A recommended method, system and medium for handling power distribution network faults

By constructing a multi-graph topology structure and a deep learning model, the shortcomings of the distribution network fault handling recommendation algorithm in topology modeling and text semantic mining are solved. This enables efficient fault handling recommendation under sparse and unbalanced data, thereby improving the intelligence level of distribution network operation and maintenance.

CN123068985APending Publication Date: 2026-11-10STATE GRID JIANGSU ELECTRIC POWER CO LTD RESEARCH INSTITUTE +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611273953.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing recommendation algorithms for fault handling in power distribution networks suffer from limitations such as simplistic topology modeling, insufficient text semantic mining, poor adaptability to sparse data, and inadequate training performance with imbalanced samples, making them unsuitable for the needs of complex power grid scenarios.

Method used

We construct a fault-disposal bipartite graph, a fault-fault collaboration graph, and a disposal-disposal collaboration graph that integrate multiple graph topologies. We use the Sentence-BERT pre-trained language model to extract initial semantic embedding features, combine graph convolutional networks and bidirectional graph attention networks for feature propagation and aggregation, use multi-head attention mechanism for feature fusion, and calculate and rank the matching probability of candidate disposal measures.

Benefits of technology

By comprehensively exploring the hidden common patterns of fault occurrence and the collaborative logic of handling procedures, the accuracy and robustness of fault handling recommendations are improved, adapting to sparse and unbalanced data, and outputting the optimal solution that fits the on-site operation and maintenance process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068985A_ABST
    Figure CN123068985A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, and medium for recommending fault handling measures in power distribution networks, belonging to the field of intelligent operation and maintenance technology for power systems. The method includes: constructing a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph; using a pre-trained language model to complete semantic encoding of fault text and handling text, obtaining initial semantic embeddings of entities; using a graph convolutional network with residual connections to propagate homogeneous node features in the fault-fault collaboration graph and the handling-handling collaboration graph, and using a bidirectional graph attention network to adaptively allocate neighbor node weights to aggregate heterogeneous node features in the bipartite heterogeneous graph; introducing a multi-head attention mechanism to adaptively fuse heterogeneous features from multiple graphs to obtain a unified entity feature representation; pooling and aggregating the fault set features, calculating the matching probability of each handling measure using a bilinear matching function, and outputting a fault handling recommendation scheme sorted in descending order of matching probabilities, thereby achieving accurate recommendation of fault handling measures in power distribution networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of power systems, and specifically relates to a recommended method, system and medium for handling distribution network faults. Background Technology

[0002] Efficient and accurate decision-making in distribution network fault handling can quickly locate grid faults, match the optimal repair plan, effectively shorten line outage time, and comprehensively improve the power supply stability of the regional power grid. At present, there are still many shortcomings in distribution network operation and maintenance. Traditional fault handling plans rely heavily on the on-site work experience of front-line operation and maintenance personnel and fixed expert rule bases. When faced with complex grid scenarios of multiple concurrent faults and coupled compound faults, problems such as incorrect matching of handling plans, chaotic repair steps, and omission of key handling procedures are prone to occur, resulting in insufficient intelligence.

[0003] Current intelligent recommendation algorithms for distribution network faults still suffer from three major technical shortcomings: First, power maintenance work orders are characterized by dense technical terminology and complex entity relationships, making it difficult for traditional coding methods to uncover deep semantic information in fault and handling texts, resulting in poor entity representation. Second, historical work order data for power grid maintenance is extremely sparse, and a single graph structure cannot simultaneously model cross-entity relationships between faults and handling, common-cause relationships within faults, and collaborative relationships within handling, leading to incomplete topological relationship modeling. Third, distribution network faults are typical multi-label classification scenarios with extremely uneven fault sample distribution, making it impossible for conventional loss functions to account for sparse fault samples. Model training tends to favor common faults, resulting in extremely poor recommendation performance for rare faults. Therefore, existing technologies cannot adapt to the characteristics of real distribution network maintenance data and the needs of complex fault scenarios. Summary of the Invention

[0004] This invention provides a method, system, and medium for recommending fault handling in power distribution networks. It aims to extract initial semantic embedding features of fault entities and handling entities by constructing a multi-graph topology structure, extracting isomorphic and heterogeneous graph features respectively, and adaptively fusing them to obtain fault node features and handling node features. The matching probability of each candidate handling measure is calculated, and a fault handling recommendation scheme sorted in descending order of matching probability is output. This addresses the problems of existing power distribution network fault handling recommendation algorithms, such as single topology modeling, insufficient text semantic mining, poor adaptability to sparse data, and poor training effect of imbalanced samples.

[0005] To solve the above problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides a recommended method for handling power distribution network faults, comprising:

[0007] Obtain power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability;

[0008] The data processing method for the fault handling recommendation model includes:

[0009] A topology graph structure is constructed based on historical maintenance records of the power distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph.

[0010] A pre-trained language model was used to semantically encode the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph, and to extract the initial semantic embedding features of the fault entities and the handling entities.

[0011] Based on the initial semantic embedding features, a graph convolutional network with residual connections is used to propagate the homogeneous node features of the fault-fault collaboration graph and the disposal-disposal collaboration graph to obtain heterogeneous features of the fault graph and the disposal graph. A bidirectional graph attention network is used to adaptively allocate the weights of neighbor nodes to aggregate the heterogeneous node features of the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph.

[0012] A multi-head attention mechanism is set up to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain the fault node features and the handling node features.

[0013] Based on the characteristics of the fault node and the handling node, the matching probability of each candidate handling measure is calculated, and the recommended fault handling scheme is output after the matching probability is sorted in descending order.

[0014] Furthermore, the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph are used to respectively characterize the direct association between faults and handling, the common cause association between faults, and the collaborative association between handling measures. The edge sets corresponding to the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph are respectively represented as follows:

[0015] ;

[0016] ;

[0017] ;

[0018] In the formula, For the edge set of the fault-handling bipartite graph, For faulty entities, In order to dispose of the entity, For the first The set of faults corresponding to each historical maintenance record. For the first The set of actions corresponding to each historical maintenance record. This is a collection of historical maintenance records for the power distribution network. For the edge set of the fault-fault collaboration graph, , The first The fault entity corresponding to the historical maintenance record, the first The fault entity corresponding to each historical maintenance record For the set of faulty entities, In order to handle the collection of entities, The threshold for co-occurrence of faulty entities. For the edge set of the disposal-disposal collaboration graph, , The first The corresponding handling entity for each historical maintenance record, the first The entity responsible for handling each historical maintenance record. To handle entity co-occurrence thresholds, This is a conditional statement symbol, indicating "existence".

[0019] Furthermore, the pre-trained language model is used to perform semantic encoding on the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph, and to extract the initial semantic embedding features of the fault entity and the handling entity, including:

[0020] Based on the Sentence-BERT pre-trained language model, semantic encoding is performed on the fault-handling bipartite graph, fault-fault collaboration graph, and handling-handling collaboration graph, and deep semantic information of the text is automatically mined to obtain the initial embedding features of fault entities and handling entities.

[0021] Furthermore, before the step of using a graph convolutional network with residual connections to propagate homogeneous node features to the fault-fault collaboration graph and the disposal-disposal collaboration graph based on the initial semantic embedding features to obtain heterogeneous features of the fault graph and the disposal graph, and using a bidirectional graph attention network to adaptively allocate neighbor node weights to aggregate heterogeneous node features to obtain heterogeneous features of the bipartite graph, the following steps are also included:

[0022] Based on linear transformation, the feature dimensions of fault entities and handling entities in fault-handling bipartite graphs, fault-fault collaboration graphs, and handling-handling collaboration graphs are unified using nonlinear activation functions.

[0023] Furthermore, the calculation formula for the aggregation of heterogeneous node features is expressed as follows:

[0024] ;

[0025] In the formula, For the first Layer The feature vector obtained by aggregating heterogeneous nodes It is a non-linear activation function. For the first The set of heterogeneous nodes containing each heterogeneous node. For the first The heterogeneous node and the first Adaptive attention weights among neighboring nodes. For the first The learnable weight matrix of the layer, For the first The first layer The feature vectors of neighboring nodes.

[0026] Furthermore, the multi-head attention mechanism adaptively fuses the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain fault node features and handling node features, including:

[0027] A multi-head attention mechanism is adopted to construct the first query vector by adding the heterogeneous features of the fault map and the heterogeneous features of the bipartite map element by element. Then, the first key vector and the first value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the first query vector, the first key vector and the first value vector, the attention weight is calculated in each independent semantic subspace. The outputs of the multi-head attention in each independent semantic subspace are concatenated and then subjected to linear projection dimensionality reduction to obtain the fault node features.

[0028] A multi-head attention mechanism is adopted. The heterogeneous features of the processing graph and the heterogeneous features of the bipartite graph are added element by element to construct a second query vector. Then, the second key vector and the second value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the second query vector, the second key vector and the second value vector, the attention weight is calculated in each independent semantic subspace. The outputs of the multi-head attention in each independent semantic subspace are concatenated and then subjected to linear projection dimensionality reduction to obtain the processing node features.

[0029] Furthermore, the step of calculating the matching probability of each candidate handling measure based on the characteristics of the fault node and the handling node, and outputting the fault handling recommendation scheme after sorting the matching probabilities in descending order, includes:

[0030] Based on the fault node features and the handling node features, the global features under multiple fault scenarios are aggregated by summation pooling to obtain a comprehensive representation vector of the fault set.

[0031] The comprehensive characterization vector and the characterization vectors of each candidate treatment measure are input into a bilinear matching function to calculate the bidirectional correlation score between the fault set and each treatment measure. After Softmax normalization, the matching probability of each candidate treatment measure is obtained.

[0032] The candidates are sorted in descending order according to their matching probabilities, and the candidate handling measures with the highest matching probability are selected as the optimal fault handling recommendation scheme.

[0033] Wherein, the bilinear matching function is pre-adopted by... Sorting loss and The joint loss function of the loss combination is obtained through supervised training.

[0034] Furthermore, the joint loss function is expressed as:

[0035] ;

[0036] In the formula, The total loss of the joint loss function, for loss and Ranking loss The balancing hyperparameters.

[0037] Secondly, the present invention provides a power distribution network fault handling recommendation system, comprising:

[0038] The data acquisition module is used to acquire power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability.

[0039] The data processing method of the fault handling recommendation model is implemented by the following modules:

[0040] The topology graph construction module is used to construct a topology graph structure based on historical maintenance record data of the distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph.

[0041] The semantic embedding module is used to perform semantic encoding on the fault-disposal bipartite graph, the fault-fault collaboration graph, and the disposal-disposal collaboration graph using a pre-trained language model, and to extract the initial semantic embedding features of fault entities and disposal entities.

[0042] The heterogeneous feature extraction module is used to obtain heterogeneous features of the fault map and the disposal map by using a graph convolutional network with residual connections to propagate homogeneous node features on the fault-fault collaboration graph and the disposal-disposal collaboration graph based on the initial semantic embedding features. It also uses a bidirectional graph attention network to adaptively allocate the weights of neighbor nodes to complete the aggregation of heterogeneous node features on the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph.

[0043] The multi-graph feature fusion module is used to set up a multi-head attention mechanism to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the disposal map and the bipartite map, to obtain the fault node features and the disposal node features.

[0044] The disposal recommendation module is used to calculate the matching probability of each candidate disposal measure based on the characteristics of the fault node and the disposal node, and output the fault disposal recommendation scheme after the matching probability is sorted in descending order.

[0045] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the recommended method for handling power distribution network faults as described in the first aspect.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention constructs a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph, which fully covers the isomorphic collaborative relationships and heterogeneous bipartite relationships between fault entities and handling entities throughout the entire process of power distribution network operation and maintenance. The synergistic and complementary structure of the three types of graphs effectively alleviates the problem of insufficient feature representation caused by the sparsity of operation and maintenance work order data. Compared with a single topology structure, it can comprehensively explore the implicit co-occurrence patterns of faults and the collaborative logic of handling procedures. It enriches the structured semantic information of faults and handling links at the graph structure level, and provides complete graph structure support for pre-trained language models to extract initial semantic embedding features, graph convolutional networks with residual connections and bidirectional graph attention networks to extract heterogeneous features of multiple graphs, and adaptive fusion of multi-head attention mechanisms to obtain fault node features and handling node features.

[0048] 2. This invention constructs three complementary topological graphs: a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph. These graphs respectively characterize the direct association between faults and handling, the common cause association between faults, and the collaborative association between handling measures. Based on co-occurrence thresholds, three types of edge sets are defined. This invention comprehensively covers the isomorphic collaborative association and heterogeneous bipartite association between fault entities and handling entities throughout the entire process of power distribution network operation and maintenance at the topological level. This effectively alleviates the problem of insufficient feature representation caused by the sparsity of operation and maintenance work order data, and provides complete graph structure support for pre-trained language models to extract initial semantic embedding features and for subsequent multi-graph feature fusion.

[0049] 3. This invention employs a Sentence-BERT pre-trained language model to semantically encode three types of topological graphs and extract initial semantic embedding features of fault entities and handling entities. Based on this, a bidirectional graph attention network is used to aggregate heterogeneous node features in the fault-handling bipartite graph. A graph convolutional network with residual connections is used to propagate isomorphic node features in the fault-fault collaboration graph and the handling-handling collaboration graph. Then, a multi-head attention mechanism is used to adaptively fuse the heterogeneous features of the fault graph with the heterogeneous features of the bipartite graph, and the heterogeneous features of the handling graph with the heterogeneous features of the bipartite graph, respectively, to obtain fault node features and handling node features. This fully explores the deep semantic information of fault text and handling text, and overcomes the impact of incomplete information in a single graph structure and the difficulty in effectively fusing heterogeneous features on recommendation accuracy.

[0050] 4. This invention employs a joint loss function combining ranking loss and cross-entropy loss during the model training phase to supervise the training of the bilinear matching function. Softmax normalization is used to calculate the matching probability of each candidate handling measure, and the optimal fault handling recommendation is selected and output based on the descending order of matching probabilities. By utilizing the synergistic constraint of ranking loss and cross-entropy loss, the dominant role of easily classified negative samples in gradient updates is reduced, allowing model training to focus on difficult-to-classify samples, effectively improving the accuracy and robustness of fault handling recommendations under imbalanced sample conditions. Simultaneously, ranking loss optimizes the output order of handling measures, ensuring that the recommendation results align with the emergency repair workflow of on-site maintenance personnel. The model exhibits stronger generalization ability under real sparse and imbalanced power grid data, demonstrating practical engineering value. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the data processing flow of the fault handling recommendation model provided in this embodiment of the invention;

[0052] Figure 2 This is a schematic diagram of the technical architecture of the recommended method for handling power distribution network faults provided in the embodiments of the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0054] Example 1

[0055] like Figure 1 As shown in the figure, this embodiment introduces a recommended method for handling distribution network faults, including:

[0056] Obtain power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability;

[0057] The data processing method for the fault handling recommendation model includes:

[0058] Step 1: Construct a topology graph structure based on the historical maintenance records of the power distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph.

[0059] This embodiment constructs a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph to depict the direct correlation between faults and handling, the common cause correlation between faults, and the collaborative correlation between handling measures, respectively. It fully covers the isomorphic collaborative correlation and heterogeneous bipartite correlation between fault entities and handling entities in the entire process of power distribution network operation and maintenance. Relying on the collaborative structure of the three types of complementary topological graphs, it effectively alleviates the problem of insufficient feature representation caused by the sparsity of operation and maintenance work order data, and provides complete graph structure support for the extraction of initial semantic embedding features by the pre-trained language model and the subsequent fusion of multi-graph features. It overcomes the defect that a single topological structure is difficult to fully explore the implicit co-occurrence rules of faults and the collaborative logic of handling procedures.

[0060] Step 2: Using a pre-trained language model, semantic encoding is performed on the fault-disposal bipartite graph, the fault-fault collaboration graph, and the disposal-disposal collaboration graph, and the initial semantic embedding features of the fault entities and disposal entities are extracted.

[0061] This embodiment employs the Sentence-BERT pre-trained language model to semantically encode the fault-disposal bipartite graph, fault-fault collaboration graph, and disposal-disposal collaboration graph. This automatically mines the deep semantic information of the fault text and disposal text, and extracts the initial semantic embedding features of the fault entity and disposal entity. Compared with traditional static word vector methods such as bag-of-words or Word2Vec, Sentence-BERT can fully capture the contextual semantic relationships in the fault description and disposal description, improve the representational ability of the initial embedding features, and provide high-quality semantic input for the feature propagation and aggregation of subsequent graph convolutional networks and bidirectional graph attention networks.

[0062] Step 3: Based on the initial semantic embedding features, a graph convolutional network with residual connections is used to propagate the homogeneous node features of the fault-fault collaboration graph and the disposal-disposal collaboration graph to obtain heterogeneous features of the fault graph and the disposal graph. A bidirectional graph attention network is used to adaptively allocate the weights of neighbor nodes to aggregate the heterogeneous node features of the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph.

[0063] This implementation uses a graph convolutional network with residual connections to propagate isomorphic node features to the fault-fault collaboration graph and the handling-handling collaboration graph based on the initial semantic embedding features. The residual connections are used to alleviate the gradient vanishing problem in deep network training and fully aggregate the collaborative association information between entities of the same type to obtain heterogeneous features of the fault graph and the handling graph. At the same time, a bidirectional graph attention network is used to adaptively allocate the weights of neighbor nodes to complete the aggregation of heterogeneous node features in the fault-handling bipartite graph. The bidirectional graph attention network can adaptively learn the bidirectional interaction weights between fault nodes and handling nodes through a self-attention mechanism without the need for preset neighborhood aggregation rules. Even under the condition of sparse operation and maintenance work order data, it can still effectively aggregate heterogeneous neighborhood information to obtain heterogeneous features of the bipartite graph.

[0064] Step 4: Set up a multi-head attention mechanism to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain the fault node features and the handling node features.

[0065] This embodiment employs a multi-head attention mechanism to adaptively fuse heterogeneous features of the fault map and bipartite map, as well as heterogeneous features of the treatment map and bipartite map. The query vector is constructed by adding the heterogeneous features of the fault map and bipartite map element-wise, and the key and value vectors are concatenated. Multiple independent semantic subspaces are then divided to calculate attention weights for each, and the fault node features are obtained through concatenation and linear projection dimensionality reduction. Similarly, the treatment node features are obtained. This multi-head attention mechanism captures the correlation weights between heterogeneous features of the fault and treatment sides from different semantic perspectives through multiple independent semantic subspaces, adaptively determining the contribution of each map feature in the fusion process. This overcomes the technical challenge of directly fusing heterogeneous map features and improves the expressive richness and discriminative power of the fused features.

[0066] Step 5: Based on the characteristics of the fault node and the handling node, calculate the matching probability of each candidate handling measure, and output the recommended fault handling scheme after sorting the matching probabilities in descending order.

[0067] This embodiment uses the features of fault nodes and treatment nodes to aggregate global features under multiple fault scenarios through summation pooling to obtain a comprehensive representation vector of the fault set. The comprehensive representation vector and the representation vector of each candidate treatment measure are input into a bilinear matching function to calculate the bidirectional association score. After Softmax normalization, the matching probability of each candidate treatment measure is obtained. The optimal fault treatment recommendation scheme is selected and output according to the descending order of the matching probability. The bilinear matching function fully captures the bidirectional semantic association between the fault set and the treatment measures through bilinear transformation. The joint loss function of ranking loss and cross-entropy is used to coordinately constrain the optimization direction of the model during the training phase, so that the recommendation result is close to the emergency repair workflow of on-site operation and maintenance personnel in terms of matching accuracy and output order, effectively improving the accuracy and robustness of fault treatment recommendation under imbalanced sample conditions.

[0068] Example 2

[0069] Based on the same inventive concept as Embodiment 1, this embodiment describes the implementation steps of a recommended method for handling distribution network faults, such as... Figure 2 As shown, it includes:

[0070] Obtain power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability.

[0071] The data processing method for the fault handling recommendation model includes:

[0072] Step 1: Construct a topology diagram based on historical maintenance records of the power distribution network.

[0073] In this embodiment, the topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph. The fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph are used to respectively depict the direct relationship between faults and handling, the common cause relationship between faults, and the collaborative relationship between handling measures.

[0074] The fault-handling bipartite graph uses fault entities and handling entities as heterogeneous nodes. Connection edges are established between fault nodes and handling nodes that co-occur within the same historical maintenance record data of the distribution network, resulting in the edge set of the fault-handling bipartite graph, represented as:

[0075] ;

[0076] The fault-fault collaboration graph uses only faulty entities as network nodes. When the number of times two sets of faulty entities co-occur reaches a preset threshold, a connection edge is established, resulting in the edge set of the fault-fault collaboration graph, represented as:

[0077] ;

[0078] The disposal-disposal collaboration graph uses only disposal entities as network nodes. When the number of times two disposal measures occur together reaches a preset threshold, a connection edge is established to obtain the edge set of the disposal-disposal collaboration graph, represented as:

[0079] ;

[0080] In the formula, For the edge set of the fault-handling bipartite graph, For faulty entities, In order to dispose of the entity, For the first The set of faults corresponding to each historical maintenance record. For the first The set of actions corresponding to each historical maintenance record. This is a collection of historical maintenance records for the power distribution network. For the edge set of the fault-fault collaboration graph, , The first The fault entity corresponding to the historical maintenance record, the first The fault entity corresponding to each historical maintenance record For the set of faulty entities, In order to handle the collection of entities, The threshold for co-occurrence of faulty entities. For the edge set of the disposal-disposal collaboration graph, , The first The corresponding handling entity for each historical maintenance record, the first The entity responsible for handling each historical maintenance record. To handle entity co-occurrence thresholds, This is a conditional statement symbol, indicating "existence".

[0081] This embodiment uses the fault entity co-occurrence threshold. Co-occurrence threshold with disposal entity All are set to 3.

[0082] Step 2: Using a pre-trained language model, semantic encoding is performed on the fault-disposal bipartite graph, the fault-fault collaboration graph, and the disposal-disposal collaboration graph, and the initial semantic embedding features of the fault entity and the disposal entity are extracted.

[0083] This embodiment uses the Sentence-BERT pre-trained language model to perform semantic encoding on the fault-disposal bipartite graph, fault-fault collaboration graph, and disposal-disposal collaboration graph, and automatically mines deep semantic information of the text to obtain the initial embedding features of fault entities and disposal entities. The specific network structure and training process of the Sentence-BERT pre-trained language model can adopt known schemes in the field, and this invention does not limit them.

[0084] Step 3: Based on the linear transformation, unify the feature dimensions of fault entities and handling entities in the fault-handling bipartite graph, fault-fault collaboration graph, and handling-handling collaboration graph using a nonlinear activation function.

[0085] Step 4: Based on the initial semantic embedding features, a graph convolutional network with residual connections is used to propagate the homogeneous node features of the fault-fault collaboration graph and the disposal-disposal collaboration graph to obtain heterogeneous features of the fault graph and the disposal graph. A bidirectional graph attention network is used to adaptively allocate the weights of neighbor nodes to aggregate the heterogeneous node features of the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph.

[0086] In this embodiment, the calculation formula for the aggregation of heterogeneous node features is expressed as follows:

[0087] ;

[0088] In the formula, For the first Layer The feature vector obtained by aggregating heterogeneous nodes It is a non-linear activation function. For the first The set of heterogeneous nodes containing each heterogeneous node. For the first The heterogeneous node and the first Adaptive attention weights among neighboring nodes. For the first The learnable weight matrix of the layer, For the first The first layer The feature vectors of neighboring nodes.

[0089] Step 4: Set up a multi-head attention mechanism to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain the fault node features and the handling node features.

[0090] Step 4.1: Using a multi-head attention mechanism, the heterogeneous features of the fault map and the heterogeneous features of the bipartite map are added element by element to construct the first query vector. Then, the first key vector and the first value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the first query vector, the first key vector, and the first value vector, the attention weights are calculated in each independent semantic subspace. The outputs of each independent semantic subspace of the multi-head attention are concatenated and then subjected to linear projection dimensionality reduction to obtain the fault node features.

[0091] Step 4.2: Using a multi-head attention mechanism, the heterogeneous features of the processing graph and the heterogeneous features of the bipartite graph are added element by element to construct a second query vector. Then, the second key vector and the second value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the second query vector, the second key vector, and the second value vector, the attention weights are calculated in each independent semantic subspace. The outputs of each independent semantic subspace of the multi-head attention are concatenated and then subjected to linear projection dimensionality reduction to obtain the processing node features.

[0092] Step 5: Based on the characteristics of the fault node and the handling node, calculate the matching probability of each candidate handling measure, and output the fault handling recommendation scheme after the matching probability is sorted in descending order.

[0093] Step 5.1: Based on the fault node features and the handling node features, the global features under multiple fault scenarios are aggregated using a summation pooling method to obtain a comprehensive representation vector of the fault set;

[0094] Step 5.2: Input the comprehensive representation vector and the representation vector of each candidate treatment measure into the bilinear matching function to calculate the bidirectional correlation score between the fault set and each treatment measure. After Softmax normalization, the matching probability of each candidate treatment measure is obtained.

[0095] Step 5.3: Sort the candidates in descending order according to their matching probabilities, and select the candidate with the highest matching probability as the optimal fault handling recommendation solution.

[0096] In this embodiment, the bilinear matching function is pre-trained using a joint loss function consisting of a ranking loss and a loss function. The joint loss function is expressed as:

[0097] ;

[0098] In the formula, The total loss of the joint loss function, for loss and Ranking loss The balancing hyperparameters. This embodiment will... Set it to 0.025.

[0099] Example 3

[0100] Based on the same inventive concept as other embodiments, this embodiment introduces a power distribution network fault handling recommendation system, including:

[0101] The data acquisition module is used to acquire power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability.

[0102] The data processing method of the fault handling recommendation model is implemented by the following modules:

[0103] The topology graph construction module is used to construct a topology graph structure based on the historical maintenance records of the distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph.

[0104] The semantic embedding module is used to perform semantic encoding on the fault-disposal bipartite graph, the fault-fault collaboration graph, and the disposal-disposal collaboration graph using a pre-trained language model, and to extract the initial semantic embedding features of fault entities and disposal entities.

[0105] The heterogeneous feature extraction module is used to obtain heterogeneous features of the fault map and the disposal map by using a graph convolutional network with residual connections to propagate homogeneous node features on the fault-fault collaboration graph and the disposal-disposal collaboration graph based on the initial semantic embedding features. It also uses a bidirectional graph attention network to adaptively allocate the weights of neighbor nodes to complete the aggregation of heterogeneous node features on the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph.

[0106] The multi-graph feature fusion module is used to set up a multi-head attention mechanism to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the disposal map and the bipartite map, to obtain the fault node features and the disposal node features.

[0107] The disposal recommendation module is used to calculate the matching probability of each candidate disposal measure based on the characteristics of the fault node and the disposal node, and output the optimal fault disposal recommendation scheme.

[0108] The specific functions of each module described above are explained in the relevant content of the methods in Embodiment 1 or 2, and will not be repeated here.

[0109] Example 4

[0110] Based on the same inventive concept as other embodiments, this embodiment describes a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the methods of Embodiment 1 or 2 described above.

[0111] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

[0112] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0114] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A recommended method for handling power distribution network faults, characterized in that, include: Obtain power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability; The data processing method for the fault handling recommendation model includes: A topology graph structure is constructed based on historical maintenance records of the power distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph. A pre-trained language model was used to semantically encode the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph, and to extract the initial semantic embedding features of the fault entities and the handling entities. Based on the initial semantic embedding features, a graph convolutional network with residual connections is used to propagate the homogeneous node features of the fault-fault collaboration graph and the disposal-disposal collaboration graph to obtain heterogeneous features of the fault graph and the disposal graph. A bidirectional graph attention network is used to adaptively allocate the weights of neighbor nodes to aggregate the heterogeneous node features of the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph. A multi-head attention mechanism is set up to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain the fault node features and the handling node features. Based on the characteristics of the fault node and the handling node, the matching probability of each candidate handling measure is calculated, and the recommended fault handling scheme is output after the matching probability is sorted in descending order.

2. The recommended method for handling distribution network faults according to claim 1, characterized in that, The fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph are used to respectively characterize the direct correlation between faults and handling, the common cause correlation between faults, and the collaborative correlation between handling measures. The edge sets corresponding to the fault-handling bipartite graph, the fault-fault collaboration graph, and the handling-handling collaboration graph are respectively represented as follows: ; ; ; In the formula, For the edge set of the fault-handling bipartite graph, For faulty entities, In order to dispose of the entity, For the first The set of faults corresponding to each historical maintenance record. For the first The set of actions corresponding to each historical maintenance record. This is a collection of historical maintenance records for the power distribution network. For the edge set of the fault-fault collaboration graph, , The first The fault entity corresponding to the historical maintenance record, the first The fault entity corresponding to each historical maintenance record For the set of faulty entities, In order to handle the collection of entities, The threshold for co-occurrence of faulty entities. For the edge set of the disposal-disposal collaboration graph, , The first The corresponding handling entity for each historical maintenance record, the first The entity responsible for handling each historical maintenance record. To handle entity co-occurrence thresholds, This is a conditional statement symbol, indicating "existence".

3. The recommended method for handling distribution network faults according to claim 1, characterized in that, The process employs a pre-trained language model to perform semantic encoding on the fault-handling bipartite graph, fault-fault collaboration graph, and handling-handling collaboration graph, and extracts initial semantic embedding features of fault entities and handling entities, including: Based on the Sentence-BERT pre-trained language model, semantic encoding is performed on the fault-handling bipartite graph, fault-fault collaboration graph, and handling-handling collaboration graph, and deep semantic information of the text is automatically mined to obtain the initial embedding features of fault entities and handling entities.

4. The recommended method for handling distribution network faults according to claim 1, characterized in that, Before the step of using a graph convolutional network with residual connections to propagate isomorphic node features to the fault-fault collaboration graph and the disposal-disposal collaboration graph based on the initial semantic embedding features to obtain heterogeneous features of the fault graph and the disposal graph, and using a bidirectional graph attention network to adaptively allocate neighbor node weights to aggregate heterogeneous node features to obtain heterogeneous features of the bipartite graph, the following steps are also included: Based on linear transformation, the feature dimensions of fault entities and handling entities in fault-handling bipartite graphs, fault-fault collaboration graphs, and handling-handling collaboration graphs are unified using nonlinear activation functions.

5. The recommended method for handling distribution network faults according to claim 1, characterized in that, The calculation formula for the aggregation of heterogeneous node features is expressed as follows: ; In the formula, For the first Layer The feature vector obtained by aggregating heterogeneous nodes It is a non-linear activation function. For the first The set of heterogeneous nodes containing each heterogeneous node. For the first The heterogeneous node and the first Adaptive attention weights among neighboring nodes. For the first The learnable weight matrix of the layer, For the first The first layer The feature vectors of neighboring nodes.

6. The recommended method for handling distribution network faults according to claim 1, characterized in that, The multi-head attention mechanism adaptively fuses the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the handling map and the bipartite map, to obtain fault node features and handling node features, including: A multi-head attention mechanism is adopted to construct the first query vector by adding the heterogeneous features of the fault map and the heterogeneous features of the bipartite map element by element. Then, the first key vector and the first value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the first query vector, the first key vector and the first value vector, the attention weight is calculated in each independent semantic subspace. The outputs of the multi-head attention in each independent semantic subspace are concatenated and then subjected to linear projection dimensionality reduction to obtain the fault node features. A multi-head attention mechanism is adopted. The heterogeneous features of the processing graph and the heterogeneous features of the bipartite graph are added element by element to construct a second query vector. Then, the second key vector and the second value vector are constructed by concatenating them. Multiple independent semantic subspaces are divided. Based on the second query vector, the second key vector and the second value vector, the attention weight is calculated in each independent semantic subspace. The outputs of the multi-head attention in each independent semantic subspace are concatenated and then subjected to linear projection dimensionality reduction to obtain the processing node features.

7. The recommended method for handling distribution network faults according to claim 1, characterized in that, The process of calculating the matching probability of each candidate handling measure based on the characteristics of the fault node and the handling node, and outputting the fault handling recommendation scheme after sorting the matching probabilities in descending order, includes: Based on the fault node features and the handling node features, the global features under multiple fault scenarios are aggregated by summation pooling to obtain a comprehensive representation vector of the fault set. The comprehensive characterization vector and the characterization vectors of each candidate treatment measure are input into a bilinear matching function to calculate the bidirectional correlation score between the fault set and each treatment measure. After Softmax normalization, the matching probability of each candidate treatment measure is obtained. The candidates are sorted in descending order according to their matching probabilities, and the candidate handling measures with the highest matching probability are selected as the optimal fault handling recommendation scheme. Wherein, the bilinear matching function is pre-adopted by... Sorting loss and The joint loss function of the loss combination is obtained through supervised training.

8. The recommended method for handling distribution network faults according to claim 7, characterized in that, The joint loss function is expressed as: ; In the formula, The total loss of the joint loss function, for loss and Ranking loss The balancing hyperparameters.

9. A power distribution network fault handling recommendation system, characterized in that, include: The data acquisition module is used to acquire power distribution network maintenance record data, input it into a pre-trained fault handling recommendation model, and output fault handling recommendation schemes sorted in descending order of matching probability. The data processing method of the fault handling recommendation model is implemented by the following modules: The topology graph construction module is used to construct a topology graph structure based on historical maintenance record data of the distribution network. The topology graph structure includes a fault-handling bipartite graph, a fault-fault collaboration graph, and a handling-handling collaboration graph. The semantic embedding module is used to perform semantic encoding on the fault-disposal bipartite graph, the fault-fault collaboration graph, and the disposal-disposal collaboration graph using a pre-trained language model, and to extract the initial semantic embedding features of fault entities and disposal entities. The heterogeneous feature extraction module is used to obtain heterogeneous features of the fault map and the disposal map by using a graph convolutional network with residual connections to propagate homogeneous node features on the fault-fault collaboration graph and the disposal-disposal collaboration graph based on the initial semantic embedding features. It also uses a bidirectional graph attention network to adaptively allocate the weights of neighbor nodes to complete the aggregation of heterogeneous node features on the fault-disposal bipartite graph to obtain heterogeneous features of the bipartite graph. The multi-graph feature fusion module is used to set up a multi-head attention mechanism to adaptively fuse the heterogeneous features of the fault map and the bipartite map, as well as the heterogeneous features of the disposal map and the bipartite map, to obtain the fault node features and the disposal node features. The disposal recommendation module is used to calculate the matching probability of each candidate disposal measure based on the characteristics of the fault node and the disposal node, and output the fault disposal recommendation scheme after the matching probability is sorted in descending order.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the recommended method for handling power distribution network faults as described in any one of claims 1 to 8.