Power grid dispatching operation ticket checking method for intelligently matching power failure and power transmission measures

Abstract

The invention belongs to the technical field of power grid dispatching, and particularly relates to a power grid dispatching operation ticket checking method for intelligently matching power cut and power transmission measures, and the method comprises the steps: obtaining operation instruction information in an operation ticket and power cut and power transmission measure information in a maintenance application form through a large language model; the power failure and transmission measure information is matched with the operation instruction information by constructing an expert rule knowledge graph, and the correctness of the operation order instruction sequence from top to bottom is checked according to the matching result. Whether the operation order operation instruction is correct or not can be logically verified from the bottom layer based on the overhaul application power failure and power transmission measure reply, the checking depth and breadth are improved, and the change of the checking rule can be flexibly adapted.

Description

Technical Field

[0001] The present application belongs to the technical field of power grid dispatching, and specifically relates to a method for verifying power grid dispatching operation tickets for intelligently matching power outage measures. Background Art

[0002] Operation process verification is an important part of the operation instruction review. The dispatcher must not only verify the contents of the instructions item by item, but also verify the connection, logical sequence, and rationality between the various operation instructions according to the business regulations of the primary and secondary equipment and the power outage measures approved by the power grid maintenance application to ensure that the entire process of the operation instruction is correct.

[0003] For complex power outage tasks, in addition to the conventional "maintenance-cold standby-hot standby-operation" equipment state conversion, it is necessary to combine equipment line disconnection, secondary circuit changes and other work, comprehensively consider line phase verification, protection measurement and other power outage requirements, and ensure that the primary and secondary equipment operations are coordinated. Therefore, executing a successful complex power outage task requires taking into account many aspects, and it also tests the dispatcher's understanding of the operating instructions.

[0004] Dispatchers are bound to make omissions or misunderstandings when faced with complex power outage verification tasks. The existing error prevention system is unable to match the operation ticket with the power grid maintenance application form, and cannot combine the power outage measures approved in the maintenance application with the operation instructions, and thus cannot verify whether the operation ticket is correct from the underlying logic. For example, the existing error prevention system only performs one-time equipment error prevention judgment, and cannot carry out coordinated verification of primary and secondary equipment operation instructions. Summary of the invention

[0005] The present application provides a method for verifying power grid dispatching operation tickets for intelligently matching power outage measures, which can verify whether the operation instructions of the operation ticket are correct from the bottom-level logic based on the approval of the maintenance application.

[0006] The technical solution of this application is as follows: A method for verifying a power grid dispatching operation ticket for intelligently matching power outage measures comprises the following steps: S1. Construct a rule knowledge graph based on the rule knowledge text, which specifically includes the following steps: S101, using devices, operation rules of devices, and location relationships between devices as entities, attributes, and relationships of a rule knowledge graph, respectively; S102, collecting rule knowledge text and preprocessing it, and performing sentence processing on the preprocessed text to divide it into a number of text paragraphs with overlapping texts; S103, extracting entities, attributes, and relationships in the text paragraph using a large language model, and generating text analysis; S104, using a large language model to merge the extracted repeated entities and their attributes, and parsing the merged text to construct a data set; S105, parsing the data set into a CSV file and storing it; S2. Obtain the power grid operation ticket to be verified, and use the large language model to extract information from it to obtain operation instructions with a logical order in the power grid operation ticket, wherein the operation instructions act on the operation object; S3, using the large language model to use the operation object as a search field in the power grid maintenance application database to obtain the maintenance application form containing the operation object; S4. Use a large language model to extract and parse the text data of the power outage / power transmission measures approval in the maintenance application form. After parsing, the primary equipment state change, secondary equipment startup and shutdown logic, and dispatching operation terms contained in the text of the power outage / power transmission measures approval are obtained. During the parsing process, the rule knowledge graph is called to semantically understand the text of the power outage / power transmission measures approval; S5. Use the large language model to match the primary equipment state change, secondary equipment startup and shutdown logic, and dispatch operation terms analyzed in step S4 with the operation instructions in the operation ticket in sequence. If the power outage / power supply measures and the operation instructions can be successfully matched in sequence, the operation instruction verification passes; otherwise, the operation instruction verification fails.

[0007] Furthermore, in step S103, a scheduling professional term list library is constructed for the large language model to understand entities, attributes, and relationships in text paragraphs; the scheduling professional term list library includes device name terms, abbreviations of device name terms, synonymous names, and device status terms, abbreviations of device status terms, and description statements of device status.

[0008] Furthermore, in step S103, a scheduling operation prompt word template is constructed for the large language model to extract entities, attributes, and relationships in the text paragraph; the prompt word template includes the extraction sources of device name terms and device status terms, conversion examples and restriction examples of device status, and the operation sequence, operation restrictions, and operation conditions of the device.

[0009] Furthermore, in step S2, when the large language model extracts information from the power grid operation ticket to be verified, the rule knowledge graph is called.

[0010] Furthermore, in step S2, when extracting information from the power grid operation ticket to be verified using the large language model, the following steps are included: S201, structurally representing the operation instruction of the power grid operation ticket as a combination of an operation object, an operation action, and an operation state; S202, using a large language model to identify and extract operation instructions based on a combination of an operation object and an operation action or a combination of an operation object and an operation state; S203: parse and store the extracted operation instructions.

[0011] Furthermore, the process of calling the rule knowledge graph is as follows: SA1. The large language model locates the initial entities it contains according to the query question and sets the maximum number of iterations; SA2, obtain the set of initial entities, take the initial entity as the starting point of the search path, and thus obtain the set consisting of all initial entities; SA3. Iterate the relationship search and entity search when the maximum number of iterations has not been reached. After each iteration, ask the large language model whether the current retrieval path satisfies the query. If so, use the retrieval path as the call result of the large language model. If not, prune the entity and continue to iterate until the maximum number of iterations is reached.

[0012] Furthermore, in step SA3, the relation search groups all relations of the current entity into a candidate relation set, and the large language model calculates the contribution of each relation in the candidate relation set based on the query question, retaining the top relations. N relations and prunes other relations to complete the current relation search iteration; where N is the number of initial entities; In step SA3, the entity search will be based on the current N The corresponding candidate entity set is obtained by the large language model, and the contribution of each entity in the candidate entity set is completed, and the previous N entities and prune other entities to complete the current entity search iteration, and update the corresponding N A search path.

[0013] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows: 1. The operation ticket has standard rules in the process of drafting the ticket. This application takes advantage of the standardized characteristics of the operation ticket, uses a large language model to extract the operation instruction information contained therein, and uses the operation instruction as a benchmark to seek verification from the power outage measures in the underlying maintenance application form. This top-down verification process can comprehensively discover the source of errors and improve the accuracy of verification.

[0014] 2. This application uses a large language model to verify the operation ticket, which realizes the verification of the logical accuracy of the operation ticket. It not only avoids errors in the content and form of the operation ticket, but also can deeply verify its logical errors. It helps the reviewers to learn the internal logic of the draft ticket, accumulate draft ticket experience, and form a closed loop of "learning-feedback-iteration".

[0015] 3. This application will establish a rule knowledge graph and use the knowledge graph in the large language model to make the application scenario closer to reality and help dispatchers adapt to learning. At the same time, the model is also made more flexible through expert experience. On the one hand, the wording in the application for maintenance is more flexible. On the other hand, when the verification rules change, the knowledge graph can be modified to flexibly adapt to the verification rules.

[0016] 4. The present application realizes the coordinated verification of primary and secondary equipment. The traditional anti-error system only has the verification of primary equipment. In the technical solution of the present application, the large language model can simultaneously identify the operating status and operating actions of the primary and secondary equipment, and match the recognition results with the operation tickets, further improving the depth of the verification.

[0017] 5. The present invention adopts the collaborative working mode of knowledge graph and large language model, extracts diverse and multi-step retrieval paths from the knowledge graph as the basis for large language model reasoning, enhances the deep reasoning ability of the large language model in knowledge-intensive tasks, and at the same time, the clear reasoning path improves the interpretability of the large language model reasoning process, and enables the source of the model output to be tracked and corrected. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0019] Figure 1 A flow chart of a method for checking operation instructions of a power grid operation ticket provided in this application; Figure 2 This is a flowchart of the construction process of the rule knowledge graph; Figure 3 It is a schematic diagram of extracting entities, relations and attributes of a large model based on a prompt word template; Figure 4 This is a flow chart of the integration of rule knowledge graph and large language model in this application; Figure 5 Flowchart for extracting operation ticket information for the large language model in this application; Figure 6 This is a schematic diagram of the large language model algorithm architecture in this application. DETAILED DESCRIPTION

[0020] Based on the background technology, this application provides a method for checking the operation instructions of a power grid operation ticket. Figure 1 , including the following steps: S1, construct a rule knowledge graph based on the rule knowledge text, see the attached Figure 2 , specifically including the following steps: S101, using devices, operation rules of devices, and location relationships between devices as entities, attributes, and relationships of a rule knowledge graph, respectively; S102, collecting rule knowledge text and preprocessing it, and performing sentence processing on the preprocessed text to divide it into a number of text paragraphs with overlapping texts; S103, extracting entities, attributes, and relationships in the text paragraph using a large language model, and generating text analysis; S104, using a large language model to merge the extracted repeated entities and their attributes, and parsing the merged text to construct a data set; S105: Parse the data set into a CSV file and store it.

[0021] See attached Figure 6 The large language model (LLM) in this application adopts the publicly available Transformer algorithm architecture, which captures long dependencies in text by using a self-attention mechanism in the encoder and decoder.

[0022] This paper is based on a pre-trained large language model with a parameter volume of 72B, and uses domestic GPU servers to train and create a professional large model in the field of power grid control and operation. Model training involves four key steps: data preparation, multiple rounds of professional knowledge training, task-oriented fine-tuning, evaluation verification and deployment: Data preparation. The main tasks include the collection of high-quality training corpora, data cleaning and annotation, and the construction of a domain-specific corpus. First, the corpus collection included more than 100 documents such as industry standards and technical specifications in the field of power grid control and operation, more than 3,000 dispatching professional terms, more than 800 correct historical operation tickets, and more than 2,200 power grid maintenance application forms. Based on the correct historical operation tickets, more than 2,000 incorrect operation tickets were constructed, with a total of more than 10,000 operation instructions. Second, in the data cleaning and annotation part, the denoising and segmentation of unstructured text data such as industry standards and technical specifications were completed, and some semi-structured operation tickets and maintenance application data were annotated for professional training of large models. Third, a domain-specific corpus was constructed. Through the above data preparation and processing, a training corpus was established, and a dispatching professional terminology library was constructed at the same time, which included special terms and technical terminology in the field of power grid control to enhance the model's understanding of professional terminology.

[0023] Multiple rounds of professional knowledge training. By continuously adding proprietary corpus in the field of power grid control operation tickets, the model's ability to understand specific areas of the power grid is further optimized on the basis of the existing pre-trained model, so that the model can better capture the characteristics and operating specifications of the power system.

[0024] Task-oriented fine-tuning and continuous learning mechanism. For the specific task of checking the operation ticket, the model training and fine-tuning are used to accurately analyze the operation instructions. With the help of the set prompt word template, through interactive feedback iteration with the dispatcher, the expert experience is continuously absorbed to improve the model analysis ability. Secondly, a continuous learning mechanism for the model is constructed, and the incremental learning method is adopted to enable the model to continuously absorb new typical operation cases and adapt the model to professional scenarios.

[0025] Evaluation, verification and deployment of model capabilities. By deploying a local GPU server, the fully trained and verified model is deployed to the actual working environment of power grid control operation ticket verification, realizing real-time verification of operation tickets.

[0026] Based on the above steps, a professional large language model can be created that can be applied to the power grid dispatching operation ticket verification business scenario.

[0027] See attached Figure 2 The rule knowledge text is a text representation of the dispatching procedures, specifications, expert experience, five-prevention rules, and general rules related to the operation ticket dispatching. The knowledge graph of this application is constructed based on these rule knowledge texts. The equipment in S101 includes primary equipment and secondary equipment. The primary equipment includes transformers, busbars, lines, switches and auxiliary high-voltage equipment. The secondary equipment refers to relay protection devices, safety automatic devices, and stability control devices, such as automatic backup power supply devices, reclosing switches, etc. The equipment has operating rules, that is, the operating regulations for the equipment in the rule knowledge text. The relationship is the positional relationship between the devices. The position has a sequence, so the positional relationship can indicate the application order of the equipment. The rule knowledge text preprocessing in S102 includes removing repeated and erroneous redundant data. When processing sentences, it is necessary to ensure that there are overlapping parts between the sentences. The purpose is to enable the large model to recognize the contextual relationship between the sentences. As a preferred embodiment of the present application. In S103, a scheduling professional term list library is constructed for the large language model to understand entities, attributes, and relationships in the text paragraph; the scheduling professional term list library includes device name terms, abbreviations of device name terms, synonymous names, and device status terms, abbreviations of device status terms, and description statements of device status. The scheduling professional term list library constructed in this embodiment is shown in Table 1.

[0028] Table 1 Scheduling terminology list The scheduling professional terminology list library lists the terms and abbreviations of entities, as well as their common names or common descriptions in the professional field. The large language model has the ability to understand the scheduling professional terminology list library and can extract entities, attributes, and relationships by calling the scheduling professional terminology list library.

[0029] To further improve the accuracy of information extraction from large language models, see the attached Figure 3 In a specific embodiment of the above implementation, a scheduling operation prompt word template is constructed for the large language model to extract entities, attributes, and relationships in the text paragraph; the prompt word template includes the extraction sources of equipment name terms and equipment status terms, equipment status conversion examples and restriction examples, equipment operation sequence, operation restrictions, and operation conditions. Constructing a list library of scheduling professional terms expands the understanding scope of the large language model, and constructing scheduling operation prompt words guides the large language model on how to use the expanded list library of scheduling professional terms. The scheduling operation prompt words of this embodiment are as follows: You are an expert in the field of power grid scheduling. You need to extract the three elements of entities, relationships, and attributes used to construct a knowledge graph based on the provided documentation in the scheduling field. The entities, relationships, and attributes are: 1. Entity: represents a specific object or system component in the rule. For example: equipment components, equipment status, both of which are extracted according to the "List of Dispatching Professional Terms". 2. Relationship: represents the dynamic interaction or dependency between entities. Prompt examples such as, state transition: represents the transition between different states, such as "from operation to hot standby". Restriction relationship, "cross-state commands are not allowed" 3. Attribute: describes the characteristics or conditions of an entity. Prompt examples such as, operation sequence: such as "close the switch on the power supply side first, and then close the switch on the load side when power is supplied". Operation restriction: such as "cross-state commands are not allowed". Operation conditions: such as "after all sides of the line have switched to cold standby, finally order the line to be switched to maintenance".

[0030] The purpose of S104 is to disambiguate the fusion of rule knowledge. The large language model quickly understands the entities referenced by the nodes by calling the rule knowledge graph. After repeating this process for all nodes, a data set of text parsing can be obtained. In S105, the data set is parsed into independent CSV files corresponding to different nodes and relationships. The CSV file is a format file used for knowledge graph storage. The CSV file is imported into Neo4j for storage, and the rule knowledge graph is established.

[0031] S2. Obtain the power grid operation ticket to be verified, and use the large language model to extract information from it to obtain operation instructions with a logical sequence in the power grid operation ticket, wherein the operation instructions act on the operation object.

[0032] There are sequential logic operation instructions in the power grid operation ticket, which include operation objects, operation actions and / or operation status. The operation objects include primary equipment and secondary equipment. The operation status of primary equipment includes four states: "running", "hot standby", "cold standby" and "maintenance". The operation status of secondary equipment includes "on" and "off". Secondary equipment is generally a control device, a safety automatic device, and a stability control device. The steps for extracting operation ticket information using the large language model are as follows: S201, the operation instruction of the power grid operation ticket is structured as a combination of the operation object, the operation action, and the operation state. For example, an operation instruction is: Station B switches the 500kV AA-BⅠ line 5022 and 5023 switches from operation to hot standby. The above instruction is structured as shown in Table 2.

[0033] Table 2 Structured representation of operation instructions S202: Use a large language model to identify and extract operation instructions based on a combination of an operation object and an operation action or a combination of an operation object and an operation state.

[0034] See attached Figure 5 ,Recognizing and extracting operation instructions means extracting information from the operation instructions by combining "operation object + operation action" or "operation object + operation status". In the process of recognition and extraction, denoising and encoding are required. Denoising means removing unnecessary text. Input encoding needs to be segmented and position encoding is added. Position encoding is used to process the position information in the input sequence. Through position encoding, the large language model can understand the relative order of the operation instructions in the input.

[0035] S203: parse and store the extracted operation instructions. The text parsing is the operation instructions learned by the large language model. The large language model can learn the contextual relationship between different instructions through a multi-layer self-attention mechanism.

[0036] In S2, the large language model calls on the rule knowledge graph to assist in identifying and extracting information.

[0037] S3. Using the operation object as a search field in the power grid maintenance application database, obtain a maintenance application form containing the operation object.

[0038] S4. Use the large language model to extract and parse the text data of the power outage / power transmission measures approval in the maintenance application form. After parsing, the primary equipment state transformation, secondary equipment commissioning and decommissioning logic, and dispatching operation terms contained in the text of the power outage / power transmission measures approval are obtained. During the parsing process, the rule knowledge graph is called to semantically understand the text of the power outage / power transmission measures approval. In S4, the large language model calls the rule knowledge graph to assist in semantic understanding and text parsing.

[0039] It should be understood by those skilled in the art that the maintenance application form contains fixed fields, and the text of the equipment power outage / power transmission measures approval can be obtained from the fixed fields according to different types of maintenance applications. After the text of the equipment power outage / power transmission measures approval is input into the large language model, the model can understand the context and extract information. Finally, the power outage measures text is parsed to obtain the primary equipment state change, the secondary equipment input and output logic, and the dispatching operation terms. For example: the power outage / power transmission measures in the maintenance application form are: A. "Station A uses 5033 switch to power the 500kV A and B line and the 5022 and 5023 switches of Station B"; B. "Station B opens 500kV and opens 5022 and 5023 switches of A and B I line". The power transmission measures corresponding to the operation ticket are: "Station A switches the 500kV A and B line 5033 from hot standby to operation and charging"; "Station B switches the 500kV A and B I line 5022 and 5023 switches from operation to hot standby".

[0040] For the above-mentioned power transmission measures, the large language model divides the power transmission measures into logical segments based on contextual understanding. The logical segments are: Station A uses switch 5033 to transmit power to 500kV line A and B and switches 5022 and 5023 of station B", and station B opens switches 5022 and 5023 of line A and BⅠ at 500kV. The large language model identifies and extracts information from the logical segments according to the principle of "operation object + operation action" or "operation object + operation status". For the logical segment "station B opens switches 5022 and 5023 of line A and BⅠ at 500kV", the large model extracts entities as: station B, line A and BⅠ, switch 5022, switch 5023, and the operation action is: "open". This process requires calling the knowledge rule graph to assist the large model in understanding the context of the dispatching terms (actions) and entities in the power transmission measures, capturing contextual information, and ensuring that the specific meaning of the action of "open" is understood as "operation to hot standby".

[0041] In steps S2 and S4, the large language model needs to call the rule knowledge graph to assist in understanding the semantics.

[0042] See attached Figure 4 , the process of calling the rule knowledge graph is as follows: SA1. The large language model locates the initial entities it contains according to the query question and sets the maximum number of iterations; SA2, obtain the set of initial entities, and use the initial entity as the starting point of the search path; SA3. Iterate the relationship search and entity search when the maximum number of iterations has not been reached. After each iteration, ask the large language model whether the current retrieval path satisfies the query. If so, use the retrieval path as the call result of the large language model. If not, prune the entity and continue to iterate until the maximum number of iterations is reached.

[0043] In step SA3, the relation search groups all relations of the current entity into a candidate relation set, and the large language model calculates the contribution of each relation in the candidate relation set based on the query question, retaining the top N relations and prune other relations to complete the current relation search iteration; among them, N is the number of initial entities; In step SA3, the entity search will be based on the current N The corresponding candidate entity set is obtained by the large language model, and the contribution of each entity in the candidate entity set is completed, and the previous N entities and prune other entities to complete the current entity search iteration, and update the corresponding N A search path.

[0044] Suppose the query question is q , the large language model is based on q The initial entity located is N The initial entity set is express, N The initial entity is also N The starting point of a search path. The search path is the combination of relations and entities. The collection of search paths is When locating the initial entity, the prompt word method can be used for identification.

[0045] For i iterations, the previous ( i- 1) Iteration results, each search path consists of ( i- 1) It consists of three tuples, which are head entity, relation, and tail entity. The relation set of all search paths is Indicates that the tail entity set is represented by express.

[0046] For relational search, E i-1 arrive R i The search process, in i In the first iteration, link each search path to the tail entity The candidate relationship set middle, Represents the tail entity A collection of all related relationships. Get the search path Then, we use the large language model based on the question q and tail relation set Select a new front N Search Path P , and obtain the new tail entity set accordinglyR i .

[0047] For entity search, it is from R i arrive E i The search process, in i In the iteration, the N Search Path P and the new tail entity set R i , get the candidate relationship set , so as to further obtain the search path P The tail entity set, and then use the large language model based on the question q and the candidate relation set , before reselecting N Search Path P The search path here starts with the tail entity set E i Finish.

[0048] After searching for relationships and entities, a new frontend is rebuilt. N The length of the search path for an entity increases by one level, that is, a triple.

[0049] Through the above method, the hallucination problem of large language models can be effectively improved.

[0050] S5. Use the large language model to match the primary equipment state change, secondary equipment launch and retreat logic, and dispatch operation terms analyzed in step S4 with the operation instructions in the operation ticket. If the match is successful, the operation instruction verification is passed; otherwise, the operation instruction verification fails. The matching of operation instructions includes the matching of primary equipment (main transformer, busbar, line, switch, knife switch, etc.) state change (operation, hot standby, cold standby, maintenance) and the matching of secondary equipment (relay protection, safety automatic device, etc.) launch and retreat requirements. The specific matching rules are as follows: Operation object matching: Match the entities in the power transmission measures identified by the large model with the entities in the identified operation ticket one by one. If they are consistent, the operation object matching is successful.

[0051] Operation action matching: The action instructions recognized by the large model, such as "pull out the XX switch", are semantically converted into the scheduling operation term "run to hot standby" with the help of the rule knowledge graph, and matched with the standard operation actions in the operation ticket. If they are consistent, the operation action matching is successful.

[0052] Operation status matching: The action instruction recognized by the large model, such as "turn on switch 5022 reclosing", is checked for consistency with the device status in the operation ticket. If they are consistent, the operation status is successfully matched. The operation status check can be divided into the status check of primary equipment and secondary equipment.

[0053] Anything not described in this application can be achieved by adopting or drawing on existing technologies. The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for verifying power grid dispatching operation tickets with intelligent matching of power outage measures, characterized in that: The following steps are involved: S1. Construct a rule knowledge graph based on the rule knowledge text, which specifically includes the following steps: S101, using devices, operation rules of devices, and location relationships between devices as entities, attributes, and relationships of a rule knowledge graph, respectively; S102, collecting rule knowledge text and preprocessing it, and performing sentence processing on the preprocessed text to divide it into a number of text paragraphs with overlapping texts; S103, extracting entities, attributes, and relationships in the text paragraph using a large language model, and generating text analysis; S104, using a large language model to merge the extracted repeated entities and their attributes, and parsing the merged text to construct a data set; S105, parsing the data set into a CSV file and storing it; S2. Obtain the power grid operation ticket to be verified, and use the large language model to extract information from it to obtain operation instructions with a logical order in the power grid operation ticket, wherein the operation instructions act on the operation object; S3, using the large language model to use the operation object as a search field in the power grid maintenance application database to obtain the maintenance application form containing the operation object; S4. Use a large language model to extract and parse the text data of the power outage / power transmission measures approval in the maintenance application form. After parsing, the primary equipment state change, secondary equipment startup and shutdown logic, and dispatching operation terms contained in the text of the power outage / power transmission measures approval are obtained. During the parsing process, the rule knowledge graph is called to semantically understand the text of the power outage / power transmission measures approval; S5. Use the large language model to match the primary equipment state change, secondary equipment startup and shutdown logic, and dispatch operation terms analyzed in step S4 with the operation instructions in the operation ticket in sequence. If the power outage / power supply measures and the operation instructions can be successfully matched in sequence, the operation instruction verification passes; otherwise, the operation instruction verification fails.

2. A method for checking power grid dispatching operation tickets for intelligently matching power outage measures according to claim 1, characterized in that: In step S103, a scheduling professional term list library is constructed for the large language model to understand entities, attributes, and relationships in text paragraphs; the scheduling professional term list library includes device name terms, abbreviations of device name terms, synonymous names, and device status terms, abbreviations of device status terms, and description statements of device status.

3. A method for checking power grid dispatching operation tickets for intelligently matching power outage measures according to claim 2, characterized in that: In step S103, a scheduling operation prompt word template is constructed for the large language model to extract entities, attributes, and relationships in the text paragraph; the prompt word template includes the extraction sources of device name terms and device status terms, device status conversion examples and restriction examples, device operation sequence, operation restrictions, and operation conditions.

4. A method for checking power grid dispatching operation tickets for intelligently matching power outage measures according to claim 1, characterized in that: In step S2, the large language model extracts information from the power grid operation ticket to be verified and calls the rule knowledge graph.

5. The method for verifying power grid dispatching operation tickets for intelligently matching power outage measures according to claim 1, characterized in that: In step S2, the large language model is used to extract information from the power grid operation ticket to be verified, including the following steps: S201, structurally representing the operation instruction of the power grid operation ticket as a combination of an operation object, an operation action, and an operation state; S202, using a large language model to identify and extract operation instructions based on a combination of an operation object and an operation action or a combination of an operation object and an operation state; S203: parse and store the extracted operation instructions.

6. A method for checking power grid dispatching operation tickets for intelligently matching power outage measures according to any one of claims 4 or 5, characterized in that: The process of calling the rule knowledge graph is as follows: SA1. The large language model locates the initial entities it contains according to the query question and sets the maximum number of iterations; SA2, obtain the set of initial entities, take the initial entity as the starting point of the search path, and thus obtain the set consisting of all initial entities; SA3. Iterate the relationship search and entity search when the maximum number of iterations has not been reached. After each iteration, ask the large language model whether the current retrieval path satisfies the query. If so, use the retrieval path as the call result of the large language model. If not, prune the entity and continue to iterate until the maximum number of iterations is reached.

7. A method for checking power grid dispatching operation tickets for intelligently matching power outage measures according to claim 6, characterized in that: In step SA3, the relation search groups all relations of the current entity into a candidate relation set, and the large language model calculates the contribution of each relation in the candidate relation set based on the query question, retaining the top N relations and prune other relations to complete the current relation search iteration; among them, N is the number of initial entities; In step SA3, the entity search will be based on the current N The corresponding candidate entity set is obtained by the large language model, and the contribution of each entity in the candidate entity set is completed, and the previous N entities and prune other entities to complete the current entity search iteration, and update the corresponding N A search path.

Images