An artificial intelligence agent interaction management and control method and system for power distribution network services

Abstract

The application discloses an artificial intelligence agent interaction management and control method and system for power distribution network services, relates to the technical field of power distribution network digitization, and comprises the following steps: constructing a power distribution network global data correlation system and a professional knowledge base, generating a standardized data pool, a four-level correlation intermediate table and knowledge index data; receiving a user natural language query text, completing intent recognition and keyword extraction through a large language model; performing integrity checking on the keyword set, converting the parameter into a structured API request parameter after iterative completion of the parameter through guided supplementary inquiry; querying data according to the classification of business scenarios, generating a natural language reply through analysis by the large language model; performing five-dimensional interaction management and control throughout the interaction and generating whole-process management and control data; and finally synchronizing the reply content and the management and control data to a visual application layer for rendering and display. The application can realize global data penetration, multi-scene general adaptation and whole-process safe control of the power distribution network.

Description

Technical Field

[0001] This invention relates to the field of digital distribution network technology, and in particular to an artificial intelligence-based interactive management and control method and system for distribution network operations. Background Technology

[0002] With the continuous advancement of the construction of new power systems, the business coverage of distribution networks is constantly expanding. Massive amounts of multi-source business data have been accumulated in various aspects such as operation, maintenance, and marketing. Improving the efficiency of data query and business analysis by relying on intelligent interactive technology has become an important development direction for the digital operation and maintenance of distribution networks.

[0003] Currently, most intelligent business interaction systems in the power distribution network field adopt a single-scenario customized development model. They achieve user query response based on preset fixed keyword matching rules. Relevant business data are stored in independent business systems such as electricity information collection, power supply service, and dispatch monitoring. The interaction execution process does not establish a unified end-to-end control mechanism, and the query results are mostly output in independent text form, which is independent of the power distribution network visualization display module.

[0004] Existing technologies cannot simultaneously achieve multi-source data association and connection, flexible natural language interaction, full-process security control and visual collaborative display under a unified architecture, making it difficult to adapt to the actual needs of integrated intelligent interactive control in multiple business scenarios of power distribution networks. Summary of the Invention

[0005] To address the technical challenges of existing technologies in simultaneously achieving multi-source data interconnection, flexible natural language interaction, end-to-end security control, and visualized collaborative display under a unified architecture, this invention provides an AI-based interactive control method and system for power distribution network operations.

[0006] The technical solution adopted in this invention is:

[0007] The first aspect of this application provides an AI-based agent interaction management method for distribution network operations, including the following:

[0008] Step 1: Construct a data association system and professional knowledge base for the entire distribution network. Integrate, clean, associate and standardize the multi-source business data of the distribution network to generate a standardized data pool and a four-level association intermediate table. Based on the standardized data pool and the four-level association intermediate table, integrate the professional knowledge data of the distribution network to generate knowledge index data.

[0009] Step 2: Receive the natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set;

[0010] Step 3: Perform integrity verification on the keyword set. If the integrity verification fails, generate guided supplementary query text based on the keyword set and the knowledge index data. Push the guided supplementary query text to the user and receive the user's supplementary reply text to the guided supplementary query text. Update the keyword set based on the supplementary reply text. Iterate until the integrity verification passes. Convert the verified keyword set into a structured API request parameter set.

[0011] Step 4: Classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool based on the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text;

[0012] Step 5: Execute five-dimensional interactive control during the interaction process from Step 2 to Step 4 to generate full-process control data;

[0013] Step 6: Synchronize the natural language response text with the full-process management data to the visualization application layer, and render and display it in the visualization application layer.

[0014] The second aspect of this application provides an AI agent interaction management and control system for distribution network services, applied to the aforementioned AI agent interaction management and control method for distribution network services, and further includes:

[0015] The data knowledge construction module is used to build a data association system and professional knowledge base for the entire distribution network. It integrates, cleans, associates and standardizes multi-source business data of the distribution network, generates a standardized data pool and a four-level association intermediate table, and integrates distribution network professional knowledge data based on the standardized data pool and the four-level association intermediate table to generate knowledge index data.

[0016] The natural language parsing module is used to receive natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set;

[0017] The parameter completion and conversion module is used to perform integrity verification on the keyword set. When the integrity verification fails, it generates guided supplementary query text based on the keyword set and the knowledge index data. After pushing the guided supplementary query text to the user, it receives the user's supplementary reply text in response to the guided supplementary query text, updates the keyword set based on the supplementary reply text, and iterates until the integrity verification passes. Finally, it converts the verified keyword set into a structured API request parameter set.

[0018] The data query and response module is used to classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool according to the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text.

[0019] The five-dimensional interactive control module is used to perform five-dimensional interactive control during the interaction process from step 2 to step 4, and generate full-process control data.

[0020] The visualization rendering module is used to synchronize the natural language response text and the full-process management data to the visualization application layer, and render and display them in the visualization application layer.

[0021] The beneficial effects of the present invention are at least one of the following:

[0022] This invention constructs a data association system and professional knowledge base covering the entire power distribution network, integrates, cleans, associates, and standardizes multi-source business data, generates a standardized data pool and a four-level association intermediate table, and forms knowledge index data based on data and documents. This can unify the indicator caliber and hierarchical association relationship of multi-source data, alleviate the data silo problem between various business systems of the power distribution network, improve data consistency and traceability, and provide standardized data and knowledge support for subsequent interactive analysis.

[0023] This invention uses a large language model to perform intent recognition and keyword extraction on the natural language query text input by the user. Combined with keyword integrity verification and guided completion mechanism, it can improve the rigidity of interaction under the traditional fixed rule-driven approach. When the query information is incomplete, it guides the user to supplement the necessary parameters, improves the smoothness of interaction, and lowers the usage threshold for front-line operation and maintenance personnel.

[0024] This invention categorizes structured request parameters according to business scenarios, calls the corresponding authorized data interface to query a standardized data pool, and then parses the query results using a large language model to generate a natural language response. This facilitates adaptation to various power distribution network business scenarios and improves the scenario versatility of the solution and the readability of the output results.

[0025] This invention performs five-dimensional interactive control and generates full-process control data during the entire interactive process, which can standardize the interactive process and data access behavior, improve the security and auditability of system operation, and help meet the security and compliance requirements of the core power system.

[0026] This invention synchronizes natural language response text with full-process management and control data to the visualization application layer for rendering and display. This can alleviate the problem of disconnect between intelligent agents and power distribution network visualization platforms, improve the intuitiveness of business status and interaction results, and facilitate maintenance personnel to quickly locate and obtain relevant information. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0028] Figure 2 This is a schematic diagram of the natural language interaction interface according to Embodiment 1 of the present invention;

[0029] Figure 3 This is a schematic diagram of the early warning visualization linkage interface of Embodiment 1 of the present invention;

[0030] Figure 4 This is a schematic diagram of the report list interface in Embodiment 1 of the present invention. Detailed Implementation

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] Example 1

[0033] It should be noted that this embodiment is based on the LangGraph engine to build three interconnected backend sub-services. The three interconnected backend sub-services include a natural language query service, a power distribution network early warning service, and an automatic reporting service. The natural language query service is used for natural language queries and hierarchical guided responses; the power distribution network early warning service is used for hierarchical early warning reminders across the entire domain; and the automatic reporting service is used for the automatic generation of standardized reports.

[0034] All sub-services adopt a four-node serial basic workflow, which includes the classify node, the extract_entities node, the run_handlers node, and the analyze node. The functionality is implemented based on JSON configuration files, Python execution scripts, vector retrieval, large language model inference, and rule engine. It provides standardized REST / SSE interfaces to the outside world and supports local deployment and virtual environment operation.

[0035] It should be noted that the LangGraph engine refers to a large language model application orchestration framework based on a directed graph state machine. In this embodiment, it is used to orchestrate workflow nodes and state transitions for intelligent agent interactions.

[0036] In the four-node serial basic workflow: the classify node performs semantic parsing and intent recognition; the extract_entities node performs keyword entity extraction and integrity verification; the run_handlers node performs parameter transformation, scene classification, and authorization data query; and the analyze node performs large language model result parsing and natural language response generation.

[0037] This embodiment provides an AI-based agent interaction management method for distribution network services, such as... Figure 1 As shown, it includes the following:

[0038] Step 1: Construct a data association system and professional knowledge base for the entire distribution network. Integrate, clean, associate and standardize multi-source business data of the distribution network to generate a standardized data pool and a four-level association intermediate table. Based on the standardized data pool and the four-level association intermediate table, integrate professional knowledge data of the distribution network to generate knowledge index data.

[0039] This step utilizes the data platform interface to acquire multi-source data in real time, including user voltage data and transformer area operation data from the electricity information acquisition system, emergency repair work orders, complaint work orders, and opinion work orders from the power supply service system, substation bus voltage and line operation data from the dispatch monitoring system, transformer area files, line files, and power supply station jurisdiction data from the power grid resource business platform, and township and village administrative division data from the geographic information system. A combination of incremental and full-volume acquisition is used to ensure data real-time performance and integrity. During the data acquisition process, special early warning indicator fields and report statistical fields are simultaneously distinguished and stored according to the classification and identification of early warning ledgers and report ledgers.

[0040] The collected raw data is cleaned to remove duplicate, abnormal, and null data, and data format errors are corrected. Preprocessing refers to the unified format conversion of voltage values, duration values, transformer area numbers, user numbers, and administrative division codes, and the unified measurement units of indicators in accordance with national standard early warning parameters, and the matching of field format standards for various report statistical specifications.

[0041] The formula for determining outliers during data cleaning is as follows:

[0042] or

[0043] in, For distribution network voltage data, This is the minimum threshold voltage for the distribution network. The highest voltage threshold of the distribution network. and According to national standards, voltage data outside this range are considered abnormal and are removed.

[0044] The unified core business indicators for the distribution network include low voltage coefficient, number of low voltage users, duration of low voltage user service, number of work orders, number of transformer substations, and average voltage. The unified time statistics dimensions are minute, hour, day, month, quarter, and grade. The unified data storage format is a structured data table format, generating a standardized data pool.

[0045] In one possible implementation, the fourth-level association intermediate table is generated based on the multi-source business data of the distribution network as follows:

[0046] Determine the corresponding jurisdictional relationships between villages, townships, and districts / counties for each transformer substation area;

[0047] An automated verification based on the accuracy of four-level association matching is performed on the determined administrative correspondence between transformer substations and villages / communities, townships, and districts / counties. When there is a transformer substation association that fails the verification, an association correction process is triggered. The verified transformer substation association is integrated with the transformer substation number, transformer substation name, power supply company name, work group name, township administrative division code, village / community administrative division code, and geographic coordinate information to obtain the four-level association intermediate table. The four-level association intermediate table is used to establish a one-to-one correspondence between the four levels of transformer substations, villages / communities, townships, and districts / counties.

[0048] It should be noted that the jurisdictional correspondence is determined through manual on-site investigation. The power supply company's operation and maintenance personnel visit the transformer substations to compare the geographical coordinates of the substations with the administrative boundaries of townships and villages, and record the correspondence between the substations and the administrative regions.

[0049] The accuracy rate of the fourth-level association matching is the ratio of the number of verified station associations to the total number of station areas, calculated using the following formula:

[0050]

[0051] in, For the accuracy of level 4 association matching, To verify the number of successfully associated transformer substations, This refers to the total number of stations / regions.

[0052] when When the accuracy rate is below 100%, the association correction process is triggered, marking the failed station associations as pending correction and pushing them to the management terminal for manual review and correction.

[0053] In one possible implementation, the knowledge index data generated based on the power distribution network expertise includes the following:

[0054] The knowledge blocks are vector-encoded using a word embedding model, and a knowledge index is constructed based on vector similarity. Dedicated index labels are created for warning threshold items, guiding response script items, and report compilation specification items to obtain the knowledge index data.

[0055] It should be noted that the professional knowledge materials on distribution networks cover all scenarios of low voltage management, power supply reliability, power quality and work order handling. They include technical specifications for voltage quality management, manuals for handling distribution network faults, standards for closed-loop management of work orders, investment specifications for distribution network projects, technical documents on rural power grid management, case studies of on-site management experience, as well as national standards for voltage over-limit, management methods for early warning of distribution network indicators, guidelines for compiling operation and maintenance reports, and specifications for natural interactive guidance.

[0056] Imported documents are formatted uniformly, invalid information is removed, professional knowledge points are segmented and organized, long documents are manually divided into blocks to ensure that each knowledge block contains complete professional knowledge points, and warning threshold items, guiding response script items, and report preparation specification items are separately separated as independent knowledge units.

[0057] The vector encoding calculation formula is as follows:

[0058]

[0059] in, Vector data for knowledge blocks. For word embedding encoding function, This refers to the text content of the knowledge block.

[0060] Word embedding encoding function The text-embedding-ada-002 model is used, and the output vector dimension is 1536. (Vector data) It is a 1536-dimensional floating-point vector.

[0061] The retrieval success rate of the knowledge index data is the ratio of the number of retrieval queries that successfully retrieve the target knowledge block to the total number of retrieval queries, calculated using the following formula:

[0062]

[0063] in, To improve the success rate of knowledge retrieval, To successfully retrieve the target knowledge block, the number of search queries is required. To represent the total number of search queries, this embodiment has undergone multiple rounds of recall testing and optimization to achieve a higher knowledge retrieval success rate. No less than 95%.

[0064] The knowledge points in the knowledge base are verified based on a standardized data pool and a four-level association intermediate table to ensure that the knowledge content conforms to the actual situation of the power distribution network. A dynamic update mechanism for the knowledge base is established to regularly import new technical specifications and governance experience to ensure the timeliness of the knowledge base.

[0065] This step involves constructing a four-level interconnected intermediate table and a standardized data pool at the distribution area, village / community, township, and district / county levels. This unifies the indicator definitions, data formats, and hierarchical relationships of multi-source data, breaking down data silos among various business systems in the distribution network and improving the accuracy of overall data positioning and causal analysis. Simultaneously, it constructs a distribution network professional knowledge base and vector index to provide standardized and traceable professional data and knowledge support for subsequent intelligent agent interaction, early warning judgment, and report generation.

[0066] Step 2: Receive the natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set.

[0067] In one possible implementation, the intent recognition and keyword extraction are performed based on the natural language query text to obtain the keyword set, which includes the following:

[0068] The semantic parsing results are compared with the distribution network business intent database to determine the user's core intent. The distribution network business intent database is divided into data query, early warning verification, and report generation categories.

[0069] Calculate the intent recognition confidence score. When the intent recognition confidence score is not lower than a preset threshold, the intent recognition result is determined to be valid. When the intent recognition confidence score is lower than the preset threshold, the intent clarification process is triggered to obtain the keyword set.

[0070] It should be noted that the large language model described in this embodiment is a dedicated model fine-tuned for distribution network services, which includes five modules: input layer, semantic parsing layer, intent recognition layer, keyword extraction layer, and output layer.

[0071] The fine-tuning base uses DeepSeek-R1-32B and Qwen3-30B, with the following training parameters: learning rate 0.0001, batch size 16, number of iterations 100, and ReLU activation function. The fine-tuning dataset contains no less than 50,000 question-and-answer pairs related to power distribution network operations, covering four scenarios: low voltage management, power supply reliability, power quality, and work order processing.

[0072] The distribution network business intent database is constructed using pre-annotated distribution network business corpus, with each intent category containing typical question templates and keyword features.

[0073] The formula for calculating the confidence score of intent recognition is:

[0074]

[0075] in, To determine the confidence level for intent identification, For semantic parsing similarity, Weighting coefficients for business intent matching degree and Determined by grid search of the intent recognition sample set.

[0076] The preset threshold in this embodiment is: ,when If the intent recognition result is deemed valid, keyword extraction is performed based on the valid intent recognition result; when When the intent clarification process is triggered, an intent confirmation question is sent to the user.

[0077] Keyword extraction is based on a pre-set distribution network business keyword library. Core keywords are extracted from the semantic parsing results and categorized into five major categories: region, time, equipment, indicator, and work order. A keyword set is generated, and the structured representation of the keyword set is as follows:

[0078]

[0079] in, For a set of keywords, to For different categories of core keywords, This represents the total number of keywords.

[0080] This step uses a large language model-driven semantic parsing and intent recognition mechanism to replace the traditional fixed keyword matching rules. It can accurately understand the business needs expressed by users in natural language, support the automatic identification of multiple types of power distribution network business intents, effectively solve the problem of rigid interaction in fixed workflows, and improve the accuracy of intent recognition and the flexibility of scenario adaptation.

[0081] Step 3: Perform integrity verification on the keyword set. If the integrity verification fails, generate guided supplementary query text based on the keyword set and the knowledge index data. Push the guided supplementary query text to the user and receive the user's supplementary reply text. Update the keyword set based on the supplementary reply text. Iterate until the integrity verification passes. Convert the verified keyword set into a structured API request parameter set.

[0082] In one possible implementation, the keyword set is subjected to a completeness check. When the completeness check fails, guided supplementary query text is generated based on the keyword set and the knowledge index data.

[0083] A predefined set of required parameter categories corresponding to each business intent is provided, including region category, time category, equipment category, indicator category, and work order category.

[0084] The category labels of each keyword in the keyword set are compared one by one with the required parameter category set corresponding to the current intent to identify the missing required parameter categories and generate a list of missing parameters.

[0085] Based on the missing parameter list and the guided response script entries in the knowledge index data, the guided supplementary question text is generated. The guided supplementary question text generates a corresponding natural language question statement for each missing required parameter category in the missing parameter list, thus obtaining the guided supplementary question text.

[0086] It should be noted that the set of required parameter categories is preset according to the business scenario. Taking the low voltage data query intent as an example, the required parameter categories include region category and time category.

[0087] Taking a work order processing query as an example, the required parameter categories include region category, time category, and work order category. The guidance response script entries are standard guidance response script templates pre-stored in the knowledge base. The templates contain placeholders for missing parameter names to be filled. The system fills in the placeholders according to the list of missing parameters and then generates the final text.

[0088] The information completion similarity assessment is used to evaluate the degree of matching between the information collected in the guided completion query and the missing parameters. The calculation formula is as follows:

[0089]

[0090] in, To complete the similarity of the information, For keyword matching degree, For information integration, The weighting coefficients for knowledge support are determined through training with a complete sample set.

[0091] Complete the similarity of information Compared with the preset completion threshold, when The acquisition of valid missing information is considered successful when the threshold is not lower than the preset completion threshold. If the value is below the preset completion threshold, it is determined that no valid missing information has been obtained.

[0092] The verified keyword set is formatted according to the preset parameter mapping configuration file, mapping natural language keywords to the standard parameter names and parameter value formats required by each authorized data interface, generating a structured API request parameter set; for example, regional keywords are mapped to township administrative division code parameters, and time keywords are mapped to start and end timestamp parameters.

[0093] This step, through mandatory parameter integrity verification and a guided multi-round completion mechanism, automatically initiates targeted supplementary queries when the user's query information is incomplete, rather than directly returning an error. This significantly improves the success rate of interactive execution and lowers the usage threshold for front-line operations and maintenance personnel. At the same time, information completion similarity verification ensures the effectiveness of the completed information, ensuring the completeness and accuracy of parameters for subsequent data queries.

[0094] Step 4: Classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool based on the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text.

[0095] In one possible implementation, the structured API request parameter set is categorized according to business scenarios, and the data query results are obtained by calling the authorized data interface to query the standardized data pool based on the categorization results.

[0096] The structured API request parameter set is classified according to business scenarios, including low voltage management scenarios, power supply reliability scenarios, power quality scenarios, work order processing scenarios, and project investment analysis scenarios, and the classification results are obtained.

[0097] Based on the classification results, the authorized data interface is invoked to query the standardized data pool;

[0098] During the process of calling the authorized data interface to query the standardized data pool, the measured index values ​​are extracted from the data query results, the corresponding index limit thresholds stored in the standardized data pool are retrieved, the index deviation is calculated, and the warning level is divided into three levels: general warning, important warning and emergency warning, based on the preset value range in which the index deviation falls. The data exceeding the limit is automatically generated into structured warning raw data and stored in the management database to obtain the data query results.

[0099] It should be noted that the scenario classification is determined based on a combination of intent recognition results and indicator category labels in the keyword set. The scenario classification matching degree is used to evaluate the adaptability of the current query to various business scenarios. The calculation formula is as follows:

[0100]

[0101] in, For scene classification matching degree, For feature parameter matching degree, The weighting coefficients for scene library similarity are determined through training on scene classification sample sets.

[0102] The formula for calculating the deviation of the indicator is:

[0103]

[0104] in, This refers to the deviation of the indicator. These are the measured index values. Set thresholds for the corresponding indicators.

[0105] The direction of the value is consistent with the direction of the abnormality of the indicator: for positive indicators (power outage frequency, low voltage duration, number of work orders, etc.), This is the upper limit threshold; for reverse indicators (such as user voltage), This is the lower threshold.

[0106] For user voltage reverse indicators Take the lower voltage threshold specified in the national standard; for the positive indicator of power outage frequency... Take the maximum number of power outages allowed in the operation and maintenance management regulations.

[0107] The preset numerical ranges are set according to the industry standards corresponding to the index types: for voltage indexes, the nominal voltage is used as the benchmark, and a deviation of no more than 10% of the nominal voltage is judged as a general warning, a deviation of 10% but no more than 30% of the nominal voltage is judged as an important warning, and a deviation of 30% of the nominal voltage is judged as an emergency warning; for power outage frequency indexes, the annual allowable power outage number is used as the benchmark, and the indexes are divided into levels according to the same proportion.

[0108] The formula for calculating data query accuracy is:

[0109]

[0110] in, To improve data query accuracy, To retrieve the correct number of data rows, This represents the total number of data records retrieved.

[0111] In one possible implementation, the data query results are input into the large language model for parsing and processing to generate a natural language response text including the following:

[0112] Retrieve the report compilation specification entries and preset report templates from the knowledge index data;

[0113] The large language model automatically fills in the data query results according to the field structure of the preset report template to generate a draft report; the preset report template includes a daily report template, a weekly report template, a monthly report template, and a quarterly governance analysis template;

[0114] Based on the data query results and the initial report draft, the natural language response text is generated through parsing and processing using the large language model.

[0115] It should be noted that the report preparation specifications are the State Grid Operation and Maintenance Report Preparation Specifications pre-stored in the knowledge base, and each report template predefines fixed field structures and statistical dimensions.

[0116] Automatic population involves extracting values ​​from each statistical dimension in the data query results, matching and filling in the corresponding fields in the report template. For fields in the report template that require textual analysis, the large language model automatically writes analysis comments based on the data results.

[0117] like Figure 2 As shown, the large language model interprets the data results based on power distribution network expertise, summarizes the core conclusions, and generates natural language response text containing core data, conclusions, and professional suggestions. The response text also includes a report preview entry, allowing users to directly view the generated draft report.

[0118] The formula for calculating the accuracy of result analysis is:

[0119]

[0120] in, For the accuracy of result analysis, To improve data matching accuracy, To assess semantic accuracy, a harmonic mean function is used to evaluate the results, taking into account both data matching and semantic representation.

[0121] This step adopts a generalized architecture to adapt to various distribution network business scenarios such as low voltage management and power supply reliability, eliminating the need for repeated development for a single scenario and reducing development and promotion costs; through automatic calculation of indicator deviation and three-level early warning, abnormal risks can be identified in advance, reducing on-site handling costs; based on knowledge base templates, standardized reports are automatically generated, eliminating manual summary and table creation work, saving manpower and time and reducing the probability of human error.

[0122] Step 5: Execute five-dimensional interactive control during the interaction process from Step 2 to Step 4 to generate full-process control data.

[0123] In one possible implementation, the five-dimensional interaction control includes interaction parsing control, data access control, model reasoning control, security permission control, and audit backtracking control;

[0124] The interactive parsing control includes: preloading the semantic parsing rule base, verifying the confidence of intent recognition in real time, verifying the completeness and accuracy of keyword extraction, and verifying the semantic consistency of information completion;

[0125] The data access control includes: performing data access permission pre-verification, generating and authenticating access tokens, performing de-identification processing on core sensitive data, and performing interface call rate limiting.

[0126] The model inference control includes: setting thresholds for model inference parameters, configuring read-only permissions for raw data, real-time verification of inference nodes, and verification of the accuracy of inference results.

[0127] The security access control includes: building a four-level access control system at the city, district / county, township, and village levels; binding user roles and permissions; verifying access area permissions in real time; and blocking unauthorized access.

[0128] The audit backtracking and control includes: deploying log collection nodes across the entire domain to collect interaction logs, inference logs, data access logs, interface call logs, and permission verification logs in real time, storing log data in a structured manner, and generating an audit index.

[0129] The control results generated by the interaction parsing control, data access control, model inference control, and security permission control are integrated with session identifier, operation time, and user identifier as associated fields to obtain the full-process control data.

[0130] It should be noted that quantitative evaluation indicators are set for each dimension of control:

[0131] Interactive parsing and control dimensions include intent recognition compliance rate, keyword extraction accuracy rate, and information completion compliance rate. Among them, the intent recognition compliance rate is the ratio of the number of compliant intent recognition operations to the total number of intent recognition operations; the keyword extraction accuracy rate is the ratio of the number of correctly extracted keywords to the total number of extracted keywords; and the information completion compliance rate is the ratio of the number of compliant information completion operations to the total number of information completion operations.

[0132] Data access control dimensions include token authentication success rate and data anonymization accuracy rate; the token authentication success rate is the ratio of the number of successfully authenticated tokens to the total number of tokens generated, and the data anonymization accuracy rate is the ratio of the number of correctly anonymized data fields to the total number of data fields that need to be anonymized; data transmission is encrypted using the SM2 national cryptographic algorithm to ensure data transmission security.

[0133] The model inference control dimension includes the model inference compliance rate, which is the ratio of the number of compliant model inference operations to the total number of model inference operations.

[0134] Security access control dimensions include access control compliance rate and unauthorized access interception rate; where the access control compliance rate is the ratio of the number of compliant access verifications to the total number of access verifications, and the unauthorized access interception rate is the ratio of the number of successfully intercepted unauthorized access attempts to the total number of unauthorized access attempts.

[0135] The audit backtracking control dimension includes log collection completeness rate, which is the ratio of the number of successfully collected log entries to the total number of log entries that should have been collected.

[0136] After the data integration of the entire process is completed, the data integrity is verified and stored in the management database. The data integration integrity rate is the ratio of the number of successfully integrated management data items to the total number of management data items that should be integrated.

[0137] This step involves five-dimensional full-process control: interactive parsing, data access, model reasoning, security permissions, and audit backtracking. It constructs a four-level hierarchical permission system and a full-link log auditing mechanism to achieve core data anonymization, unauthorized access interception, and full traceability of operations. This comprehensively improves the data security and compliance of the distribution network intelligent agent operation and meets the safety operation specifications of the core power system.

[0138] Step 6: Synchronize the natural language response text with the full-process management data to the visualization application layer, and render and display it in the visualization application layer.

[0139] In one possible implementation, the visualization application layer is built on a power grid map platform; it renders heat maps of distribution network data according to four levels: city level, district / county level, township level, and village level, displays core business indicators, and automatically highlights abnormal areas in red; it supports integrated interactive operations such as regional drill-down, data viewing, intelligent questioning, and result display to obtain the rendered display.

[0140] It should be noted that core business indicators include low voltage coefficient, number of low voltage users, power supply reliability, and work order processing timeliness; the abnormal red marking is based on the early warning judgment results in step 4, and the abnormal areas of the three levels of early warning, important early warning and emergency early warning are marked with different colors respectively.

[0141] The response time for visualization is the sum of data transmission time and interface rendering time, calculated using the following formula:

[0142]

[0143] in, Total response time For data transmission time, This refers to the interface rendering time.

[0144] In this embodiment The average streaming time for SSE is approximately 200ms. The average rendering time for the front end is approximately 150ms, and the total response time is... Keep it within 400ms.

[0145] Based on the early warning judgment and automatic report generation capabilities, the front-end visualization platform simultaneously implements linkage functions:

[0146] Real-time traversal of all operational metrics within the standardized data pool, such as Figure 3 As shown, the system automatically performs early warning calculations, categorizing all early warnings by business type and placing them in the early warning card bar on the right side of the interface. Maintenance personnel can click on any early warning card to import the warning information into the natural language interaction window, automatically generating corresponding query commands. All abnormal transformer locations are automatically highlighted in red and flashing on the power grid map. Figure 4 As shown, the right side of the interface features a separate report list module that supports searching all historical reports by power supply unit, business type, and time interval, and provides online preview and one-click export functions.

[0147] This step deeply integrates the intelligent agent interaction results with the power grid visualization platform, enabling intuitive perception of the overall situation through multi-level data heat maps and automatic red highlighting of abnormal areas; it supports integrated operations such as regional drill-down, early warning linkage, and report export, helping front-line operation and maintenance personnel to quickly locate problem areas and effectively improve the response speed and visual collaborative combat capabilities of distribution network governance.

[0148] Example 2

[0149] This embodiment provides an AI agent interaction management and control system for distribution network services, applied to the aforementioned AI agent interaction management and control method for distribution network services, and further includes:

[0150] The data knowledge construction module is used to build a data association system and professional knowledge base for the entire distribution network. It integrates, cleans, associates and standardizes multi-source business data of the distribution network, generates a standardized data pool and a four-level association intermediate table, and integrates distribution network professional knowledge data based on the standardized data pool and the four-level association intermediate table to generate knowledge index data.

[0151] The natural language parsing module is used to receive natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set;

[0152] The parameter completion and conversion module is used to perform integrity verification on the keyword set. When the integrity verification fails, it generates guided supplementary query text based on the keyword set and the knowledge index data. After pushing the guided supplementary query text to the user, it receives the user's supplementary reply text in response to the guided supplementary query text, updates the keyword set based on the supplementary reply text, and iterates until the integrity verification passes. Finally, it converts the verified keyword set into a structured API request parameter set.

[0153] The data query and response module is used to classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool according to the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text.

[0154] The five-dimensional interactive control module is used to perform five-dimensional interactive control during the interaction process from step 2 to step 4, and generate full-process control data.

[0155] The visualization rendering module is used to synchronize the natural language response text and the full-process management data to the visualization application layer, and render and display them in the visualization application layer.

[0156] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for AI-powered agent interaction control in power distribution network operations, characterized in that, Includes the following: Step 1: Construct a data association system and professional knowledge base for the entire distribution network. Integrate, clean, associate and standardize the multi-source business data of the distribution network to generate a standardized data pool and a four-level association intermediate table. Based on the standardized data pool and the four-level association intermediate table, integrate the professional knowledge data of the distribution network to generate knowledge index data. Step 2: Receive the natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set; Step 3: Perform integrity verification on the keyword set. If the integrity verification fails, generate guided supplementary query text based on the keyword set and the knowledge index data. Push the guided supplementary query text to the user and receive the user's supplementary reply text to the guided supplementary query text. Update the keyword set based on the supplementary reply text. Iterate until the integrity verification passes. Convert the verified keyword set into a structured API request parameter set. Step 4: Classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool based on the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text; Step 5: Execute five-dimensional interactive control during the interaction process from Step 2 to Step 4 to generate full-process control data; Step 6: Synchronize the natural language response text with the full-process management data to the visualization application layer, and render and display it in the visualization application layer.

2. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, The method for generating the fourth-level association intermediate table based on the multi-source business data of the distribution network is as follows: Determine the corresponding jurisdictional relationships between villages, townships, and districts / counties for each transformer substation area; An automated verification based on the accuracy of the four-level association matching is performed on the determined administrative correspondence between the transformer substation and the village / community, township, and district / county. When there is a transformer substation association that fails the verification, the association correction process is triggered. The verified transformer substation association is integrated with the transformer substation number, transformer substation name, power supply company name, work group name, township administrative division code, village / community administrative division code, and geographic coordinate information to obtain the four-level association intermediate table.

3. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, Based on the aforementioned power distribution network expertise, the generated knowledge index data includes the following: The knowledge blocks are vector-encoded using a word embedding model, and a knowledge index is constructed based on vector similarity. Dedicated index labels are created for warning threshold items, guiding response script items, and report compilation specification items to obtain the knowledge index data.

4. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, Based on the natural language query text, the intent recognition and keyword extraction are performed to obtain the keyword set, which includes the following: The semantic parsing results are compared with the distribution network business intent database to determine the user's core intent. The distribution network business intent database is divided into data query, early warning verification, and report generation categories. Calculate the intent recognition confidence score. When the intent recognition confidence score is not lower than a preset threshold, the intent recognition result is determined to be valid. When the intent recognition confidence score is lower than the preset threshold, the intent clarification process is triggered to obtain the keyword set.

5. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, The method for generating guided supplementary query text based on the keyword set and the knowledge index data when the integrity check fails is as follows: A predefined set of required parameter categories corresponding to each business intent is provided, including region category, time category, equipment category, indicator category, and work order category. The category labels of each keyword in the keyword set are compared one by one with the required parameter category set corresponding to the current intent to identify the missing required parameter categories and generate a list of missing parameters. Based on the missing parameter list and the guided response script entries in the knowledge index data, the guided supplementary question text is generated. The guided supplementary question text generates a corresponding natural language question statement for each missing required parameter category in the missing parameter list, thus obtaining the guided supplementary question text.

6. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, The structured API request parameter set is categorized by business scenario. Based on the categorization results, the authorized data interface is called to query the standardized data pool, and the data query results are obtained in the following way: The structured API request parameter set is classified according to business scenarios, including low voltage management scenarios, power supply reliability scenarios, power quality scenarios, work order processing scenarios, and project investment analysis scenarios, and the classification results are obtained. Based on the classification results, the authorized data interface is invoked to query the standardized data pool; During the process of calling the authorized data interface to query the standardized data pool, the measured index values ​​are extracted from the data query results, the corresponding index limit thresholds stored in the standardized data pool are retrieved, the index deviation is calculated, and the warning level is divided into three levels: general warning, important warning and emergency warning, based on the preset value range in which the index deviation falls. The data exceeding the limit is automatically generated into structured warning raw data and stored in the management database to obtain the data query results.

7. The AI ​​agent interaction control method for distribution network services according to claim 1, characterized in that, The method of inputting the data query results into the large language model for parsing and processing, and generating natural language response text, also includes: Retrieve the report compilation specification entries and preset report templates from the knowledge index data; The large language model automatically fills in the data query results according to the field structure of the preset report template to generate a draft report; the preset report template includes a daily report template, a weekly report template, a monthly report template, and a quarterly governance analysis template; Based on the data query results and the initial report draft, the natural language response text is generated through parsing and processing using the large language model.

8. An AI-powered interactive management and control system for power distribution network operations, characterized in that: An AI agent interaction control method for power distribution network services, applied to any one of claims 1-7, includes: The data knowledge construction module is used to build a data association system and professional knowledge base for the entire distribution network. It integrates, cleans, associates and standardizes multi-source business data of the distribution network, generates a standardized data pool and a four-level association intermediate table, and integrates distribution network professional knowledge data based on the standardized data pool and the four-level association intermediate table to generate knowledge index data. The natural language parsing module is used to receive natural language query text input by the user, input the natural language query text into the big language model to perform intent recognition and keyword extraction, and generate a keyword set; The parameter completion and conversion module is used to perform integrity verification on the keyword set. When the integrity verification fails, it generates guided supplementary query text based on the keyword set and the knowledge index data. After pushing the guided supplementary query text to the user, it receives the user's supplementary reply text in response to the guided supplementary query text, updates the keyword set based on the supplementary reply text, and iterates until the integrity verification passes. Finally, it converts the verified keyword set into a structured API request parameter set. The data query and response module is used to classify the structured API request parameter set according to business scenarios, call the authorized data interface to query the standardized data pool according to the classification results, obtain the data query results, input the data query results into the large language model for parsing and processing, and generate natural language response text. The five-dimensional interactive control module is used to perform five-dimensional interactive control during the interaction process from step 2 to step 4, and generate full-process control data. The visualization rendering module is used to synchronize the natural language response text and the full-process management data to the visualization application layer, and render and display them in the visualization application layer.

Images