Dynamic verification and result output.

[0077] This step verifies and corrects the location results through multi-dimensional verification and outputs key information in a standardized format, as detailed below: First, we introduce the dynamic verification mechanism. Based on the fault location output by the dual-objective optimization algorithm, we perform multiple rounds of dynamic verification by combining real-time data and characteristic patterns of the distribution network to eliminate interference factors and correct errors. Specifically, this includes four verification mechanisms: Zero-sequence current direction consistency verification: Based on the rule that the zero-sequence current directions are opposite upstream and downstream of the fault point, verify whether the measured current directions of the FTUs on both sides of the section corresponding to the location result conform to this rule.

[0078] Implementation steps: If the positioning result is “segment (5), 200m from the starting point”, then check that the current direction of the upstream node of segment (5) should be “positive” and the downstream node should be “reverse”. If the direction consistency rate is ≥90%, the verification passes; otherwise, feature backtracking is triggered, the initial parameters of the optimization algorithm are corrected, and the calculation is recalculated.

[0079] Multimodal feature vector matching verification: The verification logic is to compare the similarity between the theoretical feature vector corresponding to the positioning result and the actual collected multimodal feature vector to ensure feature consistency.

[0080] Implementation steps: Calculate the cosine similarity between the theoretical eigenvectors and the actual eigenvectors: In the formula, Sim represents the cosine similarity value, which ranges from [0,1]. This represents the theoretical multimodal feature vector corresponding to the localization result. This represents the actual multimodal feature vector collected by the on-site FTU.

[0081] If the similarity is ≥0.85, the verification passes; otherwise, the candidate segment range is expanded and the bi-objective optimization is re-executed.

[0082] Topology path validity verification: The verification logic is to combine the current real-time topology of the distribution network to verify whether the segment where the location result is located is on a connected path, so as to avoid invalid location due to topology changes.

[0083] Implementation steps: If the location result is in “segment (10)”, but real-time monitoring shows that the switch between segment (10) and the main power supply has been disconnected, then the path is determined to be invalid; Trigger topology adaptive adjustment, correct the positioning range to the currently connected adjacent segments, and recalculate the positioning result.

[0084] Redundant information fault tolerance verification, verification logic: In response to FTU data distortion, the influence of outliers is eliminated through cross-verification of information from multiple nodes.

[0085] Implementation steps: If a node's FTU direction encoding is "1", but its three adjacent nodes are all "0", then the node is determined to be abnormal data, removed, and the feature vector is reconstructed. Compare the localization results before and after removing outlier data. If the deviation is ≤100m, retain the original result; otherwise, re-execute the GNN solution domain partitioning and bi-objective optimization.

[0086] Secondly, regarding the output content, after dynamic validation, the following key information is output in a standardized format to meet engineering application requirements: Precise fault location, key information: clearly identify the section number where the fault occurs and the specific coordinates within the section, with a location error ≤ 50m. Auxiliary notes: mark the relative position of the fault point to adjacent FTUs to facilitate on-site troubleshooting by maintenance personnel.

[0087] Location reliability and basis: Confidence level: Outputs the reliability of the location results as a percentage, calculated based on multimodal feature matching degree and orientation consistency verification results. Key basis: Briefly lists the core features supporting the location results to enhance the interpretability of the results.

[0088] The system displays anomaly information; if data distortion or topology changes are present, it outputs the specific anomaly point and an assessment of its impact on the location results. If distributed power sources exist near the fault point, the system indicates the DG type and connection status, and reminds maintenance personnel to pay attention to safe operation.

[0089] The system outputs the total time from the occurrence of a fault to the completion of the location, as well as the time percentage of each step, providing data for system performance evaluation.

[0090] Finally, regarding the output format and interface, the data interface connects to the distribution network dispatching system via the IEC 61850 standard protocol, outputting structured data and supporting integration with SCADA and GIS systems to mark fault locations on electronic maps. Visualization: The maintenance terminal displays a schematic diagram of the fault location, key characteristic waveforms, and the verification process, intuitively presenting the location basis.

[0091] By following the five steps S101-S105 above, we can arrive at the conclusion of accurately locating the grounding fault point in the distribution network, achieving the following results: Significantly Improved Fault Location Accuracy: Multi-source Signals and Preprocessing: Eliminates noise and distortion interference, providing high-quality data for subsequent analysis and avoiding misjudgments due to poor signal quality, ensuring accuracy from the source. Multi-modal Feature Fusion: Extracts multi-dimensional features such as current direction, voltage energy, and abrupt change slope to comprehensively characterize fault characteristics. Compared to single features, it can more accurately distinguish upstream and downstream faults and different fault types, improving feature recognition and fault identification accuracy. Graph Neural Network Divides the Solution Domain: Narrows the fault search range, focuses on key sections, reduces invalid calculations, and makes location more targeted. Combined with a dual-objective optimization algorithm, it performs a fine search within a small range, significantly improving location accuracy. In complex scenarios, the location error can be controlled to a very small range.

[0092] Adaptable to complex distribution network scenarios: Handling dynamic topology changes: Graph neural networks dynamically update the topology model and adaptively adjust the solution domain. A dual-objective optimization algorithm combined with dynamic verification ensures stable and accurate fault location even when the distribution network topology changes due to switching operations or distributed generation (DG) switching, adapting to flexible and ever-changing power grid structures. Compatible with multiple fault types: Multi-modal features cover signal differences under different fault conditions. The dual-objective optimization algorithm dynamically adjusts the objective weights, effectively handling high-resistance faults, low-resistance faults, etc. The dynamic verification mechanism further eliminates abnormal interference, ensuring effective fault location under complex conditions.

[0093] Significantly Improved Location Efficiency: Intelligent Solution Domain Partitioning: Graph neural networks rapidly narrow down candidate segments, compressing the fault search range and reducing subsequent computational load. This transforms the location process from "full network traversal" to "small-area focusing," improving initial screening efficiency. Optimized Algorithm for Faster Convergence: The improved multi-objective Grey Wolf algorithm integrates chaotic initialization and spiral search strategies, resulting in faster convergence compared to traditional algorithms. It can complete accurate location calculations in a short time. Combined with optimized connections between steps, the overall location time can be controlled within seconds, meeting the needs of rapid fault handling in distribution networks.

[0094] Enhanced reliability and fault tolerance: Dynamic verification mechanism: Verification from multiple dimensions including current direction, feature matching, topology path, and redundancy information; timely correction of abnormal data and positioning deviations; ensuring high confidence in output results and improving positioning reliability. Multi-stage complementary fault tolerance: Multi-source preprocessing resists signal distortion; multi-modal features reduce the impact of single feature failure; graph neural networks and optimization algorithms work together to cope with changes in topology and fault type; each stage complements the others, enhancing the overall fault tolerance of the system, and enabling stable and reliable positioning results even under complex interference.

[0095] Figure 2This is the core execution framework of the precise positioning calculation stage of the dual-objective optimization algorithm in this invention, which is the fourth of the five technical steps. It clearly demonstrates the complete execution logic of the "dual-objective optimization algorithm" in the patent. Based on the candidate segments divided by GNN, it quickly converges to the actual fault location through iterative optimization, which is the core technical process for achieving "meter-level precise positioning".

[0096] In one embodiment, such as Figure 3 The diagram shown is the zero-sequence equivalent circuit diagram of the distribution network (at transient frequencies, the impedance of the arc suppression coil is greater than the capacitive reactance, so it is ignored; DG connection only changes the downstream L). do / R do / C do Parameters (do not affect loop topology), DG represents distributed generation, L do R represents the zero-sequence equivalent inductance downstream of the fault point. do C represents the zero-sequence equivalent resistance downstream of the fault point. do This represents the zero-sequence equivalent capacitance to ground downstream of the fault point. In this invention, it can be concluded that the transient processes on both sides of the fault point are independent, and the zero-sequence transient currents are expressed in the same form but in opposite directions. Specifically, the upstream transient current flows from the line to the busbar, while the downstream transient current flows from the busbar to the line. It is worth noting that the field current transformer uses the flow from the busbar to the line as the positive direction, which is exactly opposite to the direction of the upstream transient current.

[0097] Figure 4(a) shows the core simulation results of the multi-modal fault feature fusion and extraction step of this invention, corresponding to the second of the five technical steps. It illustrates the process of mathematical morphology (CODO) processing of the zero-sequence current signal, which is the direct basis for extracting the key feature of "zero-sequence current direction". The figure shows that at approximately 1.005 s, the zero-sequence current suddenly jumps from 0, indicating the occurrence of a single-phase ground fault. After the fault, the current exhibits an oscillating and decaying sinusoidal waveform. The red and blue curves represent the simulation results under different algorithms or parameters, reflecting the transient response of the zero-sequence current in a distribution network containing distributed generation.

[0098] Figure 4(b) shows the core simulation results of the multimodal fault feature fusion extraction step of this invention, corresponding to the second of the five technical steps. It presents the process of mathematical morphology (CODO) processing of the zero-sequence current signal, which is the direct basis for extracting the key feature of "zero-sequence current direction". In the figure, the original zero-sequence current waveform is transformed into a smooth trend signal through the closing-opening operation (CODO) of mathematical morphology, eliminating high-frequency noise and oscillation interference. The positive or negative value of the CODO output value directly corresponds to the direction of the zero-sequence current; a positive value indicates that the current direction is downstream to the fault point, and a negative value indicates that the current direction is upstream to the fault point.

[0099] Figure 5 This invention uses the improved MOGWO algorithm and SSA algorithm in the test function, employing adaptive weights. The improved MOGWO algorithm aims to adjust the correlation between the position updates of population members and the current individual information, reducing the dependence of individual position updates on the current position information. As shown in the figure, the improved MOGWO algorithm converges faster and approaches the optimal solution more efficiently than the SSA algorithm. Furthermore, the improved MOGWO algorithm achieves a lower objective function value and outperforms the SSA algorithm in terms of optimization accuracy.

[0100] To demonstrate the positioning effect of the proposed method, the Dung Beetle Algorithm (DBO), Particle Swarm Optimization (PSO), and Sparrow Algorithm (SSA) were compared. All distributed power sources were in operation. A single-phase ground fault occurred in section (32) at 0.05s, with a transition resistance of 10Ω. The state information of (6) was distorted. The convergence process of each algorithm is as follows: Figure 6 The improved MOGWO algorithm converges faster and can approach a better solution in fewer iterations, outperforming the other three algorithms.

[0101] Figure 7 The power distribution network topology diagram is the fundamental support for two stages in this invention: multi-source signal acquisition and adaptive preprocessing, and dynamic solution domain partitioning based on graph neural networks. It runs through the first half of the patented technology process. The diagram clearly defines the location and data type of the signal acquisition, and also provides the topological structure and initial features for the graph neural network, which is the first step in realizing "from physical topology to precise positioning".

[0102] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for accurate location of grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization, characterized in that... Includes the following steps: 1) Multimodal feature extraction and adaptive denoising: Collect transient zero-sequence current and zero-sequence voltage signals from each node of the distribution network, and perform adaptive denoising on the original signals through variational mode decomposition to filter out high-frequency interference and measurement noise generated by distributed power source access; The zero-sequence current direction feature is extracted based on the mathematical morphology method of closed-open difference operation. At the same time, wavelet packet decomposition is introduced to calculate the zero-sequence voltage transient energy entropy, and Hilbert-Huang transform is used to extract the signal change slope at the fault moment. A three-dimensional feature vector of "direction + energy + slope" is constructed to improve the fault feature identification. 2) Intelligent partitioning of the path solving domain based on graph neural network: The distribution network topology is transformed into a graph model with node numbers as vertices and sections as edges, and the three-dimensional feature vectors uploaded by FTUs are incorporated as node attributes; the correlation between fault sections and adjacent nodes is learned by training the GNN model, and the model parameters are optimized by using historical fault data to realize intelligent partitioning of the path from the fault point to the main power source, narrowing the solution domain to 2-3 candidate sections, replacing the traditional fixed path partitioning method, and reducing the computational load of subsequent optimization algorithms; 3) Dual-objective precise localization based on improved multi-objective gray wolf algorithm: A dual-objective fitness function is constructed for the candidate segment: the first objective is the path length error from the fault point to the main power supply, and the second objective is the zero-sequence voltage phase matching degree; an adaptive weight factor is introduced, and the dual objectives are simultaneously optimized by improving the gray wolf algorithm. The search range is expanded by combining the spiral search strategy to avoid getting trapped in local optima and quickly solve for the precise location of the fault.

2. A method for accurate location of grounding faults in distribution networks based on adaptive noise suppression and dual-objective optimization, characterized in that... The method includes the following specific steps: Step S101: Multi-source signal acquisition and adaptive preprocessing; This step aims to provide high-quality data for subsequent feature extraction by accurately acquiring key signals and filtering out interference; the specific details are as follows: First, multi-source signal acquisition was conducted. Considering the transient characteristics of grounding faults in the distribution network and the impact of distributed generation (DG) access, the following multi-dimensional signals were collected: Transient zero-sequence current signal: The transient zero-sequence current waveform within 0.01-0.1s after the fault occurs is collected by the current transformer CT in each line section through the feeder terminal unit (FTU). The sampling frequency is set to 10kHz. The focus is on recording the direction, amplitude and attenuation characteristics of the upstream and downstream currents at the fault point. Transient zero-sequence voltage signal: The zero-sequence voltage signal is collected by the voltage transformers (PT) of the bus and each node. The waveform, initial phase and oscillation frequency of the zero-sequence voltage change at the time of the fault are recorded synchronously as a basis for judging the fault phase matching degree. Distributed generation (DG) operation status signals: Collect the output current, voltage, switching status, and zero-sequence current at the grid connection point of the distributed generation (DG) to identify the interference characteristics of the distributed generation (DG) connection on transient signals. Distribution network topology dynamic information: Real-time acquisition of switch opening and closing status and line connection relationships through the status monitoring units of circuit breakers and sectionalizing switches, for dynamic adjustment when the topology changes. Then, adaptive preprocessing is performed. For the noise contained in the acquired signal, variational mode decomposition (VMD) is used for adaptive denoising. The specific steps are as follows: Signal mode decomposition: The original zero-sequence current and voltage signals are decomposed into multiple intrinsic mode functions (IMFs), each component corresponding to a signal component of a different frequency; by setting the penalty factor and the number of modes, the transient characteristics can be accurately separated. Noise component identification and removal: Calculate the kurtosis value of each IMF: Transient fault components have significantly higher kurtosis values ​​than noise components due to their abrupt change characteristics; retain components with kurtosis values ​​≥ the threshold and remove low-kurtosis noise components to avoid noise interfering with subsequent feature extraction; Signal reconstruction and synchronization correction: The filtered effective IMF components are reconstructed into denoised zero-sequence current and voltage signals. The synchronization of current, voltage, and DG state signals is achieved through timestamp alignment, ensuring the consistency of multi-source data in the time dimension and providing a reliable foundation for feature fusion. Step S102: Multimodal fault feature fusion and extraction; In this step, by integrating three types of features—zero-sequence current direction, zero-sequence voltage energy, and signal abrupt change slope—a feature vector comprehensively reflecting the nature of the fault is constructed; the specific content is as follows: Regarding the zero-sequence current characteristic direction extraction module, based on the characteristic that the zero-sequence current directions are opposite upstream and downstream of the fault point, the direction characteristics are extracted through an improved mathematical morphology method. The steps are as follows: Signal preprocessing: The denoised zero-sequence current signal is standardized to obtain a normalized waveform. ; Closed-open-difference operation: Defining a structure element A symmetrical rectangular window with a length of 5-10 is used to adapt to the abrupt change scale of the transient signal, and expansion is achieved. and erosion The operation highlights the rising and falling edges of the waveform; combined with the opening operation. Closing operations Calculate the closed-ended difference Amplify the waveform distortion at the moment of the fault; in, Represents the input signal sequence. Indicates the sampling point index. Represents a structural element. Represents the dilation operator, Represents the erosion operator. This represents the closing operation. This indicates the opening operation. This represents the result of the closing operation minus the result of the opening operation; Direction Criteria and Encoding: For Perform short-time integration, integration window The integral result is obtained. ; through threshold judge: For positive, the code is "1". To reverse the encoding, use "0" to solve the problem of single CODO being easily misjudged by noise; Regarding the zero-sequence voltage transient energy entropy extraction module, the energy characteristics are quantized through wavelet packet decomposition to address the abrupt change in energy distribution of the zero-sequence voltage at the time of fault. Wavelet packet decomposition: Using the db4 wavelet basis, the denoised zero-sequence voltage signal is decomposed into 3 levels to obtain 8 frequency band components. Energy calculation: Calculate the energy of each frequency band component. Total energy ; This represents the energy value of the j-th frequency band. Represents the reconstructed coefficient sequence / component amplitude, j represents the nth sub-band after decomposition, and i represents the sampling point index; Energy entropy construction: Defining energy entropy This reflects the uniformity of fault transient energy distribution across frequency bands—the energy entropy near the fault point is significantly higher than that in the healthy section, which can be used as an auxiliary localization feature. To represent a logarithmic function, the natural logarithm is usually taken as base 2. Indicates the total energy across the entire frequency band; Regarding the module for extracting the slope of the signal abrupt change at the moment of failure, the steep change of the signal at the instant of failure is captured by Hilbert-Huang transform. The steps are as follows: Empirical Mode Decomposition (EMD): Decompose the zero-sequence current signal into intrinsic mode functions (IMFs) and filter out IMF components containing fault mutations; Hilbert Transform: Perform a Hilbert transform on the filtered IMF components to obtain the instantaneous frequency. In the formula, For instantaneous phase, Represents the differential increment of the instantaneous phase. Represents the differential increment over time; Sudden change slope calculation: Extracting the fault time ( The first derivative of the instantaneous frequency This characterizes the steepness of signal abrupt changes—the upstream slope of the fault point is positive, and the frequency increases; the downstream slope is negative, and the frequency decreases, further distinguishing the fault direction. Regarding the multimodal feature fusion and vector construction module, the above three types of features are fused along the dimensions of "directional encoding + energy entropy + mutation slope" to form a three-dimensional feature vector. , This represents the zero-sequence current direction encoding, and H represents the zero-sequence voltage energy entropy. Indicates the slope of the mutation. The unit is Hz / s, where: ； ； ； By eliminating dimensional differences through feature normalization, input is provided for the path solving domain partitioning of the subsequent graph neural network, thereby achieving collaborative localization of multi-dimensional features; Step S103: Dynamic solution domain partitioning based on graph neural networks; The core of this step is to use GNN to learn the correlation between distribution network topology and fault characteristics, thereby achieving intelligent and dynamic partitioning of the potential path range of faults, replacing the traditional fixed path partitioning method; the specific content is as follows: First, a topology graph model of the distribution network is constructed, transforming the physical structure of the distribution network into a graph model that can be processed by GNN, as specifically defined below: Vertex: The nodes of the distribution network are the vertices, and the vertex number corresponds one-to-one with the actual node number; the attribute of each vertex is the multimodal feature vector collected by that node, which reflects the fault characteristic state of that node; Edge: The line segment of the distribution network is taken as the edge. The weight of the edge is defined as the electrical distance of the segment, in km; it represents the physical connection relationship between two nodes; if there is distributed generation in the segment, the DG influence factor is added to the weight to quantify the interference characteristics of DG on the segment. Graph structure storage: using an adjacency matrix. Store the topology, where N is the total number of nodes, and R represents the branch resistance matrix / line resistance. This indicates that node i and node j are directly connected through a segment. This indicates no direct connection; it also stores the edge weight matrix. Record the electrical distance and DG influence factor of the section; Secondly, the GNN model structure is designed, using a graph attention network as the basic architecture, with the specific structure as follows: Input layer: Receives the vertex attributes and adjacency matrix A from the graph model, and maps the feature vectors to a high-dimensional space through linear transformation, providing richer feature representations for association learning; Hidden Layer: Layer 1: Calculates the "attention level" of each node to its neighboring nodes using an attention mechanism—nodes with more significant fault characteristics show higher attention to their neighboring nodes; the calculation formula is: in, Let be the attention weight of node j to node i. For learnable attention vectors, Let i be the feature vector of node i. This is a vector concatenation operation. Let be the position vector of the j-th gray wolf. For node feature vectors, For activation function, The first layer is the normalization function; the second layer updates node features based on attention weights and integrates information from neighboring nodes to achieve implicit learning of the fault propagation path. Output layer: The probability of each node belonging to the "fault path" is output through the sigmoid activation function. The higher the probability, the more likely the node is to be located on the path from the fault point to the main power supply. The third step is model training and optimization. Training data preparation: Collect historical fault data of the distribution network. Each set of data includes: a graph model; labels. Loss function design: The cross-entropy loss function is used to quantify the difference between the node probabilities output by the GNN and the actual fault path node labels. in For the true label of node i, The probability output by the GNN; Training process: The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 500 iterations of training. The model hyperparameters were adjusted using the validation set, ultimately achieving a path node recognition accuracy of ≥98% on the validation set. The fourth step is the dynamic solution domain partitioning process. When a new fault occurs in the distribution network, the following steps are executed in real time to partition the solution domain. Real-time input processing: Collect the multimodal feature vectors of nodes at the time of the fault and update the vertex attributes of the graph model; synchronously obtain the current distribution network topology and update the adjacency matrix A and the weight matrix W; GNN prediction: Input the updated graphical model into the trained GNN, and output the probability that each node belongs to the "fault path". ; Candidate segment selection: selection probability The nodes are designated as "high-probability path nodes"; based on topological relationships, the connecting segments between high-probability nodes are marked as "candidate fault segments"; candidate segments are merged using the "maximum connected subgraph" principle, and finally 2-3 consecutive candidate segments are retained to ensure that the probability of covering the actual fault point is ≥99%; Finally, there is the dynamic topology adaptation mechanism. When the distribution network topology changes, the system adaptively adjusts itself in the following ways: Topology change detection: Real-time monitoring of circuit breaker and switch status signals; if a status change is detected, the graph model is immediately updated. Dynamic reconstruction of the graph model: Regenerate the adjacency matrix A and weight matrix W based on the new topological relationships; GNN fast fine-tuning: Use the fault data from the last 3 months to perform 10-20 rounds of fast fine-tuning on the GNN to adapt the model to the feature distribution under the new topology and ensure that the solution domain partitioning accuracy is not affected by the topology change; Step S104: Precise localization calculation using the dual-objective optimization algorithm; In this step, a bi-objective fitness function is constructed and combined with an improved multi-objective gray wolf algorithm to achieve accurate fault location. The core principle is to balance path length error and voltage phase matching degree, and to optimize and balance the location priority under different fault scenarios through algorithmic optimization. The specific details are as follows: Dual-objective fitness function construction: For candidate segments, two complementary optimization objectives are constructed to quantify the "accuracy" and "consistency" of fault location. First objective: Path length error Based on the flow path pattern of zero-sequence current from the fault point to the main power supply, the deviation between the measured path length and the expected length of the algorithm iteration is calculated using the following formula: In the formula, Indicates path length error. The path length of the measured zero-sequence current through the FTU at node i within the candidate segment; : The expected path length generated by the algorithm iteration; Physical meaning: The smaller the value, the closer the path predicted by the algorithm matches the actual zero-sequence current flow path, and the smaller the positioning error at the path level. Second objective: Zero-sequence voltage phase deviation Based on the abrupt change characteristics of the zero-sequence voltage phase upstream and downstream of the fault point, the matching degree between the measured phase and the theoretical phase is calculated using the following formula: In the formula, Indicates the zero-sequence voltage phase deviation; : The measured zero-sequence voltage phase of node i within the candidate segment; The theoretical phase, calculated based on the zero-order network equivalent model, is given by the following formula: in, The angular frequency representing the zero-sequence transient component. , Let i be the zero-sequence inductance and resistance from node i to the main power supply; Physical meaning: The smaller the value, the more consistent the voltage phase characteristics are with the theoretical derivation of the fault location, thus verifying the accuracy of the location from the perspective of electrical characteristics; Dual-objective fusion and constraints: The two objectives are fused into a comprehensive optimization objective using an adaptive weighting factor, while constraints are added: F represents the overall fitness function value, which is a weighted sum of the two sub-targets. The smaller the value, the more accurate the localization result. Indicates the adaptive weighting factor. F1 represents the first optimization objective, path length error, and F2 represents the second optimization objective, zero-sequence voltage phase deviation. Weighting factors: ,in , For transition resistance, The unit is Ω; when Increase Reduce When the impedance increases, the location is primarily determined by phase deviation; conversely, when the impedance decreases, the path length error is primarily relied upon for low-resistance faults. Constraints: The fault location must be within the candidate segment divided by the GNN, and the zero-sequence current direction must be consistent with the multi-mode characteristics. Improved optimization strategy for the multi-objective gray wolf algorithm: To quickly find the optimal solution of the bi-objective function, the traditional MOGWO algorithm is improved as follows to enhance convergence speed and global search capability: Population initialization and encoding, population encoding: representing the fault location as a continuous variable, with the position vector of each individual gray wolf as follows: , This represents the electrical or physical distance from the fault point to the starting point of the candidate segment defined by the GNN; Initialization strategy: Chaotic mapping is used to initialize and generate an initial population with a size of 30 to ensure that individuals are evenly distributed within the candidate segment and avoid local optima caused by initial clustering; The hunting mechanism has been optimized by drawing inspiration from the "surround prey-hunt-attack" mechanism of the gray wolf algorithm and combining it with a spiral search strategy to expand the search range. Encircle the prey: Calculate the distance between an individual and the current optimal solution. α represents the individual with the best fitness in the current population, β represents the individual with the second best fitness, and γ represents the individual with the third best fitness. A dynamic adjustment coefficient 'a' is used; this is a global control coefficient that linearly decreases from 2 to 0 to dynamically adjust the algorithm's search range and step size, balancing global exploration with local exploitation capabilities. Controlling the encirclement range: Where D represents the distance between the current individual and the target prey, reflecting the difference between the individual and the optimal solution; C represents the random weight coefficient, ranging from 0 to 2; and X represents the position vector of the gray wolf individual to be updated. This represents the current position vector of the prey. This represents the position update result, where A represents the dynamic step size coefficient. ; A random number in the range [0,1]. A random number in the range [0,1]. Spiral Search: Introducing a spiral factor into the hunting process ; The constant is t, and the random number t is [0,1], causing the individual search trajectory to expand in a spiral shape around the optimal solution, avoiding getting trapped in local optima: The distance between the individual and the optimal solution; Dynamic step size adjustment: The search step size is adjusted according to the iteration progress to balance global exploration and local development. In the initial stage of iteration: the step size coefficient is set to a larger value to expand the search range and explore potential optimal solutions within the candidate segment; In the later stages of iteration: the step size coefficient is linearly decreased to focus on a fine search around the optimal solution, thus accelerating convergence; Non-dominated ranking and elite preservation: a dual objective for individuals in a population. , Non-dominated sorting is performed to select elite individuals in the Pareto optimal solution set, accounting for 20% of the population; crowding distance is used to maintain the diversity of the solution set, avoid the optimal solution being concentrated in local areas, and ensure that the algorithm can still find the global optimal solution under complex failure scenarios. Precise localization solution: Initialization: Input candidate segment range, multimodal characteristic parameters, and transition resistance. Initialize the improved MOGWO population and the maximum number of iterations; calculate fitness: for each individual, calculate its corresponding... and The population position is updated by adaptive weight fusion into F; Algorithm iteration: The encirclement, hunting, and spiral search operations of the gray wolf algorithm are executed to update the population position; Non-dominated sorting is performed every 5 iterations to retain elite individuals; Convergence judgment: When the change of the optimal solution in 5 consecutive iterations is ≤0.5m, or the maximum number of iterations is reached, the iteration stops; Result output: The individual with the smallest F in the Pareto optimal solution set is selected as the precise fault location; Step S105: Dynamic verification and result output This step verifies and corrects the location results through multi-dimensional verification and outputs key information in a standardized format, as detailed below: First, we introduce the dynamic verification mechanism. Based on the fault location output by the dual-objective optimization algorithm, we perform multiple rounds of dynamic verification by combining real-time data and characteristic patterns of the distribution network to eliminate interference factors and correct errors. Specifically, this includes four verification mechanisms: Zero-sequence current direction consistency verification: Based on the rule that the zero-sequence current directions are opposite upstream and downstream of the fault point, verify whether the measured current directions of the FTUs on both sides of the section corresponding to the location result conform to this rule. Implementation steps: If the positioning result is "segment (5), 200m from the starting point", then check that the current direction of the upstream node of segment (5) should be "positive" and the current direction of the downstream node should be "reverse". If the direction consistency rate is ≥90%, the verification passes; otherwise, feature backtracking is triggered, the initial parameters of the optimization algorithm are corrected, and the calculation is recalculated. Multimodal feature vector matching verification: The verification logic is to compare the similarity between the theoretical feature vector corresponding to the positioning result and the actual collected multimodal feature vector to ensure feature consistency. Implementation steps: Calculate the cosine similarity between the theoretical eigenvectors and the actual eigenvectors: In the formula, Sim represents the cosine similarity value, which ranges from [0,1]. This represents the theoretical multimodal feature vector corresponding to the localization result. This represents the actual multimodal feature vectors acquired by the on-site FTU; If the similarity is ≥0.85, the verification passes; otherwise, the candidate segment range is expanded and the bi-objective optimization is performed again. Topology path validity verification: The verification logic is to combine the current real-time topology of the distribution network to verify whether the segment where the location result is located is on a connected path, so as to avoid invalid location due to topology changes. Implementation steps: If the location result is in "segment (10)", but real-time monitoring shows that the switch between segment (10) and the main power supply has been disconnected, then the path is determined to be invalid; Trigger topology adaptive adjustment, correct the positioning range to the currently connected adjacent segments, and recalculate the positioning result; Redundancy information fault tolerance verification, verification logic: In response to FTU data distortion, the influence of outliers is eliminated through cross-verification of information from multiple nodes; Implementation steps: If a node's FTU direction encoding is "1", but its three adjacent nodes are all "0", then the node is determined to be abnormal data, removed, and the feature vector is reconstructed. Compare the localization results before and after removing outlier data. If the deviation is ≤100m, retain the original result; otherwise, re-execute the GNN solution domain partitioning and bi-objective optimization. Secondly, regarding the output content, after dynamic validation, the following key information is output in a standardized format to meet engineering application requirements: Precise fault location, core information: clearly define the section number where the fault is located and the specific coordinates within the section, with a positioning error ≤50m; auxiliary explanation: mark the relative position of the fault point to the adjacent FTU to facilitate on-site troubleshooting by maintenance personnel; Location reliability and basis: Confidence: Outputs the reliability of the location results as a percentage, calculated based on multimodal feature matching degree and orientation consistency verification results; Key basis: Briefly lists the core features supporting the location results to enhance the interpretability of the results; The system provides error message alerts. If data distortion or topology changes are present, the system will output the specific anomaly point and an assessment of its impact on the location results. If there are distributed power sources near the fault point, the system will indicate the DG type and access status, and remind maintenance personnel to pay attention to safe operation. Location time and process indicators, outputting the total time from the occurrence of the fault to the completion of the location, as well as the time ratio of each step, to provide data for system performance evaluation; Finally, regarding the output format and interface, the data interface is connected to the distribution network dispatching system via the IEC 61850 standard protocol, outputting structured data and supporting linkage with SCADA and GIS systems to mark fault locations on electronic maps; the visualization display shows the fault location diagram, key characteristic waveforms, and verification process on the operation and maintenance terminal, intuitively presenting the location basis.

Images