Power system risk management and control method, device and system and storage medium

Abstract

The invention discloses an electric power system risk management and control method, device and system and a storage medium. The method comprises the following steps: receiving inquiry data which is input by a user and is related to electric power system risk management and control; inputting the inquiry data into a language model for semantic analysis to obtain a semantic vector, and inputting the semantic vector into a classification model to determine a target intention category of the user; extracting scene keywords related to a power grid system from the target intention category, mapping the scene keywords into a preset professional knowledge base to obtain retrieval fields, performing retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management and control, and performing risk management and control on the risk knowledge fragments. Filling the risk knowledge fragment into a preset answer template to generate a risk management and control scheme; and managing and controlling the risk prevention and control process of the power system based on the risk management and control scheme. The accuracy of risk management and control of the power system can be improved based on the professional knowledge base.

Description

Technical Field

[0001] This invention relates to the field of power system risk management, and in particular to a power system risk management method, apparatus, system and storage medium. Background Technology

[0002] With the rapid development of science and technology and information technology, the demand for intelligent and automated solutions in various industries continues to rise, with the field of smart grid management in the power system being particularly prominent. This field, due to its complex system structure, diverse equipment types, and direct impact on the safe and stable operation of the power grid and personal safety, places extremely high demands on the accuracy, timeliness, and reliability of risk control.

[0003] To address the aforementioned pain points, power system risk management systems based on professional knowledge bases are gradually being applied in the power sector. These systems integrate professional knowledge, historical data, and real-time information to provide scientific management support for power grid management, improving management efficiency to some extent. However, existing systems still have several shortcomings and cannot fully meet actual needs: First, the knowledge base lacks comprehensiveness, failing to fully cover national, industry, and enterprise-level regulations, standards, and safety procedures, and its low degree of knowledge structuring hinders the accuracy of management; second, its intelligent correlation capabilities are insufficient, resulting in low accuracy in matching risk points with risk management solutions, easily leading to omissions or mismatches, and consequently, low management accuracy; third, its intent recognition and question-answering capabilities are limited, making it difficult to accurately understand the core demands of users' natural language needs and quickly locate risk points and corresponding safety measures. Summary of the Invention

[0004] This invention provides a method, apparatus, system, and storage medium for power system risk management, which can improve the accuracy of power system risk management based on a professional knowledge base.

[0005] In a first aspect, an embodiment of the present invention provides a power system risk management method, comprising:

[0006] Receive user-input query data related to power system risk management;

[0007] The query data is input into a language model for semantic analysis to obtain a semantic vector. The semantic vector is then input into a classification model to determine the user's target intent category.

[0008] The system extracts scenario keywords related to the power grid system from the target intent category, maps the scenario keywords to a preset professional knowledge base to obtain retrieval fields, and performs a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management. The risk knowledge fragments are then filled into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and the initial professional knowledge base is updated based on the analysis results.

[0009] The risk control scheme is used to manage the risk prevention and control process of the power system.

[0010] In this way, by receiving user inquiries directly related to power system risk management, a precise demand data source is provided for subsequent solution formulation, ensuring the accuracy of subsequent management. Inputting the inquiry data into a language model for semantic analysis to generate semantic vectors allows for precise parsing of power-related semantics, avoiding deviations in demand understanding that could lead to off-target knowledge base retrieval, thus improving management accuracy. Determining the user's target intent category through a classification model avoids misjudgments of demand due to general semantic misunderstandings, accurately locating the user's true needs from the demand analysis level, ensuring the accuracy of risk management. Extracting keywords related to scenario characteristics from intent categories, rather than generalized keywords, provides precise positioning basis for subsequent retrieval, avoiding irrelevant risk knowledge retrieval due to keyword generalization, thus improving the accuracy of risk management. Mapping scenario keywords with structured knowledge from the professional knowledge base to generate retrieval fields ensures the direction of the retrieval. Precision is ensured to avoid irrelevant or redundant knowledge fragments retrieved due to vague keywords. From the knowledge retrieval entry point, the retrieved risk knowledge is highly matched to user needs, guaranteeing the accuracy of subsequent risk management. Retrieving knowledge fragments from a professional knowledge base "updated after user feedback analysis" based on search fields ensures that the knowledge fragments are authoritative, timely, and adapted to current risk management needs. Simultaneously, filling these knowledge fragments into preset answer templates guarantees a standardized and complete risk management solution structure, reducing errors caused by information omissions or logical inconsistencies, and ensuring a more accurate output risk management solution. The generated risk management solution is used to manage the power system risk prevention and control process, aligning with actual user management needs and complying with power system risk regulations, ensuring that the risk prevention and control process is strictly implemented according to the professional standards in the knowledge base. Ultimately, this achieves improved accuracy in power system risk management based on the professional knowledge base.

[0011] Furthermore, the step of inputting the semantic vector into a classification model to determine the user's target intent category specifically includes:

[0012] The semantic vector is input into the classification model to calculate probability values, wherein the probability values ​​include risk query probability values, prevention and control strategy consultation probability values, and strategy legal basis query probability values.

[0013] The initial intent category is determined by comparing the probability values ​​of risk queries, consultations on prevention and control strategies, and queries on the legal basis of the strategies.

[0014] If the probability value corresponding to the initial intent category is greater than or equal to a preset threshold, then the user's target intent category is determined based on the initial intent category.

[0015] By using a classification model to determine the user's target intent category, we can avoid misjudgment of needs due to biases in general semantic understanding, accurately locate the user's true needs from the perspective of needs analysis, and ensure the accuracy of risk management.

[0016] Furthermore, the construction of the initial professional knowledge base specifically includes:

[0017] The first learning data corresponding to each release time point is collected in real time. The first learning data is prioritized to obtain several second learning data. The first learning data comes from different data sources related to power system risk management.

[0018] Determine whether there is a content conflict among the second learning data. If there is a content conflict, determine whether the conflicting content is at the same priority. If the conflicting content is not at the same priority, retain the conflicting content with the highest priority as the third learning data. If the conflicting content is at the same priority, compare the release time points corresponding to the conflicting content and retain the conflicting content with the most recent release time point as the fourth learning data.

[0019] The third learning data, the fourth learning data, and the second learning data that do not have content conflicts are integrated and processed to obtain the initial professional knowledge base.

[0020] By constructing an initial professional knowledge base through multi-source data collection, priority classification, and conflict classification, the distortion of knowledge base information caused by messy data sources and content conflicts can be avoided. This ensures the authority and timeliness of risk management-related data in the initial knowledge base, laying a precise data foundation for subsequent risk knowledge retrieval and management plan generation based on the professional knowledge base, thereby improving the accuracy of power system risk management.

[0021] Furthermore, the process of integrating the third learning data, the fourth learning data, and the second learning data that do not have content conflicts to obtain the initial professional knowledge base specifically includes:

[0022] Entity recognition is performed on the second learning data, the third learning data, and the fourth learning data to obtain several entities related to power system risk management;

[0023] Syntactic dependency structures are obtained by analyzing the second, third, and fourth learning data respectively. Risk relationships are extracted from the syntactic dependency structures using a semantic role labeling method. The risk relationships are the relationships between risk points and prevention and control measures related to the power system.

[0024] A consistency check is performed on the entities and the risk relationships to obtain a first entity risk set, and conflict elimination processing is performed on the first entity risk set to obtain a second entity risk set.

[0025] The second entity risk set is transformed into structured knowledge represented in the form of triples, and the structured knowledge is stored in a preset database to obtain the initial professional knowledge base.

[0026] By processing learning data through entity identification, risk relationship extraction, verification and elimination, and structured storage, the core entities and key risk relationships of power system risk management can be accurately identified, eliminating problems such as contradictions in entity definitions and logical conflicts in risk relationships. At the same time, the structured format ensures the standardization of knowledge storage, ensuring that matching risk knowledge can be quickly located when searching based on the professional knowledge base. This reduces deviations in management plans caused by fragmented knowledge and logical confusion, thereby improving the accuracy of power system risk management.

[0027] Furthermore, the step of retrieving risk knowledge fragments related to risk management from the professional knowledge base based on the search fields specifically includes:

[0028] A multi-database joint search is performed on the search field in the professional knowledge base to obtain several first search results;

[0029] Each of the first search results is filtered to obtain several second search results;

[0030] By integrating the second search results through a preset logical structure, risk knowledge fragments related to risk management are obtained.

[0031] By retrieving knowledge fragments from a professional knowledge base that is updated based on user feedback analysis using search fields, it is ensured that the knowledge fragments are authoritative, real-time, and adapted to current risk management needs, thus ensuring the accuracy of the risk management solutions generated subsequently.

[0032] Furthermore, the step of filling the risk knowledge fragments into a preset answer template to generate a risk management plan specifically includes:

[0033] The risk knowledge fragments are broken down and classified to obtain classified knowledge units;

[0034] The knowledge unit is subjected to template matching processing to obtain a target template that matches the scenario of the knowledge unit, and the knowledge unit is integrated and filled into the target template to obtain a preliminary solution.

[0035] The risk points and prevention and control steps of the preliminary plan are verified to obtain a risk management plan.

[0036] By filling knowledge fragments into a pre-set answer template to generate answers, the risk management plan can be structurally standardized and complete, reducing errors caused by missing information or logical confusion, and ensuring that the output risk management plan is more accurate.

[0037] Furthermore, the professional knowledge base is obtained by analyzing and processing user feedback data on historical risk management solutions, and updating the initial professional knowledge base based on the analysis results. Specifically, this includes:

[0038] Collect several initial feedback data points from users regarding the historical risk management plan;

[0039] The first feedback data is structured and transformed to obtain several second feedback data. The second feedback data is then classified into question categories to obtain several third feedback data with category labels. The third feedback data includes knowledge base content and knowledge-related content.

[0040] Knowledge extraction is performed on the knowledge base content to obtain supplementary and correction information, and association calculation is performed on the knowledge-related content to obtain association deviation parameters;

[0041] The initial professional knowledge base is supplemented according to the supplementary information to obtain a first professional knowledge base. The first professional knowledge base is then corrected according to the correction information to obtain a second professional knowledge base. The knowledge association strength of the second professional knowledge base is adjusted using the association deviation parameter to obtain an updated professional knowledge base.

[0042] This process of optimizing the professional knowledge base through feedback data structuring, classification analysis, and targeted updates allows for targeted supplementation of missing content, correction of errors, and optimization of knowledge association logic. It avoids deviations in risk knowledge retrieval caused by outdated or loosely connected knowledge base content, ensuring that the professional knowledge base always remains accurate and adaptable. This provides reliable knowledge support for the subsequent generation of risk management solutions tailored to actual needs, ultimately improving the accuracy of power system risk management based on the professional knowledge base.

[0043] Secondly, an embodiment of the present invention provides a power system risk management and control device, including a first module, a second module, a third module and a fourth module;

[0044] The first module is used to receive user-input query data related to power system risk management;

[0045] The second module is used to input the query data into a language model for semantic analysis to obtain a semantic vector, and then input the semantic vector into a classification model to determine the user's target intent category.

[0046] The third module is used to extract scenario keywords related to the power grid system from the target intent category, map the scenario keywords to a preset professional knowledge base to obtain retrieval fields, perform a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management, and fill the risk knowledge fragments into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and updating the initial professional knowledge base based on the analysis results.

[0047] The fourth module is used to manage the risk prevention and control process of the power system based on the risk management scheme.

[0048] The first module receives user inquiries directly related to power system risk management, providing a precise data source of demand for subsequent solution development and ensuring the accuracy of subsequent management. The second module inputs the inquiry data into a language model for semantic analysis to generate semantic vectors, accurately parsing the semantics of the power field and avoiding deviations in demand understanding that could lead to off-target knowledge base retrieval, thus improving the accuracy of management. A classification model determines the user's target intent category, avoiding misjudgments of demand due to general semantic misunderstandings, accurately locating the user's true needs from the demand analysis level, and ensuring the accuracy of risk management. The third module extracts keywords related to scenario features from the intent category, rather than generalized keywords, providing precise positioning basis for subsequent retrieval and avoiding irrelevant risk knowledge retrieved due to keyword generalization, thus improving the accuracy of risk management. The mapping of scenario keywords to structured knowledge in the professional knowledge base generates retrieval fields, ensuring the accuracy of retrieval. The search direction is precise, avoiding irrelevant or redundant knowledge fragments retrieved due to vague keywords. This ensures that the acquired risk knowledge highly matches user needs from the knowledge retrieval entry point, guaranteeing the accuracy of subsequent risk management. Retrieving knowledge fragments from a professional knowledge base "updated after user feedback analysis" based on search fields ensures that the knowledge fragments are authoritative, timely, and adapted to current risk management needs. Simultaneously, filling these knowledge fragments into preset answer templates guarantees a standardized and complete risk management solution structure, reducing errors caused by information omissions or logical inconsistencies, and ensuring a more accurate output risk management solution. The fourth module manages the power system risk prevention and control process based on the generated risk management solution, aligning with both actual user management needs and power system risk standards, ensuring that the risk prevention and control process is strictly implemented according to the professional standards in the knowledge base. Ultimately, this achieves improved accuracy in power system risk management based on the professional knowledge base.

[0049] Thirdly, another embodiment of the present invention provides a terminal device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0050] The memory is used to store at least one executable instruction that causes the processor to perform an operation of a power system risk management method.

[0051] Fourthly, a computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device or apparatus containing the computer-readable storage medium to perform a power system risk management method. Attached Figure Description

[0052] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating one embodiment of a power system risk management method provided in this application;

[0054] Figure 2 This is a flowchart illustrating steps S201 to S203 provided in this application;

[0055] Figure 3 This is a structural schematic diagram of a power system risk management device provided in this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0058] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0059] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0060] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0061] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0062] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0063] The field of smart grid management in power systems demands extremely high precision, timeliness, and reliability in risk control due to the complexity of the systems, the diversity of equipment, and the implications for grid and personal safety. While existing risk control systems based on professional knowledge bases are being gradually implemented, significant shortcomings remain: First, the knowledge base is incomplete, failing to fully cover multi-level regulations, standards, and safety procedures, and exhibiting low levels of knowledge structuring, thus hindering accurate control. Second, intelligent correlation capabilities are weak, resulting in low accuracy in matching risk points with control solutions, easily leading to omissions or mismatches. Third, intent recognition and question-answering capabilities are limited, making it difficult to accurately understand users' core natural language requests and quickly locate risk points and corresponding safety measures.

[0064] See Figure 1 To improve the accuracy of power system risk management based on a professional knowledge base, an embodiment of the present invention provides a power system risk management method, including steps S101 to S104.

[0065] Step S101: Receive query data related to power system risk management input by the user;

[0066] In some embodiments, receiving user-inputted query data related to power system risk management specifically includes: designing an intuitive and easy-to-use user interface that combines the user's (e.g., power maintenance personnel's) work scenario (e.g., voice interaction during on-site two-handed operation, text input in an office scenario), the interface including a voice input entry and a text input entry; receiving user-inputted query data related to power system risk management from the input entry; if the query data is in voice form, converting the voice input into text using an automatic speech recognition engine (e.g., Baidu speech recognition API), and then uniformly converting the text into UTF-8 encoding; if the query data is in text form, directly performing UTF-8 encoding conversion, and segmenting, removing stop words, and standardizing terminology (e.g., unifying "high voltage line" as "high voltage line") after conversion.

[0067] By receiving user inquiries directly related to power system risk management, this provides a precise data source of demand for subsequent solution development, ensuring the accuracy of subsequent management.

[0068] Step S102: Input the query data into a language model for semantic analysis to obtain a semantic vector, and input the semantic vector into a classification model to determine the user's target intent category;

[0069] In some embodiments, the query data is input into a language model for semantic analysis to obtain a semantic vector. Specifically, this includes: first, segmenting the query data into words to obtain segmentation results; then, tagging each segmentation result with part-of-speech tags; subsequently, using a named entity recognition tool to identify key entities in the text; then, inputting the text with completed part-of-speech tagging and named entity recognition into a dependency parsing tool (such as StanfordParser) to analyze the dependency relationships between words in the text, thereby extracting the subject-verb-object core structure of the text; then, according to the input format requirements of a language model (such as a knowledge-enhanced ERNIE pre-trained model), integrating the segmentation results, part-of-speech tagging information, and subject-verb-object core structure into a text form that the model can recognize; then, inputting the integrated text into the pre-trained model, which performs sentence semantic representation of the text, transforming the semantic information of the text into a fixed-dimensional vector, and finally outputting a semantic vector.

[0070] In some embodiments, inputting the semantic vector into a classification model to determine the user's target intent category specifically includes: inputting the semantic vector into the classification model to calculate probability values, wherein the probability values ​​include risk query probability values, prevention and control strategy consultation probability values, and strategy legal basis query probability values; comparing the risk query probability values, the prevention and control strategy consultation probability values, and the strategy legal basis query probability values ​​to determine an initial intent category; if the probability value corresponding to the initial intent category is greater than or equal to a preset threshold, then determining the user's target intent category based on the initial intent category.

[0071] Specifically, the semantic vector is input into a classification model (such as the TextCNN classifier) ​​to calculate the probability value of the semantic vector belonging to a preset intent category. The sum of all probability values ​​is 1. The preset intent categories include risk query category, prevention and control strategy consultation category, and strategy legal basis query category. The risk query category includes "query device / scenario risk points" and "assess risk point level" categories. The prevention and control strategy consultation category includes "obtain risk point prevention and control plan" category, "query historical risk cases" category, and "consult emergency procedures for risk handling" category. The strategy legal basis query category includes "query the legal basis of prevention and control plan" category and "feedback knowledge base / system problem" category. In addition, the preset intent category also includes "consult system operation guidelines" category. The probability values ​​of all intent categories are compared, and the category with the highest probability value is taken as the initial intent category. Then, this probability value is compared with a preset threshold (such as 0.7). If the probability value is greater than or equal to the preset threshold, the initial intent category is determined as the user's target intent category.

[0072] For further explanation, please refer to Table 1. Table 1 is a table of preset intent categories, including the categories of "Querying Device / Scenario Risk Points", "Assessing Risk Point Levels", "Obtaining Risk Point Prevention and Control Plans", "Querying Historical Risk Cases", "Consulting on Emergency Risk Handling Procedures", "Querying the Legal Basis of Prevention and Control Plans", "Reporting Knowledge Base / System Issues", and "Consulting on System Operation Guidelines".

[0073] Table 1 Intent Category Table

[0074]

[0075]

[0076] By using a classification model to determine the user's target intent category, we can avoid misjudgment of needs due to biases in general semantic understanding, accurately locate the user's true needs from the perspective of needs analysis, and ensure the accuracy of risk management.

[0077] In some embodiments, determining the user's target intent category based on the initial intent category if the probability value corresponding to the initial intent category is greater than or equal to a preset threshold further includes: if the probability value corresponding to the initial intent category is less than a preset threshold, triggering a preset three-level progressive processing mechanism to process the query data until a preset condition is met to obtain the target intent category.

[0078] Specifically, if the probability value of the initial intent category is less than a preset threshold (e.g., 0.7), the first-level processing—intelligent guidance for users to supplement information—is triggered. This involves generating a "scenario-based guidance script" and providing prompts via pop-ups or voice prompts on the interactive interface to guide the user to supplement key information based on the query data, narrowing the scope of needs and helping the system to perform secondary intent matching. Please refer to Table 2, which is the guidance script table. The guidance script follows the principle of "customization for power scenarios" to avoid generic questions.

[0079] Table 2 Guiding Script

[0080]

[0081] If, after the user provides additional information, the probability value calculated by the classification model (such as the TextCNN classifier) ​​for a certain intent category is greater than or equal to a preset threshold, then that intent category is determined as the target intent category; if the probability value is still less than the preset threshold after providing additional information, then the process proceeds to the second stage – human expert consultation.

[0082] In the second-level processing, the system interface displays a "Manual Consultation Portal," offering two connection methods to ensure that the urgent needs of operations and maintenance personnel are responded to: First, real-time online consultation, where the interface displays a "Security Expert Online" button. Clicking this redirects to the company's internal "Security Management Expert Communication Group," which supports real-time communication via text, images, and voice. Expert response time is ≤15 minutes (working hours 8:00-22:00). Second, message feedback processing. Outside of working hours or when experts are busy, the system provides a "Message Feedback" form, requiring users to fill in "problem description, contact information, and urgency level (general / urgent)." After submission, it is automatically assigned to the corresponding expert. Urgent issues (such as handling sudden risks) are responded to within 2 hours, and general issues within 24 hours. Simultaneously, the system automatically records "unmatched issues + user supplementary information + manual answers," storing them in the "Unmatched Intent Corpus," providing data support for the third-level processing—corpus accumulation and model iteration.

[0083] In the third-level processing, the "unmatched intent corpus" will be processed periodically (e.g., once a week). The intent recognition steps will be optimized based on expert review results. The specific process is as follows: First, the corpus is labeled. Power safety experts are invited to label the "unmatched questions" to determine whether they belong to "new expressions of existing intents" (e.g., "preventing falling" belongs to "obtaining risk point prevention and control plans") or "requires adding a new intent category" (e.g., "querying risk prevention and control training materials"). If it is a "new expression of existing intents," the question and its corresponding intent label are added to the training set, and the language model is fine-tuned (e.g., knowledge augmentation). The system uses a strong ERNIE pre-trained model and a classification model (such as the TextCNN classifier) ​​to improve the ability to recognize niche expressions. If the labeling determines that "a new intent category needs to be added" (such as a power company adding a "risk prevention and control training management" requirement), a corresponding category (such as "query risk prevention and control training materials") is added to the preset intent categories, and the training set of the classification model and the interface guidance script are updated simultaneously. The fine-tuned model needs to be verified through a "test set" (selecting 100 unmatched historical questions) to ensure that the intent recognition accuracy is improved by ≥5% before it is deployed to the production environment.

[0084] This three-tiered processing mechanism addresses users' immediate needs while continuously optimizing intent recognition capabilities through data accumulation, creating a closed loop of "problem-processing-optimization" and avoiding long-term blind spots of "inability to match intent."

[0085] Step S103: Extract scenario keywords related to the power grid system from the target intent category, map the scenario keywords to a preset professional knowledge base to obtain retrieval fields, perform a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management, and fill the risk knowledge fragments into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and updating the initial professional knowledge base based on the analysis results.

[0086] Please refer to Figure 2 In some embodiments, the construction of the initial professional knowledge base includes steps S201 to S203:

[0087] Step S201: Collect the first learning data corresponding to each release time point in real time, prioritize the first learning data, and obtain several second learning data. The first learning data comes from different data sources related to power system risk management.

[0088] In some embodiments, step S201, which involves real-time collection of first learning data corresponding to each publication time point, prioritizing the first learning data to obtain several second learning data, specifically includes: firstly determining the scope of first learning data related to power system risk management; then, synchronously recording the publication time point corresponding to each data point during real-time collection of each data source through a combination of API interface calls and web crawling, forming first learning data containing data content, source identifier, and publication time; and then prioritizing the first learning data according to the priority rule of "authority first, adaptability second, and timeliness supplement" to obtain several second learning data.

[0089] Specifically, the primary learning data includes customized internal enterprise data (such as the "Operation and Maintenance Safety Management Rules," personal risk ledgers, and historical accident handling reports from the power company's OA system), national / industry professional databases (such as the State Grid Corporation's safety regulations database, the IEEE power system safety standards library, and the "Power Safety Production Operation Procedures" issued by the National Energy Administration), authoritative industry reports (such as the "Annual Analysis Report on Power Personal Safety Risks" issued by the China Electricity Council and the "White Paper on Smart Grid Operation and Maintenance Risk Prevention and Control" from the State Grid Energy Research Institute), and publicly available internet data (such as policy interpretations on the National Energy Administration's official website and technical articles from vertical media in the power industry). Then, the internal enterprise data and national / industry professional database data are accessed through official API interfaces, and publicly available internet data is crawled using compliant web scraping tools to obtain authoritative industry reports. During the data collection process, the release time of each data point is recorded synchronously to form the first learning data, which includes data content, source identifier, and release time. Then, the first learning data is prioritized according to the priority rule of "authority first, adaptability second, and timeliness supplement". Customized data within the enterprise directly matches the actual operation and maintenance scenario of the enterprise and is classified as the first priority (i.e., the highest priority). National / industry professional database data has official authority and covers general industry security requirements, so it is classified as the second priority. Authoritative industry reports contain industry practice cases and trend analysis, supplementing scenarios not covered by general rules, so they are classified as the third priority. Publicly available Internet data is only used to fill the edge scenarios not covered by the first three levels of data sources and is classified as the fourth priority (i.e., the lowest priority). Finally, several second learning data corresponding to different priorities are obtained.

[0090] It should be noted that this application also establishes a data collection frequency mechanism of "tiered scheduled collection + emergency trigger collection" for the professional knowledge base: For enterprise internal data belonging to the first priority, collection is stipulated to be carried out once a day, with the collection time set during non-peak operation and maintenance hours (such as 2:00-4:00 AM daily), covering newly added / revised procedures, ledgers, and operation and maintenance logs of the OA system on the same day, ensuring that the enterprise internal data is synchronized with actual business; For national / industry professional databases belonging to the second priority, collection is stipulated to be carried out once a week, with the collection time being Sunday morning (such as 3:00-6:00 AM Sunday), synchronizing the updated content of the State Grid Corporation's safety procedure database and the IEEE standard library (such as the State Grid Corporation's safety procedure database and the IEEE standard library). For data revisions and standard supplements, which are typically updated monthly or quarterly, weekly collection avoids missing urgent updates. For authoritative industry reports that fall under the third priority category, collection is required once per quarter, at the early morning of the last Sunday of each quarter, synchronizing with the latest reports released by institutions such as CEC and State Grid Energy Research Institute. These reports are mostly released quarterly or semi-annually, so high-frequency collection is unnecessary. For publicly available internet data that falls under the fourth priority category, collection is required once per week, staggered from the second priority collection time (3:00-5:00 AM every Wednesday). After collection, "keyword filtering + redundancy removal" (such as filtering content unrelated to "personal risk prevention") is used to retain only valid data, reducing storage pressure. In addition, when an emergency update is detected in a high-priority data source (such as the National Energy Administration issuing an "Emergency Notice on Electricity Personal Safety" or an enterprise revising its safety regulations due to an accident), the system automatically triggers "real-time data collection": the update is triggered through the "update notification callback" of the data source interface (such as a Webhook notification pushed when the State Grid Corporation of China updates its database) or by manual marking (the enterprise administrator marks "emergency update" in the system background); data collection is started within 10 minutes after triggering, conflict determination and data entry are completed first, and an "emergency data update reminder" is sent to the system administrator simultaneously to ensure that critical safety data can be accessed within 1 hour. The conflict determination here is described in step S202.

[0091] It should be noted that the professional knowledge base in this application also has a dynamic update mechanism, which can automatically update the content of the knowledge base according to new laws and regulations, enterprise safety regulations, or operational procedures and system documents. This mechanism regularly scans sources such as the official website of the National Standardization Administration and enterprise OA systems, uses GitHub Webhook to monitor document changes, identifies newly added or modified content through a Diff algorithm, and automatically triggers the knowledge extraction and integration process to ensure the real-time nature of the knowledge base. For example, when the state issues new power safety regulations, the intelligent association module can promptly identify newly added or changed risk points and prevention and control plans, and integrate this information into the existing knowledge base. The specific steps of the dynamic update mechanism include: (1) Document monitoring: Regularly scan the preset regulatory and standard release platforms (such as the official website of the National Standardization Committee and the enterprise OA system); (2) Change detection: Use GitHub Webhook or Diff algorithm to detect changes in document versions and identify new / modified content; (3) Automatic extraction: Trigger entity and relation extraction for new content; (4) Knowledge fusion: Integrate new knowledge into the existing knowledge base and perform conflict detection and resolution; (5) Version management: Record the history of knowledge base version changes and support rollback and auditing.

[0092] Step S202: Determine whether there is a content conflict among the second learning data. If there is a content conflict, determine whether the conflicting content is at the same priority. If the conflicting content is not at the same priority, retain the conflicting content with the highest priority as the third learning data. If the conflicting content is at the same priority, compare the release time points corresponding to the conflicting content and retain the conflicting content with the closest release time point as the fourth learning data. Specifically, this includes:

[0093] In some embodiments, determining whether there is content conflict among the second learning data, and if there is content conflict, determining whether the conflicting content is of the same priority; if the conflicting content is not of the same priority, retaining the conflicting content with the highest priority as the third learning data; if the conflicting content is of the same priority, comparing the publication time points corresponding to the conflicting content, and retaining the conflicting content with the most recent publication time point as the fourth learning data, specifically includes: firstly, structurally decomposing each second learning data after priority division according to three perspectives: "risk point type", "prevention and control measures", and "legal basis", and extracting the core information of each second learning data. Then, by calculating text similarity and matching key information, it is detected whether there is any inconsistency in core information between different second learning data. If there is an inconsistency, it is determined to be a content conflict (for example, in data A, "the safe distance for high-voltage line maintenance is 3 meters", while in data B, the safe distance for the same risk point is stated as "2.5 meters", so data A and B are determined to have a content conflict). If the conflicting content is of different priorities, the conflicting content with the highest priority is retained as the third learning data (for example, when corporate rules and internet articles conflict, the corporate rules with the highest priority shall prevail). If the conflicting content is of the same priority, the publication time of the conflicting content is compared, and the conflicting content with the most recent publication time is retained as the fourth learning data.

[0094] It should be noted that for some special conflict scenarios (such as key data conflicts involving personal safety red lines, such as "withstanding voltage standards for live-line working insulating tools"), an "expert review process" will be automatically triggered. The power company's safety management experts will make a ruling within 24 hours, and the ruling result will serve as the final data basis and be recorded in the "conflict handling log".

[0095] Step S203: Integrate the third learning data, the fourth learning data, and the second learning data that do not have content conflicts to obtain the initial professional knowledge base;

[0096] In some embodiments, the process of integrating the third learning data, the fourth learning data, and the second learning data without content conflicts to obtain the initial professional knowledge base specifically includes: performing entity recognition on the second learning data, the third learning data, and the fourth learning data to obtain several entities related to power system risk management; analyzing the grammatical dependency structures from the second learning data, the third learning data, and the fourth learning data respectively, and extracting risk relationships from the grammatical dependency structures using a semantic role labeling method, wherein the risk relationships are the relationships between risk points and prevention and control measures related to the power system; performing consistency verification on the entities and the risk relationships to obtain a first entity risk set, and performing conflict removal processing on the first entity risk set to obtain a second entity risk set; converting the second entity risk set into structured knowledge represented in the form of triples, and storing the structured knowledge in a preset database to obtain the initial professional knowledge base.

[0097] Specifically, an entity recognition model (such as the BERT-BiLSTM-CRF model, where BERT is a Transformer-based bidirectional encoder representation model, BiLSTM is a bidirectional long short-term memory network, and CRF is a conditional random field) is used to perform entity recognition on the second, third, and fourth learning data that do not have content conflicts. The pre-training corpus of the entity recognition model contains power industry-specific entity types, including risk point entities (such as "electric shock during high-voltage line maintenance" and "excessive toxic gas in cable trenches"), prevention and control measure entities (such as "wearing insulating gloves" and "grounding resistance testing"), legal basis entities (such as Article 3.2.1 of the "Power Safety Work Regulations" and the IEEE 1584 standard), equipment entities (such as "tower climbing tools" and "GIS equipment"), and scene entities (such as "outdoor work in rainy weather" and "indoor maintenance in substations"). After recognition, a unique entity ID is assigned to each entity (such as risk point entity ID = R001 and prevention and control measure entity ID = M003), forming a preliminary entity list.

[0098] Then, dependency parsing tools were used on the three types of data to extract the grammatical dependency structures of the text, such as subject-verb-object, attributive, adverbial, and complement (e.g., extracting the core structure "high-voltage line maintenance (subject) - need (verb) - wear insulated gloves (object)" from "high-voltage line maintenance requires wearing insulated gloves"); then, through semantic role labeling, the correspondence between "risk point - prevention and control measures" was located from the grammatical dependency structure (e.g., "high-voltage line maintenance" is a risk point, "wearing insulated gloves" is a prevention and control measure, and the two form a risk relationship of "risk point - need to be implemented - prevention and control measures"); at the same time, the corresponding legal basis entities were associated (e.g., the prevention and control measure is based on Article 3.2.1 of the "Electric Power Safety Work Regulations"), and "basis label" and "priority label" were added to the risk relationship (the priority label inherits the original priority of the corresponding data, e.g., the priority of the risk relationship corresponding to the internal data of the enterprise is 1);

[0099] Next, a consistency verification rule base is constructed, including: ① consistency of entity description (e.g., "high voltage line" must be unified as "high voltage line"); ② consistency of risk relationship logic (e.g., "prevention of falls from heights" control measures can only be associated with risk points of the "high-altitude operation" category, and cannot be associated with risk points of "low-voltage equipment maintenance"); ③ validity of legal basis (e.g., verifying whether the legal entity is the current valid version and eliminating the association of expired regulations); then, based on this rule base, the identified entities and risk relationships are verified one by one, eliminating content with inconsistent descriptions, logical contradictions, and invalid basis, forming the first entity risk set; then, the first entity risk set is conflict eliminated. If there are multiple contradictory control measures corresponding to the same risk point, the "priority-time" filtering logic introduced in step S202 is called again, and finally a conflict-free second entity risk set is obtained;

[0100] Finally, the second entity risk set is transformed into structured knowledge in the form of RDF triples (RDF, Resource Description Framework), and attribute information (such as update timestamps and priority tags) is added to each triple to form a complete structured knowledge unit. Finally, all structured knowledge units are stored in a preset database (such as Neo4j graph database or MySQL database), and the initial professional knowledge base is generated after storage is completed.

[0101] For further explanation, please refer to Table 3, which is a structured knowledge data table. The second entity risk set is not transformed into a single RDF triple, but into three types of structured data: "triple core library + entity attribute table + relation mapping table," forming a complete structured knowledge unit. All data carries standardized labels for retrieval.

[0102] Table 3 Structured Knowledge Data Table

[0103]

[0104] It should be noted that before entity recognition, the second, third, and fourth learning data that do not have content conflicts need to be cleaned (removing HTML tags and handling missing values), deduplicated (based on the SimHash algorithm), and formatted (converted to JSON format) to ensure data quality.

[0105] By processing learning data through entity identification, risk relationship extraction, verification and elimination, and structured storage, the core entities and key risk relationships of power system risk management can be accurately identified, eliminating problems such as contradictions in entity definitions and logical conflicts in risk relationships. At the same time, the structured format ensures the standardization of knowledge storage, ensuring that matching risk knowledge can be quickly located when searching based on the professional knowledge base. This reduces deviations in management plans caused by fragmented knowledge and logical confusion, thereby improving the accuracy of power system risk management.

[0106] By constructing an initial professional knowledge base through multi-source data collection, priority classification, and conflict classification, the distortion of knowledge base information caused by messy data sources and content conflicts can be avoided. This ensures the authority and timeliness of risk management-related data in the initial knowledge base, laying a precise data foundation for subsequent risk knowledge retrieval and management plan generation based on the professional knowledge base, thereby improving the accuracy of power system risk management.

[0107] In some embodiments, extracting scenario keywords related to the power grid system from the target intent category specifically includes: based on a preset intent-keyword mapping rule, using a word segmentation tool (such as HanLP) to extract keywords from the target intent category to obtain several scenario keywords related to the power grid system. Specifically, the preset intent-keyword mapping rule includes the correspondence between different intent categories and power grid system-specific terms. When the target intent category is querying equipment risk points, the equipment name entity in the query data is extracted as a scenario keyword; when the target intent category is obtaining risk point prevention and control solutions, the names of identified risk points are extracted as scenario keywords; when the target intent category is querying regulatory basis, the names of prevention and control measures are extracted as scenario keywords.

[0108] In some embodiments, the scenario keywords are mapped to a preset professional knowledge base to obtain retrieval fields. Specifically, this includes matching and mapping scenario keywords related to the power grid system with standardized tags in the professional knowledge base to obtain retrieval fields.

[0109] It should be noted that for matching mapping, the device name and risk point name can be mapped to the name tag field of the entity attribute table; the operation scenario type can be mapped to the scenario tag field of the entity attribute table; and the prevention and control measure name can be mapped to both the name tag of the entity attribute table and the relation type tag of the RDF triple core library.

[0110] In some embodiments, the step of searching the professional knowledge base based on the search field to obtain risk knowledge fragments related to risk management specifically includes: performing a multi-database joint search on the professional knowledge base based on the search field to obtain several first search results; filtering each of the first search results to obtain several second search results; and integrating each of the second search results through a preset logical structure to obtain risk knowledge fragments related to risk management. Specifically, the system uses name tags, scenario tags, etc., contained in the search fields as query conditions to perform a matching query in the entity attribute table of the professional knowledge base, obtaining the corresponding entity ID and its attribute information. The obtained entity ID is then used as the query key to retrieve all relation triples associated with that entity in the RDF triple core library. The association strength label of each relation triple is then retrieved from the relation mapping table as the result sorting criterion, resulting in several first search results with sorting criteria. Next, the first search results are filtered by entity type based on the user's target intent category; for example, when the target intent category is querying risk points, entities related to prevention and control measures are filtered out. Then, the first search results are further filtered based on the association strength label, retaining results with association strength label values ​​greater than or equal to a preset threshold (e.g., 7). When the target intent involves querying legal basis, the search results are validated for timeliness, filtering out entities marked as invalid in the entity attribute table, resulting in several refined second search results. Finally, the second search results are integrated according to different intent categories using corresponding logical structures, ultimately generating risk knowledge fragments related to risk management.

[0111] It should be noted that the preset logical structure integrates the second search results, which may include: risk prevention and control intentions adopt the structure of "risk description - list of measures - legal basis", historical case intentions adopt the structure of "background - cause - measures - value", and emergency response intentions adopt the structure of "identification - operation - attention - report".

[0112] By retrieving knowledge fragments from a professional knowledge base that is updated based on user feedback analysis using search fields, it is ensured that the knowledge fragments are authoritative, real-time, and adapted to current risk management needs, thus ensuring the accuracy of the risk management solutions generated subsequently.

[0113] In some embodiments, the step of filling the risk knowledge fragment into a preset answer template to generate a risk management plan specifically includes: splitting and classifying the risk knowledge fragment to obtain classified knowledge units; performing template matching processing on the knowledge units to obtain target templates that match the scenarios of the knowledge units, and integrating the knowledge units into the target templates to obtain a preliminary plan; and verifying the risk points and prevention and control steps of the preliminary plan to obtain a risk management plan. Specifically, the risk knowledge fragment is split into several knowledge units based on core information such as risk point descriptions, prevention and control measures, legal basis, correlation strength descriptions, and historical cases contained in the risk knowledge fragment, resulting in classified knowledge units; then, template matching processing is performed according to the user's target intent category, and the knowledge units are integrated and filled according to the field order of the target template to obtain a preliminary plan; finally, the logical compatibility of the risk points and prevention and control measures is verified, and the compliance of the prevention and control steps is verified to obtain a risk management plan.

[0114] It should be noted that, in addition to using preset answer templates, generative models can also be used to construct natural language answers and form a preliminary solution.

[0115] By filling knowledge fragments into a pre-set answer template to generate answers, the risk management plan can be structurally standardized and complete, reducing errors caused by missing information or logical confusion, and ensuring that the output risk management plan is more accurate.

[0116] In some embodiments, the professional knowledge base is obtained by analyzing and processing user feedback data on historical risk management solutions to obtain analysis results, and updating the initial professional knowledge base based on the analysis results. Specifically, this includes: collecting several first feedback data points from users regarding the historical risk management solutions; performing structured transformation on each of the first feedback data points to obtain several second feedback data points, and classifying each of the second feedback data points by issue attribution to obtain several third feedback data points with classification labels, wherein the third feedback data points include knowledge base content and knowledge-related content; extracting knowledge from the knowledge base content to obtain supplementary information and correction information; performing association calculation on the knowledge-related content to obtain association deviation parameters; supplementing the initial professional knowledge base according to the supplementary information to obtain a first professional knowledge base, and correcting the first professional knowledge base according to the correction information to obtain a second professional knowledge base; and adjusting the knowledge association strength of the second professional knowledge base using the association deviation parameters to obtain an updated professional knowledge base.

[0117] Specifically, the process begins by collecting initial feedback data (such as star ratings, text suggestions, and image annotations) from users regarding historical risk management plans through the feedback entry point on the interactive interface. This includes user satisfaction ratings (1-5 stars), suggestions for problem correction (such as "missing lightning protection measures" or "outdated legal basis"), and attachments (such as photos of work scenarios and screenshots of procedures). The initial feedback data is then cleaned to remove invalid data, including feedback of 3 stars or higher containing only ratings without accompanying text or labels, text feedback containing garbled characters or meaningless words, and attachments unrelated to historical risk management plans. The initial feedback data is then structured and transformed, mapping star ratings to quantitative indicators (5 stars = excellent, 4 stars = good, 3 stars or below = needs optimization). Text suggestions are extracted using word segmentation and entity recognition to extract key information and are transformed into second feedback data containing "feedback type, problem description, and associated plan ID." This second feedback data is then categorized by problem affiliation into "knowledge base content" (such as outdated regulations or missing risk points) and "knowledge association" (such as incorrect measure ordering or unreasonable association strength). Category tags are added to obtain the third feedback data. Next, knowledge extraction is performed on the third feedback data of the "Knowledge Base Content" category. For example, if the feedback is "missing lightning protection measures", "supplementary information: additional lightning protection grounding measures are required for tower climbing operations in rainy weather" is generated. If the feedback is "the 2022 version of the regulations has expired", "correction information: update the 2022 version of the 'Electric Power Safety Work Regulations' to the 2024 version" is generated. Association calculation is performed on the third feedback data of the "Knowledge Association" category. For example, if the feedback is "fall protection measures should take priority over lightning protection measures", the association deviation parameter is calculated by comparing the historical association strength labels (original fall protection strength 8, lightning protection strength 9). Finally, based on the supplementary information, relevant entities and relationships are added to the entity attribute table and RDF triple core library of the initial professional knowledge base to obtain the first professional knowledge base. Based on the correction information, the old regulations entities are marked as "expired" and replaced with the new regulations content to obtain the second professional knowledge base. The knowledge association strength in the second professional knowledge base is adjusted according to the association deviation parameter. If the deviation is large, the association strength is adjusted significantly, and if the deviation is small, it is finely adjusted to update the association strength label of the "relationship mapping table" to obtain the updated professional knowledge base.

[0118] This process of optimizing the professional knowledge base through feedback data structuring, classification analysis, and targeted updates allows for targeted supplementation of missing content, correction of errors, and optimization of knowledge association logic. It avoids deviations in risk knowledge retrieval caused by outdated or loosely connected knowledge base content, ensuring that the professional knowledge base always remains accurate and adaptable. This provides reliable knowledge support for the subsequent generation of risk management solutions tailored to actual needs, ultimately improving the accuracy of power system risk management based on the professional knowledge base.

[0119] Step S104: Manage the risk prevention and control process of the power system based on the risk management and control scheme; in some embodiments, the management and control of the risk prevention and control process of the power system based on the risk management and control scheme specifically includes: the operation and maintenance personnel of the power company carry out risk management and control according to the risk management and control scheme in the risk prevention and control process.

[0120] It should be noted that after users manage the risk control process of the power system based on the risk management plan, this application also includes accuracy assessment based on user feedback on the risk management plan. Through collecting user feedback, organizing expert reviews, and conducting case comparisons, the system performance is continuously optimized. User feedback mainly includes satisfaction ratings, text comments, and question type selections. The system uses this information to identify deficiencies in intent understanding or measure association, and optimizes the language model and association algorithm accordingly. Expert reviews focus on professionally evaluating the system's output answers. If semantic understanding deviations or inconsistencies between associated measures and the latest regulations are found, industry corpora are supplemented, knowledge base content is updated, and scenario matching weights in the algorithm are adjusted. Case comparison involves matching and analyzing system-recommended measures with historical cases to identify missing or redundant measures, thereby expanding scenario tags, supplementing associated knowledge, and optimizing the filtering logic. All evaluation results are triggered by scheduled tasks to update the model and algorithm, and effectiveness is verified before deployment to ensure that intent recognition accuracy, measure matching accuracy, and user satisfaction meet preset standards. The system also features version management and rollback mechanisms to ensure a stable and controllable optimization process, ultimately forming a continuously iterative "evaluation-optimization-application" closed loop. Furthermore, the user feedback design incorporates immediacy and traceability, supporting both rapid on-site evaluation and allowing for subsequent detailed explanations. All feedback data is cleaned, categorized, and prioritized, then used to drive the optimization and verification of corresponding modules based on problem type, ensuring the system maintains high accuracy and practicality in power safety risk decision-making.

[0121] In this way, by receiving user inquiries directly related to power system risk management, a precise demand data source is provided for subsequent solution formulation, ensuring the accuracy of subsequent management. Inputting the inquiry data into a language model for semantic analysis to generate semantic vectors allows for precise parsing of power-related semantics, avoiding deviations in demand understanding that could lead to off-target knowledge base retrieval, thus improving management accuracy. Determining the user's target intent category through a classification model avoids misjudgments of demand due to general semantic misunderstandings, accurately locating the user's true needs from the demand analysis level, ensuring the accuracy of risk management. Extracting keywords related to scenario characteristics from intent categories, rather than generalized keywords, provides precise positioning basis for subsequent retrieval, avoiding irrelevant risk knowledge retrieval due to keyword generalization, thus improving the accuracy of risk management. Mapping scenario keywords with structured knowledge from the professional knowledge base to generate retrieval fields ensures the direction of the retrieval. Precision is ensured to avoid irrelevant or redundant knowledge fragments retrieved due to vague keywords. From the knowledge retrieval entry point, the retrieved risk knowledge is highly matched to user needs, guaranteeing the accuracy of subsequent risk management. Retrieving knowledge fragments from a professional knowledge base "updated after user feedback analysis" based on search fields ensures that the knowledge fragments are authoritative, timely, and adapted to current risk management needs. Simultaneously, filling these knowledge fragments into preset answer templates guarantees a standardized and complete risk management solution structure, reducing errors caused by information omissions or logical inconsistencies, and ensuring a more accurate output risk management solution. The generated risk management solution is used to manage the power system risk prevention and control process, aligning with actual user management needs and complying with power system risk regulations, ensuring that the risk prevention and control process is strictly implemented according to the professional standards in the knowledge base. Ultimately, this achieves improved accuracy in power system risk management based on the professional knowledge base.

[0122] See Figure 3 Based on the above method embodiments, corresponding device embodiments are provided;

[0123] One embodiment of the present invention provides a power system risk management and control device, including a first module 100, a second module 200, a third module 300 and a fourth module 400;

[0124] The first module 100 is used to receive query data related to power system risk management input by the user;

[0125] The second module 200 is used to input the query data into a language model for semantic analysis to obtain a semantic vector, and input the semantic vector into a classification model to determine the user's target intent category;

[0126] The third module 300 is used to extract scenario keywords related to the power grid system from the target intent category, map the scenario keywords to a preset professional knowledge base to obtain retrieval fields, perform a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management, and fill the risk knowledge fragments into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and updating the initial professional knowledge base based on the analysis results.

[0127] The fourth module 400 is used to manage the risk prevention and control process of the power system based on the risk management scheme.

[0128] The first module receives user inquiries directly related to power system risk management, providing a precise data source of demand for subsequent solution development and ensuring the accuracy of subsequent management. The second module inputs the inquiry data into a language model for semantic analysis to generate semantic vectors, accurately parsing the semantics of the power field and avoiding deviations in demand understanding that could lead to off-target knowledge base retrieval, thus improving the accuracy of management. A classification model determines the user's target intent category, avoiding misjudgments of demand due to general semantic misunderstandings, accurately locating the user's true needs from the demand analysis level, and ensuring the accuracy of risk management. The third module extracts keywords related to scenario features from the intent category, rather than generalized keywords, providing precise positioning basis for subsequent retrieval and avoiding irrelevant risk knowledge retrieved due to keyword generalization, thus improving the accuracy of risk management. The mapping of scenario keywords to structured knowledge in the professional knowledge base generates retrieval fields, ensuring the accuracy of retrieval. The search direction is precise, avoiding irrelevant or redundant knowledge fragments retrieved due to vague keywords. This ensures that the acquired risk knowledge highly matches user needs from the knowledge retrieval entry point, guaranteeing the accuracy of subsequent risk management. Retrieving knowledge fragments from a professional knowledge base "updated after user feedback analysis" based on search fields ensures that the knowledge fragments are authoritative, timely, and adapted to current risk management needs. Simultaneously, filling these knowledge fragments into preset answer templates guarantees a standardized and complete risk management solution structure, reducing errors caused by information omissions or logical inconsistencies, and ensuring a more accurate output risk management solution. The fourth module manages the power system risk prevention and control process based on the generated risk management solution, aligning with both actual user management needs and power system risk standards, ensuring that the risk prevention and control process is strictly implemented according to the professional standards in the knowledge base. Ultimately, this achieves improved accuracy in power system risk management based on the professional knowledge base.

[0129] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement any of the above-described method embodiments of the present invention to provide a power system risk management method.

[0130] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0131] Based on the above-described embodiment of a power system risk management method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a power system risk management method according to any embodiment of the present invention.

[0132] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0133] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0134] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0135] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute a power system risk management method as described in any of the above-described method embodiments of the present invention.

[0136] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0137] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for risk management and control of power systems, characterized in that, include: Receive user-input query data related to power system risk management; The query data is input into a language model for semantic analysis to obtain a semantic vector. The semantic vector is then input into a classification model to determine the user's target intent category. The system extracts scenario keywords related to the power grid system from the target intent category, maps the scenario keywords to a preset professional knowledge base to obtain retrieval fields, and performs a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management. The risk knowledge fragments are then filled into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and the initial professional knowledge base is updated based on the analysis results. The risk control scheme is used to manage the risk prevention and control process of the power system.

2. The power system risk management method as described in claim 1, characterized in that, The step of inputting the semantic vector into the classification model to determine the user's target intent category specifically includes: The semantic vector is input into the classification model to calculate probability values, wherein the probability values ​​include risk query probability values, prevention and control strategy consultation probability values, and strategy legal basis query probability values. The initial intent category is determined by comparing the probability values ​​of risk queries, consultations on prevention and control strategies, and queries on the legal basis of the strategies. If the probability value corresponding to the initial intent category is greater than or equal to a preset threshold, then the user's target intent category is determined based on the initial intent category.

3. The power system risk management method as described in claim 1, characterized in that, The construction of the initial professional knowledge base specifically includes: The first learning data corresponding to each release time point is collected in real time. The first learning data is prioritized to obtain several second learning data. The first learning data comes from different data sources related to power system risk management. Determine whether there is a content conflict among the second learning data. If there is a content conflict, determine whether the conflicting content is at the same priority. If the conflicting content is not at the same priority, retain the conflicting content with the highest priority as the third learning data. If the conflicting content is at the same priority, compare the release time points corresponding to the conflicting content and retain the conflicting content with the most recent release time point as the fourth learning data. The third learning data, the fourth learning data, and the second learning data that do not have content conflicts are integrated and processed to obtain the initial professional knowledge base.

4. The power system risk management method as described in claim 3, characterized in that, The process of integrating the third learning data, the fourth learning data, and the second learning data that do not have content conflicts to obtain the initial professional knowledge base specifically includes: Entity recognition is performed on the second learning data, the third learning data, and the fourth learning data to obtain several entities related to power system risk management; Syntactic dependency structures are obtained by analyzing the second, third, and fourth learning data respectively. Risk relationships are extracted from the syntactic dependency structures using a semantic role labeling method. The risk relationships are the relationships between risk points and prevention and control measures related to the power system. A consistency check is performed on the entities and the risk relationships to obtain a first entity risk set, and conflict elimination processing is performed on the first entity risk set to obtain a second entity risk set. The second entity risk set is transformed into structured knowledge represented in the form of triples, and the structured knowledge is stored in a preset database to obtain the initial professional knowledge base.

5. The power system risk management method as described in claim 1, characterized in that, The step of searching the professional knowledge base based on the search fields to obtain risk knowledge fragments related to risk management specifically includes: A multi-database joint search is performed on the search field in the professional knowledge base to obtain several first search results; Each of the first search results is filtered to obtain several second search results; By integrating the second search results through a preset logical structure, risk knowledge fragments related to risk management are obtained.

6. The power system risk management method as described in claim 1, characterized in that, The step of filling the risk knowledge fragments into a preset answer template to generate a risk management plan specifically includes: The risk knowledge fragments are broken down and classified to obtain classified knowledge units; The knowledge unit is subjected to template matching processing to obtain a target template that matches the scenario of the knowledge unit, and the knowledge unit is integrated and filled into the target template to obtain a preliminary solution. The risk points and prevention and control steps of the preliminary plan are verified to obtain a risk management plan.

7. The power system risk management method as described in claim 1, characterized in that, The professional knowledge base is obtained by analyzing and processing user feedback data on historical risk management solutions, and updating the initial professional knowledge base based on the analysis results. Specifically, it includes: Collect several initial feedback data points from users regarding the historical risk management plan; The first feedback data is structured and transformed to obtain several second feedback data. The second feedback data is then classified into question categories to obtain several third feedback data with category labels. The third feedback data includes knowledge base content and knowledge-related content. Knowledge extraction is performed on the knowledge base content to obtain supplementary and correction information, and association calculation is performed on the knowledge-related content to obtain association deviation parameters; The initial professional knowledge base is supplemented according to the supplementary information to obtain a first professional knowledge base. The first professional knowledge base is then corrected according to the correction information to obtain a second professional knowledge base. The knowledge association strength of the second professional knowledge base is adjusted using the association deviation parameter to obtain an updated professional knowledge base.

8. A power system risk management and control device, characterized in that, It includes Module 1, Module 2, Module 3, and Module 4; The first module is used to receive user-input query data related to power system risk management; The second module is used to input the query data into a language model for semantic analysis to obtain a semantic vector, and then input the semantic vector into a classification model to determine the user's target intent category. The third module is used to extract scenario keywords related to the power grid system from the target intent category, map the scenario keywords to a preset professional knowledge base to obtain retrieval fields, perform a retrieval in the professional knowledge base based on the retrieval fields to obtain risk knowledge fragments related to risk management, and fill the risk knowledge fragments into a preset answer template to generate a risk management plan. The professional knowledge base is obtained by analyzing and processing the feedback data collected from users on historical risk management plans to obtain analysis results, and updating the initial professional knowledge base based on the analysis results. The fourth module is used to manage the risk prevention and control process of the power system based on the risk management scheme.

9. A terminal device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of a power system risk management method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device or apparatus containing the computer-readable storage medium to perform a power system risk management method as described in any one of claims 1 to 7.

Images