Large language model-based power grid regulation system semantic retrieval method based on knowledge graph

Abstract

This invention discloses a semantic retrieval method for power grid regulations based on a large-scale language model using a knowledge graph. Specifically, it includes: collecting dispatch terminology and contextual data to establish a terminology database; cleaning, standardizing, and annotating the data to form structured terminology data; comparing the current and historical terminology databases to identify newly added, drifting, and discontinued terms; constructing a terminology knowledge graph and storing triple relationships; extracting semantic features by calling a pre-trained language model; fusing scenario, equipment, and operational features to generate terminology embeddings; writing the newly added terminology embeddings into an adaptation layer to incrementally reconstruct the scenario model; and performing spatial alignment and weight fusion based on graph connections to output standardized terminology results. This invention achieves unified semantic retrieval and dynamic standardized output of power grid terminology.

Description

Technical Field

[0001] This invention relates to the field of power grid semantic retrieval technology, and in particular to a semantic retrieval method for power grid regulations based on a large language model with knowledge graphs. Background Technology

[0002] With the increasing number of power grid dispatching regulations and systems, the texts of regulations, operating procedures, and equipment instructions are characterized by their large scale, frequent updates, and diverse expressions. Traditional retrieval methods mostly adopt keyword matching techniques, relying primarily on the literal consistency of terms for retrieval. They lack the ability to effectively handle abbreviations, synonyms, and semantic differences in different contexts, easily leading to incomplete or inaccurate search results.

[0003] In recent years, semantic retrieval methods have been gradually applied to the power industry. By using word vector representations or neural network models to semantically model text, the ability to identify synonyms and semantically similar expressions has been improved to some extent. However, these methods usually focus on calculating the overall semantic similarity of the text, making it difficult to fully utilize the inherent equipment connection relationships, operational process relationships, and scenario-dependent features in power grid operations. This results in insufficient semantic discrimination capabilities in complex scheduling contexts.

[0004] Meanwhile, knowledge graph-based technologies, by constructing entities and their relationship networks, can structurally express the associations between terms, thereby enhancing the description of the relationships between terms to some extent. However, existing methods mostly rely on static graph structures, which have limited adaptability to semantic changes and new terms arising from regulatory revisions, and lack a deep understanding of the context.

[0005] In addition, large-scale pre-trained language models that have emerged in recent years can perform contextual semantic encoding on text and have shown strong capabilities in natural language understanding tasks. However, existing applications are mostly focused on general text processing and lack effective integration with power grid dispatch knowledge structure and equipment topology relationships, making them difficult to directly apply to rule and regulation semantic retrieval scenarios.

[0006] Therefore, there is an urgent need for a semantic retrieval method for power grid regulations that can integrate semantic representation capabilities, structured association capabilities, and dynamic adaptation capabilities, in order to improve the accuracy of terminology recognition and the stability of retrieval results. Summary of the Invention

[0007] One objective of this invention is to propose a semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs. This invention introduces a term association graph construction mechanism and a scenario model incremental adaptation mechanism, and performs joint gating fusion and dynamic reconstruction processing on the semantic features, scenario features, equipment association features, and operation type features of scheduling terms. This constructs a standard term semantic retrieval process for power grid regulations, which has the advantages of high term recognition accuracy, strong semantic retrieval adaptability, and good stability of standard term output.

[0008] The semantic retrieval method for power grid regulations based on a large-scale language model using a knowledge graph, according to an embodiment of the present invention, includes the following steps: Collect scheduling terminology and contextual data to form a terminology database; The terminology database is cleaned, standardized, and labeled to form structured terminology data; The current terminology database is compared with the historical terminology database to identify newly added terms, semantically drifted terms, and discontinued terms. A terminology knowledge graph is constructed based on structured terminology data. Term nodes, device nodes, operation nodes, and scenario nodes are defined, and synonym relationships, hierarchical relationships, device connection relationships, and operation dependency relationships are established. The terminology association relationships are transformed into a triple structure for storage. The pre-trained language model is invoked to perform contextual semantic encoding on term content and contextual data, generating semantic representation vectors. The semantic representation vectors are then enhanced with contextual fragments to extract semantic features. Based on the terminology knowledge graph, scene features, device association features and operation type features are extracted, and gating fusion is performed on semantic features, scene features, device association features and operation type features. The terminology embedding is generated by combining co-occurrence frequency, semantic similarity, physical topology correlation and instruction temporal span. The terms corresponding to the newly added terms are embedded into the scene model adaptation layer, and incremental adjustments are made to the scene model adaptation layer to form the reconstructed scene model. The embedded features corresponding to the reconstructed scene model are extracted, and spatial alignment and dynamic weight fusion are performed based on the node connection relationship in the terminology knowledge graph to output standard terminology results.

[0009] Optionally, the formation of the terminology database specifically includes: The process involves: first, compiling scheduling regulations, operating procedures, and equipment descriptions to form a raw terminology dataset; second, dividing the raw terminology dataset into text units, extracting term fragments, and forming candidate terminology sequences; third, unifying the expression of candidate terminology sequences to generate standardized terminology sequences; fourth, establishing terminology identifiers within standardized terminology sequences, recording source type identifiers and timestamps, and forming an indexed terminology dataset; fifth, constructing context fragment sets at the corresponding positions of terminology identifiers to generate terminology context pairs; sixth, generating co-occurrence count sequences within terminology context pairs, and writing the co-occurrence count results at the terminology identifier positions; seventh, establishing semantic vector recording units at the terminology identifier positions to generate terminology semantic vectors; and finally, assembling the terminology identifiers, standardized terminology content, context fragment sets, co-occurrence count sequences, semantic vectors, and source type identifiers to form a terminology library.

[0010] Optionally, the formation of structured terminology data includes the following steps: The standardized term sequences in the terminology database are divided into fields, generating term content fields, source type fields, and time stamp fields. A field index mapping table is established at the term identifier number position, recording the corresponding field content at each index position. The context fragment set is position-labeled, generating a context index sequence at the term identifier number position. The co-occurrence count sequence is numerically normalized, generating a co-occurrence weight sequence at the term identifier number position. The semantic vector record unit is dimensionally normalized, generating a unified dimension semantic vector at the term identifier number position. The source type identifier is categorically encoded, generating a type encoding sequence at the term identifier number position. The term identifier number, term content field, field index mapping table, context index sequence, co-occurrence weight sequence, unified dimension semantic vector, and type encoding sequence are combined and mapped to form structured terminology data.

[0011] Optionally, the construction of a terminology knowledge graph based on structured terminology data specifically includes: Read the term identifier numbers from the structured terminology data to generate a term node set; extract the device name field, operation instruction field, and scene description field from the structured terminology data to generate device node sets, operation node sets, and scene node sets respectively; establish connection relationships between terms and devices based on co-occurrence positions recorded in the context index sequence between the term node set and the device node set, generating term-device relationship edges; establish association relationships between terms and operations based on the operation instruction field content between the term node set and the operation node set, generating term-operation relationship edges; establish association relationships between terms and scenes based on the source type identifier and context fragment set between the term node set and the scene node set, generating term-scene relationship edges; within the term node set, establish synonym relationship edges and hierarchical relationship edges between terms based on the similarity calculation results between semantic vectors of the same dimension. Optionally, the gating fusion of semantic features, scene features, device association features, and operation type features includes the following steps: To construct a semantic feature vector, the following steps are performed: First, read the unified-dimensional semantic vector, type encoding sequence, context index sequence, co-occurrence weight values, and semantic similarity values ​​from the connection edge set at the corresponding position in the terminology node set. Second, read the source type identifier and context fragment set at the corresponding position in the terminology node set, extract the corresponding field content from the field index mapping table, and combine and encode the source type identifier and context fragment set to construct a scene feature vector. Third, read the connection relationship between node pairs at the corresponding position in the connection edge set, extract the corresponding field content of the device name between node pairs, establish a device mapping relationship in the field index mapping table, and numerically encode the device mapping relationship to construct a device association feature vector. Fourth, read the terminology content field at the corresponding position in the terminology node set, extract operation instruction terms from the field index mapping table, classify the operation instruction terms, generate operation category codes at the terminology identifier number position, and construct an operation type feature vector. Finally, establish a feature combination matrix at the terminology identifier number position and write the semantic feature vector, scene feature vector, device association feature vector, and operation type feature vector into the corresponding positions of the feature combination matrix. A gating parameter recording unit is established in the feature combination matrix. Initial gating weight values ​​are written to the corresponding positions of each feature vector. The gating weight values ​​are normalized to generate a gating weight sequence. Weighted calculations are performed on each feature vector in the feature combination matrix according to the gating weight sequence, and a fused feature vector is generated at the term identifier number position. During the generation of the fused feature vector, joint constraint calculations are performed on the gating weight values ​​and the co-occurrence weight values ​​in the connection edge set, and the gating weight values ​​are updated at the term identifier number position. The fused feature vector is output at the term identifier number position, forming the gating fusion result.

[0012] Optionally, the step of generating term embeddings by combining co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span specifically includes: The system extracts co-occurrence weight values ​​at corresponding positions in the connection edge set, extracts fusion feature vectors at corresponding positions in the term node set, calculates semantic similarity values ​​between node pairs, and generates a semantic similarity sequence. It also extracts device name, device connection, and device level field content from the field index mapping table, performs corresponding matching on these fields to generate a device connection path sequence, and performs level conversion, adjacency conversion, and path length conversion on the device connection path sequence to generate a physical topology correlation sequence. Finally, it extracts operation instruction term sequences and time stamp fields from the context fragment set, accumulates the time interval values ​​between adjacent operation instructions, and generates an instruction temporal span sequence. An embedded factor record matrix is ​​established at the term identifier number position. The co-occurrence weight value, semantic similarity sequence, physical topological correlation sequence, and instruction temporal span sequence are written into the corresponding positions of the embedded factor record matrix. The scale uniform processing of each factor value in the embedded factor record matrix is ​​performed to generate a standard factor sequence. A factor combination order is established at the corresponding position of the standard factor sequence, and a combined factor vector is formed according to the arrangement order of co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span. A fusion feature vector is introduced at the corresponding position of the combined factor vector. The combined factor vector and the fusion feature vector are concatenated and mapped to generate a term embedding input vector. An embedding parameter recording unit is established at the term identifier number position. The term embedding input vector is written into the embedding parameter recording unit. Linear transformation calculation and nonlinear compression calculation are performed on the term embedding input vector to generate an initial embedding vector. Adjacency propagation calculation is performed on the initial embedding vector in combination with the node connection relationship in the edge set to generate an association propagation vector. A weighted combination calculation is performed on the initial embedding vector and the association propagation vector to form the term embedding result.

[0013] Optionally, the incremental adjustment of the scene model adaptation layer includes the following steps: Extract the term embedding results corresponding to the newly added terms to form an incremental embedding sequence; establish an incremental write position table in the scene model adaptation layer to record the correspondence between scene category identifiers and incremental embedding sequences; write the incremental embedding sequence to the corresponding position in the scene model adaptation layer to generate incremental feature recording units; extract the original scene parameters at the corresponding position of the incremental feature recording unit and calculate the parameter offset sequence between the incremental embedding sequence and the original scene parameters; map the parameter offset sequence to generate a local update parameter sequence; overwrite and update the local update parameter sequence to generate an incremental parameter sequence; write the incremental parameter sequence to the scene model adaptation layer to form the reconstructed scene model.

[0014] Optionally, the extraction of the embedded features corresponding to the reconstructed scene model specifically includes: In the reconstructed scene model, locate the parameter positions corresponding to each scene category identifier and generate a feature extraction position sequence; extract incremental parameter sequences and original scene parameters at the corresponding positions of the feature extraction position sequence to form a scene parameter combination sequence; expand the scene parameter combination sequence dimensionally to generate a scene feature matrix; extract term identifier numbers and incremental embedding sequences at the corresponding positions of the scene feature matrix to form a scene-related feature sequence; combine and map the scene feature matrix and the scene-related feature sequence to generate an initial embedding feature sequence; perform dimensional normalization on the initial embedding feature sequence to form a unified dimensional embedding feature sequence; output the unified dimensional embedding feature sequence to form the embedding features corresponding to the reconstructed scene model.

[0015] The beneficial effects of this invention are: This invention forms a terminology database by collecting scheduling terminology data and contextual data, and performs comparison processing in conjunction with the historical terminology database. It can simultaneously identify newly added terms, semantically drifting terms, and discontinued terms, thereby solving the problems of asynchronous terminology updates, mixed use of old and new words, and inconsistent retrieval criteria in power grid regulations and documents, and making the terminology management process more continuous and adaptable.

[0016] This invention establishes a term association graph and jointly extracts semantic features, scene features, device association features, and operation type features. By performing gated fusion processing on multiple types of features, it can break through the retrieval method that relies solely on literal matching, enhance the ability to express the meaning of term context, device relationships, and operational context, thereby improving the accuracy of term recognition and semantic retrieval in complex business scenarios.

[0017] This invention generates term embeddings by combining co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span. The corresponding representation results of newly added terms are written into the scenario model adaptation layer to perform incremental adjustments. This can maintain the consistency between the model representation results and the actual business context even when regulations and terminology are continuously revised and expanded, thereby improving the stability of standard terminology output and the reliability of retrieval results. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs, as proposed in this invention. Figure 2This is a schematic diagram illustrating the terminology multi-feature gating fusion and embedding generation process of the semantic retrieval method for power grid regulations based on a large-scale language model using a knowledge graph, as proposed in this invention. Figure 3 This diagram illustrates the incremental adaptation of the scenario model and the semantic retrieval output process of the knowledge graph-based large-scale language model power grid regulations semantic retrieval method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs includes the following steps: Collect scheduling terminology and contextual data to form a terminology database; The terminology database is cleaned, standardized, and labeled to form structured terminology data; The current terminology database is compared with the historical terminology database to identify newly added terms, semantically drifted terms, and discontinued terms. A terminology knowledge graph is constructed based on structured terminology data. Term nodes, device nodes, operation nodes, and scenario nodes are defined, and synonym relationships, hierarchical relationships, device connection relationships, and operation dependency relationships are established. The terminology association relationships are transformed into a triple structure for storage. The pre-trained language model is invoked to perform contextual semantic encoding on term content and contextual data, generating semantic representation vectors. The semantic representation vectors are then enhanced with contextual fragments to extract semantic features. Based on the terminology knowledge graph, scene features, device association features and operation type features are extracted, and gating fusion is performed on semantic features, scene features, device association features and operation type features. The terminology embedding is generated by combining co-occurrence frequency, semantic similarity, physical topology correlation and instruction temporal span. The terms corresponding to the newly added terms are embedded into the scene model adaptation layer, and incremental adjustments are made to the scene model adaptation layer to form the reconstructed scene model. The embedded features corresponding to the reconstructed scene model are extracted, and spatial alignment and dynamic weight fusion are performed based on the node connection relationship in the terminology knowledge graph to output standard terminology results.

[0021] In this embodiment, forming the terminology database specifically includes: The process involves: first, compiling scheduling regulations, operating procedures, and equipment descriptions to form a raw terminology dataset; second, dividing the raw terminology dataset into text units, extracting term fragments, and forming candidate terminology sequences; third, unifying the expression of candidate terminology sequences to generate standardized terminology sequences; fourth, establishing terminology identifiers within standardized terminology sequences, recording source type identifiers and timestamps, and forming an indexed terminology dataset; fifth, constructing context fragment sets at the corresponding positions of terminology identifiers to generate terminology context pairs; sixth, generating co-occurrence count sequences within terminology context pairs, and writing the co-occurrence count results at the terminology identifier positions; seventh, establishing semantic vector recording units at the terminology identifier positions to generate terminology semantic vectors; and finally, assembling the terminology identifiers, standardized terminology content, context fragment sets, co-occurrence count sequences, semantic vectors, and source type identifiers to form a terminology library.

[0022] Specifically, the unified processing of candidate term sequences includes: The process involves several steps: First, performing character format checks on each term fragment in the candidate term sequence to remove differences in spaces, full-width / half-width characters, capitalization, and invalid separators, resulting in a uniformly formatted term sequence. Second, performing synonym replacement on each term fragment in the uniformly formatted term sequence to map abbreviations, full names, alternative names, and common spellings with consistent meanings to uniform term content, resulting in a synonym-merged term sequence. Third, performing word form correction on each term fragment in the synonym-merged term sequence to unify the spelling of numbers, symbols, units of measurement, and equipment numbers, resulting in a rule-unified term sequence. Fourth, performing word order correction on each term fragment in the rule-unified term sequence to rearrange term fragments with inconsistent prefix / suffix order or modifier positions, resulting in a word-order-unified term sequence. Fifth, performing ambiguity checks on each term fragment in the word-order-unified term sequence to distinguish term fragments with multiple meanings from contextual fragment sets, generating semantically marked term fragments. Sixth, performing duplicate merging on semantically marked term fragments to generate a unique term entry when both the term content and semantic markings are consistent. The unique term entries are written into the standard term record unit in the order of term identifier number to form a standardized term sequence.

[0023] Specifically, constructing a set of contextual fragments at the corresponding positions of term identifiers includes: The process involves: 1) Reading the text unit content corresponding to each term identifier number from the original terminology dataset, locating the occurrence position of the term content within the text unit, and generating a term occurrence position sequence; 2) Extracting pre-defined length preceding and following text fragments before and after each term occurrence position to generate local context fragments; 3) Performing boundary checks on the local context fragments, deleting invalid content exceeding the start or end position of the text unit to form valid context fragments; 4) Cleaning up irrelevant symbols, duplicate separators, and blank placeholders in the valid context fragments to form standardized context fragments; 5) Retaining device name terms, operation instruction terms, status description terms, and time description terms that are directly related to the term content in the standardized context fragments to form associated context fragments; 6) Categorizing and arranging the associated context fragments corresponding to the same term identifier number according to the source text type and time stamp to generate a context fragment sequence; 7) Performing deduplication on the context fragment sequence, merging context fragments with consistent content and source type to form a context fragment set corresponding to the term identifier.

[0024] Specifically, establishing a semantic vector record unit at the term identifier number position includes: The process involves: retrieving standardized term content and contextual fragment sets corresponding to each term identifier number from the terminology database to generate a term semantic input sequence; performing term segmentation and sequential encoding on the standardized term content in the term semantic input sequence to form a term content representation sequence; parsing each context fragment in the context fragment set to extract device name terms, operation instruction terms, status description terms, and time description terms to form a context feature sequence; assembling the term content representation sequence and the context feature sequence to generate a term semantic combination sequence; performing unified dimension mapping on the term semantic combination sequence to form an initial semantic vector; performing length correction and numerical normalization on the initial semantic vector to form a standard semantic vector; establishing semantic vector recording units at the term identifier number positions and writing the standard semantic vectors into the corresponding recording positions; and performing index association on the semantic vector recording results corresponding to each term identifier number to form a one-to-one correspondence between term identifier numbers and semantic vectors.

[0025] In this embodiment, forming structured terminology data includes the following steps: The standardized term sequences in the terminology database are divided into fields, generating term content fields, source type fields, and time stamp fields. A field index mapping table is established at the term identifier number position, recording the corresponding field content at each index position. The context fragment set is position-labeled, generating a context index sequence at the term identifier number position. The co-occurrence count sequence is numerically normalized, generating a co-occurrence weight sequence at the term identifier number position. The semantic vector record unit is dimensionally normalized, generating a unified dimension semantic vector at the term identifier number position. The source type identifier is categorically encoded, generating a type encoding sequence at the term identifier number position. The term identifier number, term content field, field index mapping table, context index sequence, co-occurrence weight sequence, unified dimension semantic vector, and type encoding sequence are combined and mapped to form structured terminology data.

[0026] Specifically, the location marking of the set of context fragments includes: The process involves: reading the set of contextual fragments corresponding to each terminology identifier number; extracting the start and end character positions of each contextual fragment within the original text unit to form a fragment position sequence; recording the text unit number, paragraph number, and sentence number corresponding to each contextual fragment in the fragment position sequence to form a position identifier sequence; locating the position of the terminology content in each contextual fragment to generate a terminology center position value; calculating the forward interval between the start position of the contextual fragment and the terminology center position, and the backward interval between the end position of the contextual fragment and the terminology center position, based on the terminology center position value, to form a relative position value sequence; assembling the position identifier sequence and the relative position value sequence to generate contextual position recording results; establishing a contextual index record unit at the terminology identifier number position and writing the contextual position recording results to the corresponding position; and arranging the contextual position recording results corresponding to the same terminology identifier number in the order of text unit number, paragraph number, sentence number, and start character position to generate a contextual index sequence.

[0027] Specifically, the dimensionality normalization of semantic vector recording units includes: Read the semantic vector records corresponding to each term identifier number, extract the dimension length values ​​of each semantic vector, and form a dimension length sequence; determine a uniform dimension value in the dimension length sequence and write the uniform dimension value into the dimension parameter recording unit; for semantic vectors with a dimension length less than the uniform dimension value, add zero-value components to the end of the vector according to the dimension difference to generate an extended semantic vector; for semantic vectors with a dimension length greater than the uniform dimension value, delete redundant components to the end of the vector according to the dimension difference to generate a truncated semantic vector; perform numerical range correction on the extended semantic vectors and truncated semantic vectors to form a regularized semantic vector; establish a regularized vector recording unit at the term identifier number position and write the regularized semantic vector into the corresponding position; perform sequential index update on the set of regularized semantic vectors to form a uniform dimension semantic vector sequence.

[0028] In this embodiment, constructing a terminology knowledge graph based on structured terminology data specifically includes: The process involves: reading term identifiers from structured terminology data to generate term node sets; extracting device name, operation instruction, and scene description fields from the structured terminology data to generate device node sets, operation node sets, and scene node sets, respectively; establishing connections between terms and devices based on co-occurrence positions recorded in the context index sequence between the term node sets and device node sets, generating term-device relationship edges; establishing associations between terms and operations based on operation instruction field content between the term node sets and operation node sets, generating term-operation relationship edges; establishing associations between terms and scenes based on source type identifiers and context fragment sets between the term node sets and scene node sets, generating term-scene relationship edges; and establishing synonym and hierarchical relationship edges between terms within the term node sets based on similarity calculations between semantic vectors of a unified dimension. Specifically, semantic features, scene features, device association features, and operation types include: Semantic features are composed of a unified-dimensional semantic vector and semantic similarity values. The unified-dimensional semantic vector is derived from semantic vector record units in the terminology database, and its dimension has been normalized during the structuring process; the semantic similarity values ​​are derived from the semantic similarity calculation results between node pairs in the terminology association graph.

[0029] Scene features consist of a source type identifier and a context index sequence. The source type identifier reflects the text category to which the term belongs, including scheduling regulations, operating procedures, equipment instructions, etc.; the context index sequence reflects the contextual position of the term in the original text.

[0030] The device association feature is composed of node connection relationships and device field content. Node connection relationships are derived from the set of connection edges in the terminology association graph and are used to represent the association structure between terms; device field content is derived from the device name, device connection relationships, and device hierarchy information recorded in the field index mapping table.

[0031] The operation type feature is composed of a term content field and a type encoding sequence. The term content field is derived from the term content field in the structured term data; the type encoding sequence is derived from the operation category identifier obtained after classifying the term.

[0032] In this embodiment, gating fusion of semantic features, scene features, device association features, and operation type features includes the following steps: To construct a semantic feature vector, the following steps are performed: First, read the unified-dimensional semantic vector, type encoding sequence, context index sequence, co-occurrence weight values, and semantic similarity values ​​from the connection edge set at the corresponding position in the terminology node set. Second, read the source type identifier and context fragment set at the corresponding position in the terminology node set, extract the corresponding field content from the field index mapping table, and combine and encode the source type identifier and context fragment set to construct a scene feature vector. Third, read the connection relationship between node pairs at the corresponding position in the connection edge set, extract the corresponding field content of the device name between node pairs, establish a device mapping relationship in the field index mapping table, and numerically encode the device mapping relationship to construct a device association feature vector. Fourth, read the terminology content field at the corresponding position in the terminology node set, extract operation instruction terms from the field index mapping table, classify the operation instruction terms, generate operation category codes at the terminology identifier number position, and construct an operation type feature vector. Finally, establish a feature combination matrix at the terminology identifier number position and write the semantic feature vector, scene feature vector, device association feature vector, and operation type feature vector into the corresponding positions of the feature combination matrix. A gating parameter recording unit is established in the feature combination matrix. Initial gating weight values ​​are written to the corresponding positions of each feature vector. The gating weight values ​​are normalized to generate a gating weight sequence. Weighted calculations are performed on each feature vector in the feature combination matrix according to the gating weight sequence, and a fused feature vector is generated at the term identifier number position. During the generation of the fused feature vector, joint constraint calculations are performed on the gating weight values ​​and the co-occurrence weight values ​​in the connection edge set, and the gating weight values ​​are updated at the term identifier number position. The fused feature vector is output at the term identifier number position, forming the gating fusion result.

[0033] The specific calculation process is as follows: For feature input, suppose we have four feature vectors: semantic features, scene features, device-related features, and operation type features. Semantic features s , Scene feature c , Device association feature e , Operation type feature o Where d represents the dimension of the feature vector, Gating mechanism calculation: Each feature vector will be weighted through a gating mechanism. Assume the gating weight vector for each feature is: Semantic feature gating weights , Scene feature gating weights , Device-related feature gating weight , Operation type feature gating weight , The gating mechanism weights each feature to obtain a weighted feature. : ; ; ; ; in, The activation function for the gated weights is typically the sigmoid function, which is: ; This represents the element-wise multiplication operation. Feature-weighted fusion: After the gating operation, the weighted features are fused to generate the final fused feature vector. f=

[0034] in, represents the hyperparameters, indicating the importance or weight of each feature. These weights can be obtained through training to adaptively adjust the contribution of different features to the final fused vector.

[0035] The final fusion result output: After gating and weighted combination, the resulting fusion feature vector f represents the comprehensive features of the term, which can fully reflect the multidimensional information of the term in terms of semantics, scenario, device association, and operation type. This feature vector can be used as input for subsequent tasks (such as semantic retrieval or knowledge reasoning).

[0036] Specifically, the combination and encoding of source type identifiers and context fragment sets to construct scene feature vectors includes: Read the source type identifier corresponding to each term identifier number and map the source type identifier to a fixed-length type encoding vector; extract the source text type and time stamp corresponding to each context fragment from the context fragment set, perform corresponding encoding on the source text type and time stamp to generate a context attribute encoding sequence; perform term parsing on each context fragment in the context fragment set to extract device name terms, operation instruction terms, and status description terms to form a context content feature sequence; perform corresponding assembly on the context attribute encoding sequence and the context content feature sequence to generate a context encoding vector at the corresponding position of each context fragment; perform sequential arrangement and merging on each context encoding vector corresponding to the same term identifier number to form a context combination vector; perform concatenation processing on the type encoding vector and the context combination vector under the same dimensional structure to generate an initial scene feature vector; perform numerical normalization and dimensional regularization on the initial scene feature vector to form a standard scene feature vector; write the standard scene feature vector at the term identifier number position to complete the construction of the scene feature vector.

[0037] Specifically, the joint constraint calculation of the gating weight values ​​and the co-occurrence weight values ​​in the connection edge set during the fusion feature vector generation process includes: Read the gate weight values ​​corresponding to each feature vector in the feature combination matrix and the co-occurrence weight values ​​of the corresponding nodes in the edge set to form a weight pair sequence; perform difference and product calculations on the weight pair sequence at the corresponding positions to generate difference and product sequences; perform addition calculations on the difference and product sequences at the corresponding positions to generate a weight correction value sequence; write the weight correction value sequence into the corresponding positions of the gate weight values ​​to generate a corrected gate weight sequence; perform sum calculations on the corrected gate weight sequence, divide each value by the sum value to generate a normalized gate weight sequence; read the normalized gate weight sequence from the feature combination matrix, perform weighted summation calculations on each feature vector to generate a fused feature vector.

[0038] In this embodiment, the generation of term embeddings by combining common occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span specifically includes: The system extracts co-occurrence weight values ​​at corresponding positions in the connection edge set, extracts fusion feature vectors at corresponding positions in the term node set, calculates semantic similarity values ​​between node pairs, and generates a semantic similarity sequence. It also extracts device name, device connection, and device level field content from the field index mapping table, performs corresponding matching on these fields to generate a device connection path sequence, and performs level conversion, adjacency conversion, and path length conversion on the device connection path sequence to generate a physical topology correlation sequence. Finally, it extracts operation instruction term sequences and time stamp fields from the context fragment set, accumulates the time interval values ​​between adjacent operation instructions, and generates an instruction temporal span sequence. An embedded factor record matrix is ​​established at the term identifier number position. The co-occurrence weight value, semantic similarity sequence, physical topological correlation sequence, and instruction temporal span sequence are written into the corresponding positions of the embedded factor record matrix. The scale uniform processing of each factor value in the embedded factor record matrix is ​​performed to generate a standard factor sequence. A factor combination order is established at the corresponding position of the standard factor sequence, and a combined factor vector is formed according to the arrangement order of co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span. A fusion feature vector is introduced at the corresponding position of the combined factor vector. The combined factor vector and the fusion feature vector are concatenated and mapped to generate a term embedding input vector. An embedding parameter recording unit is established at the term identifier number position. The term embedding input vector is written into the embedding parameter recording unit. Linear transformation calculation and nonlinear compression calculation are performed on the term embedding input vector to generate an initial embedding vector. Adjacency propagation calculation is performed on the initial embedding vector in combination with the node connection relationship in the edge set to generate an association propagation vector. A weighted combination calculation is performed on the initial embedding vector and the association propagation vector to form the term embedding result.

[0039] Specifically, the process of performing hierarchical conversion, adjacency conversion, and path length conversion on the device connection path sequence to generate a physical topology correlation sequence includes: The process involves several steps: First, reading the node sequence corresponding to each path in the device connection path sequence. This involves extracting the device hierarchy number corresponding to the starting and ending nodes of each path, calculating the difference between these hierarchy numbers, and generating a hierarchy difference sequence. Then, performing a reciprocal calculation on the hierarchy difference sequence at the corresponding positions to generate a hierarchy association value sequence. Next, reading the number of adjacent node pairs corresponding to each path in the device connection path sequence, counting the number of directly connected node pairs in the path, and generating an adjacency count sequence. Then, performing normalization on the adjacency count sequence at the corresponding positions to generate an adjacency association value sequence. Finally, reading the total number of nodes corresponding to each path in the device connection path sequence, calculating the path length value by subtracting one from the number of nodes in the path, and generating a path length sequence. Then, performing a reciprocal calculation on the path length sequence at the corresponding positions to generate a path length association value sequence. Finally, performing weighted summation on the hierarchy association value sequence, the adjacency association value sequence, and the path length association value sequence at the corresponding positions to generate a physical topology association degree sequence. Finally, performing numerical range normalization on the physical topology association degree sequence to form a physical topology association degree sequence with a uniform scale. The specific calculation process is as follows: Assume the set of device nodes in the power grid is V={ , where each node An edge represents a device, and devices are connected by edges E, representing the physical connection between devices. The weight of an edge represents the connection strength or quality between devices, which can be determined by physical distance and the collaborative working ability between network bandwidths.

[0040] Define a topology graph G(V,E), where V represents the set of device nodes and E represents the set of device connection edges. the term Corresponding equipment The formula for calculating the physical topological correlation factor is as follows: . in, It is a term The device Physical topological correlation factor; It is equipment , The connection weight between devices is usually related to factors such as the physical connection strength between devices, network bandwidth, and failure rate. It is equipment The physical distance between them It is a hyperparameter used to control the degree to which distance affects physical topological correlation. It is equipment The hierarchical relationship between factors increases with the size of the factor. To ensure that the calculated physical topology correlation factor is adaptable to power grid systems of different sizes, it is usually necessary to normalize the results. The maximum correlation is set to... Normalized physical topological correlation factor The calculation formula is as follows: ; It is the maximum physical topology correlation factor for all terminology device nodes, used to ensure that the physical topology correlation of different terms is within the same range.

[0041] Specifically, performing adjacency propagation calculations on the initial embedding vector and the node connection relationships in the edge set includes: Read the initial embedding vectors corresponding to each term node and form an initial embedding vector sequence according to the term identifier number; read the start node number and end node number corresponding to each connection edge in the connection edge set to form a node connection relationship table; count the number of adjacent nodes corresponding to each term node in the node connection relationship table to generate an adjacency count sequence; perform a dimension-wise summation calculation on the initial embedding vectors of adjacent nodes corresponding to each term node to generate an adjacency cumulative vector; divide the adjacency cumulative vector by the corresponding adjacency count value to generate an adjacency average vector; read the co-occurrence weight value and semantic similarity value corresponding to each connection edge in the connection edge set, perform a product calculation on the co-occurrence weight value and semantic similarity value to generate a connection weight sequence; multiply the initial embedding vectors corresponding to each adjacent node by the corresponding connection weight value and perform a dimension-wise summation calculation to generate a weighted propagation vector; divide the weighted propagation vector by the corresponding sum of connection weights to generate an adjacency propagation result vector; write the adjacency propagation result vector into the corresponding position of each term identifier number to form an association propagation vector.

[0042] In this embodiment, performing incremental adjustments to the scene model adaptation layer includes the following steps: Extract the term embedding results corresponding to the newly added terms to form an incremental embedding sequence; establish an incremental write position table in the scene model adaptation layer to record the correspondence between scene category identifiers and incremental embedding sequences; write the incremental embedding sequence to the corresponding position in the scene model adaptation layer to generate incremental feature recording units; extract the original scene parameters at the corresponding position of the incremental feature recording unit and calculate the parameter offset sequence between the incremental embedding sequence and the original scene parameters; map the parameter offset sequence to generate a local update parameter sequence; overwrite and update the local update parameter sequence to generate an incremental parameter sequence; write the incremental parameter sequence to the scene model adaptation layer to form the reconstructed scene model.

[0043] The mapping of parameter offset sequences to generate locally updated parameter sequences specifically includes: reading the incremental embedding sequences corresponding to each newly added term in the incremental feature recording unit, and the original scene parameters at the corresponding positions, performing subtraction calculations at the corresponding positions to generate parameter offset sequences; performing absolute value calculations on each parameter offset value in the parameter offset sequences to generate offset magnitude sequences; performing positive and negative direction determination on each parameter offset value in the parameter offset sequences to generate offset direction sequences; multiplying the offset magnitude sequences and offset direction sequences at the corresponding positions to generate direction-preserving offset sequences; reading the parameter values ​​in the original scene parameters, performing addition calculations on the direction-preserving offset sequences and the original scene parameters at the corresponding positions to generate locally updated parameter sequences; performing parameter boundary checks on the locally updated parameter sequences, writing parameter values ​​exceeding the preset upper limit as upper limit values, and writing parameter values ​​below the preset lower limit as lower limit values, forming a restricted locally updated parameter sequence; and writing the restricted locally updated parameter sequence into the locally updated parameter recording position to form a locally updated parameter sequence.

[0044] In this embodiment, extracting the embedding features corresponding to the reconstructed scene model specifically includes: In the reconstructed scene model, locate the parameter positions corresponding to each scene category identifier and generate a feature extraction position sequence; extract incremental parameter sequences and original scene parameters at the corresponding positions of the feature extraction position sequence to form a scene parameter combination sequence; expand the scene parameter combination sequence dimensionally to generate a scene feature matrix; extract term identifier numbers and incremental embedding sequences at the corresponding positions of the scene feature matrix to form a scene-related feature sequence; combine and map the scene feature matrix and the scene-related feature sequence to generate an initial embedding feature sequence; perform dimensional normalization on the initial embedding feature sequence to form a unified dimensional embedding feature sequence; output the unified dimensional embedding feature sequence to form the embedding features corresponding to the reconstructed scene model.

[0045] Example 1: To verify the feasibility of this invention in practice, it was applied to a semantic retrieval scenario for power grid regulations. In this scenario, dispatching regulations, operating procedures, equipment descriptions, and historical revisions coexist, and different texts often use different names for the same equipment, operation, or state. When searching, users frequently input habitual terms, abbreviations, or older terms instead of the standard terminology in the regulations, easily leading to problems such as no search results, scattered search results, and inaccurate returned clauses. Especially in scenarios involving equipment switching, operation mode adjustment, and maintenance isolation, the terminology not only differs literally but is also related to the upstream and downstream relationships of equipment, the order of operations, and the context. Traditional keyword retrieval struggles to take all these factors into account simultaneously, thus making it difficult to achieve accurate and stable returns of standard terminology.

[0046] In this implementation, multiple types of institutional texts are first collected, candidate terms are extracted, and a unified expression processing is performed on the candidate terms, merging abbreviations, full names, old names, and conventional spellings into a unified expression to form a terminology database. Subsequently, the terminology database undergoes field segmentation, context labeling, co-occurrence weight calculation, semantic vector normalization, and type encoding to form structured terminology data. The current terminology database is then compared with historical terminology databases to identify newly added terms, semantically drifting terms, and discontinued terms. Next, a terminology association graph is built based on the structured terminology data, extracting semantic features, scenario features, equipment association features, and operation type features, and fusion is completed through gating. Further, terminology embeddings are generated by combining co-occurrence frequency, semantic similarity, physical topological association, and instruction temporal span. The representation results of newly added terms are written into the scenario model adaptation layer, and incremental adjustments are performed on the scenario model adaptation layer to form a reconstructed scenario model. After users input search content, the system extracts the embedding features corresponding to the reconstructed scenario model, completes spatial alignment and dynamic weight fusion, and outputs the standard terminology results and the corresponding institutional clause positions.

[0047] In practical verification, a total of 1260 documents, including regulatory texts, procedure texts, equipment instruction texts, and historical texts, were selected. 18420 candidate terms were extracted, and after unifying their expressions, 8016 standardized terms were formed. 392 newly added terms, 241 semantically drifting terms, and 176 discontinued terms were identified. 900 sets of retrieval test samples were constructed, including 580 sets of non-standard expression samples and 320 sets of standard expression samples. Compared with traditional keyword retrieval methods, this invention significantly improves the hit rate of standard terms, the accuracy of first-result results, the accuracy of equipment-related terminology identification, and the success rate of new terminology adaptation. Simultaneously, the average retrieval time and manual review time are significantly reduced. This demonstrates that this invention can effectively solve the problems of inconsistent terminology expression, semantic changes with version variations, complex equipment relationships, and the difficulty in timely adaptation of newly added terms in power grid regulations, and has good practicality.

[0048] Table 1: Comparison of the Implementation Effects of Semantic Retrieval of Power Grid Regulations and Rules

[0049] As can be seen from Table 1, this invention outperforms traditional keyword retrieval methods in several core indicators. First, in terms of terminology organization, although the traditional method also processed 1260 texts and 18420 candidate terms, it only resulted in 7058 standardized terms, while this invention resulted in 8016, an increase of 958, indicating that this invention has a stronger ability to unify synonyms, abbreviations, and old names.

[0050] In terms of terminology change identification, this invention identifies 392 newly added terms, 241 semantically drifting terms, and 176 discontinued terms, which is 274, 172, and 118 more than the traditional method, respectively, indicating that this invention responds more fully to terminology changes brought about by regulatory revisions.

[0051] In terms of search performance, the hit rate of standard terms increased from 79.1% to 93.2%, the accuracy of the first result increased from 73.6% to 90.7%, and the coverage of the top three results increased from 82.4% to 96.5%, indicating that the present invention can not only retrieve results, but also rank the correct results in a higher position.

[0052] For equipment association identification, which is particularly critical in power grid operations, the accuracy rate of this invention reaches 89.9%, 19.7 percentage points higher than traditional methods; the accuracy rate for operation type discrimination reaches 91.8%, an improvement of 17.0 percentage points. Meanwhile, the success rate of new terminology adaptation increased from 57.4% to 91.3%, indicating that this invention can complete model adaptation more quickly when new terms enter the terminology database. The false return rate of deactivated terms decreased from 17.6% to 4.9%, a reduction of 12.7 percentage points, indicating that the interference of invalid old terms on search results is significantly reduced.

[0053] In terms of search efficiency, the average search time decreased from 1.79 seconds to 0.98 seconds, a reduction of 0.81 seconds; the manual review time decreased from 24.8 seconds to 10.6 seconds, a reduction of 14.2 seconds; and the number of redirects to erroneous entries decreased from 129 to 27, a reduction of 102.

[0054] The stable output rate after continuous revisions reached 93.8%, an improvement of 24.4 percentage points compared to traditional methods. These data demonstrate that this invention has significant advantages in terms of terminology standardization, dynamic adaptation, retrieval accuracy, and output stability.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs, characterized in that: Includes the following steps: Collect scheduling terminology and contextual data to form a terminology database; The terminology database is cleaned, standardized, and labeled to form structured terminology data; The current terminology database is compared with the historical terminology database to identify newly added terms, semantically drifted terms, and discontinued terms. A terminology knowledge graph is constructed based on structured terminology data. Term nodes, device nodes, operation nodes, and scenario nodes are defined, and synonym relationships, hierarchical relationships, device connection relationships, and operation dependency relationships are established. The terminology association relationships are transformed into a triple structure for storage. The pre-trained language model is invoked to perform contextual semantic encoding on term content and contextual data, generating semantic representation vectors. Then, the semantic representation vectors are enhanced with contextual fragments to extract semantic features. Based on the terminology knowledge graph, scene features, device association features and operation type features are extracted, and gating fusion is performed on semantic features, scene features, device association features and operation type features. The terminology embedding is generated by combining co-occurrence frequency, semantic similarity, physical topology correlation and instruction temporal span. The terms corresponding to the newly added terms are embedded into the scene model adaptation layer, and incremental adjustments are made to the scene model adaptation layer to form the reconstructed scene model. The embedded features corresponding to the reconstructed scene model are extracted, and spatial alignment and dynamic weight fusion are performed based on the node connection relationship in the terminology knowledge graph to output standard terminology results.

2. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 1, characterized in that, The formation of the terminology database specifically includes: The process involves: first, compiling scheduling regulations, operating procedures, and equipment descriptions to form a raw terminology dataset; second, dividing the raw terminology dataset into text units, extracting term fragments, and forming candidate terminology sequences; third, unifying the expression of candidate terminology sequences to generate standardized terminology sequences; fourth, establishing terminology identifiers within standardized terminology sequences, recording source type identifiers and timestamps, and forming an indexed terminology dataset; fifth, constructing context fragment sets at the corresponding positions of terminology identifiers to generate terminology context pairs; sixth, generating co-occurrence count sequences within terminology context pairs, and writing the co-occurrence count results at the terminology identifier positions; seventh, establishing semantic vector recording units at the terminology identifier positions to generate terminology semantic vectors; and finally, assembling the terminology identifiers, standardized terminology content, context fragment sets, co-occurrence count sequences, semantic vectors, and source type identifiers to form a terminology library.

3. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 2, characterized in that, The formation of structured terminology data includes the following steps: The standardized term sequences in the terminology database are divided into fields, generating term content fields, source type fields, and time stamp fields. A field index mapping table is established at the term identifier number position, recording the corresponding field content at each index position. The context fragment set is position-labeled, generating a context index sequence at the term identifier number position. The co-occurrence count sequence is numerically normalized, generating a co-occurrence weight sequence at the term identifier number position. The semantic vector record unit is dimensionally normalized, generating a unified dimension semantic vector at the term identifier number position. The source type identifier is categorically encoded, generating a type encoding sequence at the term identifier number position. The term identifier number, term content field, field index mapping table, context index sequence, co-occurrence weight sequence, unified dimension semantic vector, and type encoding sequence are combined and mapped to form structured terminology data.

4. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 3, characterized in that, The construction of a terminology knowledge graph based on structured terminology data specifically includes: The process involves: reading term identifiers from structured terminology data to generate term node sets; extracting device name, operation instruction, and scene description fields from the structured terminology data to generate device node sets, operation node sets, and scene node sets, respectively; establishing connections between terms and devices based on co-occurrence positions recorded in the context index sequence between the term node sets and device node sets, generating term-device relationship edges; establishing associations between terms and operations based on operation instruction field content between the term node sets and operation node sets, generating term-operation relationship edges; establishing associations between terms and scenes based on source type identifiers and context fragment sets between the term node sets and scene node sets, generating term-scene relationship edges; establishing synonym and hierarchical relationship edges between terms within the term node sets based on similarity calculations between unified-dimensional semantic vectors; performing triple transformation on the term node sets, device node sets, operation node sets, scene node sets, and each relationship edge to represent each node and relationship as a "node-relation-node" triple structure; and performing unified mapping on all triple structures to form a terminology knowledge graph.

5. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 4, characterized in that, The gating fusion of semantic features, scene features, device association features, and operation type features includes the following steps: To construct a semantic feature vector, the following steps are performed: First, read the unified-dimensional semantic vector, type encoding sequence, context index sequence, co-occurrence weight values, and semantic similarity values ​​from the connection edge set at the corresponding position in the terminology node set. Second, read the source type identifier and context fragment set at the corresponding position in the terminology node set, extract the corresponding field content from the field index mapping table, and combine and encode the source type identifier and context fragment set to construct a scene feature vector. Third, read the connection relationship between node pairs at the corresponding position in the connection edge set, extract the corresponding field content of the device name between node pairs, establish a device mapping relationship in the field index mapping table, and numerically encode the device mapping relationship to construct a device association feature vector. Fourth, read the terminology content field at the corresponding position in the terminology node set, extract operation instruction terms from the field index mapping table, classify the operation instruction terms, generate operation category codes at the terminology identifier number position, and construct an operation type feature vector. Finally, establish a feature combination matrix at the terminology identifier number position and write the semantic feature vector, scene feature vector, device association feature vector, and operation type feature vector into the corresponding positions of the feature combination matrix. A gating parameter recording unit is established in the feature combination matrix. Initial gating weight values ​​are written to the corresponding positions of each feature vector. The gating weight values ​​are normalized to generate a gating weight sequence. Weighted calculations are performed on each feature vector in the feature combination matrix according to the gating weight sequence, and a fused feature vector is generated at the term identifier number position. During the generation of the fused feature vector, joint constraint calculations are performed on the gating weight values ​​and the co-occurrence weight values ​​in the connection edge set, and the gating weight values ​​are updated at the term identifier number position. The fused feature vector is output at the term identifier number position, forming the gating fusion result.

6. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 5, characterized in that, The specific method for generating term embeddings by combining co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span includes: The system extracts co-occurrence weight values ​​at corresponding positions in the connection edge set, extracts fusion feature vectors at corresponding positions in the term node set, calculates semantic similarity values ​​between node pairs, and generates a semantic similarity sequence. It also extracts device name, device connection, and device level field content from the field index mapping table, performs corresponding matching on these fields to generate a device connection path sequence, and performs level conversion, adjacency conversion, and path length conversion on the device connection path sequence to generate a physical topology correlation sequence. Finally, it extracts operation instruction term sequences and time stamp fields from the context fragment set, accumulates the time interval values ​​between adjacent operation instructions, and generates an instruction temporal span sequence. An embedded factor record matrix is ​​established at the term identifier number position. The co-occurrence weight value, semantic similarity sequence, physical topological correlation sequence, and instruction temporal span sequence are written into the corresponding positions of the embedded factor record matrix. The scale uniform processing of each factor value in the embedded factor record matrix is ​​performed to generate a standard factor sequence. A factor combination order is established at the corresponding position of the standard factor sequence, and a combined factor vector is formed according to the arrangement order of co-occurrence frequency, semantic similarity, physical topological correlation, and instruction temporal span. A fusion feature vector is introduced at the corresponding position of the combined factor vector. The combined factor vector and the fusion feature vector are concatenated and mapped to generate a term embedding input vector. An embedding parameter recording unit is established at the term identifier number position. The term embedding input vector is written into the embedding parameter recording unit. Linear transformation calculation and nonlinear compression calculation are performed on the term embedding input vector to generate an initial embedding vector. Adjacency propagation calculation is performed on the initial embedding vector in combination with the node connection relationship in the edge set to generate an association propagation vector. A weighted combination calculation is performed on the initial embedding vector and the association propagation vector to form the term embedding result.

7. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 6, characterized in that, The incremental adjustment of the scene model adaptation layer includes the following steps: Extract the term embedding results corresponding to the newly added terms to form an incremental embedding sequence; establish an incremental write position table in the scene model adaptation layer to record the correspondence between scene category identifiers and incremental embedding sequences; write the incremental embedding sequence to the corresponding position in the scene model adaptation layer to generate incremental feature recording units; extract the original scene parameters at the corresponding position of the incremental feature recording unit and calculate the parameter offset sequence between the incremental embedding sequence and the original scene parameters; map the parameter offset sequence to generate a local update parameter sequence; overwrite and update the local update parameter sequence to generate an incremental parameter sequence; write the incremental parameter sequence to the scene model adaptation layer to form the reconstructed scene model.

8. The semantic retrieval method for power grid regulations based on a large-scale language model using knowledge graphs according to claim 7, characterized in that, The specific embedded features extracted from the reconstructed scene model include: In the reconstructed scene model, locate the parameter positions corresponding to each scene category identifier and generate a feature extraction position sequence; extract incremental parameter sequences and original scene parameters at the corresponding positions of the feature extraction position sequence to form a scene parameter combination sequence; expand the scene parameter combination sequence dimensionally to generate a scene feature matrix; extract term identifier numbers and incremental embedding sequences at the corresponding positions of the scene feature matrix to form a scene-related feature sequence; combine and map the scene feature matrix and the scene-related feature sequence to generate an initial embedding feature sequence; perform dimensional normalization on the initial embedding feature sequence to form a unified dimensional embedding feature sequence; output the unified dimensional embedding feature sequence to form the embedding features corresponding to the reconstructed scene model.

Images