Power grid knowledge graph dynamic updating method and device based on AI incremental learning

Abstract

The invention provides a power grid knowledge graph dynamic updating method and device based on AI incremental learning, and relates to the technical field of power grids. According to the method, the structured operation data and the unstructured text data are obtained in real time, the AI incremental analysis model is combined, online recognition and confidence evaluation of the new entity and the new relation are achieved, an updating proposal is generated only for the changed part in the knowledge graph, and high expenditure of full-amount reconstruction is avoided. And local updating and rationality verification of a vector space are carried out through a dynamic graph embedding model, so that the consistency of newly added knowledge in semantic and physical levels is ensured. The whole updating process is automatic, light and highly credible, the period from knowledge extraction from data to atlas landing is remarkably shortened, efficient updating and maintenance of the power grid knowledge atlas are achieved, the lag problem existing in a power grid knowledge atlas full-amount updating mode is solved, the response speed and the supporting capacity of the knowledge atlas are greatly improved, and the reliability of the power grid knowledge atlas is improved. And the timeliness and the accuracy of the power grid knowledge graph are enhanced.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and in particular to a method and apparatus for dynamically updating a power grid knowledge graph based on AI incremental learning. Background Technology

[0002] With the deepening of smart grid construction, the data on power grid equipment, operating status, topology connections, and user data are growing massively and changing dynamically. Knowledge graphs, as an important technology for representing and organizing power grid knowledge, can effectively model power grid entities (such as transformers, lines, and switches) and their complex relationships (such as connections, affiliations, and fault impacts), providing a reliable semantic knowledge foundation for advanced applications such as fault diagnosis, intelligent scheduling, and load forecasting.

[0003] However, current methods for constructing and maintaining power grid knowledge graphs mostly employ a full-update approach, which involves periodically rebuilding the entire power grid knowledge graph. Faced with the real-time, streaming growth of power grid data, this method consumes enormous computational resources and has a long update cycle, resulting in the knowledge graph lagging significantly behind the actual state of the power grid and making it difficult to guarantee its timeliness. Summary of the Invention

[0004] This invention provides a method and apparatus for dynamic updating of a power grid knowledge graph based on AI incremental learning, which solves the lag problem in the full update mode of the power grid knowledge graph, realizes efficient updating and maintenance of the power grid knowledge graph, and improves the timeliness and accuracy of the power grid knowledge graph.

[0005] In a first aspect, the present invention provides a method for dynamically updating a power grid knowledge graph based on AI incremental learning. The method includes: acquiring multi-source incremental data of the power grid, including structured operational data and unstructured text data; based on the multi-source incremental data and a pre-set AI incremental analysis model, performing new entity identification, new relationship extraction, and confidence assessment of newly added items to determine update proposals for the power grid knowledge graph; based on the update proposals and a pre-set dynamic graph embedding model, performing local dynamic updates and verification of the vector space to determine the verification results of the power grid knowledge graph, including the updated vector space and the probability of its reasonableness; and dynamically updating the power grid knowledge graph based on the verification results.

[0006] Secondly, embodiments of the present invention provide an AI-based incremental learning-based dynamic update device for a power grid knowledge graph. This update device includes a communication module and a processing module. The communication module is used to acquire multi-source incremental data of the power grid, including structured operational data and unstructured text data. The processing module is used to, based on the multi-source incremental data and a preset AI incremental analysis model, perform new entity identification, new relationship extraction, and confidence assessment of newly added items to determine an update proposal for the power grid knowledge graph. Based on the update proposal and a preset dynamic graph embedding model, it performs local dynamic updates and verification of the vector space to determine the verification result of the power grid knowledge graph, including the updated vector space and the probability of its reasonableness. Based on the verification result of the power grid knowledge graph, it dynamically updates the power grid knowledge graph.

[0007] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0008] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] This invention provides a method and apparatus for dynamically updating a power grid knowledge graph based on AI incremental learning. By acquiring structured operational data and unstructured text data in real time and combining it with an AI incremental analysis model, this invention achieves online identification and confidence assessment of new entities and relationships. Update proposals are generated only for the changed parts of the knowledge graph, avoiding the high overhead of full reconstruction. Furthermore, a dynamic graph embedding model is used for local updates and rationality verification of the vector space, ensuring the semantic and physical consistency of newly added knowledge. The entire update process is automated, lightweight, and highly reliable, significantly shortening the cycle from knowledge extraction from data to graph implementation. This enables efficient updating and maintenance of the power grid knowledge graph, solves the lag problem inherent in full-scale knowledge graph updates, and significantly improves the response speed and support capabilities of the knowledge graph, enhancing its timeliness and accuracy. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a method for dynamically updating a power grid knowledge graph based on AI incremental learning, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an AI-based incremental learning power grid knowledge graph dynamic update device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0013] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0014] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0015] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, this embodiment of the invention provides a method for dynamically updating a power grid knowledge graph based on AI incremental learning. The method includes steps S101-S104.

[0018] S101. Obtain multi-source incremental data of the power grid.

[0019] In this embodiment of the application, the multi-source incremental data includes structured runtime data and unstructured text data.

[0020] In some embodiments, structured operational data is real-time / near real-time streaming data; originating from monitoring systems, it is typically timestamped numerical values ​​or status signals and is the primary source triggering updates to entity status attributes in the knowledge graph. For example, SCADA (Supervisory Control and Data Acquisition) system data: real-time measurements of various devices in the power grid, such as active power, reactive power, current, and voltage values ​​of lines / transformers, and the open / closed status (0 / 1) of switches. This is used to update dynamic attributes such as currentLoad and status (running / disconnected) of relevant device entities (e.g., circuit breaker CB01) in the knowledge graph. WAMS / PMU (Wide Area Measurement System / Phasor Measurement Unit) data: higher-precision synchronous phasor data (voltage, phase angle) used to monitor the dynamic behavior of the power grid. Updating precise status attributes such as voltage and phase angle of key nodes can be used for subsequent knowledge reasoning regarding stability. OMS (Outage Management System) data: fault alarms, protection action signals, and user outage complaint information. New "fault" entities are created in the knowledge graph and associated with tripped switches, power-deprived lines, affected users, and other entities.

[0021] In some embodiments, unstructured text data, containing a wealth of implicit knowledge, is crucial for enriching the details of the knowledge graph and discovering new events, and needs to be extracted using NLP techniques (such as LLM). For example, dispatch logs and operation tickets contain dispatcher-issued operation instructions (e.g., "Close circuit breaker 1011 on line L1") and records of the operation reasons. The operation events are parsed, equipment status is updated, and the operation intent (e.g., reason: eliminate overload) is recorded. Inspection reports and maintenance work orders contain descriptions of equipment defects, fault handling processes, and maintenance measures recorded by maintenance personnel (e.g., "Oil leakage found in transformer A, flange tightened"). "Defect" or "Maintenance" event entities are created and associated with the equipment, updating the equipment's healthIndex. Emergency plans and technical procedures are rule texts describing how to operate under specific faults. The rule logic is extracted to construct or update "rule" entities and their associations in the knowledge graph, supporting reasoning applications.

[0022] In some embodiments, multi-source incremental data also includes structured models and topology data, used to describe the static model and connection relationships of the power grid. These change less frequently, but when they do, they represent fundamental topology updates. Examples include: PMS (Production Management System) / Asset Management data: equipment ledger information (such as transformer model, commissioning date, maintenance history), new equipment commissioning, and old equipment decommissioning. This involves adding / decommissioning equipment entities and updating their static attributes (such as lastMaintenanceDate). GIS (Geographic Information System) data: spatial coordinates of equipment and path information for overhead lines / cable corridors. This updates the spatialLocation attribute of entities in the knowledge graph (such as transmission line TL01), enriching their spatial relationship knowledge. Graph model data (CIM / G format power grid model): a topology model describing how substations, lines, switches, etc., are connected. This directly triggers the addition, deletion, and modification of topology relationships in the knowledge graph (e.g., a new connection is created between bus B and transformer T1).

[0023] S102. Based on multi-source incremental data and a pre-set AI incremental analysis model, new entity identification, new relationship extraction, and confidence assessment of new items are performed to determine the update proposal for the power grid knowledge graph.

[0024] In some embodiments, the AI ​​incremental analysis model is based on a composite artificial intelligence model combining deep learning and incremental learning techniques. Its core function is to act as the system's "perception and understanding brain," responsible for processing continuous, multimodal incremental power grid data and outputting structured knowledge proposals that can be subsequently verified and updated. The AI ​​incremental analysis model employs a multi-task learning architecture with shared underlying parameters, including a shared underlying encoder and multiple sub-models. These sub-models include a new entity recognition model, a new relationship recognition model, and a relationship evaluation model.

[0025] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1027.

[0026] S1021. Based on multi-source incremental data and AI incremental analysis model, feature extraction is performed to determine the basic features of incremental data.

[0027] In some embodiments, the basic features of incremental data include textual semantic features, numerical and state features, temporal and dynamic features, and graph context features.

[0028] For example, text semantic features include: entity-level embedding vectors (such as "main transformer A", "switch 102"), sentence / paragraph-level semantic vectors (such as "execute operation to switch from running to cold standby"), relation trigger word features (such as the encoding and location information of words such as "connect", "trip", "cut off"), and semantic role labeling features.

[0029] For example, numerical and status characteristics include: equipment status quantities (such as the open / closed status of equipment such as switches and disconnectors), operating measurement values ​​(such as the normalized values ​​or statistical characteristics of analog quantities such as active power, reactive power, current, and voltage of lines / transformers), event flag bits (such as binary event flags such as protection action signals, fault recording start signals, and alarm signals), and spatiotemporal characteristics (such as the timestamp of data generation and the spatial location code of the equipment).

[0030] For example, time series and dynamic features include: short-term time series patterns (such as the moving average, trend, volatility, etc. of measurements within a recent time window), comparison of features before and after an event (such as the amount and rate of change of relevant electrical quantities before and after an event), sequence embedding, etc.

[0031] For example, knowledge graph context features: These associate new data with existing knowledge graphs, providing important contextual information. Examples include: Neighborhood entity features: Encoding the types of neighboring entities and their relationships directly connected to the devices involved in the current data within the knowledge graph. Topological connectivity features: Characteristics representing the location of a device in the power grid topology (e.g., whether it is a hub node, degree centrality, etc.). Entity attribute features: Encoding the inherent attributes of relevant devices retrieved from the knowledge graph (e.g., rated capacity, commissioning date, model).

[0032] For example, embodiments of the present invention can use a shared underlying encoder in an AI incremental analysis model to process multi-source incremental data in parallel to obtain the basic features of the incremental data.

[0033] S1022. Based on the basic characteristics of incremental data and the AI ​​incremental analysis model, new entity identification is performed to determine the newly added power grid equipment entities.

[0034] For example, in embodiments of the present invention, incremental data basic features can be input into a new entity recognition model, and new power grid equipment entities can be identified through a conditional random field decoding layer. The entity type and boundary are then output to determine the new power grid equipment entities.

[0035] S1023. Based on the newly added power grid equipment entities and the entity list of the power grid knowledge graph, perform matching links and entity disambiguation to determine the linked entities of the newly added power grid equipment entities in the power grid knowledge graph.

[0036] For example, in embodiments of the present invention, the similarity between the identified new entities and existing entities in the knowledge graph can be calculated; based on the similarity, candidate linking entities are determined. Then, a disambiguation algorithm based on semantic embedding and topological structure dual constraints is used to identify ambiguous entities; for ambiguous entities, disambiguation decisions are made by combining contextual features and graph topology, and finally, the linking entities of the newly added power grid equipment entities are obtained.

[0037] S1024. Based on the basic characteristics of incremental data, add the linked entities of the new power grid equipment entities, and the AI ​​incremental analysis model to identify new relationships and determine the new associations.

[0038] For example, embodiments of the present invention can input incremental data basic features, newly added power grid equipment entities and linked entities into a new relationship recognition model, use a relationship-aware attention mechanism to capture semantic associations between entities, predict relationship types through a multi-class classifier, and determine the newly added association relationship.

[0039] S1025. Based on the existing relationships in the power grid knowledge graph, perform reliability calculations on newly added relationships to determine the confidence probability of the newly added relationships.

[0040] For example, in embodiments of the present invention, newly added relationships can be input into the relationship evaluation model, combined with existing relationships in the power grid knowledge graph, to perform reliability calculations and determine the confidence probability of the newly added relationships.

[0041] S1026. Based on the basic features of incremental data, new power grid equipment entities and new relationships are added. An attention mechanism is used to perform knowledge transfer and collaborative optimization between entities and relationships, and to determine the optimized new power grid equipment entities, new relationships and confidence probabilities.

[0042] For example, embodiments of the present invention can utilize an attention mechanism to achieve knowledge transfer and collaborative optimization between entities and relationships based on incremental data features, newly added power grid equipment entities, and newly added relationships. This process first constructs a cross-task attention fusion layer, using entity features as query vectors and relationship features as key-value vectors. Interaction weights are calculated through a multi-head attention mechanism to establish deep semantic associations between entities and relationships. Based on this, bidirectional knowledge transfer is achieved: forward transfer enhances the semantic integrity of relationship representations using entity features, while backward transfer optimizes the contextual relevance of entity representations using relationship features.

[0043] The collaborative optimization process employs a multi-task joint learning strategy. By designing a composite objective function that includes entity-relation consistency loss, cross-task contrastive loss, and orthogonality constraint loss, it ensures that the feature spaces of different tasks maintain both independence and consistency. Attention weights are used to dynamically reweight features, strengthening important feature dimensions and suppressing noise, while residual connections preserve the integrity of the original information. Finally, a multilayer perceptron generates the optimized output, including entity vectors with rich contextual information, relation vectors with precise semantic relationships, and a comprehensive confidence score fused from multiple dimensions.

[0044] The entire process also includes an internal consistency verification step, which performs logical consistency checks on the optimized entity types and relationship types, verifies the rationality of the confidence probability distribution, performs adaptive calibration based on the verification results, and finally outputs the optimized new power grid equipment entities, new associations and their corresponding confidence probabilities.

[0045] S1027. Based on the optimized newly added power grid equipment entities, newly added relationships, and confidence probabilities, determine the update proposal for the power grid knowledge graph.

[0046] S103. Based on the update proposal and the preset dynamic graph embedding model, perform local dynamic updates and verification of the vector space to determine the verification results of the power grid knowledge graph.

[0047] In some embodiments, the verification result includes the updated vector space and the probability of reasonableness of the updated vector space.

[0048] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1034.

[0049] S1031, Extract the proposal type of the update proposal.

[0050] In some embodiments, proposal types include new entity proposals, new relationship proposals, and entity attribute update proposals.

[0051] S1032. Based on the proposal type and the dynamic graph embedding model, the vector space of the power grid knowledge graph is dynamically updated to obtain the updated vector space.

[0052] For example, step S1032 can be specifically implemented as steps A1-A5.

[0053] A1. Based on the proposal type and the updated proposal, determine the adjustment items for the vector space.

[0054] In some embodiments, the adjustment project includes the static core vector and / or dynamic relation vector of the entity to be updated.

[0055] For example, embodiments of the present invention can analyze the content and type of update proposals to determine which parts of the vector space require minimal modifications.

[0056] The proposal type is "New Entity Proposal": The adjustments mainly include the static core vector of the new entity itself, as well as the relation vectors of several new relationships established due to its introduction. Fine-tuning of the vectors of existing entities directly connected to it may also be necessary.

[0057] The proposal type is "New Relationship Proposal": The main adjustment to the project is the relation vector of the new relation. It is usually also necessary to fine-tune the static core vectors of the head and tail entities at both ends of the relation so that the new relation (h,r,t) satisfies the scoring function (e.g., h+r≈t) in the vector space.

[0058] The proposal type is "Entity Attribute Update Proposal": An attribute can be considered a relationship. The main adjustments are to the static core vector of the entity (as the head entity) and the relation vector representing the attribute. If the attribute value is also an entity (the tail entity), then the vector of that tail entity may also need to be adjusted.

[0059] A2. Based on the updated proposal and the adjusted project, the vector space is dynamically adjusted locally to obtain the locally adjusted vector space.

[0060] For example, to avoid the huge overhead of global retraining, incremental learning or online learning strategies are used to optimize only the vectors involved in the adjustment project. This invention keeps the vectors of most existing entities and relations in the fixed vector space unchanged. Only newly added entities are randomly initialized with an initial vector, or an initial vector is generated based on their text description, type, or other information. Newly added relations are initialized with a relation vector.

[0061] Then, with the goal of "achieving a high reasonableness score for the triples in the new proposals," several rounds of gradient descent optimization are performed on a very small local subgraph. The optimization targets are limited to these newly initialized vectors and a small number of existing vectors that require fine-tuning. This process is efficient and fast, achieving "local dynamic adjustment" of the vector space.

[0062] A3. Based on the preset loss function, perform physical consistency verification on the locally adjusted vector space and determine the verification result.

[0063] In some embodiments, the preset loss function includes a regularization term composed of physical rules and logical constraints in the power grid domain.

[0064] For example, simply ensuring the new triplet holds is insufficient; it is also necessary to guarantee that the updated vector space does not violate the physical laws and logical constraints of the power grid domain. This is achieved through a carefully designed loss function.

[0065] Loss function composition: Standard graph embedding loss; Loss based on scoring functions, such as marginal ranking loss, to ensure that the score of correct triples is higher than that of incorrect triples.

[0066] Domain rule regularization: Power grid knowledge is transformed into mathematical constraints and added as penalty terms to the loss function. For example: Physical rules: Generator and transformer entities should satisfy certain mathematical relationships (such as difference constraints) on the relation vectors representing "voltage levels". Logical constraints: Connection relationships should be symmetric (if A is connected to B, then B is also connected to A) or inverse relationships (transmitting power to and receiving power from should be inverse relationships). Type constraints: Substation type entities cannot be directly connected to generator user type entities through containment relationships.

[0067] Verification process: Calculate the value of the locally adjusted vector space on this composite loss function. If the loss value decreases significantly and falls below a threshold, the verification result is "meets requirements"; if the loss value is still high, especially if the regularization term contributes significantly, it indicates that the update may have violated domain rules, and the verification result is "does not meet requirements".

[0068] A4. If the verification result meets the requirements, the locally adjusted vector space will be output as the updated vector space.

[0069] For example, if the verification result meets the requirements, it indicates that this local update not only satisfies data-driven similarity but also conforms to domain knowledge logic. The system formally merges the result of this local adjustment into the global vector space as the final output of this update—the "updated vector space".

[0070] A5. If the verification result does not meet the requirements, the locally adjusted vector space is fine-tuned based on the locally adjusted vector space and the preset loss function to determine the fine-tuned vector space. Based on the fine-tuned vector space, the physical consistency verification and fine-tuning are repeated until the verification result meets the requirements.

[0071] For example, if the verification result does not meet the requirements, this invention will enter an iterative optimization loop. It will not directly reject the proposal, but rather consider the current vector adjustment direction or magnitude unsatisfactory. Starting with the current locally adjusted vector, this invention continues gradient descent optimization with the goal of minimizing the aforementioned composite loss function (especially reducing the regularization term). This process may slightly relax the requirements for the new triplet score in exchange for stronger rule consistency. After each fine-tuning, the loss function is recalculated for verification. This loop may have a maximum number of iterations. If the verification still does not meet the requirements after multiple iterations, it may be determined that the update proposal fundamentally conflicts with the existing knowledge system and domain rules, thereby triggering a subsequent alarm mechanism.

[0072] Thus, this invention describes an efficient vector space adjustment strategy that updates only the affected parts by utilizing "locality" and "domain consistency". By embedding domain knowledge into the loss function, it ensures that the incremental update is not only mathematically sound, but also physically and logically correct, thereby greatly improving the quality and reliability of knowledge graph updates.

[0073] S1033. Based on the updated vector space, a link prediction algorithm is used to calculate the probability of the updated vector space being reasonable.

[0074] For example, step S1033 can be specifically implemented as steps B1-B3.

[0075] B1. Extract the ternary vector corresponding to the updated proposal from the updated vector space.

[0076] In some embodiments, the ternary vector includes a head entity vector, a relation vector, and a tail entity vector.

[0077] For example, embodiments of the present invention can accurately find the vector representations corresponding to the triples involved in the updated global vector space based on the specific content of the update proposal (e.g., the proposal is to add a new relation (Device A, Connected to, Device B)). If the proposal is to add a new entity, then the entity and its relationship with the existing entity constitute the triples that need to be verified. If the proposal is to add a new relation, then the relation itself defines an explicit triple (h, r, t). If the proposal is to update entity attributes, the attribute can be regarded as a special relation (e.g., (Entity, Has Voltage Rating, 500kV)), and therefore also constitutes a triple. The system searches in the vector space index using the unique identifiers (IDs) of the entity and relation, and extracts the vector vec_h of the head entity h, the vector vec_r of the relation r, and the vector vec_t of the tail entity t, respectively.

[0078] B2. Based on the ternary vector and the preset scoring function, the rationality score is calculated.

[0079] In some embodiments, the scoring function includes the translation model TransE, the bilinear model DistMult, or the complex embedding model ComplEx.

[0080] For example, embodiments of the present invention can use the scoring function corresponding to the selected graph embedding model to measure the logical rationality (i.e., "score") of the triple (h,r,t) in the vector space.

[0081] The Translation Equation (TransE) model's core idea is that h + r ≈ t. The scoring function is typically defined as f(h, r, t) = -||vec_h + vec_r - vec_t||_{L1 / L2}. A higher score (i.e., a smaller absolute value of the negative norm) indicates that h + r is closer to t, and the triplet is more reasonable.

[0082] The bilinear model (DistMult) treats the relation r as a diagonal matrix and scores it by performing a bilinear transformation on the vector. The scoring function is f(h,r,t) =<vec_h,Mr,vec_t> , where Mr is a diagonal matrix, and <,·> denote the inner product. A higher score indicates a more reasonable triplet.

[0083] Complex embedding models (ComplEx): embedding in complex space allows for better handling of complex relationships such as symmetry and antisymmetry. Its scoring function is f(h,r,t)=Re(<vec_h,Mr, conjugate(vec_t)> ), where Re represents taking the real part and conjugate represents the conjugate complex number. Similarly, a higher score indicates a more reasonable answer.

[0084] In this case, the three vectors vec_h, vec_r, and vec_t extracted in step B1 are substituted into the selected scoring function f(h,r,t) to calculate a real-valued score.

[0085] B3. Based on the rationality score, normalization is performed to obtain the rationality probability of the updated vector space.

[0086] For example, in embodiments of the present invention, the score value obtained directly from the scoring function has an uncertain range and lacks probabilistic meaning. It needs to be mapped to the (0,1) interval using an activation function (usually the sigmoid function) and interpreted as a probability value. The calculation formula is typically: probability = σ(score) = 1 / (1 + exp(-score)). Probability is the "reasonable probability," which intuitively represents how likely the newly added or updated triple is to be true under the current vector space model. The closer the probability is to 1, the better the consistency between the update proposal and the overall structure of the existing knowledge graph, and the more reliable the update. The closer the probability is to 0, the more likely the proposal is noise or an error, conflicting with existing knowledge.

[0087] S1034. Based on the updated vector space and the probability of rationality of the updated vector space, generate the verification results of the power grid knowledge graph.

[0088] For example, step S1034 can be specifically implemented as steps C1-C3.

[0089] C1. Based on the probability of the updated vector space being reasonable and a preset threshold, determine whether the updated vector space is reasonable.

[0090] For example, embodiments of the present invention compare the probability of reasonableness with a predefined threshold to make a binary judgment. Threshold definition: Typically, a high confidence threshold is set, such as 0.95. This threshold is high, meaning that only when a new item is highly consistent with the existing knowledge graph structure will it be directly adopted. This judgment is the first and most important automated checkpoint based on the mathematical representation of the vector space model. A high probability means that the update proposal is "credible".

[0091] C2. If the updated vector space is reasonable, then based on the updated vector space and the probability of its reasonableness, generate the verification result of the power grid knowledge graph.

[0092] For example, when a proposal is deemed reasonable, the system generates a positive verification result. The verification result is a structured data object, including: Final decision: adoption; Updated vector space: providing data support for subsequent knowledge graph entity / relationship update operations; Reasonableness probability value: recording the confidence level of this update, stored as metadata for this knowledge item, for subsequent querying, tracing, or weight calculation; Proposal ID and type: associated with the original update proposal; Flow: This positive verification result is sent to the knowledge graph update execution module, triggering the actual graph write operation.

[0093] C3. If the updated vector space is unreasonable, then based on the multi-source incremental data, update the proposal, and generate alarm information based on the updated vector space and the probability of reasonableness.

[0094] For example, when a proposal is deemed unreasonable, the system does not fail silently but generates a negative alert, triggering subsequent processing. The alert aims to provide operations personnel or upstream systems with sufficient context for diagnosis, including: Core alert: Update verification failed; Raw data: Multi-source incremental data that triggered the update (e.g., which log record, which document paragraph), for tracing the source; Failed proposal: The specific content of the update proposal (i.e., what was attempted to be added or modified); Verification details: The calculated probability of reasonableness, and possible cause analysis (e.g., which existing entity's vector conflict is greatest).

[0095] Thus, this invention achieves "closed-loop management" of the update process, which ensures automation efficiency under high confidence conditions and ensures system reliability and traceability through an alarm mechanism under low confidence conditions, preventing erroneous knowledge from contaminating the graph.

[0096] S104. Based on the verification results of the power grid knowledge graph, dynamically update the power grid knowledge graph.

[0097] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1045.

[0098] S1041. Based on the updated vector space, extract the static core vector and / or dynamic relation vector to be updated.

[0099] In some embodiments, the present invention parses the validated updated vector space from the positive validation results. Then, based on the type of update proposal, it precisely locates and extracts from this new global vector space those vectors that need to be actually written into the knowledge graph database. These vectors are the "final version" vectors that have been optimized and verified during the previous local dynamic update and validation process. They are indexed and queried using vector IDs (bound to entity IDs and relation IDs). For example, extracting the static_vector of the newly added entity E_new, the relation_vector of the newly added relation R_new, etc.

[0100] S1042. If the proposal type of the update proposal is a new entity proposal, then add an entity node in the power grid knowledge graph based on the static core vector to be updated.

[0101] S1043. Based on the dynamic relationship vector to be updated, add new entity nodes and the power grid knowledge graph, establish the association between the new entity nodes and the power grid knowledge graph, and realize the dynamic update of the power grid knowledge graph.

[0102] In some embodiments, a new entity node is created: In a graph database (such as Neo4j) or a triple database (such as RDF storage), a new node is created. This node is assigned a unique identifier (URI or ID), a label (such as Generator, Substation), and properties (such as name: "Xinfeng Wind Turbine", capacity: "2MW"). A key operation is to store the node's static core vector as a special attribute (such as embedding_vector) or in a dedicated vector index for subsequent graph computation, semantic search, and reasoning.

[0103] In some embodiments, establishing relationships involves iterating through all new relationships between the new entity and existing entities in the update proposal. For each relationship (e.g., (E_new, connected_to, E_existing)): Locate the head entity node E_new and the tail entity node E_existing in the database. Create an edge between them, specifying the relationship type (e.g., CONNECTED_TO). Key operation: Append the dynamic relationship vector of this relationship to this edge as an embedding attribute of the relationship.

[0104] S1044. If the proposal type of the update proposal is a new relationship proposal, then based on the dynamic relationship vector to be updated and the power grid knowledge graph, establish the new association relationship in the power grid knowledge graph.

[0105] In some embodiments, proposing a new relationship does not involve creating new nodes; it simply establishes a new relationship between existing entities. For the new relationship (E_existing_A, new_relation, E_existing_B) in the proposal: Existing nodes E_existing_A and E_existing_B are retrieved from the graph database. A new edge is created between them, specifying the relationship type. Key operation: The extracted dynamic relationship vector corresponding to this relationship is appended to this newly created edge. Although the entities are existing, their static core vectors may have been fine-tuned in previous local dynamic updates. These updated entity vectors are stored in the updated vector space and will be synchronously updated to the knowledge graph when needed.

[0106] S1045. If the proposal type of the update proposal is an entity attribute update proposal, then based on the static core vector to be updated, update the attribute information of the node corresponding to the entity attribute update proposal in the power grid knowledge graph.

[0107] For example, in a knowledge graph, attributes typically exist as triples of (entity, attribute name, attribute value). Updating an attribute can be viewed as correcting a fact of knowledge. The target node that needs updating is found based on the proposal. Update operations: Update attribute value: Directly modify the value of the corresponding attribute name for that node in the database. Update vector representation: Write the extracted updated static core vector of the entity back to the database, overwriting the old vector. This new vector encodes the entity's new attribute information and its new relationships with other elements in the graph. If the attribute is a relationship pointing to another entity (e.g., voltage_level might link to an entity representing voltage levels), this operation may be transformed into "adding a relationship" or modifying an existing relationship.

[0108] This invention provides a dynamic update method for a power grid knowledge graph based on AI incremental learning. By acquiring structured operational data and unstructured text data in real time and combining it with an AI incremental analysis model, it achieves online identification and confidence assessment of new entities and relationships. Update proposals are generated only for the changed parts of the knowledge graph, avoiding the high overhead of full reconstruction. Furthermore, a dynamic graph embedding model is used for local updates and rationality verification of the vector space, ensuring the semantic and physical consistency of newly added knowledge. The entire update process is automated, lightweight, and highly reliable, significantly shortening the cycle from knowledge extraction from data to graph implementation. This enables efficient updating and maintenance of the power grid knowledge graph, solves the lag problem inherent in full-scale knowledge graph updates, and significantly improves the response speed and support capabilities of the knowledge graph, enhancing its timeliness and accuracy.

[0109] Optionally, the AI-based incremental learning method for dynamically updating the power grid knowledge graph provided in this embodiment of the invention further includes steps S201-S206.

[0110] S201. Obtain multiple error samples from the error sample library.

[0111] In some embodiments, erroneous samples include multi-source incremental data determined to be unreasonable in the updated vector space, update proposals, updated vector space, and reasonableness probability.

[0112] In some embodiments, this invention maintains an error sample library, the data source of which is the alarm information generated above. Whenever an update proposal is determined to be "unreasonable" by the verification module and an alarm is generated, the system automatically packages the relevant data of this failed case and stores it as an error sample in the library. Sample content: Multi-source incremental data: The original data that triggered this update (such as a piece of text, a SCADA record), which is the "source" of the model error. Update proposal: The (incorrect) proposal generated by the AI ​​incremental analysis model based on the original data, which is the "direct manifestation" of the model error. Updated vector space & reasonableness probability: The calculation result of the verification module, providing feedback signals and contextual information on why the model output was determined to be incorrect.

[0113] S202. Based on multiple erroneous samples and historical positive samples, an incremental training sample set is constructed using a dynamic memory replay mechanism.

[0114] In some embodiments, this step aims to address the problem of catastrophic forgetting. To prevent the model from forgetting previously learned correct knowledge (positive samples) when learning new errors (negative samples), a dynamic memory replay strategy is employed to construct a balanced training set. Historical positive samples: Randomly selected positive samples (i.e., correct update proposals and their data) that have been successfully validated and adopted in the past are drawn from the historical log. Construction strategy: A certain number of the latest error samples are taken from the error sample pool. A comparable number of positive samples are taken from the historical positive sample pool. This "dynamic" aspect is reflected in the fact that the ratio of positive to negative samples can be adaptively adjusted according to the error rate to ensure a balanced training set. These two sets of samples are then combined to form a mini-batch for incremental training.

[0115] S203. Based on the incremental training sample set, perform incremental training on the AI ​​incremental analysis model. During the incremental training process, adjust the importance weights of the parameters in the AI ​​incremental analysis model through the elastic weight consolidation algorithm to obtain the first model.

[0116] In some embodiments, incremental training: instead of training from scratch, additional training is performed using a built-in sample set on top of an existing, deployed AI incremental analytics model.

[0117] Elastic Weight Consolidation Algorithm: Calculating Parameter Importance: Before incremental training begins, the Fisher information matrix for each parameter (weight) in the model network is calculated on a large amount of historical data (or data from the previous stable version of the model). This matrix quantifies the "importance" of each parameter to the learned task. Important parameters have a significant impact on the final output and should not be easily changed. Modifying the Loss Function: An EWC regularization term is added to the standard training loss function (such as cross-entropy loss). This term penalizes significant modifications to "important parameters." The new loss function takes the form: L_total = L_new + λ*Σ_i[F_i * (θ_i - θ*_i)^2], where: L_new is the loss on the new task; F_i is the importance (Fisher information) of parameter i; θ*_i is the value of parameter i in the old model; and λ is a hyperparameter controlling the strength of the constraints.

[0118] The first model is obtained through training: the model is trained using this constrained loss function. The model learns to correct new errors while retaining as much old knowledge as possible. The model obtained after training is called the "first model".

[0119] S204. Determine the distribution characteristics of multiple erroneous samples.

[0120] In some embodiments, this invention performs error analysis to achieve more refined model adjustments. Statistical analysis is performed on the error samples of the current batch: Error type distribution: From which sub-model do these errors mainly originate? Are there more entity recognition errors, relation extraction errors, or inaccurate confidence assessments? Data source distribution: Are errors more likely to originate from unstructured text (higher difficulty in NLP) or structured data (parsing errors)? Entity / relation type distribution: Are certain types of devices or relations (such as "relay protection," "power flow calculation") more prone to errors? S205. Based on the distribution characteristics of multiple erroneous samples, adaptively adjust the weight coefficients of the loss functions of each sub-model in the first model to obtain the updated AI incremental analysis model.

[0121] In some embodiments, the sub-models include a new entity recognition model, a new relationship recognition model, and a relationship evaluation model.

[0122] In some embodiments, the AI ​​incremental analysis model is typically a pipeline or end-to-end model containing multiple sub-modules. Total Loss Function: The total loss of the entire model is usually a weighted sum of the losses of each sub-model: L_total = α * L_entity + β * L_relation + γ * L_confidence. Adaptive Adjustment: These weight coefficients (α, β, γ) are dynamically adjusted based on the analyzed error distribution characteristics. For example, if analysis reveals a significantly higher proportion of entity recognition errors in the current batch of erroneous samples, then in subsequent training epochs, α (the weight of the entity recognition loss term) is increased, forcing the model to focus more on correcting entity recognition errors. Using the adjusted loss function, the "first model" is further fine-tuned, ultimately resulting in a more balanced and robust updated AI incremental analysis model.

[0123] S206. Based on the updated AI incremental analysis model, dynamically update the power grid knowledge graph.

[0124] In some embodiments, this iteratively optimized new model is deployed to the production environment, replacing the original model. Subsequent dynamic knowledge graph update processes will use this more capable new model, which is less prone to similar errors, to generate update proposals, thereby forming a self-iteratory, self-improving closed-loop learning system.

[0125] Thus, this embodiment of the invention provides an advanced model self-optimization closed loop. It not only handles failure cases, but more importantly, utilizes these failure cases as "negative feedback" signals. Through a series of advanced technologies such as dynamic memory replay (preventing forgetting), elastic weight consolidation (preserving prior knowledge), and error analysis-based adaptive loss adjustment (precise optimization), it continuously and automatically improves the core capabilities of the AI ​​incremental analysis model, enabling the entire knowledge graph dynamic update system to possess powerful online learning and self-evolutionary characteristics.

[0126] Optionally, the AI-based incremental learning method for dynamically updating the power grid knowledge graph provided in this embodiment of the invention further includes steps S301-S308.

[0127] S301. Obtain the last update time and current update time of the target to be updated in the update proposal of the power grid knowledge graph.

[0128] For example, this invention maintains metadata for each entity and relationship in a knowledge graph, a key element of which is the last update timestamp. Current update time: i.e., the timestamp at which this update proposal was generated.

[0129] Last update time: The system queries the metadata database of the knowledge graph based on the "target to be updated" (i.e., entity ID or relation ID) in the update proposal to obtain the most recent last update timestamp of the target.

[0130] S302. Calculate the update time difference of the target to be updated based on the last update time and the current update time of the target to be updated.

[0131] S303. Based on the update time difference, calculate the confidence score of the target to be updated; the confidence score of the target to be updated is used to characterize the probability that the entity or relationship to be updated is a real update event.

[0132] For example, the longer a knowledge item remains stable, the higher its confidence level; the more frequently it is updated, the greater the likelihood that the update is noise or error. The calculation function: the confidence score is typically a monotonically increasing function based on the time difference Δt. A common implementation uses the inverse of the exponential decay function or the sigmoid function: S_confidence = 1 - exp(-λ*Δt) (where λ is the decay coefficient, controlling the rate of score increase) or S_confidence = 2 / (1 + exp(-k*Δt)) - 1.

[0133] The smaller Δt is (frequent updates), the closer S_confidence is to 0, indicating low confidence and the update may be questionable. The larger Δt is (first update after long-term stability), the closer S_confidence is to 1, indicating high confidence and the update is more likely to reflect real-world changes.

[0134] S304. Based on the confidence score of the target to be updated, determine whether the target to be updated needs secondary verification.

[0135] For example, if the confidence score is greater than or equal to the confidence threshold, the confidence level is high and no secondary verification is required; if the confidence score is less than the confidence threshold, the confidence level is low and secondary verification is required.

[0136] S305. If the target to be updated requires secondary verification, then based on the relevant data of the target to be updated in the update proposal and the dynamic graph embedding model, perform local dynamic updates and verifications in the vector space to determine the secondary verification result of the target to be updated.

[0137] In some embodiments, the secondary verification result includes the updated vector space and the secondary rationality probability.

[0138] S306. If the probability of secondary reasonableness is greater than the threshold, the target to be updated is dynamically updated and the confidence score of the target to be updated is reset.

[0139] S307. If the probability of secondary rationality is in the middle range, an update request for the target to be updated is generated and the update request is pushed to the manual review queue.

[0140] S308. If the probability of secondary rationality is lower than the elimination threshold, then the target to be updated in the power grid knowledge graph remains unchanged.

[0141] In some embodiments, the update proposal is ultimately determined to be highly unreliable, likely an error or noise. The update is rejected, and the content of the target in the knowledge graph remains unchanged. This case is likely to be recorded in the error sample library for subsequent model optimization.

[0142] Thus, this invention introduces a timestamp-based confidence management mechanism, adding a crucial temporal dimension to risk control in the update process. It effectively identifies and handles the abnormal behavior of frequent updates, preventing the system from being "attacked" by a series of erroneous update proposals. Through a hierarchical decision-making mechanism of secondary verification and manual review, it significantly improves the prudence in handling suspicious update events and the overall robustness of the system. The confidence score reset mechanism ensures the fairness and effectiveness of the process, avoiding a vicious cycle of "one suspicious, every suspicious." This allows the entire method to no longer process each update in isolation, but rather to view each element in the knowledge graph historically and dynamically, giving it a memory and recognition capability similar to an "immune system."

[0143] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0144] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0145] Figure 2 This diagram illustrates the structure of an AI-based incremental learning-based dynamic update device for a power grid knowledge graph, according to an embodiment of the present invention. The update device 400 includes a communication module 401 and a processing module 402.

[0146] The communication module 401 is used to acquire multi-source incremental data of the power grid, including structured operation data and unstructured text data.

[0147] The processing module 402 is used to identify new entities, extract new relationships, and assess the confidence of newly added items based on multi-source incremental data and a preset AI incremental analysis model, and determine the update proposal of the power grid knowledge graph; based on the update proposal and a preset dynamic graph embedding model, perform local dynamic updates and verification of the vector space, and determine the verification result of the power grid knowledge graph, including the updated vector space and the probability of its rationality; and dynamically update the power grid knowledge graph based on the verification result of the power grid knowledge graph.

[0148] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the above-described method embodiments. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above-described device embodiments.

[0149] For example, the computer program 503 may be divided into one or more modules / units, which are stored in the memory 502 and executed by the processor 501 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 503 in the electronic device 500.

[0150] The processor 501 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0151] The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or memory of the electronic device 500. The memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 500. Furthermore, the memory 502 can include both internal and external storage units of the electronic device 500. The memory 502 is used to store the computer program and other programs and data required by the terminal. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0152] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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, and should all be included within the protection scope of the present invention.

Claims

1. A method for dynamic updating of a power grid knowledge graph based on AI incremental learning, characterized in that, include: Acquire multi-source incremental data of the power grid, including structured operation data and unstructured text data; Based on the multi-source incremental data and the preset AI incremental analysis model, new entity identification, new relationship extraction and new item confidence assessment are performed to determine the update proposal of the power grid knowledge graph. Based on the update proposal and the preset dynamic graph embedding model, the vector space is locally dynamically updated and verified to determine the verification result of the power grid knowledge graph. The verification result includes the updated vector space and the probability of rationality of the updated vector space. Based on the verification results of the power grid knowledge graph, the power grid knowledge graph is dynamically updated.

2. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 1, characterized in that, Based on the multi-source incremental data and a preset AI incremental analysis model, new entity identification, new relationship extraction, and confidence assessment of new items are performed to determine the update proposal for the power grid knowledge graph, including: Based on the multi-source incremental data and the AI ​​incremental analysis model, feature extraction is performed to determine the basic features of the incremental data. Based on the basic characteristics of the incremental data and the AI ​​incremental analysis model, new entity identification is performed to determine the newly added power grid equipment entities; Based on the newly added power grid equipment entities and the entity list of the power grid knowledge graph, matching links and entity disambiguation are performed to determine the linked entities of the newly added power grid equipment entities in the power grid knowledge graph. Based on the basic characteristics of the incremental data, the linked entities of the newly added power grid equipment entities, and the AI ​​incremental analysis model, new relationships are identified and the newly added associations are determined. Based on the existing relationships in the power grid knowledge graph, the reliability of the newly added relationships is calculated to determine the confidence probability of the newly added relationships. Based on the incremental data features, the newly added power grid equipment entities and the newly added relationships, an attention mechanism is used to perform knowledge transfer and collaborative optimization between entities and relationships, and to determine the optimized newly added power grid equipment entities, newly added relationships and confidence probabilities. Based on the optimized new power grid equipment entities, new associations, and confidence probabilities, update proposals for the power grid knowledge graph are determined.

3. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 1, characterized in that, The process of performing local dynamic updates and verifications of the vector space based on the update proposal and a preset dynamic graph embedding model to determine the verification result of the power grid knowledge graph includes: Extract the proposal type of the update proposal, which includes new entity proposal, new relationship proposal, and entity attribute update proposal; Based on the proposed type and the dynamic graph embedding model, the vector space of the power grid knowledge graph is dynamically updated to obtain the updated vector space. Based on the updated vector space, a link prediction algorithm is used to calculate the probability of the rationality of the updated vector space. Based on the updated vector space and the probability of its rationality, the verification results of the power grid knowledge graph are generated.

4. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 3, characterized in that, The process of dynamically updating the vector space of the power grid knowledge graph based on the proposal type and the dynamic graph embedding model to obtain the updated vector space includes: Based on the proposal type and the update proposal, the adjustment items for the vector space are determined; the adjustment items include the static core vector and / or dynamic relation vector of the entity to be updated. Based on the update proposal and the adjustment items, the vector space is dynamically adjusted locally to obtain the locally adjusted vector space; Based on a preset loss function, a physical consistency check is performed on the locally adjusted vector space to determine the check result; the preset loss function includes a regularization term composed of physical rules and logical constraints in the power grid field. If the verification result meets the requirements, the locally adjusted vector space will be output as the updated vector space. If the verification result does not meet the requirements, the locally adjusted vector space is fine-tuned based on the locally adjusted vector space and the preset loss function to determine the fine-tuned vector space. Based on the fine-tuned vector space, the physical consistency verification and fine-tuning are repeated until the verification result meets the requirements.

5. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 3, characterized in that, The step of calculating the reasonableness probability of the updated vector space using a link prediction algorithm based on the updated vector space includes: From the updated vector space, extract the ternary vector corresponding to the update proposal, wherein the ternary vector includes a head entity vector, a relation vector, and a tail entity vector; Based on the ternary vector and the preset scoring function, a reasonableness score is calculated; the scoring function includes the translation model TransE, the bilinear model DistMult, or the complex embedding model ComplEx. Based on the rationality score, normalization is performed to obtain the rationality probability of the updated vector space.

6. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 3, characterized in that, The verification result of generating the power grid knowledge graph based on the updated vector space and the rationality probability of the updated vector space includes: Based on the probability of the updated vector space being reasonable and a preset threshold, it is determined whether the updated vector space is reasonable. If the updated vector space is reasonable, then based on the updated vector space and the probability of its reasonableness, a verification result of the power grid knowledge graph is generated. If the updated vector space is unreasonable, an alarm message is generated based on the multi-source incremental data, the update proposal, the updated vector space, and the reasonableness probability.

7. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 1, characterized in that, The dynamic updating of the power grid knowledge graph based on the verification results includes: Based on the updated vector space, extract the static core vector and / or dynamic relation vector to be updated; If the proposal type of the update proposal is a new entity proposal, then an entity node is added in the power grid knowledge graph based on the static core vector to be updated; Based on the dynamic relationship vector to be updated, the newly added entity node and the power grid knowledge graph are used to establish the association between the newly added entity node and the power grid knowledge graph, thereby realizing the dynamic updating of the power grid knowledge graph. If the proposal type of the update proposal is a new relationship proposal, then based on the dynamic relationship vector to be updated and the power grid knowledge graph, a new association relationship is established in the power grid knowledge graph; If the proposal type of the update proposal is an entity attribute update proposal, then based on the static core vector to be updated, the attribute information of the node corresponding to the entity attribute update proposal is updated in the power grid knowledge graph.

8. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 6, characterized in that, The method further includes: Obtain multiple error samples from the error sample library. The error samples include multi-source incremental data that is determined to be unreasonable after the updated vector space, update proposals, updated vector space, and reasonableness probability. Based on the aforementioned multiple erroneous samples and historical positive samples, an incremental training sample set is constructed using a dynamic memory replay mechanism. Based on the incremental training sample set, the AI ​​incremental analysis model is incrementally trained. During the incremental training process, the importance weights of the parameters in the AI ​​incremental analysis model are adjusted through the elastic weight consolidation algorithm to obtain the first model. Determine the distribution characteristics of the multiple erroneous samples; Based on the distribution characteristics of the multiple erroneous samples, the weight coefficients of the loss function of each sub-model in the first model are adaptively adjusted to obtain the updated AI incremental analysis model; the sub-models include a new entity recognition model, a new relationship recognition model, and an association relationship evaluation model; The power grid knowledge graph is dynamically updated based on the updated AI incremental analysis model.

9. The method for dynamic updating of power grid knowledge graph based on AI incremental learning according to claim 1, characterized in that, The method further includes: Obtain the last update time and current update time of the target to be updated in the update proposal of the power grid knowledge graph; Calculate the update time difference of the target to be updated based on the last update time and the current update time of the target to be updated; Based on the update time difference, the confidence score of the target to be updated is calculated; the confidence score of the target to be updated is used to characterize the probability that the entity or relationship to be updated is a real update event; Based on the confidence score of the target to be updated, determine whether the target to be updated needs secondary verification; If the target to be updated requires secondary verification, then based on the relevant data of the target to be updated in the update proposal and the dynamic graph embedding model, the vector space is locally dynamically updated and verified to determine the secondary verification result of the target to be updated. The secondary verification result includes the updated vector space and the secondary rationality probability. If the probability of secondary reasonableness is greater than the threshold, the target to be updated is dynamically updated and the confidence score of the target to be updated is reset. If the probability of secondary rationality is in the middle range, an update request for the target to be updated is generated and pushed to the manual review queue. If the probability of secondary rationality is lower than the elimination threshold, the target to be updated in the power grid knowledge graph remains unchanged.

10. A device for dynamically updating a power grid knowledge graph based on AI incremental learning, characterized in that, The updating device includes: A communication module is used to acquire multi-source incremental data of the power grid, including structured operation data and unstructured text data. The processing module is used to perform new entity identification, new relationship extraction, and new item confidence assessment based on the multi-source incremental data and a preset AI incremental analysis model to determine the update proposal of the power grid knowledge graph; based on the update proposal and a preset dynamic graph embedding model, perform local dynamic updates and verification of the vector space to determine the verification result of the power grid knowledge graph, the verification result including the updated vector space and the probability of reasonableness of the updated vector space; and dynamically update the power grid knowledge graph based on the verification result of the power grid knowledge graph.

Images