A power distribution network fault diagnosis method and system based on a collaborative sparse graph network

Abstract

The application discloses a power distribution network fault diagnosis method and system based on a cooperative sparse graph network and belongs to the technical field of power system state monitoring. In view of the deficiencies of the prior art, by collecting multi-dimensional time sequence operation state data of the power distribution network, an initial physical graph topology and a node characteristic matrix are constructed; a sparse initialization strategy based on an ERK distribution is used to perform binary mask distribution on model weights of a graph neural network; a composite physical similarity model is constructed in combination with node characteristic correlation and electrical admittance information, edge pruning, edge growth and maximum spanning tree loop checking of the initial topology are performed with the aid of a hybrid candidate sampling strategy, and a noise-free reconstructed topology meeting the radial operation constraint is generated; according to the real-time memory occupation state of the edge gateway, dynamic pruning and adaptive sparse training of the model weights are performed, effective control of the model complexity and resource consumption is realized; and the lightweight model is deployed on the edge terminal, millisecond-level online fault diagnosis and auxiliary decision-making are realized.

Description

Technical Field

[0001] This invention belongs to the field of power system condition monitoring technology, specifically relating to a method and system for fault diagnosis of distribution networks based on cooperative sparse graph networks. Background Technology

[0002] In recent years, with the continuous development of social technology, human production and daily life have become increasingly reliant on electricity, and the scale of the power grid has expanded rapidly. However, this has also brought new challenges: in a large-scale distribution network, if a short circuit or ground fault occurs, and fault location and switch disconnection cannot be completed within milliseconds, the fault will spread rapidly, even causing widespread power outages. At the same time, the complex operating environment of the distribution network, aging communication equipment, and electromagnetic interference often result in missing and noisy sensor data, further increasing the difficulty of fault location. Therefore, the current development trend is to utilize artificial intelligence models...

[0003] Edge computing nodes, such as feeder terminal units (FTUs) and distribution terminal units (DTUs), are deployed to enable rapid on-site decision-making and real-time response. Utilizing graph neural networks (GNNs) for refined condition assessment and fault diagnosis of distribution network nodes can significantly improve the reliability of power grid supply.

[0004] However, directly applying traditional graph neural network models for distribution network fault diagnosis presents several serious problems: traditional graph neural networks typically rely on high computational resources, while the computing power and storage capacity of edge devices in distribution networks are limited, making it difficult to support the efficient operation of dense models. Various sensors (such as FTU and DTU terminals) are prone to communication delays or damage in harsh environments, leading to a large number of false alarms in the switch status received by the master station dispatch system. If this physical topology containing "structural noise" is directly input into the graph neural network, the network will aggregate information along incorrect paths, resulting in a precipitous drop in fault diagnosis accuracy.

[0005] To address the high computational complexity and storage overhead of graph neural networks (Graph Neural Networks), existing techniques have proposed Graph Neural Network Sparsification methods. These methods selectively prune network connections, node features, or model parameters to reduce redundancy and improve computational efficiency. To address topology noise, existing techniques have proposed Graph Structure Learning (GSL) methods, which regenerate the topology by calculating the similarity between all node pairs. However, these methods typically rely on pure mathematical principles, and direct application can ignore Kirchhoff's laws and radial physical constraints of power distribution networks, easily generating "pseudo-connections" that do not conform to physical reality. Furthermore, the computational complexity of existing methods is on the global quadratic order of magnitude. Furthermore, fully parameterized dense graph neural networks (DNNs) are extremely large. Distribution networks require millisecond-level fault clearing, and the computing power and memory of field edge computing gateways are extremely limited, making it impossible to support online real-time inference for such large models. Currently, there is no reasonable and effective method for distribution network fault diagnosis that balances high accuracy, adherence to physical laws, and extremely low edge computing overhead. Summary of the Invention

[0006] To address the problems of insufficient diagnostic accuracy, poor real-time performance, and difficulty in adapting to edge device deployment in existing power distribution network fault diagnosis methods under complex operating environments, this invention provides a power distribution network fault diagnosis method and system based on collaborative sparse graph networks.

[0007] To achieve the above objectives, the present invention employs the following technical solutions:

[0008] A method for fault diagnosis of distribution networks based on cooperative sparse graph networks, the method comprising the following steps:

[0009] Step 1: Obtain multi-dimensional time-series operation status data of the power distribution network and construct the initial physical graph topology and node feature matrix;

[0010] Step 1.1: Extract the area using a high-frequency sampling rate via a synchronous phasor measurement unit (PMU) or a smart distribution terminal (FTU / DTU). Each measurement node within a specific time window The multidimensional electrical operation characteristics within;

[0011] Step 1.2: The extracted feature dimensions are Includes: three-phase voltage amplitude Three-phase current amplitude Active power reactive power ;

[0012] Step 1.3: Concatenate the above features to form a node feature matrix. ;

[0013] Step 1.4: Extract the switch status between each node from the Distribution Management System (DMS); if the system displays the node... With nodes If the switch is physically closed, then the initial connection exists, and the initial physical topology is constructed accordingly. ,in, For a set of nodes, This is the initial physical edge set; due to communication delay, the initial edge set... The topology naturally contains 5% to 15% faulty connections, i.e., topological noise.

[0014] Step 2: Construct a graph neural network model. Use a sparse initialization strategy based on ERK distribution to assign binary masks to the weights of the graph neural network model, thus completing the sparse initialization of the graph neural network model.

[0015] In step 2, a sparse initialization strategy based on ERK distribution is used to perform binary mask allocation on the model weights of the graph neural network, preserving the parameter density. The calculation formula is: ;

[0016] in, The first character of the graph neural network is represented by the second character. layer, Representing the The density of parameters retained by the layer. Indicates the first The proportion of the layer that is cut off; This represents the number of input feature channels (number of neurons) for this layer. This represents the number of output feature channels; and These represent the spatial dimensions of the convolution kernel. It indicates a direct proportional relationship.

[0017] Step 3: Use the graph encoder in the graph neural network to extract features from the node feature matrix to obtain the hidden layer feature vector of each node. Combine the node feature correlation with the electrical admittance information of the distribution network bottom layer to construct a composite physical similarity model. Based on the composite physical similarity model and the hybrid candidate sampling strategy, perform edge pruning and edge growth on the initial physical graph topology. Perform loop detection on the graph topology after edge pruning and edge growth. If a closed loop exists, disconnect the redundant branches according to the preset physical topology constraint rules to generate a noiseless reconstructed topology that meets the radial operation constraints.

[0018] Step 3.1: Update the parameters of the graph neural network model based on the noiseless topology reconstruction, and simultaneously extract arbitrary nodes using the graph encoder in the graph neural network. and Hidden layer feature vectors and A composite physical similarity model is constructed, and the calculation formula for the composite physical similarity model is as follows: ;

[0019] in, Represents a node With nodes The composite physical similarity between them; The temporal Pearson correlation coefficient represents the features of the hidden layer of two nodes; Represents the node admittance matrix of the distribution network. With nodes The absolute value of the admittance element between them serves as a physical prior mask to restrict non-physical connectivity associations; and These are the normalized balance coefficients;

[0020] This composite physical similarity model uses the physical admittance between nodes as a priori physical mask. If there is no low-impedance physical path between two nodes (admittance approaches 0), the composite similarity is suppressed, thereby effectively filtering out mathematical pseudo-correlation connections caused by disturbances from the same main power source. The following situations may occur:

[0021] Scenario 1: High data correlation + High physical admittance + Initially no connection:

[0022] If node With nodes The Pearson correlation coefficient of the hidden layer features is greater than a set threshold, and the absolute value of its node admittance is... Greater than the admittance threshold, but in the initial physical graph topology If there is no connection between the two nodes, it is determined that there is a hidden electrical connection or a missed connection that was not reported by manual closure in the distribution network. The node pair is then used as a high-scoring candidate edge, and the "edge growth" operation is performed to add the edge in the reconstructed graph.

[0023] Scenario 2: High data correlation + low physical admittance:

[0024] If the temporal Pearson correlation coefficient of the hidden layer features of two nodes is extremely high, but the absolute value of their physical admittance is low... Approaching 0, meaning physically far apart or spanning different main transformers; handling method: determine that the high correlation is a "mathematical pseudo-correlation" caused by disturbances in the same upstream power system, refuse to perform edge growth, maintain the non-connected state of the two, and prevent violation of Kirchhoff's current law;

[0025] Scenario 3: Low data correlation + initial connection:

[0026] If node With nodes The correlation coefficient of the hidden layer features is lower than the discard threshold. However, in the initial physical topology If the system records that there is a closed connection between the two, then the initial connection recorded by the system is determined to be "structural topology noise" caused by sensor failure or communication delay. The "edge trimming" operation is then performed to forcibly delete the edge in the reconstructed graph.

[0027] Scenario 4: Low data correlation + no initial connections:

[0028] If the correlation coefficient between two nodes is extremely low and there is no connection in the initial topology, then it is determined that the two nodes are not directly related in terms of physical and electrical aspects, and the status quo is maintained without any processing.

[0029] Step 3.2: Generate candidate edge set based on hybrid candidate sampling strategy The formula is: ;

[0030] in, This represents the search set of the local feeder area topology boundary, which is limited to the 2-hop neighborhood of the target node belonging to the same transformer area. This represents a cross-regional targeted harmonic sampling set, which, based on the historical power quality operation logs of the distribution network, selectively extracts pairs of distant nodes that have historically experienced harmonic resonance or cascading tripping as candidate nodes across the entire global scope.

[0031] Remove composite physical similarity from the initial physical graph topology Below the set threshold The communication pseudo-edges are identified, and candidate edges are grouped and ranked by composite physical similarity. Adding (Top-k) new edges to the physical graph topology completes the graph topology reconnection in linear complexity;

[0032] Step 3.3: Perform loop detection on the reconnected graph topology, calling the Depth-First Search (DFS) algorithm to check whether a closed loop has formed in the reconstructed graph; if a closed loop exists, disconnect redundant branches according to the preset physical graph topology constraint rules, as follows:

[0033] Scenario 1: The closed loop is entirely composed of the initial physical edges:

[0034] If the DFS algorithm detects a cycle, and all edges in the cycle belong to the initial physical edge set. When the power grid itself has a special case of physical loop operation, the composite physical similarity score of all edges on the loop is retrieved, the physical edge with the lowest composite physical similarity score is selected, and it is temporarily disconnected in the graph structure. This forces the graph neural network to transmit fault features in a radial manner, avoiding excessive smoothing of local information and dead loops.

[0035] Scenario 2: The closed loop is composed of a mixture of "initial physical edges" and "newly added implicit edges":

[0036] If the loop contains both the original physical connections of the power distribution management system and the implicit electrical edges added by the graph neural network model through adaptive reconstruction, the composite similarity weights of all edges in the loop are compared. If the weight of the newly added implicit edge is greater than that of a certain physical edge, the newly added implicit edge is retained and the physical edge with the lowest score is disconnected. If the weight of the newly added implicit edge is less than that of all physical edges, the newly added implicit edge is directly removed as a redundant branch.

[0037] Scenario 3: The closed loop is entirely composed of "newly added implicit edges":

[0038] If the loop is entirely composed of newly added implicit electrical edges through adaptive reconstruction by the graph neural network model, it is determined that the graph neural network model has over-captured common-mode interference features in a local region. Strictly following the descending order of the maximum spanning tree, the newly added edge with the lowest score in the local loop is removed until the local subgraph is restored to a radial tree structure.

[0039] After the above-mentioned pruning, growth, and radial physical safety checks, the final output graph structure, which has removed structural noise and strictly meets the safety constraints of the distribution network, is the noise-free reconstructed topology.

[0040] Step 4: Update the parameters of the graph neural network model based on the noiseless reconstructed topology, and simultaneously obtain the hardware resource status of the target edge computing gateway in real time. Adjust the target sparse density of the graph neural network weights dynamically according to the hardware resource status, and use a model pruning strategy to remove redundant weights whose absolute values ​​are lower than a preset threshold. Alternately iterate through steps 3 and 4 until the graph neural network model converges to obtain a lightweight fault diagnosis model as the final output.

[0041] In step 4, the target sparse density of the graph neural network weights is dynamically adjusted based on the hardware resource status. The specific formula for calculating the target sparse density is as follows: ;

[0042] in, For the current training round, This is the starting round for dynamic pruning. This is the final round of dynamic pruning; For the initial parameter density, The final parameter density of the baseline target; This represents the real-time available memory occupancy of the edge computing gateway; This represents the memory security threshold limit for edge computing gateways; As a regulating factor; This is a hardware resource penalty function for edge computing gateways.

[0043] when Approaching or exceeding Hardware resource penalty function When the value decreases, an amplitude pruning strategy is triggered to remove redundant weights whose absolute weight value is lower than a preset threshold.

[0044] In the training process of dynamically adjusting the target sparse density of graph neural network weights Based on the real-time available memory occupancy of the edge computing gateway Adaptive adjustments are made; the resource penalty function and parameter pruning are handled as follows, taking into account different hardware load conditions of the target edge computing gateway:

[0045] Scenario 1: Sufficient hardware resources:

[0046] If the available memory occupancy is read in real time Less than the set memory safety threshold limit At that time, the edge computing gateway hardware resource penalty function The value approaches 1, and the graph neural network model achieves its final parameter density according to the preset baseline target. Perform normal linear decay pruning to preserve fault feature extraction parameters to the greatest extent while ensuring computational efficiency;

[0047] Scenario 2: Hardware resources are critically strained:

[0048] If the available memory occupancy is read in real time Approaching or equal to the memory safety threshold At that time, the edge computing gateway hardware resource penalty function Activated and its value rapidly decreases, forcibly reducing the target's sparse density. The graph neural network model immediately accelerates magnitude pruning in the current round, increasing the proportion of neurons with smaller absolute weights to be pruned, and forcibly releasing memory usage to avoid device crashes.

[0049] Scenario 3: Hardware resources are severely overloaded / Other high-level tasks suddenly occur:

[0050] If the available memory occupancy is read in real time Exceeding the safety threshold indicates that the edge gateway is currently executing other high-priority protection trip commands. At this point, the graph neural network model immediately suspends the parameter iteration and topology update operations of this round, directly locks the effective sparse graph neural network model weights and topology structure of the previous converged round, releases the underlying hardware computing resources of the edge computing gateway, and ensures the absolute priority and security of the bottom-level physical relay protection actions of the power grid.

[0051] Step 5: Deploy the noiseless reconstructed topology and the lightweight fault diagnosis model on the edge computing gateway, receive real-time electrical data, perform forward inference, and output fault diagnosis results.

[0052] Step 5.1: Combine the lightweight fault diagnosis model and noiseless reconstructed topology optimized by the joint and alternating optimization in Steps 3 and 4. This information is then fixed and distributed to the edge computing gateways deployed at the power distribution network site.

[0053] Step 5.2: During the data plane synchronization phase when a fault occurs, the edge computing gateway directly calls the already established reconstructed graph topology. As a sparse adjacency matrix, real-time electrical fault features are extracted, and a lightweight fault diagnosis model is used to perform sparse matrix forward inference to output fault diagnosis results.

[0054] A distribution network fault diagnosis system based on a collaborative sparse graph network is provided. The system is used to implement a distribution network fault diagnosis method based on a dual sparse graph neural network. The system includes a cloud training unit and a field edge online prediction unit.

[0055] The cloud training unit is deployed in the cloud or the main station backend to obtain historical operation data of the distribution network, asynchronously execute sparse initialization of the graph neural network model, adaptive reconstruction and security verification of the graph topology, and dynamic pruning of hardware resources to obtain a noiseless reconstructed topology and a lightweight fault diagnosis model.

[0056] The field edge online prediction unit is deployed in the field edge terminal device. It is used to receive and solidify the noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit. Based on the real-time collected electrical sampling data, it uses the reconstructed topology to perform sparse forward inference and outputs the distribution network node fault state classification results to assist decision-making.

[0057] The cloud-based training unit includes: a data acquisition and feature construction module, a sparse initialization module, a topology adaptive reconstruction module, and a dynamic sparse training module.

[0058] The data acquisition and feature construction module is used to acquire multi-dimensional time-series electrical operation data and historical remote signaling switch status of the distribution network, construct the initial physical diagram topology and node feature matrix, and pass the results to subsequent modules;

[0059] The sparse initialization module is used to assign initial parameters to the graph neural network model using a sparse initialization strategy based on the network dimension, and the results are passed to the training process.

[0060] The topology adaptive reconstruction module is used to calculate composite physical similarity, combine a hybrid candidate sampling strategy to prune and grow the topology, perform loop removal verification, and output the reconstructed topology.

[0061] The dynamic sparse training module is used to read the real-time memory usage status of the target edge terminal and dynamically adjust the pruning rate to remove redundant parameters of the graph neural network model;

[0062] The on-site edge online prediction unit includes: a noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit, a real-time electrical data receiving module, a sparse forward inference calculation module, and a fault classification and location output module.

[0063] The noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit are used to communicate and interact with the cloud training unit. The noiseless reconstructed topology and lightweight fault diagnosis model that has completed dual sparse training is received and locally solidified, providing a static network structure and parameter benchmark for real-time inference at the field edge.

[0064] The real-time electrical data receiving module is used to acquire real-time electrical sampling data from each field measurement node during the synchronous real-time phase when a fault occurs in the distribution network.

[0065] The sparse forward inference computation module is used to call the noiseless reconstructed topology and lightweight fault diagnosis model issued and solidified by the cloud training unit, and to perform sparse matrix forward propagation computation on the real-time electrical sampling data using the noiseless reconstructed topology as a sparse adjacency matrix.

[0066] The fault classification and location output module is used to output the state classification probability distribution of faults at each node of the distribution network based on the results of the sparse forward inference calculation, so as to assist the underlying equipment in making local fault isolation decisions.

[0067] Compared with the prior art, the present invention has the following advantages:

[0068] (1) Better topology noise resistance and physical interpretability: This invention breaks away from the black-box mode of traditional graph learning that relies solely on data correlation. It innovatively introduces "electrical admittance" as a physical prior mask and combines it with "maximum spanning tree" for radial safety verification. This mechanism not only effectively filters out communication pseudo-edges caused by sensor false alarms, but also avoids mathematical pseudo-correlations that violate Kirchhoff's laws caused by disturbances in the same transformer, ensuring that the noiseless reconstructed topology is more in line with the physical power flow law and safety bottom line of the power system.

[0069] (2) Exponential dimensionality reduction in computational complexity: Traditional graph structure learning requires traversing the entire world when constructing candidate edges, which has a complexity of up to 100%. This invention pioneers a hybrid sampling strategy that combines local co-feeder transformer area topology boundary search with cross-regional targeted harmonic sampling, successfully reducing the topology evolution complexity to linear and significantly shortening the algorithm training time, thus making dynamic graph reconstruction a reality in large-scale distribution networks.

[0070] (3) Effectively alleviates the bottleneck of limited computing power in edge terminals: This invention proposes a collaborative sparse architecture of graph topology and model weight parameters. By completely decoupling the control plane and data plane, and by introducing an innovative "hardware resource penalty function", the redundant computing power of the originally huge graph neural network model with huge parameters is accurately compressed by more than 80%, effectively preventing layer collapse. This enables cutting-edge deep graph neural network models to be smoothly deployed in existing network-based old terminal devices with extremely limited computing power and memory, achieving true millisecond-level local fault diagnosis and response without increasing hardware modification costs. Attached Figure Description

[0071] Figure 1 This is a flowchart of a power distribution network fault diagnosis method based on a dual sparse graph neural network in a specific embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of the adaptive reconstruction and verification process of the electrical semantic topology of the distribution network in a specific embodiment of the present invention;

[0073] Figure 3 This is a schematic diagram of the dual sparse cooperative evolution and dynamic pruning logic in a specific embodiment of the present invention;

[0074] Figure 4 This is a diagram illustrating the control plane and data plane architecture of the power distribution network fault diagnosis system in a specific embodiment of the present invention.

[0075] Figure 5 This is a schematic diagram of the matrix cooperative evolution principle at the bottom layer of the dual sparsity mechanism in a specific embodiment of the present invention;

[0076] Figure 6 This invention provides a performance comparison of different fault diagnosis methods on a standard simulation measurement dataset in a specific embodiment. Detailed Implementation

[0077] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0078] A method for fault diagnosis of distribution networks based on cooperative sparse graph networks is described in the appendix. Figure 1 As shown, it includes the following steps:

[0079] Step 1: Obtain multi-dimensional time-series operation status data of the power distribution network and construct the initial physical graph topology and node feature matrix;

[0080] Step 1.1: Select the IEEE 33-node distribution system as the target area. This system comprises 33 nodes (numbered 1 to 33), 32 normally closed sectionalizing lines, and 5 normally open tie lines. Intelligent measurement terminals are used to extract data for each node within a continuous time window. The multidimensional electrical operation characteristics within the system, denoted as _____. Specifically, this includes: three-phase voltage amplitude Three-phase current amplitude Active power reactive power and voltage phase angle The time-series data from the above 33 nodes are concatenated and normalized to form a dimension of [dimension number missing]. Basic feature matrix dataset ;

[0081] Step 1.2: Extract the current switch remote signaling status recorded by the Distribution Management System (DMS) and construct the initial physical topology. Due to the complex electromagnetic environment and aging communication equipment at the site, the initial physical topology was affected. There will inevitably be errors, and the following situations may occur:

[0082] Scenario 1: The state is correctly recorded; the connection between nodes 2 and 3 is actually closed, and the remote signaling report also indicates closure. The system sets the corresponding elements of the adjacency matrix... ;

[0083] Scenario 2: False communication alarms generate pseudo-edges / structural noise. The switch between nodes 7 and 8 is actually open, but the sensor falsely reports it as "closed." The system incorrectly instructs... (as attached) Figure 2 (as shown)

[0084] Scenario 3: Missed edge / feature reports due to communication misreporting. A normally open handshake switch exists between node 18 and node 33. This switch has been manually closed on-site to balance the load, but the scheduling system has not updated it in a timely manner. System command... (as attached) Figure 2 (As shown).

[0085] To simulate realistic "structural topology noise," 10%–15% of false alarms and missed alarms were randomly injected into the simulation data. Virtual intelligent distribution terminals deployed at nodes recorded time-series electrical data before and after a short-circuit fault at a high-frequency sampling rate. The data was organized according to the distribution network dataset label format and divided into training, validation, and test sets, covering four categories: "normal operation," "single-phase ground fault," "two-phase short circuit," and "three-phase short circuit." The dataset distribution is shown in Table 1.

[0086] Table 1 Distribution of Datasets for Distribution Network Faults

[0087] Normal operation 4200 1200 600 6000 Single-phase ground fault 3500 1000 500 5000 Two-phase short circuit 2100 600 300 3000 Three-phase short circuit 1400 400 200 2000 Scale Summary 11200 3200 1600 16000

[0088] Step 2: Use an ERK-based sparse initialization strategy to assign binary masks to the model weights of the graph neural network to generate an initial sparse network;

[0089] To adapt to the low memory constraints of edge devices, network parameters must be sparsified at the beginning of graph neural network model construction. However, using traditional uniform random pruning can easily lead to the interruption of key feature extraction layers with few channels. This invention uses the Erdos-Renyi-Kernel (ERK) distribution formula to perform binary masking on the graph neural network model parameters. Its parameter retention density... The calculation formula is: ;

[0090] in, The first character of the graph neural network is represented by the second character. layer, Representing the The density of parameters retained by the layer. Indicates the first The proportion of the layer that is cut off; This represents the number of input feature channels (number of neurons) for this layer. This represents the number of output feature channels; and These represent the spatial dimensions of the convolution kernel. It indicates a direct proportional relationship.

[0091] The numerator of this formula is the sum of dimensions, and the denominator is the multiplication of dimensions. For large-scale fully connected hidden layers, the denominator grows exponentially, resulting in... The retention rate is extremely low; however, for critical input layers with narrow input channels, the denominator is small, and the ERK distribution adaptively allocates a higher retention rate. This mechanism ensures the efficient transmission of short-circuit fault characteristics from a mathematical perspective.

[0092] Step 3: Extract the hidden layer feature vectors of each node using the graph encoder of the graph neural network, and construct a composite physical similarity model by combining the node feature correlation with the electrical admittance information of the distribution network bottom layer; based on the composite physical similarity model and the hybrid candidate sampling strategy, perform edge pruning and edge growth on the initial physical graph topology; perform loop detection on the graph topology after edge pruning and edge growth, and if a closed loop exists, disconnect redundant branches according to the preset physical topology constraint rules to generate a noiseless reconstructed topology that meets the radial operation constraints;

[0093] Step 3.1: Extract arbitrary nodes in the IEEE 33-node system using a graph encoder based on a graph neural network. and Hidden layer feature vectors and The formula for calculating composite physical similarity is: ;

[0094] in, Represents a node With nodes The composite physical similarity between them; The temporal Pearson correlation coefficient represents the features of the hidden layer of two nodes; Represents the node admittance matrix of the distribution network. With nodes The absolute value of the admittance element between them serves as a physical prior mask to restrict non-physical connectivity associations; and These are the normalized balance coefficients;

[0095] By using physical admittance as a priori mask, the composite similarity score is high only when the data fluctuations of the two nodes are highly consistent and the physical topology allows the current to flow directly. This effectively avoids the "mathematical pseudo-correlation that violates Kirchhoff's laws" that is easy to produce by pure data algorithms.

[0096] Step 3.2: Generate candidate edge set based on hybrid candidate sampling strategy The formula is: ;

[0097] in, This represents the search set of the local feeder area topology boundary, which is limited to the 2-hop neighborhood of the target node belonging to the same transformer area. This represents a cross-regional targeted harmonic sampling set, which, based on the historical power quality operation logs of the distribution network, selectively extracts pairs of distant nodes that have historically experienced harmonic resonance or cascading tripping as candidate nodes across the entire global scope.

[0098] Set edge retention threshold In this embodiment, a value of 0.2 is preferred. In practical applications, a reasonable range for this threshold is 0.15 to 0.3. Based on the composite similarity score... From the initial connection state Topology correction may involve the following situations:

[0099] Scenario 1: Low data correlation + initial connection, calculation reveals nodes 7 and 8... Determine the system records. This is a "communication pseudo-edge". Perform "edge pruning" to forcibly delete this connection in the reconstructed adjacency matrix (see attached). Figure 2 (as shown)

[0100] Scenario 2: High correlation + high physical admittance + no initial connection; data from nodes 18 and 33 are highly resonant, and the connection between them provides sufficient admittance flux. ,Score The graph neural network model determined that a true "latent electrical correlation" existed. An "edge growth" operation was performed, adding a data connection between nodes 18 and 33.

[0101] Scenario 3: High correlation + low physical admittance. If nodes 18 and 22, located on different branches, are interfered with by the same upstream main transformer harmonic source, the data correlation coefficient is extremely high. However, since there is no direct physical connection between the two, The graph neural network model determined that it was a "mathematical pseudo-correlation caused by common-mode interference" and rejected the connection.

[0102] Scenario 4: Low data correlation + no initial connection, significant differences in node characteristics and no physical connection, maintaining a disconnected state.

[0103] Step 3.3: Perform loop detection on the reconnected graph topology, calling the Depth-First Search (DFS) algorithm to check whether a closed loop has formed in the reconstructed graph; if a closed loop exists, disconnect redundant branches according to the preset physical graph topology constraint rules, as follows:

[0104] Scenario 1: The closed loop is entirely composed of the initial physical edges:

[0105] If the DFS algorithm detects a cycle, and all edges in the cycle belong to the initial physical edge set. When the power grid itself has a special case of physical loop operation, the composite physical similarity score of all edges on the loop is retrieved, the physical edge with the lowest composite physical similarity score is selected, and it is temporarily disconnected in the graph structure. This forces the graph neural network to transmit fault features in a radial manner, avoiding excessive smoothing of local information and dead loops.

[0106] Scenario 2: The closed loop is composed of a mixture of "initial physical edges" and "newly added implicit edges":

[0107] If the loop includes both existing physical connections from the power distribution management system and newly added implicit electrical edges through adaptive reconstruction by the graph neural network model, the composite similarity weights of all edges in the loop are compared. If the weight of the newly added implicit edge is greater than that of a physical edge, the newly added implicit edge is retained, and the physical edge with the lowest score is disconnected. If the weight of the newly added implicit edge is less than that of all physical edges, the newly added implicit edge is directly removed as a redundant branch. Assuming that after adding implicit edges at nodes 18 and 33, a large loop is formed (containing the newly added edges from 18 to 33, and existing physical edges such as 32 to 33), the processing method is to compare the weights of each edge in the loop. Because 18-33 has an extremely high score ( ), while the original physical edge 32-33 scored lower ( The actual electrical connection in section 32-33 is determined to be extremely weak. At this point, the system retains the implicit edge 18-33 and forcibly disconnects the physical segment edge 32-33 (as shown in the attached diagram). Figure 2 (As shown), until a strictly radial tree structure is restored, ultimately outputting a safe and noiseless reconstructed topology. ;

[0108] Scenario 3: The closed loop is entirely composed of "newly added implicit edges":

[0109] If the loop is entirely composed of newly added implicit electrical edges through adaptive reconstruction by the graph neural network model, it is determined that the graph neural network model has over-captured common-mode interference features in a local region. Strictly following the descending order of the maximum spanning tree, the newly added edge with the lowest score in the local loop is removed until the local subgraph is restored to a radial tree structure.

[0110] Step 4: Update the parameters of the graph neural network model based on the noiseless reconstructed topology, and simultaneously obtain the hardware resource status of the target edge computing gateway in real time. Adjust the target sparse density of the graph neural network weights dynamically according to the hardware resource status, and use a model pruning strategy to remove redundant weights whose absolute values ​​are lower than a preset threshold. Alternately iterate through steps 3 and 4 until the graph neural network model converges to obtain a lightweight fault diagnosis model.

[0111] In step 4, the target sparse density of the graph neural network weights is dynamically adjusted based on the hardware resource status. The specific formula for calculating the target sparse density is as follows: ;

[0112] in, For the current training round, This is the starting round for dynamic pruning. This is the final round of dynamic pruning; For the initial parameter density, The final parameter density of the baseline target; This represents the real-time available memory occupancy of the edge computing gateway; This represents the memory security threshold limit for edge computing gateways; As a regulating factor; Hardware resource penalty function for edge computing gateways: , To preset a memory safety threshold, its value is preferably between 75% and 90%, and in this embodiment it is set to 80%.

[0113] In the training process of dynamically adjusting the target sparse density of graph neural network weights Based on the real-time available memory occupancy of the edge computing gateway Adaptive adjustments may be made; depending on the different hardware load conditions of the target edge computing gateway, the following situations may occur (refer to the appendix). Figure 3 ):

[0114] Scenario 1: Sufficient hardware resources:

[0115] If the available memory occupancy is read in real time Less than the set memory safety threshold limit At that time, that is Hardware resource penalty function for edge computing gateway The value approaches 1, and the graph neural network model achieves its final parameter density according to the preset baseline target. Perform normal linear decay pruning to preserve fault feature extraction parameters to the greatest extent while ensuring computational efficiency;

[0116] Scenario 2: Hardware resources are critically strained:

[0117] If the available memory occupancy is read in real time Approaching or equal to the memory safety threshold At that time, that is Hardware resource penalty function for edge computing gateway Activated and its value rapidly decreases, forcibly reducing the target's sparse density. The graph neural network model immediately accelerates magnitude pruning in the current round, increasing the proportion of neurons with smaller absolute weights to be pruned, and forcibly releasing memory usage to avoid device crashes.

[0118] Scenario 3: Hardware resources are severely overloaded / Other high-level tasks suddenly occur:

[0119] If the available memory occupancy is read in real time Exceeding the security threshold means that the edge computing gateway is currently executing other high-priority protection trip commands, i.e. At this point, the graph neural network model immediately stops the parameter iteration and topology update operation of this round, directly locks the effective sparse graph neural network model weights and topology structure of the previous round of convergence, releases the underlying hardware computing resources of the edge computing gateway, and ensures the absolute priority and security of the bottom-level physical relay protection action of the power grid.

[0120] To further explain the inherent coupling relationship between the graph topology reconstruction (step 3) and the weight pruning of the graph neural network model (step 4) from a mathematical perspective, in conjunction with the appendix... Figure 5 The underlying matrix cooperative evolution principle diagram of the dual sparsity mechanism shown is discussed in depth as follows:

[0121] As attached Figure 5 As shown in the initial state above, the system decouples the optimization process into two parallel paths, left and right. In the structural sparsification path on the right, the input initial adjacency matrix... It has high density and contains a large amount of structural topological noise (manifested as an excessive number of non-zero 1s) due to sensor false alarms; in the parameter sparsification path on the left, the initial weight matrix After ERK distribution processing, it exhibits a collapse-resistant semi-sparse state (characterized by non-zero parameters * and 0 doping). Traditional graph neural network optimization often severs the relationship between topology and parameters, leading to the mutual amplification of errors.

[0122] This invention innovatively proposes a dual-sparseness cooperative evolution architecture. (See attached diagram) Figure 5 The dual-channel processing path and cross-feedback dashed line in the middle are shown below:

[0123] On the one hand, in the right-hand structural path, the topology reconstruction operation based on composite physical similarity and hybrid sampling continuously modifies the adjacency matrix. Perform discretization updates to generate a reconstructed sparse adjacency matrix. Mathematically, this process is equivalent to eliminating pseudo-connection noise; as shown by the dashed feedback line crossing to the left in the figure, the noiseless prior Laplacian matrix can guide the gradient descent process of the parameter path more accurately.

[0124] On the other hand, in the left-hand parameter path, the dynamic decay operation based on the hardware resource penalty function continuously affects the weight matrix. Apply amplitude pruning to generate a minimal weight matrix. As shown by the dashed feedback line extending to the right in the figure, a highly sparse weight matrix can filter out a large amount of feature noise and extract purer hidden layer features. This, in turn, feeds back into and improves the calculation accuracy of the composite physical similarity on the right side. (See attached image.) Figure 5 The final convergence state shown below represents the adjacency matrix after multiple joint training iterations. With weight matrix The non-zero elements in the model are significantly reduced, and double sparsity is achieved through mutual cross-feedback. This enables the algorithm model finally sent to the edge device to have a real physical information propagation path (sparse topology) at the macro level, and to be stripped of complex computational parameters (sparse weights) at the micro level, thus mathematically satisfying the stringent requirements of edge gateways for both high precision and low overhead.

[0125] Step 5: Deploy the noiseless reconstructed topology and the lightweight fault diagnosis model on the edge computing gateway, receive real-time electrical data, perform forward inference, and output fault diagnosis results.

[0126] Step 5.1: Combine the lightweight fault diagnosis model and noiseless reconstructed topology optimized by the joint and alternating optimization in Steps 3 and 4. This information is then fixed and distributed to the edge computing gateways deployed at the power distribution network site.

[0127] Step 5.2: During the data plane synchronization phase when a fault occurs, the edge computing gateway directly calls the already established reconstructed graph topology. As a sparse adjacency matrix, real-time electrical fault features are extracted, and a lightweight fault diagnosis model is used to perform sparse matrix forward inference to output fault diagnosis results.

[0128] As attached Figure 4 As shown, this system is strictly divided into a "control plane / cloud training unit" and a "data plane / on-site edge online prediction unit". Steps 1 to 4 are all completed asynchronously in the background control plane, and then the solidified noiseless reconstructed topology is... A lightweight fault diagnosis model, retaining only a few weighted parameters, is then deployed to the edge gateway. This completes the full execution flow of the distribution network fault diagnosis method of this invention. To further illustrate the practical application effects of steps 1 to 5 above, specific performance verification and data analysis are provided below using standard simulation measurement datasets.

[0129] Step 6: Performance comparison of different fault diagnosis methods on the dataset:

[0130] The experiment simulated a low-power ARM architecture edge computing gateway with a main frequency of 800MHz. By measuring the average classification accuracy and macro-F1 score (used to evaluate overall performance under class imbalance) for each fault type, the model of this invention was quantitatively compared with other benchmark models, including traditional full graph neural networks, attention-based graph networks, and traditional static sparse graph networks. The experimental results are as follows: Figure 6 As shown.

[0131] Experimental results show that the overall accuracy of the model in this invention is 2.35% higher than that of traditional full graph neural networks on datasets containing structural noise; for the "single-phase ground fault" which occurs frequently and has weak characteristics, the diagnostic accuracy is improved by 3.82%. Meanwhile, the lightweight fault diagnosis model output by this invention has only 0.48MB of parameters and a single forward inference latency of only 3.2ms, significantly meeting the timeliness requirements of distribution network protection actions. Although traditional static sparse graph networks have a smaller number of parameters, their diagnostic accuracy is significantly reduced because they cannot resist the structural noise of the initial physical graph; this invention effectively balances high accuracy with low computational cost.

[0132] To verify the effectiveness of the proposed "topology adaptive reconstruction module" and "dynamic sparse training module," module verification experiments were conducted on a standard simulation measurement dataset. The results are shown in Table 2.

[0133] Table 2 Comparison of the effectiveness verification results of dual sparse modules on the standard simulation measurement dataset.

[0134] × × 94.85 93.12 100% 18.7 × √ 92.16 90.58 20% 3.5 √ × 96.92 95.86 100% 20.4 √ √ 97.20 96.54 20% 3.2

[0135] Experimental results show that, compared with the baseline network, introducing the topology adaptive reconstruction module alone can improve the overall accuracy to 96.92%, indicating that this module effectively removes structural topological noise and improves feature aggregation capability. Introducing the dynamic sparse training module alone reduces inference latency, but the overall accuracy drops to 92.16%, indicating that direct pruning on noisy topologies exacerbates feature loss errors. When the above dual sparse modules are jointly introduced, the graph neural network model not only reduces the number of parameters by 80% and the edge inference latency to 3.2ms, but also achieves a diagnostic accuracy of 97.20% due to the effective removal of redundant noise parameters. These data fully verify the scientific validity and advancement of the "structure and parameter dual sparsity co-evolution" architecture of this invention in balancing high diagnostic accuracy with low edge computing overhead.

[0136] A distribution network fault diagnosis system based on a collaborative sparse graph network is provided. The system is used to implement a distribution network fault diagnosis method based on a dual sparse graph neural network. The system includes a cloud training unit and a field edge online prediction unit.

[0137] The cloud training unit is deployed in the cloud or the main station backend to obtain historical operation data of the distribution network, asynchronously execute sparse initialization of the graph neural network model, adaptive reconstruction and security verification of the graph topology, and dynamic pruning of hardware resources to obtain a noiseless reconstructed topology and a lightweight fault diagnosis model.

[0138] The field edge online prediction unit is deployed in the field edge terminal device. It is used to receive and solidify the noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit. Based on the real-time collected electrical sampling data, it uses the reconstructed topology to perform sparse forward inference and outputs the distribution network node fault state classification results to assist decision-making.

[0139] The cloud-based training unit includes: a data acquisition and feature construction module, a sparse initialization module, a topology adaptive reconstruction module, and a dynamic sparse training module.

[0140] The data acquisition and feature construction module is used to acquire multi-dimensional time-series electrical operation data and historical remote signaling switch status of the distribution network, construct the initial physical diagram topology and node feature matrix, and pass the results to subsequent modules;

[0141] The sparse initialization module is used to assign initial parameters to the graph neural network model using a sparse initialization strategy based on the network dimension, and the results are passed to the training process.

[0142] The topology adaptive reconstruction module is used to calculate composite physical similarity, combine a hybrid candidate sampling strategy to prune and grow the topology, perform loop removal verification, and output the reconstructed topology.

[0143] The dynamic sparse training module is used to read the real-time memory usage status of the target edge terminal and dynamically adjust the pruning rate to remove redundant parameters of the graph neural network model;

[0144] The on-site edge online prediction unit includes: a noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit, a real-time electrical data receiving module, a sparse forward inference calculation module, and a fault classification and location output module.

[0145] The noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit are used to communicate and interact with the cloud training unit. The noiseless reconstructed topology and lightweight fault diagnosis model that has completed dual sparse training is received and locally solidified, providing a static network structure and parameter benchmark for real-time inference at the field edge.

[0146] The real-time electrical data receiving module is used to acquire real-time electrical sampling data from each field measurement node during the synchronous real-time phase when a fault occurs in the distribution network.

[0147] The sparse forward inference computation module is used to call the noiseless reconstructed topology and lightweight fault diagnosis model issued and solidified by the cloud training unit, and to perform sparse matrix forward propagation computation on the real-time electrical sampling data using the noiseless reconstructed topology as a sparse adjacency matrix.

[0148] The fault classification and location output module is used to output the state classification probability distribution of faults at each node of the distribution network based on the results of the sparse forward inference calculation, so as to assist the underlying equipment in making local fault isolation decisions.

[0149] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A method for fault diagnosis of distribution networks based on cooperative sparse graph networks, characterized in that, The method includes the following steps: Step 1: Obtain multi-dimensional time-series operation status data of the power distribution network and construct the initial physical graph topology and node feature matrix; Step 2: Construct a graph neural network model. Use a sparse initialization strategy based on ERK distribution to assign binary masks to the weights of the graph neural network model, thus completing the sparse initialization of the graph neural network model. Step 3: Using a graph encoder in a graph neural network, features are extracted from the node feature matrix to obtain the hidden layer feature vectors of each node. A composite physical similarity model is constructed by combining node feature correlations with the underlying electrical admittance information of the distribution network. Based on this composite physical similarity model and a hybrid candidate sampling strategy, the physical graph topology is pruned and grown. Loop detection is performed on the pruned and grown graph topology. If a closed loop exists, redundant branches are disconnected according to preset physical topology constraint rules, generating a noiseless reconstructed topology that conforms to radial operation constraints. ; Step 4: Update the parameters of the graph neural network model based on the noiseless reconstructed topology, and simultaneously acquire the hardware resource status of the target edge computing gateway in real time. Dynamically adjust the target sparse density of the graph neural network weights according to the hardware resource status, and use a model pruning strategy to remove redundant weights whose absolute values ​​are lower than a preset threshold. Alternately iterate through steps 3 and 4 to achieve coordinated sparsification optimization of the graph topology and model weight parameters until the graph neural network model converges. Use the converged graph neural network model as the final output lightweight fault diagnosis model. Step 5: Deploy the noiseless reconstructed topology and the lightweight fault diagnosis model on the edge computing gateway, receive real-time electrical data, perform forward inference, and output fault diagnosis results.

2. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 1, characterized in that, The specific operation of step 1 is as follows: Step 1.1: Extract the region using a high-frequency sampling rate via a synchronous phasor measurement device or intelligent power distribution terminal. Each measurement node within a specific time window The multidimensional electrical operation characteristics within; Step 1.2: The extracted feature dimensions are Includes: three-phase voltage amplitude Three-phase current amplitude Active power reactive power ; Step 1.3: Concatenate the above features to form a node feature matrix. ; Step 1.4: Extract the remote signaling switch status between each node from the power distribution management system; if the system displays the node status... With nodes If the switch is in a physically closed state, then the initial connection exists, and the initial physical topology is constructed accordingly. ,in, For a set of nodes, This is the initial set of physical edges.

3. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 2, characterized in that, In step 2, a sparse initialization strategy based on ERK distribution is used to perform binary mask allocation on the weights of the graph neural network model, preserving the parameter density. The calculation formula is: ; in, The first character of the graph neural network is represented by the second character. layer, Representing the The density of parameters retained by the layer. Indicates the first The proportion of the layer that is cut off; This represents the number of input feature channels for this layer. This represents the number of output feature channels; and These represent the spatial dimensions of the convolution kernel. It indicates a direct proportional relationship.

4. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 3, characterized in that, The specific operation of step 3 is as follows: Step 3.1: Update the parameters of the graph neural network model based on the noiseless topology reconstruction, and simultaneously extract arbitrary nodes using the graph encoder in the graph neural network. and Hidden layer feature vectors and A composite physical similarity model is constructed, and the calculation formula for the composite physical similarity model is as follows: ; in, Represents a node With nodes The composite physical similarity between them; The temporal Pearson correlation coefficient represents the features of the hidden layer of two nodes; Represents the node admittance matrix of the distribution network. With nodes The absolute value of the admittance element between them serves as a physical prior mask to restrict non-physical connectivity associations; and These are the normalized balance coefficients; Step 3.2: Generate candidate edge set based on hybrid candidate sampling strategy The formula is: ; in, This represents the search set for the topological boundaries of the local feeder area. Represents a cross-regional targeted harmonic sampling set; Set edge retention threshold Remove composite physical similarities from the initial physical graph topology. Below the set edge retention threshold The communication pseudo-edges are identified, and the candidate edges are grouped and ranked by composite physical similarity. New edges are added to the physical graph topology, and the graph topology is reconnected in linear complexity; Step 3.3: Perform loop detection on the reconnected graph topology, and call the depth-first search algorithm to check whether a closed loop has been formed in the reconstructed graph; if a closed loop exists, disconnect redundant branches according to the preset physical graph topology constraint rules to generate a noiseless reconstructed topology that meets the radial operation constraints. .

5. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 4, characterized in that, Step 3.3 involves a closed loop. Redundant branches are disconnected according to preset physical graph topology constraint rules to generate a noiseless reconstructed topology that conforms to radial operation constraints. The specific method is as follows: Scenario 1: The closed loop is entirely composed of the initial physical edges: If the depth-first search algorithm detects a cycle, and all edges in the cycle belong to the initial physical edge set. When the power grid itself has a special case of physical loop operation, the composite physical similarity score of all edges on the loop is retrieved, the physical edge with the lowest composite physical similarity score is selected, and it is temporarily disconnected in the graph structure, forcing the graph neural network to transmit fault features in a radial manner. Scenario 2: The closed loop is composed of a mixture of "initial physical edges" and "newly added implicit edges": If the loop contains both the original physical connections of the power distribution management system and the implicit electrical edges added by the graph neural network model through adaptive reconstruction, the composite similarity weights of all edges in the loop are compared. If the weight of the newly added implicit edge is greater than that of a certain physical edge, the newly added implicit edge is retained and the physical edge with the lowest score is disconnected. If the weight of the newly added implicit edge is less than that of all physical edges, the newly added implicit edge is directly removed as a redundant branch. Scenario 3: The closed loop is entirely composed of "newly added hidden edges": If the loop is entirely composed of newly added implicit electrical edges by the graph neural network model through adaptive reconstruction, it is determined that the graph neural network model has over-captured common-mode interference features in the local area. Strictly following the descending order of the maximum spanning tree, the newly added edge with the lowest score in the local loop is removed until the local subgraph is restored to a radial tree structure. After the above-mentioned pruning, growth, and radial physical safety checks, the final output graph structure, which has removed structural noise and strictly meets the safety constraints of the distribution network, is the noise-free reconstructed topology. .

6. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 5, characterized in that, In step 4, the target sparse density of the graph neural network weights is dynamically adjusted based on the hardware resource status. The specific formula for calculating the target sparse density is as follows: ; in, For the current training round, This is the starting round for dynamic pruning. This is the final round of dynamic pruning; For the initial parameter density, The final parameter density of the baseline target; This represents the real-time available memory occupancy of the edge computing gateway; This represents the memory security threshold limit for edge computing gateways; As a regulating factor; This is a hardware resource penalty function for edge computing gateways.

7. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 6, characterized in that, Step 4, during the training process of dynamically adjusting the target sparse density of the graph neural network weights, describes the target sparse density... Based on the real-time available memory occupancy of the edge computing gateway Adaptive adjustments are made; the resource penalty function and parameter pruning are handled as follows, taking into account different hardware load conditions of the target edge computing gateway: Scenario 1: Sufficient hardware resources: If the available memory occupancy is read in real time Less than the set memory safety threshold limit At that time, the hardware resource penalty function of the edge computing gateway The value approaches 1, and the graph neural network model achieves its final parameter density according to the preset baseline target. Perform normal linear decay pruning; Scenario 2: Hardware resources are critically strained: If the available memory occupancy is read in real time Approaching or equal to the memory safety threshold At that time, the hardware resource penalty function of the edge computing gateway Activated and its value rapidly decreases, forcibly reducing the target's sparse density. The amplitude pruning is accelerated immediately in the current round of the graph neural network model. Scenario 3: Hardware resources are severely overloaded / Other high-level tasks suddenly occur: If the available memory occupancy is read in real time Exceeding the safety threshold indicates that the edge computing gateway is currently executing other high-priority protection trip commands. At this point, the graph neural network model immediately suspends the parameter iteration and topology update operations of this round, directly locks the weights and topology of the graph neural network model that converged in the previous round, releases the underlying hardware computing resources of the edge computing gateway, and ensures the absolute priority and security of the physical relay protection actions at the lowest level of the power grid.

8. The method for fault diagnosis of distribution networks based on cooperative sparse graph networks according to claim 7, characterized in that, The specific operation of step 5 is as follows: Step 5.1: Combine the lightweight fault diagnosis model and noiseless reconstructed topology optimized by the joint and alternating optimization in Steps 3 and 4. This information is then fixed and distributed to the edge computing gateways deployed at the power distribution network site. Step 5.2: During the data plane synchronization phase when a fault occurs, the edge computing gateway directly calls the already established reconstructed graph topology. As a sparse adjacency matrix, real-time electrical fault features are extracted, and a lightweight fault diagnosis model is used to perform sparse matrix forward inference to output fault diagnosis results.

9. A distribution network fault diagnosis system based on a cooperative sparse graph network, the system being used to implement the distribution network fault diagnosis method based on a cooperative sparse graph network as described in claims 1-8, characterized in that, The system includes: a cloud-based training unit and a field-based edge online prediction unit; The cloud training unit is deployed in the cloud or the main station backend to obtain historical operation data of the distribution network, asynchronously execute sparse initialization of the graph neural network model, adaptive reconstruction and security verification of the graph topology, and dynamic pruning of hardware resources to obtain a noiseless reconstructed topology and a lightweight fault diagnosis model. The field edge online prediction unit is deployed in the field edge terminal device. It is used to receive and solidify the noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit. Based on the real-time collected electrical sampling data, it uses the reconstructed topology to perform sparse forward inference and outputs the distribution network node fault state classification results to assist decision-making. The cloud-based training unit includes: a data acquisition and feature construction module, a sparse initialization module, a topology adaptive reconstruction module, and a dynamic sparse training module. The data acquisition and feature construction module is used to acquire multi-dimensional time-series electrical operation data and historical remote signaling switch status of the distribution network, construct the initial physical diagram topology and node feature matrix, and pass the results to subsequent modules; The sparse initialization module is used to assign initial parameters to the graph neural network model using a sparse initialization strategy based on the network dimension, and the results are passed to the training process. The topology adaptive reconstruction module is used to calculate composite physical similarity, combine a hybrid candidate sampling strategy to prune and grow the topology, perform loop removal verification, and output the reconstructed topology. The dynamic sparse training module is used to read the real-time memory usage status of the target edge terminal and dynamically adjust the pruning rate to remove redundant parameters of the graph neural network model; The on-site edge online prediction unit includes: a noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit, a real-time electrical data receiving module, a sparse forward inference calculation module, and a fault classification and location output module. The noiseless reconstructed topology and lightweight fault diagnosis model issued by the cloud training unit are used to communicate and interact with the cloud training unit. The noiseless reconstructed topology and lightweight fault diagnosis model that has completed dual sparse training is received and locally solidified, providing a static network structure and parameter benchmark for real-time inference at the field edge. The real-time electrical data receiving module is used to acquire real-time electrical sampling data from each field measurement node during the synchronous real-time phase when a fault occurs in the distribution network. The sparse forward inference computation module is used to call the noiseless reconstructed topology and lightweight fault diagnosis model issued and solidified by the cloud training unit, and to perform sparse matrix forward propagation computation on the real-time electrical sampling data using the noiseless reconstructed topology as a sparse adjacency matrix. The fault classification and location output module is used to output the state classification probability distribution of faults at each node of the distribution network based on the results of the sparse forward inference calculation, so as to assist the underlying equipment in making local fault isolation decisions.

Images