Power grid knowledge graph construction method and system based on semantic chunking and entity alignment

Abstract

The application provides a power grid knowledge graph construction method and system based on semantic segmentation and entity alignment, which comprises the following steps: forming an initial triple set according to multi-source heterogeneous power grid data; performing semantic segmentation on long text corpus in the multi-source heterogeneous power grid data to form a block-level triple set; establishing a mapping relationship between semantic blocks that meet the semantic similarity requirement; mapping the block-level triple set to a unified semantic space according to the mapping relationship, combining the mapped block-level triple set with the initial triple set to form a candidate triple set; establishing a mapping relationship between entities that meet the entity similarity requirement; establishing a mapping relationship between relationship items that meet the relationship similarity requirement; performing triple re-mapping on the triples in the candidate triple set according to the mapping relationship between entities and the mapping relationship between relationship items; and forming a power grid knowledge graph based on the re-mapped candidate triple set.

Description

Technical Field

[0001] This invention relates to the field of power grid regulation and intelligent power information processing technology, and in particular to a method and system for constructing a power grid knowledge graph based on semantic segmentation and entity alignment. Background Technology

[0002] With the continuous expansion of the power grid and the increasing automation and digitalization of control, the data sources upon which power grid operation and management rely exhibit significant heterogeneity. The current control system widely contains various types of information carriers, including semi-structured documents such as regulations, equipment ledgers, maintenance records, and protection setting sheets, as well as structured information such as measurement data, operating topology, GIS data, and time-series load data. Although these data are collected and managed within their respective systems, existing methods primarily rely on single data sources, keyword matching, or rule extraction, lacking systemic technical support for unified semantic modeling, entity alignment, and structured expression of cross-source heterogeneous power knowledge. This makes it difficult to support intelligent knowledge service capabilities in scenarios such as control-assisted decision-making, knowledge-based question answering, automatic retrieval, and regulations matching. Especially in power grid control tasks, there are problems such as large semantic spans, significant differences in expression, and mixed entity references among regulations, equipment parameters, and operating statuses. For example, the same equipment may appear as an abbreviation in the regulations, be identified using an asset code in the equipment ledger, and be identified using a real-time tag in the operating data. Existing solutions often struggle to achieve semantic-level alignment and unified indexing.

[0003] In existing technologies, some research has attempted to apply knowledge graphs to power grid equipment management and control assistance. However, these technologies primarily focus on knowledge extraction from existing scheduling plans and operational data, failing to perform semantic segmentation processing on natural language texts related to regulations, and neglecting the dynamic alignment of terminology for equipment entities from different sources. This makes them ill-suited for complex modeling scenarios involving regulations, operational status, and equipment relationships. Furthermore, some technologies focus on semantic modeling and indicator mapping of power production data to enhance the knowledge-based representation of statistical data. However, their objectives are mainly concentrated on business indicator systems, neglecting cross-modal fusion of regulations, maintenance data, and equipment models. They also lack semantic segmentation and structural mapping mechanisms for implicit entities and operational relationships within regulations, making it difficult to directly support the unified knowledge modeling needs of power grid control scenarios. Other research attempts to introduce knowledge retrieval and semantic reasoning into the power sector, aiming to achieve result ranking and inferential association through graph structure relationships. However, it does not solve the problem of semantic block-level parsing of heterogeneous source texts, lacks mechanisms for aligning terminological ambiguity, entity aliases, and cross-system identifiers, and does not yet have the ability to build a unified knowledge graph from procedure texts, equipment ledgers, and operational data.

[0004] In summary, although existing technologies have made some progress in power grid knowledge graph construction, entity extraction and semantic reasoning, they generally have the following shortcomings: (1) lack semantic layering and clause-level structured expression mechanisms for long texts such as regulations; (2) lack entity alignment methods for multi-source power grid data, making it difficult to achieve unified reference across systems, phrasing and encoding; (3) have not formed a high-quality heterogeneous knowledge graph construction process that can support semantic understanding, reasoning and retrieval of control business. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a method and system for constructing a power grid knowledge graph based on semantic segmentation and entity alignment to overcome the above problems.

[0006] This invention provides a method for constructing a power grid knowledge graph based on semantic segmentation and entity alignment, the method comprising: Entity relations are extracted from each original corpus in the pre-defined multi-source heterogeneous power grid dataset to form an initial set of triples. Semantic segmentation is performed on long text corpora in multi-source heterogeneous power grid datasets to obtain multi-granular semantic blocks; Entity relations are extracted from each semantic block to form a block-level triple set; Calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. Establish the mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements. Map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples; Entity extraction is performed on each semantic block and other raw corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora. Mapping relationships between entities are established for cross-source candidate entity pairs that meet the preset entity similarity requirements. Relationships are extracted from each semantic block and other raw data in the multi-source heterogeneous power grid dataset, excluding long text corpora. Mapping relationships are established between cross-source candidate relation pairs that meet the preset relation similarity requirements. Based on the mapping relationships between entities and between relational items, triples in the candidate triple set are remapped, and a power grid knowledge graph is formed based on the remapped candidate triple set.

[0007] Another aspect of the present invention provides a power grid knowledge graph construction system based on semantic chunking and entity alignment, the system comprising: The first knowledge extraction module is used to extract entity relations from each original corpus in the preset multi-source heterogeneous power grid dataset to form an initial set of triples. The semantic segmentation module is used to perform semantic segmentation on long text corpora in multi-source heterogeneous power grid datasets to obtain multi-granular semantic blocks; The second knowledge extraction module is used to extract entity relations for each semantic block to form a block-level triple set; The semantic block alignment module is used to calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. It also establishes a mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements. The first mapping module is used to map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and to combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples. The entity alignment module is used to extract entities from various semantic blocks and other raw corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, and to establish mapping relationships between cross-source candidate entity pairs that meet the preset entity similarity requirements. The relation alignment module is used to extract relation items from each semantic block and other raw data in the multi-source heterogeneous power grid dataset, excluding long text data, and to establish mapping relationships between relation items for cross-source candidate relation pairs that meet the preset relation similarity requirements. The second mapping module is used to remap the triples in the candidate triple set according to the mapping relationship between entities and the mapping relationship between relation items, and to form a power grid knowledge graph based on the remapped candidate triple set.

[0008] The present invention provides a method and system for constructing a power grid knowledge graph based on semantic segmentation and entity alignment. This method enables unified parsing, semantic modeling, and graph organization of multi-source heterogeneous power grid data, such as procedure texts, equipment ledgers, and operation monitoring data. By introducing semantic segmentation, cross-source entity mapping, relation extraction, and structure fusion mechanisms, it can uniformly identify power grid equipment, control actions, and operating statuses with inconsistent names and diverse expressions. This significantly improves the accuracy and completeness of knowledge fusion and avoids graph fragmentation or redundancy caused by differences in terminology. The present invention can construct a high-quality knowledge graph for power grid control scenarios, providing a computable, traceable, and scalable knowledge foundation for control knowledge management, intelligent question answering, procedure retrieval, decision support, and risk assessment.

[0009] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0010] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a flowchart of a method for constructing a power grid knowledge graph based on semantic segmentation and entity alignment, according to an embodiment of the present invention. Figure 2 This is a structural block diagram of a power grid knowledge graph construction system based on semantic segmentation and entity alignment, according to an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined.

[0013] To address the problems existing in current power grid knowledge construction processes, such as the difficulty in structuring long texts like regulations, inconsistent entity designations from heterogeneous data sources, insufficient semantic alignment accuracy, and weak knowledge fusion capabilities, this invention provides a power grid knowledge graph construction method based on semantic segmentation and entity alignment. This method enables unified parsing, semantic modeling, and graph-based organization of multi-source heterogeneous power grid data, including regulations, equipment ledgers, and operational monitoring data. Figure 1 As shown, the power grid knowledge graph construction method based on semantic segmentation and entity alignment proposed in this invention includes the following steps: S11. Extract entity relations from each original corpus in the preset multi-source heterogeneous power grid dataset to form an initial set of triples.

[0014] S12. Semantic segmentation is performed on the long text corpus in the multi-source heterogeneous power grid dataset to obtain multi-granular semantic blocks.

[0015] S13. Extract entity relations from each semantic block to form a block-level triple set.

[0016] S14. Calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. Establish the mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements.

[0017] S15. Map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples.

[0018] S16. Entity extraction is performed on each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora. The mapping relationship between entities is established for cross-source candidate entity pairs that meet the preset entity similarity requirements.

[0019] S17. Extract relation terms from each semantic block and other original corpora other than long text corpora in the multi-source heterogeneous power grid dataset, and establish mapping relationships between relation terms for cross-source candidate relation pairs that meet the preset relation similarity requirements.

[0020] S18. Perform triple remapping on the triples in the candidate triple set according to the mapping relationship between entities and the mapping relationship between relation items, and form a power grid knowledge graph based on the remapped candidate triple set.

[0021] The power grid knowledge graph construction method based on semantic segmentation and entity alignment provided in this invention introduces semantic segmentation, cross-source entity mapping, relation extraction, and structure fusion mechanisms. This enables unified identification of power grid equipment, control actions, and operating statuses with inconsistent names and diverse expressions, significantly improving the accuracy and completeness of knowledge fusion and avoiding graph fragmentation or redundancy caused by terminology differences. This invention can construct a high-quality knowledge graph for power grid control scenarios, providing a computable, traceable, and scalable knowledge foundation for control knowledge management, intelligent question answering, procedure retrieval, decision support, and risk assessment.

[0022] In this embodiment of the invention, the specific implementation process of extracting entity relations from each original corpus in the preset multi-source heterogeneous power grid dataset to form an initial set of triples in step S11 is as follows: In practical applications, dispatching procedures from the power grid control center, PDF-formatted operation management methods, Excel-formatted equipment ledgers, Word-formatted relay protection maintenance records, SCADA / EMS / WAMS real-time measurement databases, and GIS topology databases are selected as initial data sources to construct a multi-source heterogeneous power grid dataset. First, the data from different sources undergoes format unification processing. PDF documents are converted into parsable text using OCR and layout restoration, and Excel and database records are exported as CSV or JSON formats. Field encoding, date formats, and equipment numbers are standardized. Subsequently, through noise filtering, illegal symbol removal, field alignment, primary key deduplication, and unit unification, a core dataset from which semantic content can be extracted is constructed. To facilitate subsequent graph construction, the cleaned data is transformed into an initial triplet corpus in the form of head entities, relations, and tail entities, with timestamps, source, and confidence labels. Formalized, this can be expressed as: ; in , These refer to entities within the power grid sector, such as substations, transmission lines, and control commands. Indicate the relationship type, such as "connection", "monitoring", "bearer", etc. For time information, Source system identifier The initial confidence weights are defined by N, which represents the number of triples. This set can be abstracted as an entity set. and relation set ,Right now: ; After the ETL process completes the time format standardization, primary key deduplication, and outlier removal, an initial set of triples that meets the requirements of subsequent semantic parsing and alignment can be obtained. .

[0023] In this embodiment of the invention, the specific implementation process of semantic segmentation of long text corpora in the multi-source heterogeneous power grid dataset in step S12 is as follows: In processing long text corpora such as procedural documents, the complex sentence structures and strong contextual dependencies of these documents can easily lead to semantic drift and loss of key information if extracted directly. Therefore, to ensure that the extraction results have contextual boundaries and hierarchical constraints, procedural documents are treated as sequences composed of sentence units or semantic fragments. First, the entire text is initially divided into blocks based on document structural features (chapter titles, clause numbers, indentation format). Then, combining sentence segmentation algorithms, topic phrase recognition models, and semantic embedding clustering methods, the long text is refined into semantically complete sub-segments to achieve semantic blockization, and each semantic block is assigned a unique number. Let the original long text be represented as: ; Through semantic segmentation function Transform it into a set of semantic blocks: ; Where D represents long text data. This represents the text structure numbered n. This represents the segmentation parameters based on structural feature labels. B represents the semantic similarity or topic consistency threshold, and B represents the set of semantic blocks. This represents the semantic block numbered m. To enhance the expressiveness of the clauses, hierarchical annotation information and topic tag information can be added to each semantic block, for example: ; in It can represent "Chapter X - Article Y". This represents thematic categories such as "dispatch authorization," "fault clearance," and "maintenance and grid connection." This block-based process endows procedural knowledge with a basic semantic structure that allows for location, extraction, and matching.

[0024] In this embodiment of the invention, step S13, which involves extracting entity relations from each semantic block to form a block-level triplet set, specifically includes: Named entity recognition and semantic role labeling are performed on each semantic block to extract the entities contained in the semantic block and generate a semantic block entity set corresponding to each semantic block; For each semantic block, intra-block candidate relation identification is performed to generate a semantic block relation set corresponding to each semantic block; A set of block-level triples is formed based on the set of semantic block entities and the set of semantic block relations corresponding to each semantic block.

[0025] In this embodiment, after the semantic block is split, named entity recognition and semantic role labeling are performed in each semantic block by combining rule templates, word vector encoding or large language models to extract key semantic units such as equipment entities, parameter items, control actions, constraints and measures instructions, thereby obtaining the semantic block entity set and encoding representation corresponding to each semantic block.

[0026] Let semantic blocks The corresponding text sequence is Then the entity recognition function can be formalized as: ; in For domain-specific dictionaries or model parameters for power grid equipment, operation, and procedure terminology. For semantic blocks A collection of semantic block entities.

[0027] Furthermore, the semantic block relation set is extracted using predicate trigger words, dependency paths, or semantic role annotation methods: ; in These include predicates, trigger words, function words, and relational phrases, For semantic blocks The corresponding set of semantic block relations.

[0028] Based on this, the candidate statements within a block can be transformed into a set of block-level triples (i.e., block-level knowledge units): These block-level triples can be further filtered, completed, or modified in subsequent cross-source alignment and global fusion to form a structured intermediate representation.

[0029] In this embodiment of the invention, the calculation of the block-level vector representation of each semantic block in step S14 includes: The block-level vector representation of each semantic block is calculated based on the average sentence vector, average word vector, or topic vector of each semantic block.

[0030] To achieve consistent alignment between procedural blocks and equipment ledgers, control logs, and operational topology data, an embedded representation, i.e., a block-level vector representation, can be constructed for each semantic block, and similarity matching can be performed with blocks or records from other data sources. Let the set of semantic blocks be... With device / running data block set Since the vectors come from different sources, the block-level vector representation can be generated using average word vectors, sentence vectors, or topic vectors. The block-level vector generation function can be formalized as follows: .

[0031] Furthermore, in step S14, the semantic similarity of cross-source semantic block pairs is calculated based on the block-level vector representation, and a mapping relationship between semantic blocks is established for cross-source semantic block pairs that meet the preset semantic similarity requirements. The specific implementation is as follows: Calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. When the semantic similarity is higher than or equal to a preset first semantic similarity threshold, the similarity is calculated. Establish block mapping relationships for semantic blocks in cross-source semantic block pairs, when the semantic similarity is lower than the first semantic similarity threshold. And it exceeds the preset second semantic similarity threshold. The current cross-source semantic block pair is added to a preset candidate set. In this embodiment, the semantic similarity of cross-source semantic block pairs can be represented by a cosine similarity metric: ; like Above the upper threshold If it falls into the corresponding mapping, then a corresponding mapping will be automatically established; if it falls into the corresponding mapping, then a corresponding mapping will be If the interval is not specified, it will enter the candidate set to wait for consistency verification or manual sampling.

[0032] After completing the block-level mapping, the candidate knowledge unit set can be... The mapping result is projected onto a unified semantic space and compared with the original set of triples. Combine to form a set of candidate triples This provides a structural foundation for the subsequent standardization of entities and relationships.

[0033] In this embodiment of the invention, step S16 involves entity extraction from each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, and establishing mapping relationships between cross-source candidate entity pairs that meet preset entity similarity requirements. This includes the following steps, not shown in the accompanying drawings: S161. Entity extraction is performed on each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, to generate candidate entity datasets corresponding to each original corpus.

[0034] In this embodiment, before extracting candidate entities from multi-source heterogeneous data such as procedure texts, accident records, equipment ledgers, and control logs, semantic preprocessing of the original corpus is required, including sentence segmentation, word segmentation, word form restoration, noise removal, stop word filtering, and domain dictionary mapping, in order to reduce semantic redundancy and ambiguity caused by spelling differences.

[0035] S162. Map the entities in the candidate entity dataset to a unified vector space to obtain the semantic vector representation of each entity.

[0036] In this embodiment, in the obtained semantic blocks In addition to long text corpora, other original corpora in the multi-source heterogeneous power grid dataset are used to extract entities such as devices, actions, states, and parameters by combining domain dictionaries and sequence labeling models, forming candidate entity datasets corresponding to the original corpora from different sources for subsequent alignment; at the same time, each entity is mapped to a unified vector space for similarity measurement and aggregation. As shown below: ; .

[0037] S163. Calculate the entity similarity for candidate entity pairs composed of entities from different candidate entity datasets. If the entity similarity is greater than or equal to a preset first entity similarity threshold... Then, establish an entity mapping relationship for the entities in the candidate entity pair. If the entity similarity is less than the first entity similarity threshold... And greater than or equal to the preset second entity similarity threshold If so, the candidate entity pair will be added to the preset entity candidate pool.

[0038] In this embodiment, entity similarity calculation is performed on candidate entity pairs composed of entities from different candidate entity datasets, specifically including: Calculate the edit distance, sequence similarity, or Jaccard coefficient of candidate entity pairs to obtain the character similarity parameters of the candidate entity pairs. ; Calculate the vector similarity of candidate entity pairs to obtain the semantic similarity parameters of the candidate entity pairs. ; According to the preset first weighting coefficient and The entity similarity of the candidate entity pair is obtained by weighted summation of the character similarity parameters and semantic similarity parameters.

[0039] In this embodiment, entity pairs from different systems / document sources are addressed. A candidate entity pair set is constructed, and a score is calculated based on a combination of character similarity and semantic similarity. An automatic merging and candidate review threshold is set to control the risk of false merging. When the score exceeds the threshold... Establish entity mapping relationships and mark them as synonym candidates. If they fall into For intervals, the original naming is retained and the interval proceeds to the subsequent verification process. The overall similarity is defined as: ; The threshold is determined as follows: .

[0040] Furthermore, after establishing entity mapping relationships for entities in candidate entity pairs, the method further includes: constructing a bipartite graph of entities in all candidate entity datasets, with the two end nodes representing entities from different sources, the edge weight being entity similarity, generating a consistent mapping table using maximum weight matching, and verifying all entity mapping relationships based on the consistent mapping table.

[0041] Specifically, to avoid one-to-many or many-to-one conflicts, entities from different sources are constructed into a bipartite graph at the global level, with edge weights calculated based on the entity similarity. Maximum weight matching is used to generate a consistent mapping table. Low-confidence matches retain manually sampled positions to ensure that key entities are not incorrectly merged. The optimized model is as follows: ; .

[0042] Optionally, a neighborhood context-enhanced consistency judgment can be introduced. The structural similarity is measured by calculating the neighborhood entity embedding of the candidate entity pair. If the structural similarity measured by the neighborhood entity embedding and the entity similarity meet the preset consistency conditions, the mapping relationship between the entities of the candidate entity pair is confirmed. Otherwise, the candidate entity pair is added to the preset entity candidate pool for subsequent manual sampling or delayed processing to ensure that the final semantic alignment result has consistency and scalability.

[0043] The generated global mapping is denoted as: .

[0044] In this embodiment of the invention, step S17 involves extracting relational terms from each semantic block and other original corpora (excluding long text corpora) in the multi-source heterogeneous power grid dataset, and establishing mapping relationships between relational terms for cross-source candidate relation pairs that meet preset relational similarity requirements. This includes the following steps, not shown in the accompanying drawings: S171. Based on dependency parsing, semantic role labeling, regular template or remote supervision methods, relation terms are extracted from each semantic block and other original corpora in the multi-source heterogeneous power grid dataset except for long text corpora, in order to generate candidate relation datasets corresponding to each original corpus.

[0045] S172. Map the relation items in the candidate relation dataset to a unified vector space to obtain the semantic vector representation of each relation item.

[0046] After entity alignment is completed, predicates, trigger words, action words, and relational phrases involved in the procedure text, incident records, control logs, and operation specifications need to be uniformly extracted. Candidate relational items can be identified based on dependency parsing, semantic role labeling, regular expression templates, or remote supervision methods, and then mapped to a vector space.

[0047] In this embodiment, for the entity pairs identified in step S16 By performing syntactic dependency extraction, trigger word matching, or large model discrimination, candidate relation datasets corresponding to original corpora from different sources are extracted. The relation phrases are encoded into vectors to support subsequent alignment, aggregation, and conflict resolution. Defined as: ; ; Each candidate relation and its corresponding entity pair can form a preliminary triple: These can be added to the candidate triplet set to obtain a new candidate triplet set. Among them, [the triplet set is missing here]. It can remain compatible with entity embedding models (such as BERT, TransE, or sentence vector models) so that similarity and structural mapping can be calculated uniformly in the future.

[0048] S173. Calculate the relation similarity for candidate relation pairs composed of relation items from different candidate relation datasets based on the semantic vector representation of each relation item. If the relation similarity is greater than or equal to a preset first relation similarity threshold... Then, establish a relation mapping relationship for the relation items in the candidate relation pair. If the relation similarity is less than the first relation similarity threshold... And greater than or equal to the preset second relationship similarity threshold If so, the candidate relationship pair will be added to the preset relationship candidate pool.

[0049] In an optional embodiment, relation similarity is calculated for candidate relation pairs composed of relation items from different candidate relation datasets based on the semantic vector representation of each relation item, specifically including: Calculate the vector similarity of candidate relation pairs to obtain their semantic similarity parameters. ; Calculate the mean of the vector set of entities connected to each relation term in the candidate relation pair. and The vector similarity of the mean values ​​of the two vector sets is calculated. This yields the neighborhood similarity parameters for candidate relationship pairs. The semantic similarity parameter and neighborhood similarity parameter of the candidate relation pair are weighted and summed according to the preset second weight coefficient to obtain the relation similarity of the candidate relation pair.

[0050] In this embodiment, different sources contain semantically similar but differently named relation types such as "bearer / connector" and "cause / trigger," requiring merging or mapping based on semantic similarity and contextual consistency. For any relation pair... The calculated similarity of relationships is defined as follows: ; in The mean of the vector set of entities connected to this relation is given by the following statement: Greater than or equal to the upper threshold Then, a relational mapping relationship is established for automatic synonym merging. If it falls within... An interval is used for candidate verification, and a relationship mapping table is generated after applying "one-to-one" or "prohibit many-to-many" constraints: .

[0051] In another optional embodiment, relation similarity is calculated for candidate relation pairs composed of relation items from different candidate relation datasets based on the semantic vector representation of each relation item, specifically including: Calculate the edit distance, sequence similarity, or Jaccard coefficient of candidate relation pairs to obtain the character similarity parameter of the candidate relation pairs. ; Calculate the vector similarity of candidate relation pairs to obtain their semantic similarity parameters. ; The relation similarity of the candidate relation pairs is obtained by weighting and summing the character similarity parameters and semantic similarity parameters of the candidate relation pairs according to the preset second weight coefficient.

[0052] In entity mapping Relational mapping Based on this, all candidate triples are uniformly projected to form a cross-source consistent graph structure, and conflicting or duplicate items are filtered out; low-confidence items are retained in a sampling pool for manual verification. The standardized triples after mapping are expressed as: ; in and These represent the mapping functions between entities and relations. If multiple relations correspond to the same entity and there are conflicts, a confidence ranking or manual annotation mechanism is used for pruning, ultimately resulting in a structured, reasonable, and consistent knowledge graph. .

[0053] The power grid knowledge graph construction method based on semantic segmentation and entity alignment provided in this invention has the following beneficial effects: (1) The semantic segmentation and entity embedding joint modeling method proposed in this invention can extract unified semantic units from multi-source procedure texts, operation records and equipment data; its significance lies in breaking down data barriers of different formats and different departments, and laying a highly consistent semantic foundation for knowledge graph construction.

[0054] (2) The cross-source entity alignment and similarity fusion mechanism proposed in this invention can uniformly identify power grid equipment, control actions and operating status with inconsistent names and diverse expressions; its significance lies in significantly improving the accuracy and completeness of knowledge fusion and avoiding graph fragmentation or redundancy caused by differences in terminology.

[0055] (3) The relational semantic alignment and structural constraint verification method introduced in this invention can perform topological consistency mapping on power grid regulation relationships such as “monitoring-control-impact-disposal”; its significance lies in ensuring the correctness and reasonability of the graph relational network, and supporting subsequent retrieval, question answering and chain inference.

[0056] (4) The standardized heterogeneous power grid knowledge graph formed by the present invention can directly provide an aligned knowledge base for large model (LLM) and support RAG enhancement; its significance lies in improving the interpretability and professionalism of tasks such as intelligent question answering, decision support, and control reasoning, and meeting the high reliability knowledge support requirements in power grid control scenarios.

[0057] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0058] This invention provides a power grid knowledge graph construction system based on semantic segmentation and entity alignment. The system includes functional modules for implementing the power grid knowledge graph construction method based on semantic segmentation and entity alignment as described in any of the preceding claims. Figure 2 This illustration schematically shows a structural diagram of a power grid knowledge graph construction system based on semantic chunking and entity alignment provided by an embodiment of the present invention. (Refer to...) Figure 2 The system described in this embodiment of the invention includes: The first knowledge extraction module 201 is used to extract entity relations from each original corpus in the preset multi-source heterogeneous power grid dataset to form an initial set of triples. Semantic segmentation module 202 is used to perform semantic segmentation on long text corpora in multi-source heterogeneous power grid datasets to obtain multi-granularity semantic blocks; The second knowledge extraction module 203 is used to extract entity relations for each semantic block to form a block-level triple set; The semantic block alignment module 204 is used to calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation, and establish a mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements. The first mapping module 205 is used to map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and to combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples. The entity alignment module 206 is used to extract entities from each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, and to establish mapping relationships between entities for cross-source candidate entity pairs that meet the preset entity similarity requirements. The relation alignment module 207 is used to extract relation items from each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, and to establish mapping relationships between relation items for cross-source candidate relation pairs that meet the preset relation similarity requirements. The second remapping module 208 is used to remap the triples in the candidate triple set according to the mapping relationship between entities and the mapping relationship between relation items, and to form a power grid knowledge graph based on the remapped candidate triple set.

[0059] As the system implementation is basically similar to the method implementation, the description is relatively simple, and relevant parts can be found in the description of the method implementation.

[0060] In addition, another embodiment of the present invention provides a computer program product on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps in the above embodiment of the power grid knowledge graph construction method based on semantic segmentation and entity alignment.

[0061] Furthermore, another embodiment of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps described in the above embodiment of the power grid knowledge graph construction method based on semantic segmentation and entity alignment.

[0062] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, any of the claimed embodiments can be used in any combination.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a power grid knowledge graph based on semantic segmentation and entity alignment, characterized in that, The method includes: Entity relations are extracted from each original corpus in the pre-defined multi-source heterogeneous power grid dataset to form an initial set of triples. Semantic segmentation is performed on long text corpora in multi-source heterogeneous power grid datasets to obtain multi-granular semantic blocks; Entity relations are extracted from each semantic block to form a block-level triple set; Calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. Establish the mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements. Map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples; Entity extraction is performed on each semantic block and other raw corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora. Mapping relationships between entities are established for cross-source candidate entity pairs that meet the preset entity similarity requirements. Relationships are extracted from each semantic block and other raw data in the multi-source heterogeneous power grid dataset, excluding long text corpora. Mapping relationships are established between cross-source candidate relation pairs that meet the preset relation similarity requirements. Based on the mapping relationships between entities and between relational items, triples in the candidate triple set are remapped, and a power grid knowledge graph is formed based on the remapped candidate triple set.

2. The method according to claim 1, characterized in that, Entity relations are extracted from each semantic block to form a block-level triple set, including: Named entity recognition and semantic role labeling are performed on each semantic block to extract the entities contained in the semantic block and generate a semantic block entity set corresponding to each semantic block; For each semantic block, intra-block candidate relation identification is performed to generate a semantic block relation set corresponding to each semantic block; A set of block-level triples is formed based on the set of semantic block entities and the set of semantic block relations corresponding to each semantic block.

3. The method according to claim 1, characterized in that, The calculation of the block-level vector representation of each semantic block includes: The block-level vector representation of each semantic block is calculated based on the average sentence vector, average word vector, or topic vector of each semantic block.

4. The method according to claim 1, characterized in that, Entity extraction is performed on each semantic block and other raw corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora. Mapping relationships are established between cross-source candidate entity pairs that meet preset entity similarity requirements, including: Entity extraction is performed on each semantic block and other original corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, to generate candidate entity datasets corresponding to each original corpus. The entities in the candidate entity dataset are mapped to a unified vector space to obtain the semantic vector representation of each entity; Entity similarity is calculated for candidate entity pairs composed of entities from different candidate entity datasets. If the entity similarity is greater than or equal to a preset first entity similarity threshold, an entity mapping relationship is established for the entities in the candidate entity pair. If the entity similarity is less than the first entity similarity threshold but greater than or equal to a preset second entity similarity threshold, the candidate entity pair is added to a preset entity candidate pool.

5. The method according to claim 4, characterized in that, Entity similarity calculation is performed on candidate entity pairs composed of entities from different candidate entity datasets, including: Calculate the edit distance, sequence similarity, or Jaccard coefficient of candidate entity pairs to obtain the character similarity parameters of candidate entity pairs; Calculate the vector similarity of candidate entity pairs to obtain the semantic similarity parameters of the candidate entity pairs; The entity similarity of the candidate entity pair is obtained by weighting and summing the character similarity parameters and semantic similarity parameters of the candidate entity pair according to the preset first weight coefficient.

6. The method according to claim 4, characterized in that, After establishing entity mapping relationships for entities in candidate entity pairs, the method further includes: The entities in all candidate entity datasets are constructed into a bipartite graph, with the two end nodes representing entities from different sources and the edge weights representing entity similarity. A consistent mapping table is generated using maximum weight matching, and all entity mapping relationships are verified based on the consistent mapping table.

7. The method according to claim 1, characterized in that, Relationships are extracted from each semantic block and other raw corpora (excluding long text corpora) in the multi-source heterogeneous power grid dataset. Mapping relationships are established between cross-source candidate relation pairs that meet preset relation similarity requirements, including: Relation terms are extracted from each semantic block and other original corpora (excluding long text corpora) in the multi-source heterogeneous power grid dataset based on dependency parsing, semantic role labeling, regular templates or remote supervision methods to generate candidate relation datasets corresponding to each original corpus. Map the relation items in the candidate relation dataset to a unified vector space to obtain the semantic vector representation of each relation item; Based on the semantic vector representation of each relation item, the relation similarity of candidate relation pairs composed of relation items from different candidate relation datasets is calculated. If the relation similarity is greater than or equal to the preset first relation similarity threshold, a relation mapping relationship is established for the relation items in the candidate relation pair. If the relation similarity is less than the first relation similarity threshold but greater than or equal to the preset second relation similarity threshold, the candidate relation pair is added to the preset relation candidate pool.

8. The method according to claim 7, characterized in that, Based on the semantic vector representation of each relation term, relation similarity is calculated for candidate relation pairs composed of relation terms from different candidate relation datasets, including: Calculate the vector similarity of candidate relation pairs to obtain the semantic similarity parameters of the candidate relation pairs; Calculate the mean of the vector set of entities connected to each relation item in the candidate relation pair, and calculate the vector similarity of the means of the two vector sets to obtain the neighborhood similarity parameter of the candidate relation pair. The semantic similarity parameter and neighborhood similarity parameter of the candidate relation pair are weighted and summed according to the preset second weight coefficient to obtain the relation similarity of the candidate relation pair.

9. The method according to claim 7, characterized in that, Based on the semantic vector representation of each relation term, relation similarity is calculated for candidate relation pairs composed of relation terms from different candidate relation datasets, including: Calculate the edit distance, sequence similarity, or Jaccard coefficient of the candidate relation pairs to obtain the character similarity parameters of the candidate relation pairs; Calculate the vector similarity of candidate relation pairs to obtain the semantic similarity parameters of the candidate relation pairs; The relation similarity of the candidate relation pairs is obtained by weighting and summing the character similarity parameters and semantic similarity parameters of the candidate relation pairs according to the preset second weight coefficient.

10. A power grid knowledge graph construction system based on semantic segmentation and entity alignment, characterized in that, The system includes: The first knowledge extraction module is used to extract entity relations from each original corpus in the preset multi-source heterogeneous power grid dataset to form an initial set of triples. The semantic segmentation module is used to perform semantic segmentation on long text corpora in multi-source heterogeneous power grid datasets to obtain multi-granular semantic blocks; The second knowledge extraction module is used to extract entity relations for each semantic block to form a block-level triple set; The semantic block alignment module is used to calculate the block-level vector representation of each semantic block, and calculate the semantic similarity of cross-source semantic block pairs based on the block-level vector representation. It also establishes a mapping relationship between semantic blocks for cross-source semantic block pairs that meet the preset semantic similarity requirements. The first mapping module is used to map the set of block-level triples corresponding to semantic blocks with mapping relationships to a unified semantic space, and to combine the mapped set of block-level triples with the initial set of triples to form a set of candidate triples. The entity alignment module is used to extract entities from various semantic blocks and other raw corpora in the multi-source heterogeneous power grid dataset, excluding long text corpora, and to establish mapping relationships between cross-source candidate entity pairs that meet the preset entity similarity requirements. The relation alignment module is used to extract relation items from each semantic block and other raw data in the multi-source heterogeneous power grid dataset, excluding long text data, and to establish mapping relationships between relation items for cross-source candidate relation pairs that meet the preset relation similarity requirements. The second mapping module is used to remap the triples in the candidate triple set according to the mapping relationship between entities and the mapping relationship between relation items, and to form a power grid knowledge graph based on the remapped candidate triple set.

Images