An adaptive control method, system, equipment, medium, and product for power grid maintenance.

By cleaning and standardizing multi-source heterogeneous data, and combining graph embedding vectors and decision models, the problems of data fusion and scenario adaptation in power grid maintenance and management are solved, and efficient power grid maintenance and management is achieved.

CN123118074APending Publication Date: 2026-11-17GUANGDONG POWER GRID CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611560813.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing power grid maintenance and control devices suffer from insufficient multi-source data fusion capabilities, lack of scenario adaptability, and rigid functional configurations, resulting in low control efficiency.

Method used

By acquiring heterogeneous data from multiple sources, cleaning and standardizing it, a standardized maintenance dataset is generated. Maintenance business entities, control function entities, and operation scenario entities are mapped to the same vector space to generate graph embedding vectors. Combined with a decision model, the fit score is calculated, and the display and strategy of functional modules are dynamically adjusted.

Benefits of technology

It realizes semantic association of multi-source data and scene adaptive scheduling, improves the efficiency and adaptability of power grid maintenance and management, and forms a closed-loop mechanism of perception-reasoning-decision-presentation-feedback.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118074A_ABST
    Figure CN123118074A_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive control method, system, equipment, medium, and product for power grid maintenance, relating to the field of power grid dispatch automation technology. It acquires multi-source heterogeneous data, performs cleaning, time-series alignment, and standardization; extracts maintenance business entities, control function entities, and operational scenario entities, mapping them to the same vector space to generate graph embedding vectors; extracts semantic and state features based on the text description and flow status of the current maintenance business, constructing scenario state vectors; inputs the scenario state vectors and graph embedding vectors into a decision model to calculate the fit score between the control function entities and the current maintenance business entities; determines the target control function and its display layout based on the score, generating a control and dispatch strategy; renders the interactive interface according to the dispatch strategy and collects user operation data; updates the decision model parameters based on the operation data and returns for execution. Through a data-driven dynamic matching mechanism, the response efficiency of maintenance business control is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power grid dispatch automation technology, and in particular to an adaptive control method, system, equipment, medium and product for power grid maintenance. Background Technology

[0002] Under the background of the construction of new power systems, the complexity of power grid maintenance has increased significantly. As the core platform supporting power grid dispatching, the provincial and local integrated power grid operation management system (OMS) needs to meet the management and control requirements of diverse scenarios such as daily operation and maintenance, fault repair, and major power supply guarantee. Although the current power dispatching system based on OMS has realized the online flow of maintenance business, the efficiency of management and control directly affects the reliability of power supply and emergency response capabilities.

[0003] Existing maintenance control devices generally adopt a static configuration mode, deployed on OMS dispatch server clusters. They complete basic operations such as maintenance order circulation and plan approval through fixed functional modules. All positions share a unified control interface, and dispatchers need to manually switch function windows to find the required information. Specifically, existing devices have three technical shortcomings: First, they lack multi-source data fusion capabilities, only enabling independent storage and display of structured data, and failing to establish semantic relationships between heterogeneous data such as equipment topology, operation logs, and circulation status. Second, they lack scenario-adaptive control capabilities, unable to proactively identify different operating scenarios, and unable to dynamically adjust the priority and control strategies of functional modules. Third, their functional configurations are static and rigid, lacking autonomous optimization capabilities, and unable to adapt to differentiated control interfaces based on job roles and operating habits, easily leading to the burying of critical information and limited response efficiency in emergency fault scenarios.

[0004] At the level of publicly available patents, related research mainly falls into three categories: scheduling and interaction collaboration technologies focus on collaborative communication between multiple entities, without designing adaptive scheduling mechanisms for maintenance operations; maintenance plan orchestration technologies only optimize timing scheduling, without extending to the human-computer interaction level of control functions; and maintenance data structuring technologies only extract and store unstructured data, without incorporating control functions into a knowledge system to achieve semantic matching. In summary, existing maintenance control devices and technologies generally suffer from common defects such as data semantic fragmentation, lack of proactive scenario perception, and fixed functional configurations without a closed-loop evolution mechanism, making them difficult to adapt to the control needs of new power systems characterized by high complexity of maintenance operations, rapid scenario changes, and high timeliness requirements for handling. Summary of the Invention

[0005] This invention provides a method, system, equipment, medium, and product for adaptive power grid maintenance management, which solves the technical problem that existing power grid maintenance management cannot achieve data fusion reasoning and scenario-adaptive scheduling, resulting in low efficiency in power grid maintenance management.

[0006] The first aspect of this invention provides an adaptive control method for power grid maintenance, comprising: Acquire multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data to generate a standardized maintenance dataset; Based on the standardized maintenance dataset, maintenance business entities, control function entities, and operation scenario entities are extracted, and the three types of entities are mapped to the same vector space to generate a graph embedding vector. Based on the text description of the current maintenance business, extract text semantic features and flow status features from the flow status data to construct a scenario status vector; Input the scenario state vector and the graph embedding vector into the decision model, and calculate the fit score between the control function entity and the current maintenance business entity based on the graph embedding vector; Based on the adaptation score, the target control functions and their display layout are determined, and a control and scheduling strategy is generated. The maintenance control interaction interface is rendered and output according to the control and scheduling strategy, and user operation data on the interaction interface is collected. The parameters of the decision model are updated based on the user operation data, and the process returns to the step of inputting the scene state vector and the graph embedding vector into the decision model.

[0007] Optionally, the step of extracting maintenance business entities, control function entities, and operation scenario entities based on the standardized maintenance dataset, mapping the three types of entities to the same vector space, and generating a graph embedding vector includes: Extract maintenance business entities, control function entities, and operation scenario entities from the standardized maintenance dataset; Multiple entity pairs are extracted from the maintenance business entity, the control function entity, and the operation scenario entity, and the association relationship between each entity pair is obtained. Each entity pair and its corresponding association relationship are formed into a triplet, and each triplet includes a head entity, the association relationship, and a tail entity. Differentiated pre-coding is performed on the maintenance business entity, the control function entity, and the operation scenario entity respectively to generate initial embedding vectors for each type of entity; The initial embedding vector of the head entity is transformed sequentially by the type calibration matrix corresponding to the type of the head entity and the relation projection matrix corresponding to the association relationship to obtain the first projection vector; The initial embedding vector of the tail entity is transformed sequentially by the type calibration matrix corresponding to the type of the tail entity and the relation projection matrix corresponding to the association relationship to obtain the second projection vector; The energy value of the triple is obtained by calculating the distance between the sum of the first projection vector and the relationship vector of the association relationship and the second projection vector; The model is trained based on the energy values ​​of each triple, and the trained model is used to map all the initial embedding vectors to generate a graph embedding vector.

[0008] Optionally, the step of training the model based on the energy values ​​of each triplet includes: Based on the types of the head and tail entities in the triplet, negative sample triplets are constructed using the corresponding negative sampling constraint rules. Based on the priority of power grid maintenance services, determine the differentiated weights corresponding to the training samples; Based on the difference between the energy value of each triplet and the energy value of the corresponding negative sample triplet, a total loss function is constructed, which includes a marginal loss term, a weight regularization term, and a temporal smoothing regularization term. The temporal smoothing regularization term is used to constrain the semantic continuity of entity embedding in different stages of the same maintenance order. With the goal of minimizing the total loss function, the parameters of the type calibration matrix, the parameters of the relation projection matrix, and the parameters of the relation vector of the association relationship are optimized to obtain the trained model.

[0009] Optionally, the step of extracting textual semantic features from the textual description of the current maintenance business and extracting flow status features from the flow status data to construct a scenario state vector includes: The text description is encoded, and its semantic features are extracted. Within a preset time window, the backlog, obstruction, and return rate of maintenance documents are statistically analyzed as raw indicators of circulation status. The original indicators of the circulation status are normalized to obtain normalized circulation status indicators, and the slope of the change of each indicator between adjacent time windows is extracted. The normalized flow state index is concatenated with the corresponding change slope and then mapped by a multilayer perceptron to obtain the flow state features. A gating fusion mechanism is used to adaptively weight and fuse the text semantic features and the flow state features, and the fusion result is mapped into a scene state vector through a classification layer.

[0010] Optionally, the step of inputting the scenario state vector and the graph embedding vector into the decision model, and calculating the fit score between the control function entity and the current maintenance business entity based on the graph embedding vector, includes: Based on the graph attention mechanism, multi-hop association paths between the current maintenance business entity and each of the control function entities are aggregated from the graph embedding vector to obtain the path aggregation features of each of the control function entities. The prior usage probability of each of the control function entities in the current scenario is obtained based on the scenario state vector. Based on the aggregation features of each path and the corresponding prior usage probability, the fit score between each control function entity and the current maintenance business entity is calculated.

[0011] Optionally, the step of determining the target control function and its display layout based on the adaptability score, and generating a control scheduling strategy, includes: Obtain the current user role vector, and concatenate the scene state vector, the current user role vector, and the current function display state vector into a state vector; Based on the preset set of power grid maintenance and control function modules, the function selection instruction, priority sorting instruction, and display area allocation instruction are defined as the action space; Using the state vector as the state input of the reinforcement learning agent, and taking the multi-objective composite reward function as the optimization objective, the reinforcement learning agent selects the optimal action from the action space. The optimal action is converted into a control and scheduling strategy that includes target control functions, their display priorities, and display area allocation schemes.

[0012] A second aspect of the present invention provides an adaptive control system for power grid maintenance, comprising: The data perception module is used to acquire multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data, and generate a standardized maintenance dataset. The knowledge graph reasoning module is used to extract maintenance business entities, control function entities and operation scenario entities based on the standardized maintenance dataset, map the three types of entities to the same vector space, and generate a graph embedding vector. The scene recognition module is used to extract text semantic features from the text description of the current maintenance business and extract flow status features from the flow status data to construct a scene status vector; The adaptability calculation module is used to input the scenario state vector and the graph embedding vector into the decision model, and calculate the adaptability score between the control function entity and the current maintenance business entity based on the graph embedding vector. The decision-making and scheduling module is used to determine the target control function and its display layout based on the adaptability score, and to generate control and scheduling strategies. The interactive output module is used to render and output the maintenance control interactive interface according to the control and scheduling strategy, and to collect user operation data on the interactive interface. The feedback optimization module is used to update the parameters of the decision model based on the user operation data, and return to execute the step of inputting the scene state vector and the graph embedding vector into the decision model.

[0013] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the adaptive control method for power grid maintenance as described above.

[0014] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the adaptive control method for power grid maintenance as described above.

[0015] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer performs the power grid maintenance adaptive control method as described above.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: This invention acquires multi-source heterogeneous data and performs cleaning, time-series alignment, and standardization, providing a unified data foundation for subsequent semantic association of cross-domain data. This overcomes the shortcomings of existing technologies that cannot perform linked analysis due to inconsistent data formats and time bases. Based on this, by extracting maintenance business entities, control function entities, and operational scenario entities and mapping them to the same vector space to generate a graph embedding vector, the originally discrete maintenance business documents, power grid equipment topology, and control function modules are incorporated into a unified semantic space. This upgrades the data from "storable and displayable" to "associative and reasonable," providing knowledge support for functional scheduling decisions and thus solving the problem of semantic fragmentation in multi-source heterogeneous data. At the scenario perception level, by extracting textual semantic features and flow state features and fusing them to construct a scenario state vector, proactive identification of the current operational scenario is achieved without manual triggering of switching, overcoming the shortcomings of traditional solutions in lacking scenario perception capabilities. Subsequently, by inputting the scene state vector and the knowledge graph embedding vector into the decision model and calculating the fit score, the functional scheduling decision receives dual support from scene situational awareness and knowledge graph semantic reasoning capabilities. This helps to dynamically match and adapt control functions to different scenarios. Based on this, the target control function and its display layout are determined according to the fit score, and a control scheduling strategy is generated. This allows the function configuration to adaptively adjust with changes in the scene, overcoming the limitations of static and rigid function configurations in traditional solutions that cannot dynamically adapt to business situations. By rendering and outputting the interactive interface according to the control scheduling strategy and collecting user operation data, the actual usage feedback of the dispatcher can be continuously perceived. Finally, the decision model parameters are updated through user operation data and returned for execution, forming a complete perception-reasoning-decision-presentation-feedback closed loop. This allows the control strategy to be continuously optimized with changes in business needs and user habits, overcoming the deficiency of traditional solutions lacking a closed-loop evolution mechanism. The above steps are interconnected and work synergistically to jointly solve the technical problem that existing power grid maintenance control cannot achieve data fusion reasoning and scene adaptive scheduling, resulting in low control efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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.

[0018] Figure 1 A flowchart illustrating the steps of an adaptive control method for power grid maintenance provided in an embodiment of the present invention; Figure 2 This invention provides an intelligent maintenance adaptive control device embedded in a provincial-level integrated OMS power supply dispatching system. Figure 3 A structural block diagram of a power grid maintenance adaptive control system provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] This invention provides a method, system, equipment, medium, and product for adaptive power grid maintenance management, which addresses the technical problem of low efficiency in existing power grid maintenance management due to the inability to achieve data fusion reasoning and scenario-adaptive scheduling.

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of an adaptive control method for power grid maintenance provided in Embodiment 1 of the present invention.

[0022] like Figure 2 As shown, the adaptive control method and system for power grid maintenance provided in this embodiment of the invention can be applied to the provincial-local integrated OMS power supply dispatch system. It reuses the existing Platform as a Service (PaaS) cloud platform dispatch server hardware cluster and dispatch data bus resources of OMS, and relies on the dispatch data bus to interconnect with the power grid operation acquisition terminal, maintenance business management terminal, and dispatcher human-machine interaction workstation.

[0023] From a system architecture perspective, the power grid maintenance adaptive control system comprises four core functional units: a multi-source data perception unit, a knowledge graph reasoning unit, an intelligent decision-making and orchestration unit, and an adaptive interactive terminal unit. The intelligent decision-making and orchestration unit is responsible for scenario perception and scheduling decisions, while the adaptive interactive terminal unit is responsible for interface presentation and interactive feedback. These units operate collaboratively to achieve fully automated processing of maintenance operations, from data perception to intelligent control.

[0024] The multi-source data sensing unit, serving as the system's data entry point, is deployed on the OMS data access server node. It accesses both structured and unstructured maintenance-related data in parallel via the scheduling data bus. Structured data includes annual maintenance orders, monthly maintenance orders, day-ahead maintenance orders, operation mode change orders, power grid operation section data, and business process flow indicators, originating from the Optical Character Recognition (OCR) module and the Supervisory Control and Data Acquisition (SCADA) system. Unstructured data includes scheduling logs, shift handover notes, and maintenance drawings, which are standardized and converted using the built-in text parsing and OCR modules. The unit also incorporates a dynamic time-series alignment and data cleaning module to unify the format and remove outliers from data from different sources and sampling frequencies, outputting a standardized maintenance dataset. This addresses the shortcomings of existing devices where multi-source data is discrete and cannot be analyzed in a coordinated manner.

[0025] The knowledge graph reasoning unit is the core of the system's knowledge base, deployed on the OMS graph database server node and running based on graph database storage. This unit uniformly maps all elements of maintenance operations to graph entity nodes, breaking down semantic barriers between maintenance data and control functions by establishing a three-element semantic relationship of "business-function-scenario". At the same time, it incorporates graph reasoning algorithms to uncover deep implicit relationships between business objects and control functions in different scenarios, providing knowledge support for upper-level decision-making units and solving the problem of insufficient flexibility in traditional devices that rely solely on hard-coded rules to adapt to scenarios.

[0026] The intelligent decision-making and orchestration unit is the core control hub of the system, deployed on the OMS computing server node. It includes a scene intent recognition submodule and a reinforcement learning decision-making submodule. The scene intent recognition submodule integrates text semantics and flow status data from both channels to accurately determine the scene type and urgency of the current maintenance task. The reinforcement learning decision-making submodule combines the current scene status, user role, and historical operating habits to automatically calculate the optimal maintenance control function scheduling strategy, outputting function start / stop, display priority, and layout scheme to downstream terminal units, driving the terminals to complete interface adaptive adjustments.

[0027] The adaptive interactive terminal unit corresponds to the human-computer interaction workstation terminal on the dispatcher's side and is the system's business output end. This unit receives dispatch strategies issued by the decision-making unit, dynamically calls the corresponding maintenance and control function modules, and automatically adjusts the visibility, display order, and display area of ​​the function modules. It generates an adapted maintenance and control workbench for dispatchers in different positions and scenarios, realizing intelligent front-end of key functions and automatic folding of redundant information, reducing the information retrieval cost for dispatchers. At the same time, the terminal unit collects user operation behavior data in real time and sends it back to the decision-making unit for model iteration and optimization, forming a complete control closed loop of "perception-reasoning-decision-presentation-feedback".

[0028] This invention provides an adaptive control method for power grid maintenance, comprising: Step 101: Obtain multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data to generate a standardized maintenance dataset.

[0029] In this embodiment of the invention, multi-source heterogeneous data is acquired through a multi-source data sensing unit. Specifically, the multi-source data sensing unit is deployed on the OMS data access server node, and accesses both structured and unstructured maintenance-related data in parallel through a scheduling data bus.

[0030] Structured data includes annual maintenance orders, monthly maintenance orders, day-ahead maintenance orders, operation mode change orders, power grid operation section data, and business process flow indicators, sourced from existing OMS business modules and SCADA data acquisition terminals. Unstructured data includes dispatch logs, shift handover notes, and maintenance drawings, which are standardized and converted using built-in text parsing and OCR modules.

[0031] After data access, the dynamic time-series alignment and data cleaning module built into the multi-source data sensing unit performs unified processing on data from different sources and with different sampling frequencies. Specifically, it performs time-series alignment on data with different timestamp formats, interpolates or downsamples data with different sampling frequencies, and removes or fills in outliers and missing values, ultimately achieving format unification and outlier removal. After the above processing, a standardized maintenance dataset is output, which has a unified data format, a consistent time base, and a complete field structure.

[0032] Step 102: Extract maintenance business entities, control function entities, and operation scenario entities based on the standardized maintenance dataset, map the three types of entities to the same vector space, and generate a graph embedding vector.

[0033] Further, step 102 includes the following sub-steps: S11. Extract maintenance business entities, control function entities, and operation scenario entities from the standardized maintenance dataset.

[0034] In this embodiment of the invention, the knowledge graph reasoning unit extracts three types of core entities from the standardized maintenance dataset to construct a unified multimodal knowledge graph for power grid maintenance. The maintenance business entities include three sub-entities: (1) Document entities, which extract business documents such as annual, monthly, and day-ahead maintenance orders, operation mode change orders, power outage work tickets, and start-up plans. These documents are derived from structured data in the OMS business system and have structured fields such as document number, work type, planned duration, approval node, and flow status. They belong to time-series process entities. (2) Power grid equipment entities, which extract power grid equipment such as substations, main transformers, transmission lines, circuit breakers, and busbars. These documents are derived from the SCADA / OCS power grid model and equipment ledger system and have physical attributes such as voltage level, topology connection relationship, operating status, and historical defect records. They belong to topology-related entities. (3) Unstructured extraction entities, which use natural language processing technology to process unstructured text data. Through the power field-customized named entity recognition and relationship extraction algorithm, semantic tags such as fault phenomena, defect levels, rollback reasons, and risk warnings are extracted from dispatch logs, shift handover notes, and process approval notes. These tags are then linked to the corresponding equipment or document nodes in the business sub-graph to enrich the semantic dimension of the graph. Within the business subgraph, two types of associated edges are established: physical topology relationships and business logic relationships, to fully depict the entity association network of the maintenance business.

[0035] The control function entities include two sub-entities: (1) business operation functions, including process modules such as maintenance order filling, approval flow, plan verification, and method order issuance; (2) analysis and monitoring functions, including data modules such as maintenance calendar, real-time flow monitoring, fault waveform analysis, power outage impact statistics, and workload statistics. Both types of entities are derived from the OMS system function module list and have metadata attributes such as function code, business domain, applicable position, data dependency, and visualization form. They belong to capability node type entities.

[0036] The operational scenario entity definition includes typical operational scenario entities such as daily maintenance, accident handling, monthly plan preparation, peak summer power supply, power supply guarantee for major events, and typhoon / rainstorm disaster prevention. These entities are derived from dispatch business management specifications and historical scenario annotation data. They have tagged constraint attributes such as emergency level, core control objectives, response time requirements, and scope of participating positions, and belong to rule-constrained entities.

[0037] S12. Extract multiple entity pairs from the extracted maintenance business entities, control function entities, and operation scenario entities, and obtain the association relationships between each entity pair. Form a triplet by combining each entity pair and its corresponding association relationship.

[0038] In this embodiment of the invention, multiple entity pairs are extracted from the three types of entities, and the relationships between each entity pair are obtained. For example, "500kV A main transformer" is extracted from the business entity and "typhoon disaster repair" is extracted from the scenario entity to form an entity pair, and their relationship is "trigger"; "maintenance order" is extracted from the business entity and "fault waveform analysis" is extracted from the functional entity to form an entity pair, and their relationship is "adaptation".

[0039] Each entity pair and its corresponding relationship constitute a triple. Each triple includes a head entity, a relationship, and a tail entity. For example, the triple (500kV A main transformer, trigger, typhoon disaster emergency repair) represents an equipment failure triggering an emergency repair scenario; the triple (maintenance order, adaptation, fault recording analysis) represents a maintenance order adapting to the fault recording analysis function.

[0040] S13. Perform differentiated pre-coding on the maintenance business entity, control function entity, and operation scenario entity respectively to generate the initial embedding vectors of each type of entity.

[0041] In this embodiment of the invention, a dedicated pre-encoding path is designed to address the differences in data form and attributes of the three types of entities, and to uniformly map heterogeneous features to an initial vector space of the same dimension.

[0042] Business entity encoding path: One-hot encoding and numerical normalization are used to generate basic business feature vectors for structured attribute fields, and the Bidirectional Encoder Representations from Transformers (BERT) model is used to extract semantic feature vectors for unstructured descriptive text. The two types of features are weighted and concatenated through a feature attention fusion layer, and compressed through a fully connected layer to obtain the initial embedding vector of the business entity, taking into account both structured business rules and unstructured semantic information.

[0043] Functional entity encoding path: Multi-label embedding is used to generate attribute feature vectors for functional metadata (business domain, applicable positions, data dependencies, etc.), and a one-dimensional convolutional neural network is used to extract core capability features from the functional description text; the initial embedding vector of the functional entity is obtained through fully connected mapping, highlighting the business capability boundaries and applicable scope of the functional module.

[0044] Scene entity encoding path: Hierarchical label encoding is used for discrete attributes such as scene urgency level and control objectives, and mean pooling is used to generate semantic representation vectors for typical business sample sets of the scene; the two are fused to obtain the initial embedding vector of scene entities, which enhances the business differentiation and constraint features between different scenes.

[0045] S14. The initial embedding vector of the head entity is transformed sequentially by the type calibration matrix corresponding to the type of the head entity and the relation projection matrix corresponding to the association relationship to obtain the first projection vector.

[0046] In this embodiment of the invention, in order to achieve accurate semantic association calculation of heterogeneous entities, a heterogeneous entity hierarchical projection mechanism is introduced on the Translation on Relation-specific Mapping Matrices (TransR) framework. A type calibration matrix is ​​added for the three types of entities to ensure the semantic comparability of different types of entities in the relation space.

[0047] For any triplet (head entity) Relationship Tail entity First, obtain the head entity. The initial embedding vector is then sequentially processed by the type calibration matrix corresponding to the type of the head entity. and related relationships Corresponding relation projection matrix Transform to obtain the first projection vector Among them, the type calibration matrix The relation projection matrix is ​​used to adjust the initial embedding vector of the head entity from the native vector space to a unified relation space. Used to project type-calibrated entity vectors onto the semantic subspace associated with the correspondence.

[0048] S15. The initial embedding vector of the tail entity is transformed sequentially by the type calibration matrix corresponding to the type of the tail entity and the relation projection matrix corresponding to the association relationship to obtain the second projection vector.

[0049] In this embodiment of the invention, the tail entity is obtained. The initial embedding vector is then sequentially processed by the type calibration matrix corresponding to the type of the tail entity. and related relationships Corresponding relation projection matrix Transform to obtain the second projection vector .

[0050] S16. Calculate the distance between the sum of the first projection vector and the relation vector of the association relationship and the second projection vector to obtain the energy value of the triplet.

[0051] In this embodiment of the invention, for triples The energy function is defined as: ; in, , These are the initial embedding vectors for the head and tail entities, respectively, with dimension 1. d ; For association The semantic feature vector, with dimension . This represents the direction of semantic constraints for this type of relationship; For association The corresponding projection matrix has dimensions of ; , These are the type calibration matrices corresponding to the types of the head entity and the tail entity, respectively. The distance is the L-norm distance, with the L2 norm used by default to measure the semantic difference between projected entity pairs.

[0052] Calculate the first projection vector Relationship vector with association The sum of these is then calculated, and the sum is compared with the second projection vector. The energy value of the triple is obtained by measuring the distance between them. The basic training objective is to minimize the energy value of positive sample triples and maximize the energy value of negative sample triples, thereby learning implicit semantic relationships between entities.

[0053] S17. Train the model based on the energy values ​​of each triplet, and use the trained model to map all the initial embedding vectors to generate the spectral embedding vector.

[0054] Further, step S17 includes the following sub-steps: S171. Based on the types of the head entity and tail entity in the triplet, construct negative sample triplets using the corresponding negative sampling constraint rules.

[0055] In this embodiment of the invention, the completely random negative sampling method is abandoned. Instead, the power grid business rules are followed, and negative sample triplets are constructed according to the types of the head and tail entities in the triplets, using corresponding negative sampling constraint rules. Specifically: negative sampling between equipment entities strictly follows the power grid physical topology rules, constructing negative samples only between equipment entities without direct electrical connections. For example, for a transformer entity, circuit breakers with which it has a direct electrical connection are not selected as the tail entity of the negative sample; instead, remote equipment without topological connections is selected. Negative sampling between business and function prioritizes non-compatible functions within the same business domain as negative samples to enhance the differentiation of similar functions. For example, for the "maintenance order filling" business, functions from different business domains, such as "method order issuance," are not selected as negative samples; instead, the "maintenance calendar" function, which belongs to the "planning management" business domain but is not compatible, is selected. Negative sampling between scenario and function selects scenario-function pairs with low matching degrees to construct negative samples, while ensuring that the same function corresponds to at least one highly compatible scenario and two low-compatible scenarios to balance the sample distribution.

[0056] S172. Based on the priority of power grid maintenance business, determine the differential weights corresponding to the training samples.

[0057] In this embodiment of the invention, differentiated weights are set for different triplet samples: At the scenario level, the triplet weight corresponding to high-emergency scenarios such as emergency repair and critical power supply is set to 1.5, while the weight for routine maintenance scenarios is set to 1.0. At the equipment level, the triplet weight corresponding to core equipment such as 500kV and above main transformers, hub substations, and inter-regional transmission lines is set to 1.3, while the weight for ordinary power distribution equipment is set to 1.0. At the relationship level, the weight for core relationships such as "business-trigger-scenario" and "function-adaptation-scenario" is set to 1.2, while the weight for ordinary attribute relationships is set to 1.0.

[0058] S173. Based on the difference between the energy value of each triplet and the energy value of the corresponding negative sample triplet, construct a total loss function including a marginal loss term, a weight regularization term, and a temporal smoothing regularization term.

[0059] In this embodiment of the invention, a total loss function is constructed based on the difference between the energy values ​​of each positive sample triplet and the corresponding energy values ​​of each negative sample triplet. This total loss function includes a marginal loss term, a weight regularization term, and a temporal smoothing regularization term. The total loss function is defined as follows: ; in, This is the marginal loss term; T For training the triplet sample set; Negative sample triples generated for negative sampling; The business weight coefficient for the current triple is determined by the above-mentioned hierarchical weighting rules; This is the marginal margin parameter, which controls the threshold for the energy difference between positive and negative samples. The default value is 1.0. For weight regularization terms, The set of all trainable parameters for the model; This is the L2 regularization coefficient, used to prevent overfitting of the model; the default value is 0.001. For time-series smoothing regularization, Calculate the embedding vector distance between adjacent process states of the same business entity; This is the weighting coefficient for the time series regularization term, with a default value of 0.1.

[0060] Among them, the temporal smoothing regularization term is used to constrain the semantic continuity of entity embedding in different stages of the same maintenance order, ensuring that the entity embedding in different stages of the same maintenance order has semantic continuity and avoiding semantic jumps caused by state transitions.

[0061] S174. With the goal of minimizing the total loss function, optimize the parameters of the type calibration matrix, the relation projection matrix, and the relation vector of the association relationship to obtain the trained model.

[0062] In this embodiment of the invention, the parameters of the type calibration matrix, the relation projection matrix, and the relation vector of the association relationship are optimized with the goal of minimizing the aforementioned total loss function. During training, the energy value of positive sample triples is minimized, and the energy value of negative sample triples is maximized, allowing the model to learn implicit semantic associations between entities. After training, the trained model is obtained. The trained model is then used to map all initial embedding vectors to generate graph embedding vectors for each entity, which are used for subsequent fitness score calculations.

[0063] Step 103: Extract text semantic features and flow status features from the text description of the current maintenance business and construct a scenario status vector.

[0064] In embodiments of the present invention, such as Figure 2 As shown, the scene intent recognition submodule in the intelligent decision orchestration unit constructs a dual-channel perception model of "text semantics + flow status," integrating explicit text intent with implicit flow status to accurately identify the current power grid maintenance scenario and provide scenario basis for functional scheduling. This module captures the direct intent of the scheduling scenario through the unstructured text channel, mines the implicit operational status in the business flow through the structured flow channel, and adaptively fuses the two types of features to finally output a scene state vector.

[0065] Further, step S103 includes the following sub-steps: S21. Encode the text description and extract its semantic features.

[0066] In this embodiment of the invention, the direct intent of the scheduling scenario is often implicit in unstructured text such as automated log feedback, shift handover notes, and process rollback opinions. The scenario intent recognition submodule uses a domain-adaptive BERT model as a feature extractor, adapting to power industry terminology and abbreviations.

[0067] Specifically, the domain-adaptive BERT model, based on the general BERT pre-trained model, uses corpora such as maintenance order texts, dispatch logs, and procedure documents in the power dispatching field for domain-adaptive fine-tuning training. This enables the model to accurately understand the meaning of power professional terms and abbreviations such as "planned power outage", "fault power outage", "main transformer tripping", and "line overload", and can accurately distinguish between easily confused semantics such as "planned power outage" and "fault power outage".

[0068] Let the input text sequence be The extracted semantic feature vector is represented as This vector represents the level of business urgency and event type implied in the text, providing the device with explicit criteria for scenario judgment.

[0069] S22. Within a preset time window, calculate the backlog, obstruction, and return rate of maintenance documents as raw indicators of circulation status.

[0070] In this embodiment of the invention, for scenarios with intensive business operations and congested processes that lack explicit textual descriptions, the scenario intent recognition submodule introduces business flow indicators as a supplementary perception dimension. The system uses a 2-hour sliding time window and a 15-minute rolling step size to collect maintenance process timing indicators in real time, balancing the real-time performance and stability of situational awareness.

[0071] The definitions of the three core time-series indicators are as follows: backlog The total number of yellow slips (≤24 hours from the overdue threshold) and red slips (overdue) within the statistical time window reflects the current pressure of overdue pending documents.

[0072] resistance The average dwell time of a repair order at the three key nodes of review, countersigning, and approval within the statistical time window reflects the efficiency of the core approval process.

[0073] rollback rate The percentage of rejected and returned documents within the statistical time window out of the total number of documents processed reflects the quality of the maintenance order processing.

[0074] The above three types of indicators are used as the original indicators of circulation status.

[0075] S23. Normalize the original indicators of the circulation status to obtain normalized circulation status indicators, and extract the slope of the change of each indicator between adjacent time windows.

[0076] In this embodiment of the invention, the original indicators of the three types of circulation states are normalized respectively: backlog Max-min normalization was used, with the baseline value taken from the historical maximum backlog for the same time period over the past 30 days. and minimum value After normalization: ; The results are mapped to the interval [0, 1], with higher values ​​indicating more severe backlogs.

[0077] resistance Threshold truncation and standardization are adopted, with the benchmark being the standard processing time for each node set by the business specifications. After normalization: ; When the average dwell time reaches 1 times the standard dwell time, the normalized value is set to the upper limit of 1, which indicates severe node obstruction.

[0078] rollback rate Threshold truncation standardization and normalization are adopted, with the benchmark being the normal rollback rate threshold set by the business. (Default value is 5%), after normalization: ; When the rollback rate reaches the threshold, the normalized value is set to the upper limit of 1, which indicates abnormal flow quality.

[0079] Building upon this, we supplement the logic with time-series trend feature extraction to capture the dynamic state of congestion intensification or easing, avoiding misjudgments caused by fluctuations at a single moment. We extract the slope of indicator changes over the past three consecutive time windows: ; in For the current window i Normalized index value; These are the corresponding indicator values ​​before the three windows.

[0080] S24. The normalized flow state index is concatenated with the corresponding change slope and then mapped by a multilayer perceptron to obtain the flow state features.

[0081] In this embodiment of the invention, the three normalized indicators and three trend indicators at the current moment are concatenated to obtain the original six-dimensional flow characteristics. Then, it is transformed into a state feature vector with the same dimension as the text features through a multilayer perceptron (MLP) nonlinear mapping layer: ; In the formula, The output flow status feature vector captures internal system congestion and abnormal signals, making up for the information blind spots in the text channel.

[0082] Specifically, an example of feature generation for a backlog of work orders: Scenario: During the peak period for monthly maintenance plan submissions, at 10:00 AM, maintenance orders have been concentrated in the window for nearly 2 hours, resulting in significant processing congestion.

[0083] Known parameters: Historical backlog baseline =2、 =35, the total number of yellow and red slips in the current window. =28; Standard duration of resistance =4 hours, average stay at the current signing node =5.2 hours; rollback rate threshold =5%, Current window back rate =3.2%; the first three window indicators all showed an upward trend, with the following slopes: =0.08、 =0.12、 =0.03.

[0084] Calculation process: Backlog normalization: ; Normalization of resistance: ; Backoff rate normalization: ; The concatenated 6-dimensional flow original feature vector is obtained as [0.788, 1.0, 0.64, 0.08, 0.12, 0.03]. This feature vector is then nonlinearly mapped by a multilayer perceptron (MLP) and input into a gating fusion unit to participate in scene determination. In this example, the obstruction degree reaches the maximum value, the backlog degree is high and shows an upward trend, which will be determined as the "monthly plan preparation - process congestion" sub-scene, triggering the pre-scheduling of functions such as workload statistics and batch approval assistance.

[0085] S25. Adaptive weighted fusion of text semantic features and flow state features is performed using a gating fusion mechanism, and the fusion result is mapped into a scene state vector through a classification layer.

[0086] In this embodiment of the invention, to achieve intelligent weighted fusion of two types of information, the scene intent recognition submodule is designed with a gated fusion unit (GMFU) that dynamically calculates weight coefficients based on the information entropy of the input information, avoiding information mutual exclusion and noise interference caused by simple splicing. ; in, This is the gating weight vector, with values ​​in the range [0, 1], which controls the fusion ratio of text features and flow features; For the trainable weight matrix and bias terms of the gated unit; This is an operation that concatenates text feature vectors with flowing feature vectors.

[0087] The final scene state vector is obtained by weighted fusion of two types of features: ; In the formula, This is an element-wise multiplication operation. When the text contains clear high-value information such as accidents or malfunctions, the gating weights automatically shift towards the text channel; when the text is normal but the process is heavily backed up, the weights automatically shift towards the flow channel, simulating the human brain's attention allocation mechanism.

[0088] Finally, the fused scene vector is input into the fully connected classification layer, and the probability distribution of the current time belonging to each preset business scene is output through the Softmax function to complete the scene determination.

[0089] Step 104: Input the scene state vector and graph into the vector input decision model, and calculate the fit score between the control function entity and the current maintenance business entity based on the graph embedding vector.

[0090] In embodiments of the present invention, such as Figure 2 As shown, the reinforcement learning decision submodule in the intelligent decision orchestration unit receives the scene state vector output by the scene intent recognition submodule and the graph embedding vector output by the knowledge graph reasoning unit, and inputs both into the decision model. Based on the graph embedding vector, the adaptation score between each control function entity and the current maintenance business entity is calculated through a graph attention mechanism, providing a quantitative basis for subsequent function scheduling.

[0091] Further, step S104 includes the following sub-steps: S31. Based on the graph attention mechanism, aggregate the multi-hop association paths between the current maintenance business entity and each control function entity from the graph embedding vector to obtain the path aggregation features of each control function entity.

[0092] In this embodiment of the invention, when the device senses the core business object currently being processed, the system calculates the association strength between candidate modules in the function pool and that business object. Based on the graph attention mechanism, the current maintenance business entities are aggregated from the graph embedding vector. b To each control and management functional entity c The number of hops in the multi-hop associated paths should be controlled between 1 and 3, and long paths with no business significance should be filtered out.

[0093] For each associated path Calculate the semantic similarity between entities at both ends of the path. (Using the cosine similarity function), and combined with path attention weights Weighting is performed to obtain path aggregation features. .

[0094] S32. Obtain the prior usage probability of each control function entity in the current scenario based on the scenario state vector.

[0095] In this embodiment of the invention, the current scene is obtained. s Subsequent control and management functional entities c Prior use probability This represents the historical high-frequency usage probability of this functional module in the current scenario. The prior probability is calculated using Bayesian smoothing. ; in, For the scene s Next function c The cumulative number of times it has been used in history; This is the smoothing coefficient (default value is 1). N This is the total number of functional modules to avoid zero-probability issues in low-frequency scenarios.

[0096] S33. Based on the aggregation characteristics of each path and the corresponding prior usage probability, calculate the fit score between each control function entity and the current maintenance business entity.

[0097] In this embodiment of the invention, the suitability score is obtained by combining the semantic similarity of multi-hop associated paths with the prior probability of the scenario: ; in, For business entities b With candidate functional modules c The fit score ranges from [0, 1], with higher scores indicating stronger fit. This is an adjustment factor with a value range of [0, 1]. It balances the contribution ratio of semantic similarity of multi-hop paths to the prior probability of the scene, and the default value is 0.6. The Sigmoid activation function normalizes the calculation results to the [0, 1] interval.

[0098] Furthermore, the parameter iteration update mechanism is as follows: The prior probabilities of each scenario are updated daily: the click frequency and usage duration of all functional modules in each scenario are statistically analyzed offline at midnight every day, and the prior probabilities are updated smoothly using Bayesian methods.

[0099] The path attention weights are trained incrementally on a weekly basis: each week, newly added maintenance work orders and user operation data are extracted to construct a business-function matching sample set, the graph attention network is trained incrementally, the attention weights of each associated path are updated, the weights of high-frequency effective paths are strengthened, and the influence of ineffective paths is weakened.

[0100] Real-time feedback and fine-tuning: When users manually adjust the function layout or actively search for functions, this operation is regarded as negative feedback, and the corresponding business-function pair's adaptability score is adjusted in real time to achieve lightweight optimization at the hourly level.

[0101] Through this mechanism, the device can not only match explicit rule associations, but also uncover deep implicit associations, breaking through the adaptation limitations of traditional hard-coded rules and providing a precise set of candidate functions for subsequent reinforcement learning scheduling.

[0102] Specifically, a calculation example for a typhoon power outage repair scenario: Scenario setting: Under typhoon disaster weather, the current core business entity is the "500kV A main transformer fault outage maintenance work order", the current operating scenario is "typhoon disaster emergency repair", and there are 3 candidate functional modules: fault recording analysis module, outage impact statistics module, and maintenance calendar module.

[0103] Known preconditions: adjustment factor The prior probabilities of the scenarios are respectively , , There are two effective paths from each business entity to each function. The path weights and corresponding cosine similarities are shown in Table 1. Table 1

[0104] Calculation process: 1. Weighted semantic similarity of fault recording analysis module: ; Median fit: ; After activation by Sigmoid: .

[0105] 2. Weighted semantic similarity of the power outage impact statistics module: ; Median fit: ; After activation by Sigmoid: .

[0106] 3. Weighted semantic similarity of the maintenance calendar module: ; Median fit: ; After activation by Sigmoid: .

[0107] Results show that in typhoon emergency repair scenarios, the fault recording analysis and power outage impact statistics modules have significantly higher adaptability and will be scheduled to the visual hotspots on the interface; the maintenance calendar module has low adaptability and will be automatically folded and stored, which is a perfect match for actual business needs.

[0108] Step 105: Determine the target control functions and their display layout based on the adaptability score, and generate control and scheduling strategies.

[0109] In embodiments of the present invention, such as Figure 2As shown, the reinforcement learning decision submodule in the intelligent decision orchestration unit transforms the scheduling and layout problem of maintenance control functions into a Markov decision process of intelligent agents in the power grid business environment based on the adaptation scores of each control function entity calculated in step 104. It determines the target control function and its display layout through reinforcement learning algorithm and generates control scheduling strategy.

[0110] Further, step S105 includes the following sub-steps: S41. Obtain the current user role vector, and concatenate the scene state vector, the current user role vector, and the current function display state vector into a state vector.

[0111] In this embodiment of the invention, the scheduling and layout problem of maintenance and control functions is transformed into a Markov decision process for an intelligent agent in a power grid business environment. The state space is composed of three types of features: scenario, user, and current layout. ; in This is the scene state vector; Embed vectors (encoding job roles, historical operation preferences, and business permissions) into user profiles. This is the display state vector for the current functional module.

[0112] S42. Based on the preset set of power grid maintenance and control function modules, the function selection instruction, priority sorting instruction, and display area allocation instruction are defined as the action space.

[0113] In this embodiment of the invention, the preset power grid maintenance and control function module set refers to a list of all available maintenance and control function modules pre-configured in the OMS system, including business operation functions (such as maintenance order filling, approval workflow, plan verification, and method order issuance) and analysis and monitoring functions (such as maintenance calendar, real-time workflow monitoring, fault waveform analysis, power outage impact statistics, and workload statistics), totaling... N There are several candidate functional modules. The executable operations are defined as an action space, including instructions in three dimensions: function selection instructions (using...). N 3D binary vector representation, the first i A value of 1 indicates activation of the first dimension. i The functions include: a function module (0 indicates inactive), a priority sorting instruction (to set the display level and order of the selected function modules), and a display area allocation instruction (to allocate the display area and size of the terminal interface to each activated function module).

[0114] S43. Using the state vector as the state input of the reinforcement learning agent and the multi-objective composite reward function as the optimization objective, the reinforcement learning agent selects the optimal action from the action space.

[0115] In this embodiment of the invention, the reinforcement learning agent adopts a Dueling Deep Q-Network (DQN) architecture, using state vectors... s As input, the output is the Q-value corresponding to each action in the action space. The network structure consists of three parts: a shared feature extraction layer, a state value branch, and an action advantage branch. The state vector... s First, the feature representation is mapped to a high-dimensional feature representation through a shared feature extraction layer. Then, the state value branch and the action advantage branch are input separately, and the state value is output respectively. and action advantages The final Q value is obtained by combining the two. ,in, In the state s Take action below a The action value function, i.e., the Q value; s The current system state is composed of three types of features: scenario, user, and current layout. a The currently executing function scheduling action; It is a state value function that estimates the basic value of the current state, regardless of specific actions; Let the action advantage function estimate the specific action to be taken. a Advantages relative to the average level of all actions in the current state; These are parameters for the network sharing portion; Network parameters for the action-dominant branch; Network parameters for state-value branches. Multi-objective composite reward function. Used to evaluate the effect of an agent taking a certain action, where Efficiency bonus items (positive incentives will be given if the average processing time for exceeding the expected time under the new layout is lower than the historical average). A focus reward is given when core control functions in high-risk scenarios are placed in visually hot zones. This is an interaction cost penalty item (a negative feedback penalty is generated when the user manually adjusts the settings). These are the weighting coefficients for the three reward components, all of which are positive values, used to balance the priorities of different optimization objectives. The weighting coefficients are determined using a two-stage approach: "business-prior initialization + online iterative fine-tuning." In the initialization stage, initial weights are set based on the experience of power grid dispatching business experts, with default values... Prioritize efficiency in handling incidents while also considering information engagement and user experience; for high-urgency scenarios such as emergency repairs, the response time can be manually increased. Upgrade to 0.4, strengthening the reward weight for core functions.

[0116] Online fine-tuning phase: During device operation, the weight coefficients are fine-tuned online using the gradient descent method based on business KPI feedback. The weights are increased when document processing timeliness fails to meet standards. Increase when the user manually adjusts too frequently. This allows reward weights to be dynamically adapted to business needs.

[0117] Intelligent agents adopt Strategic balance exploration and utilization, in order to Probability of randomly selecting actions, in order to Probably choose the action with the highest Q value This means obtaining the optimal action.

[0118] S44. Convert the optimal action into a control and scheduling strategy that includes target control functions, their display priorities, and display area allocation schemes.

[0119] In this embodiment of the invention, the intelligent decision-making orchestration unit converts the optimal action output by the reinforcement learning agent into a specific control and scheduling strategy. a It is a parameterized representation of instructions with three dimensions. The specific transformation logic is as follows: First, the function selection vector is parsed, and the function modules corresponding to the dimensions with a value of 1 are extracted to form a target control function set; then, the priority scores are parsed, and the display level and display order of each function module are determined according to the scores from high to low; finally, the area allocation code is parsed to determine the display area (visual hot zone, regular area, edge area) and display size (large size, medium size, compact size) of each function module. After the above parsing and transformation, the system generates a complete control and scheduling strategy, including three core elements: target control functions (which function modules are activated), display priority (the display level and order of each function module), and display area allocation scheme (the display area and size of each function module on the terminal interface).

[0120] Step 106: Render and output the maintenance control interaction interface according to the control and scheduling strategy, and collect user operation data on the interaction interface.

[0121] In this embodiment of the invention, the adaptive interactive terminal unit is deployed on the dispatcher's human-computer interaction workstation and serves as the execution terminal for device control strategies. The terminal unit receives scheduling instructions from the intelligent decision-making and orchestration unit, drives the front-end rendering engine to dynamically call the corresponding maintenance and control function modules, and automatically adjusts the start / stop status, display priority, and display size of the function modules to generate a personalized maintenance and control workbench adapted to the current scenario and position for the dispatcher.

[0122] Specifically, high-priority core control functions are automatically pinned to the top and enlarged to the visual hotspot, while low-priority non-urgent functions are automatically collapsed and stored; the function set within the permission range is automatically matched for different job roles to achieve differentiated control interface adaptation.

[0123] Meanwhile, the terminal unit collects real-time data on the dispatcher's function clicks, interface adjustments, document processing, and other operational behaviors, and transmits this data back to the experience playback pool of the intelligent decision-making and orchestration unit via the scheduling data bus.

[0124] Step 107: Update the parameters of the decision model based on the user operation data, and return to execute the step of embedding the scene state vector and the graph into the vector input of the decision model.

[0125] In this embodiment of the invention, the parameters of the decision model are updated based on user operation data. Specifically, user operation data is stored as feedback samples in an experience replay pool, historical samples are sampled from the experience replay pool, and the temporal difference loss is calculated based on the sampled samples to update the network parameters of the decision model.

[0126] This closed-loop feedback mechanism provides iterative feedback for the reinforcement learning model, enabling the device's functional scheduling strategy to continuously adapt to the scheduling personnel's operating habits and business changes, forming a complete control closed loop of "perception-reasoning-decision-presentation-feedback", supporting the continuous optimization of the device's control capabilities.

[0127] Please see Figure 3 , Figure 3 This is a structural block diagram of a power grid maintenance adaptive control system provided in an embodiment of the present invention.

[0128] This invention provides an adaptive control system for power grid maintenance, comprising: The data perception module 301 is used to acquire multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data, and generate a standardized maintenance dataset. The knowledge graph reasoning module 302 is used to extract maintenance business entities, control function entities and operation scenario entities based on a standardized maintenance dataset, map the three types of entities to the same vector space, and generate graph embedding vectors. The scene recognition module 303 is used to extract text semantic features and flow status features from the text description of the current maintenance business and the flow status data to construct a scene status vector. The adaptation calculation module 304 is used to embed the scene state vector and the graph into the vector input decision model, and calculate the adaptation score between the control function entity and the current maintenance business entity based on the graph embedding vector. The decision-making and scheduling module 305 is used to determine the target control functions and their display layout based on the adaptability score, and to generate control and scheduling strategies. The interactive output module 306 is used to render and output the maintenance control interactive interface according to the control and scheduling strategy, and to collect user operation data on the interactive interface. The feedback optimization module 307 is used to update the parameters of the decision model based on user operation data and return to the step of embedding the scene state vector and the graph into the vector input of the decision model.

[0129] Since the above is a system corresponding to a power grid maintenance adaptive control method, its implementation principle is the same as that of a power grid maintenance adaptive control method. For the sake of convenience and brevity, those skilled in the art can clearly understand that the specific working process of the system and modules described above can be referred to the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0130] Furthermore, in a specific embodiment of the present invention, the deployment and implementation are based on the provincial power dispatching OMS cloud platform, reusing the dispatching server hardware cluster and dispatching data bus resources of the existing provincial integrated OMS system as an intelligent upgrade unit of the system maintenance and control module, supporting the project construction goals such as comprehensive power outage management and full-chain digital control of maintenance plans.

[0131] In the initial implementation phase, the power grid maintenance adaptive control system, following step 101, accesses multi-source heterogeneous data from the OMS system, including maintenance orders, mode change orders, SCADA equipment operation data, scheduling logs, and process flow indicators, through a multi-source data sensing unit. This data is cleaned and time-series aligned to generate a standardized maintenance dataset. Subsequently, following step 102, the knowledge graph reasoning unit extracts maintenance business entities, control function entities, and operation scenario entities from the standardized maintenance dataset based on a graph database module. This constructs a "business-function-scenario" ternary multimodal knowledge graph. The TransR model is used to embed entity vectors and train associations, generating graph embedding vectors and establishing a semantic association network for all maintenance elements.

[0132] In daily operation, the power grid maintenance adaptive control system, according to step 103, receives real-time data input from the perception unit through the scene intent recognition submodule in the intelligent decision orchestration unit: the text semantic channel parses the daily scheduling log, maintenance notes, and process rollback opinions to extract text semantic features; the flow status channel monitors the backlog of documents, the dwell time of key nodes, and the rollback rate to extract flow status features, which are then output as a scene state vector after adaptive weighted fusion by the gating fusion unit, automatically determining the current business scenario. Based on this, according to step 104, the reinforcement learning decision submodule in the intelligent decision orchestration unit embeds the scene state vector and the graph into the vector input decision model, calculating the fit score between each control function entity and the current maintenance business entity based on the graph attention mechanism. According to step 105, based on the fit score, the scheduling and layout problem of the maintenance control function is transformed into a Markov decision process of a reinforcement learning agent, determining the target control function and its display layout, and generating a control scheduling strategy. According to step 106, the adaptive interactive terminal unit renders and outputs the maintenance control interactive interface according to the control scheduling strategy.

[0133] For example, when the system detects a concentrated submission of monthly maintenance documents and a significant backlog of countersigning nodes, it identifies the scenario as "monthly plan preparation" according to step 103. The system then reinforces the learning and decision-making module by combining the responsible job roles with historical operating habits, and outputs a functional scheduling strategy according to step 105. The adaptive interactive terminal unit automatically prioritizes functional modules such as maintenance calendars, workload statistics, and order quality analysis according to step 106, while folding non-urgent fault analysis modules, assisting dispatchers in efficiently completing plan approvals. When the log text identifies keywords such as "transformer trip" or "line fault," the system immediately identifies the scenario as "accident repair," automatically highlighting core control functions such as fault waveform analysis, power-on testing management, and power outage impact statistics in the terminal's visual hotspot, while simultaneously folding non-urgent functions such as monthly plans and statistical reports, achieving second-level pre-positioning of critical handling capabilities.

[0134] During operation, the power grid maintenance adaptive control system, following step 107, collects real-time operational data from dispatchers, including function clicks, interface adjustments, and document processing, through the adaptive interactive terminal unit. This data is then transmitted back to the experience playback pool of the intelligent decision orchestration unit, driving the reinforcement learning model to continuously iterate and optimize the network parameters of the decision model. This ensures that the system's functional scheduling strategy constantly adapts to the dispatchers' operational habits and business changes, forming a complete control loop of "perception-reasoning-decision-presentation-feedback." The system architecture possesses general scalability; by replacing corresponding business entities and functional nodes, it can be expanded to adapt to more OMS business scenarios, such as operational mode adjustments and emergency dispatch command, meeting the multi-business intelligent upgrade needs of the provincial integrated dispatch system.

[0135] Compared to traditional statically configured maintenance and control modules, this invention differs significantly in data fusion and reasoning, scenario-adaptive control, and human-computer interaction optimization. As described in step 102, by constructing a multimodal knowledge graph that couples business, function, and scenario, maintenance business documents, power grid equipment topology, and control function modules can be incorporated into a unified semantic association network. This provides a knowledge system foundation for associative reasoning of heterogeneous data discretely stored in the OMS system and SCADA acquisition terminals. Furthermore, as described in step 104, by aggregating multi-hop association paths from the graph embedding vectors through a graph attention mechanism, it helps to uncover deep implicit relationships between business entities and function modules. This overcomes, to some extent, the limitations of traditional static configuration models that rely on hard-coded rules for function adaptation, providing more comprehensive data and knowledge support for maintenance and control decisions.

[0136] As described in step 103, through a dual-channel gating perception mechanism of text semantics and flow status, the system can proactively identify diverse operational scenarios such as daily maintenance, emergency repair, and power supply, reducing reliance on manual triggering for switching. Specifically, the unstructured text channel uses a domain-adaptive BERT model to capture explicit business intent from scheduling logs and shift handover notes, while the structured flow channel mines implicit operational status through time-series indicators such as backlog, obstruction, and rollback rate. Both are adaptively weighted by the gating fusion unit to output a scenario state vector. Based on this, as described in step 105, combined with a reinforcement learning adaptive scheduling algorithm, the system can dynamically adjust the priority and display strategy of control functions according to the current scenario. This facilitates the intelligent pre-positioning of key control capabilities, shortening the time for dispatchers to retrieve information and switch functions in emergency fault scenarios, and improving the efficiency of maintenance response and the accuracy of control.

[0137] As described in step 105, by concatenating the scenario state vector, user role vector, and current function display state vector into a state vector, the system can dynamically determine the target control function and its display layout based on job roles, historical operating habits, and the current business scenario. This transforms the traditional interaction mode of people searching for functions into a proactive adaptation mode of functions finding people, which can reduce the information overload and operational burden on dispatchers and reduce the risk of human error. At the same time, as described in steps 106 and 107, by collecting user operation behavior data, storing it in the experience playback pool, and sampling historical interaction samples, the network parameters of the decision model are updated with time-series differential loss. This can support the continuous adaptation of control strategies to the dynamic changes in business needs, providing a technical foundation for the digital control of the entire power grid maintenance process.

[0138] Please see Figure 4 , Figure 4 This is a structural block diagram of an electronic device provided in Embodiment 4 of the present invention.

[0139] An electronic device according to an embodiment of the present invention includes: a memory 401 and a processor 402. The memory 401 stores a computer program. When the computer program is executed by the processor 402, the processor 402 executes the power grid maintenance adaptive control method as described in the above embodiment.

[0140] Memory 401 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 401 has storage space 403 for program code 413 for performing any of the method steps described above. For example, storage space 403 for program code may include various program codes 413 for implementing the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the adaptive control method for power grid maintenance described above.

[0141] Embodiment 5 of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power grid maintenance adaptive control method as described in the above embodiments.

[0142] Embodiment 6 of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the power grid maintenance adaptive control method as described in the above embodiments.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0144] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0145] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0146] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The above 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.

Claims

1. An adaptive control method for power grid maintenance, characterized in that, include: Acquire multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data to generate a standardized maintenance dataset; Based on the standardized maintenance dataset, maintenance business entities, control function entities, and operation scenario entities are extracted, and the three types of entities are mapped to the same vector space to generate a graph embedding vector. Based on the text description of the current maintenance business, extract text semantic features and flow status features from the flow status data to construct a scenario status vector; Input the scenario state vector and the graph embedding vector into the decision model, and calculate the fit score between the control function entity and the current maintenance business entity based on the graph embedding vector; Based on the adaptation score, the target control functions and their display layout are determined, and a control and scheduling strategy is generated. The maintenance control interaction interface is rendered and output according to the control and scheduling strategy, and user operation data on the interaction interface is collected. The parameters of the decision model are updated based on the user operation data, and the process returns to the step of inputting the scene state vector and the graph embedding vector into the decision model.

2. The adaptive control method for power grid maintenance according to claim 1, characterized in that, The step of extracting maintenance business entities, control function entities, and operation scenario entities based on the standardized maintenance dataset, mapping the three types of entities to the same vector space, and generating a graph embedding vector includes: Extract maintenance business entities, control function entities, and operation scenario entities from the standardized maintenance dataset; Multiple entity pairs are extracted from the maintenance business entity, the control function entity, and the operation scenario entity, and the association relationship between each entity pair is obtained. Each entity pair and its corresponding association relationship are formed into a triplet, and each triplet includes a head entity, the association relationship, and a tail entity. Differentiated pre-coding is performed on the maintenance business entity, the control function entity, and the operation scenario entity respectively to generate initial embedding vectors for each type of entity; The initial embedding vector of the head entity is transformed sequentially by the type calibration matrix corresponding to the type of the head entity and the relation projection matrix corresponding to the association relationship to obtain the first projection vector; The initial embedding vector of the tail entity is transformed sequentially by the type calibration matrix corresponding to the type of the tail entity and the relation projection matrix corresponding to the association relationship to obtain the second projection vector; The energy value of the triple is obtained by calculating the distance between the sum of the first projection vector and the relationship vector of the association relationship and the second projection vector; The model is trained based on the energy values ​​of each triple, and the trained model is used to map all the initial embedding vectors to generate a graph embedding vector.

3. The adaptive control method for power grid maintenance according to claim 2, characterized in that, The step of training the model based on the energy values ​​of each triplet includes: Based on the types of the head and tail entities in the triplet, negative sample triplets are constructed using the corresponding negative sampling constraint rules. Based on the priority of power grid maintenance services, determine the differentiated weights corresponding to the training samples; Based on the difference between the energy value of each triplet and the energy value of the corresponding negative sample triplet, a total loss function is constructed, which includes a marginal loss term, a weight regularization term, and a temporal smoothing regularization term. The temporal smoothing regularization term is used to constrain the semantic continuity of entity embedding in different stages of the same maintenance order. With the goal of minimizing the total loss function, the parameters of the type calibration matrix, the parameters of the relation projection matrix, and the parameters of the relation vector of the association relationship are optimized to obtain the trained model.

4. The adaptive control method for power grid maintenance according to claim 1, characterized in that, The step of extracting textual semantic features from the textual description of the current maintenance business and extracting flow status features from the flow status data to construct a scenario state vector includes: The text description is encoded, and its semantic features are extracted. Within a preset time window, the backlog, obstruction, and return rate of maintenance documents are statistically analyzed as raw indicators of circulation status. The original indicators of the circulation status are normalized to obtain normalized circulation status indicators, and the slope of the change of each indicator between adjacent time windows is extracted. The normalized flow state index is concatenated with the corresponding change slope and then mapped by a multilayer perceptron to obtain the flow state features. A gating fusion mechanism is used to adaptively weight and fuse the text semantic features and the flow state features, and the fusion result is mapped into a scene state vector through a classification layer.

5. The adaptive control method for power grid maintenance according to claim 1, characterized in that, The step of inputting the scenario state vector and the graph embedding vector into the decision model, and calculating the fit score between the control function entity and the current maintenance business entity based on the graph embedding vector, includes: Based on the graph attention mechanism, multi-hop association paths between the current maintenance business entity and each of the control function entities are aggregated from the graph embedding vector to obtain the path aggregation features of each of the control function entities. The prior usage probability of each of the control function entities in the current scenario is obtained based on the scenario state vector. Based on the aggregation features of each path and the corresponding prior usage probability, the fit score between each control function entity and the current maintenance business entity is calculated.

6. The adaptive control method for power grid maintenance according to any one of claims 1-5, characterized in that, The step of determining the target control function and its display layout based on the adaptability score, and generating a control scheduling strategy, includes: Obtain the current user role vector, and concatenate the scene state vector, the current user role vector, and the current function display state vector into a state vector; Based on the preset set of power grid maintenance and control function modules, the function selection instruction, priority sorting instruction, and display area allocation instruction are defined as the action space; Using the state vector as the state input of the reinforcement learning agent, and taking the multi-objective composite reward function as the optimization objective, the reinforcement learning agent selects the optimal action from the action space. The optimal action is converted into a control and scheduling strategy that includes target control functions, their display priorities, and display area allocation schemes.

7. A power grid maintenance adaptive control system, characterized in that, include: The data perception module is used to acquire multi-source heterogeneous data, clean, time-series aligned and standardized the multi-source heterogeneous data, and generate a standardized maintenance dataset. The knowledge graph reasoning module is used to extract maintenance business entities, control function entities and operation scenario entities based on the standardized maintenance dataset, map the three types of entities to the same vector space, and generate a graph embedding vector. The scene recognition module is used to extract text semantic features from the text description of the current maintenance business and extract flow status features from the flow status data to construct a scene status vector; The adaptability calculation module is used to input the scenario state vector and the graph embedding vector into the decision model, and calculate the adaptability score between the control function entity and the current maintenance business entity based on the graph embedding vector. The decision-making and scheduling module is used to determine the target control function and its display layout based on the adaptability score, and to generate control and scheduling strategies. The interactive output module is used to render and output the maintenance control interactive interface according to the control and scheduling strategy, and to collect user operation data on the interactive interface. The feedback optimization module is used to update the parameters of the decision model based on the user operation data, and return to execute the step of inputting the scene state vector and the graph embedding vector into the decision model.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the adaptive control method for power grid maintenance as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the adaptive control method for power grid maintenance as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the power grid maintenance adaptive control method as described in any one of claims 1-6.