AI big language model intelligent interaction system oriented to power industry

Abstract

The invention relates to the technical field of electric power intelligent interaction, and discloses an AI large language model intelligent interaction system oriented to the electric power industry. The system comprises a multi-source data acquisition module for acquiring power industry multi-dimensional data from a plurality of heterogeneous data sources and integrating the power industry multi-dimensional data into a data warehouse; the semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse to obtain a semantic feature matrix; the industry knowledge graph construction module is used for constructing a power industry knowledge graph based on the semantic feature matrix; the user query analysis module is used for analyzing user query and extracting query feature vectors; the core semantic screening module is used for screening a core semantic sequence based on the semantic feature matrix and the query feature vector; and the intelligent response generation module generates an intelligent response based on the core semantic sequence and the power industry knowledge graph. The system can deeply understand user demands, generates professional and accurate responses, and is suitable for diversified intelligent interaction scenes in the power industry.

Description

Technical Field

[0001] This invention relates to the field of intelligent interaction technology in the power industry, specifically to an AI-based large language model intelligent interaction system for the power industry. Background Technology

[0002] The operation and development of the power industry generates massive and diverse data, originating from multiple heterogeneous data sources and encompassing various dimensions such as power dispatch data, equipment operation data, user electricity consumption data, power grid planning data, and fault maintenance records. The data formats, storage methods, and semantic logic of different data sources vary significantly, forming "data silos" that hinder effective correlation and sharing. Traditional power information systems, when processing this multi-source heterogeneous data, often only perform simple storage and querying, failing to conduct in-depth data integration and mining, thus failing to fully realize the value of the data. At the user interaction level, users' query needs for power industry-related issues are becoming increasingly diverse and professional. These include basic questions from ordinary users such as electricity usage inquiries and payment checks, as well as complex questions from power industry professionals regarding equipment fault diagnosis, grid optimization schemes, and professional technical specifications. However, existing interactive systems face numerous challenges in understanding user query intent. Due to the large number of technical terms and specific contexts within the power industry, traditional natural language processing technologies struggle to accurately parse the semantic information in user queries, easily leading to misunderstandings. Furthermore, these systems lack a systematic overview and structured presentation of power industry expertise. When responding to user queries, they often rely solely on pre-set question-and-answer templates or simple keyword matching, failing to combine industry knowledge for in-depth analysis and reasoning. When user queries involve cross-domain knowledge connections or complex logical reasoning, existing systems struggle to generate accurate, comprehensive, and professional responses, impacting the user experience and potentially leading to inefficient problem-solving, thus failing to meet the demands of efficient operation and service in the power industry.

[0003] With the development of artificial intelligence technology, although some fields have begun to try applying language models for intelligent interaction, there are still significant shortcomings in specialized intelligent interaction systems for the power industry. Existing models lack adaptation to the specific data and knowledge of the power industry. When dealing with professional issues in the power sector, the accuracy of semantic understanding and the professionalism of responses need to be improved, making it difficult to cope with the complex and professional interaction scenarios in the power industry. Summary of the Invention

[0004] The purpose of this invention is to address the problem that existing models lack adaptation to specific data and knowledge in the power industry. When dealing with professional issues in the power field, the accuracy of semantic understanding and the professionalism of the response need to be improved, making it difficult to cope with the complex and professional interaction scenarios in the power industry.

[0005] To address the above problems, this invention provides an AI-based large language model intelligent interaction system for the power industry, the system comprising: The multi-source data acquisition module obtains multi-dimensional data related to the power industry from multiple heterogeneous data sources and integrates the multi-dimensional data to obtain an integrated data warehouse. The semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse to obtain a semantic feature matrix. The industry knowledge graph construction module builds a knowledge graph for the power industry based on a semantic feature matrix. The user query parsing module parses the query input by the user and extracts the query feature vector; The core semantic filtering module filters core semantic sequences based on the semantic feature matrix and query feature vector; The intelligent response generation module generates intelligent responses based on core semantic sequences and a knowledge graph of the power industry.

[0006] In the preferred mode, the multi-source data acquisition module acquires multi-dimensional data related to the power industry, including: The power grid operation data includes voltage, current, frequency, power, power grid topology, dispatch instructions, historical operation records, and fault event logs; The user interaction data includes user inquiry records, complaint feedback, and electricity consumption behavior analysis data. The equipment status data includes real-time monitoring parameters, maintenance records, and life prediction data for transformers, circuit breakers, transmission lines, and power distribution equipment. The energy trading data includes medium- and long-term trading data, spot trading data, and ancillary services market data.

[0007] In the preferred approach, the multi-source data acquisition module integrates multi-dimensional data, specifically including: The multidimensional data is preprocessed, including data cleaning, standardization, and semantic mapping. The data cleaning process includes identifying and deleting duplicate records, filling missing values ​​with linear interpolation of data from adjacent time points, correcting outliers by means of a sliding window, and retaining copies of the original data during the data cleaning process. The standardization process includes converting all electrical parameters to the International System of Units (SI), converting timestamps to UTC format with millisecond precision, encoding text data into UTF-8 format, and mapping enumerated data to numerical codes. The standardization process also establishes a data conversion rule base to record the conversion logic of each data field. The semantic mapping includes transforming standardized data into semantic entities for the power industry. Based on a predefined power semantic ontology library containing four categories of semantic tags: equipment, operation, user, and market, semantic mapping is performed on equipment status data, user interaction data, power grid operation data, and energy trading data to obtain power semantic data in various dimensions. An integrated data warehouse is constructed based on power semantic data from all dimensions. The integrated data warehouse adopts a hybrid architecture of graph database and time series database. Equipment status and power grid topology relationship data are stored in the graph database, and load curves and voltage fluctuation time series data are stored in the time series database. The integrated data warehouse establishes a global indexing mechanism to achieve data integration.

[0008] In the preferred approach, the semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse, specifically including: Data in the data warehouse is separated and integrated according to dimensions, namely real-time monitoring data and historical operation logs of power grid operation dimension, consultation records and electricity consumption behavior data of user interaction dimension, sensor monitoring values ​​and maintenance reports of equipment status dimension, and market transaction records and price curves of energy trading dimension. After separating the data for each dimension, feature extraction is performed independently to obtain semantic feature vectors for each dimension; In terms of power grid operation, third-order statistics are calculated for load data to extract features, over-limit correlation features are extracted for voltage data, and fault distribution and clustering features are extracted from fault logs. Sliding window analysis is used for feature extraction. The user interaction dimension extracts consultation text features and sentiment features through NLP technology, extracts electricity fluctuation and proportion features from electricity behavior data, and extracts regional hot issue distribution features by combining spatiotemporal information. The equipment status dimension extracts temperature rise and deviation features from temperature data, and extracts operation frequency deviation features from mechanical operation data. Combined with maintenance records, it extracts lifespan and failure probability derived features. Feature extraction is introduced into the status assessment model. The correlation features between electricity price fluctuations and transaction volume are extracted from the energy trading dimension; the balance deviation features are extracted from supply and demand data; and cross-dimensional features are extracted by combining external factors. All semantic feature vectors of all dimensions are organized into a matrix structure according to the dimension number. The matrix rows represent time series and the columns represent feature dimensions, forming a spatiotemporally related semantic feature matrix, which ultimately enables semantic analysis and feature extraction of the integrated data warehouse.

[0009] In the preferred approach, the industry knowledge graph construction module constructs a power industry knowledge graph based on a semantic feature matrix, specifically including: Select one dimension from all dimensions of the semantic feature matrix and obtain the semantic feature vector of that dimension. The semantic feature vector contains feature values ​​and their spatiotemporal attributes. The feature correlation degree between each semantic feature in the semantic feature vector of the selected dimension is calculated using an improved mutual information algorithm. The algorithm considers the temporal characteristics and spatial distribution of power data, introduces a time decay factor and a spatial weight coefficient, assigns higher weight to recent data, and determines the spatial weight of key nodes according to the power grid topology. Based on the calculated feature correlation matrix, a community detection algorithm is used to identify feature communities, and feature pairs with correlation higher than a preset threshold θ are connected by edges to form an initial graph structure. The community division is adjusted by a modularity optimization algorithm to make the features within the same community closely related and the features between different communities less related. For each community, a local semantic association graph is generated, where nodes represent features and edge weights represent association strength. Derived nodes are added to the local graph to represent the overall characteristics of the community, ultimately forming a knowledge graph for the power industry.

[0010] In the preferred method, the user query parsing module parses the user-input query, specifically including: The user-input query is segmented and vectorized. The segmentation stage retains compound words in the power field, and the vectorization process adopts a domain-adaptive word embedding model. The model is pre-trained on a general corpus and then fine-tuned using power industry documents to obtain an initial query vector. The initial query vector is subjected to deep semantic encoding, and a hierarchical attention mechanism is adopted. The first layer of attention focuses on the power entities in the query, and the second layer of attention analyzes the relationships between entities, finally obtaining the query feature vector.

[0011] In the preferred approach, the core semantic filtering module filters core semantic sequences based on the semantic feature matrix and query feature vector, specifically including: Calculate the semantic dimension entropy of each semantic feature in the semantic feature matrix. In the knowledge graph of the power industry, all feature correlation degrees corresponding to each semantic feature are extracted, and the correlation degree is calculated by a spatiotemporal weighted mutual information algorithm that introduces time decay factor and spatial weight coefficient. Node connections are established between feature pairs with a correlation degree higher than a preset threshold θ to form a structured correlation relationship between features; The core semantic entropy of each semantic feature is determined by a weighted summation method. The core semantic entropy = semantic dimension entropy × 0.4 + average feature correlation degree in the knowledge graph × 0.6, where the weights are set based on the inherent importance of dimensions and cross-feature correlation value in the power data. Core semantic features are selected based on the core semantic entropy of all semantic features; A temporal correlation strategy is adopted to align the selected core semantic features by timestamps to form a multidimensional time series, namely the core semantic sequence.

[0012] In the preferred approach, the intelligent response generation module generates intelligent responses based on core semantic sequences and the power industry knowledge graph, specifically including: The core semantic sequence is matched with the knowledge graph of the power industry. A bidirectional traversal strategy is used to align each feature node in the core semantic sequence with the corresponding node in the knowledge graph. The forward search starts from the feature node to find the influencing factors of the upper-level system, and the reverse search finds the lower-level consequence nodes. The matching result forms a problem analysis tree with the query topic as the root node and branches containing the influencing factor chain and the consequence chain. Based on the problem analysis tree, an initial response draft is generated. Using a structured narrative framework, the problem analysis tree is converted into a three-part content: phenomenon description, cause analysis, and suggested measures. The draft retains data traceability markers, and each conclusion corresponds to a specific feature number and knowledge graph node ID. The initial response draft is optimized and formatted. The response optimization implements a four-layer filtering mechanism. The first layer is logical verification to detect contradictory statements and trigger a conflict resolution process to re-examine the relevance of features. The second layer is terminology standardization to convert technical descriptions into standard power terms. The third layer is context adaptation to adjust the expression according to the user's identity. The fourth layer is language fluency processing to reorganize complex sentence structures. The response formatting uses a dynamic template engine, which automatically matches templates from a template library containing accident analysis reports, operation recommendations, and user notification letters based on the query type. The system fills the reserved data slots in the templates with specific values, and the output is a multimodal response with attached data visualizations. The visualization elements are linked with the key data points in the text, resulting in an intelligent response.

[0013] In the preferred approach, a risk assessment module is also included; The risk assessment module is used to calculate a global risk index based on a semantic feature matrix. The global risk index integrates load mutation rate, equipment health index, and meteorological disaster probability parameters. The load mutation rate is derived from the feature extraction results of the semantic feature extraction module on the power grid operation dimension data; The equipment health index is derived from the feature extraction and fusion results of the semantic feature extraction module on the equipment status dimension data; The probability of meteorological disasters is derived from real-time data collected by the status monitoring module and processed by the semantic feature extraction module through cross-dimensional feature processing. When the risk assessment module generates a risk assessment value, it first obtains the key parameters mentioned above from the semantic feature extraction module, and then inputs the parameters into a preset risk assessment model that conforms to the risk assessment standards of the power industry. Combining the degree of influence of different parameters in the risk assessment of power grid operation, the module generates a risk assessment value of 0-100 points through the internal calculation logic of the model. When the risk assessment value exceeds 70, add a level 3 warning prompt at the beginning of the smart response; When the risk assessment value exceeds 85, emergency operation guidelines will be added to the intelligent response. Based on the risk assessment value, risk control logic is incorporated into the intelligent response recommendations. Low risk corresponds to daily inspection recommendations, medium risk corresponds to 72-hour special inspection requirements, and high risk corresponds to equipment shutdown decision support.

[0014] In the preferred embodiment, a status monitoring module is also included; The status monitoring module is used to monitor the power grid operation status in real time and generate status monitoring data. Through sensors, smart meters, fault recording devices and other equipment deployed at key nodes of the power grid, dynamic data such as power grid load, equipment temperature and environmental weather are collected in real time. The status monitoring module dynamically adjusts the data acquisition frequency according to the importance of the data type, acquiring temperature data at a second-level frequency and meteorological data at a minute-level frequency. After data collection, the status monitoring module performs preliminary verification of the data to remove obvious outliers, and inputs the verified data into the multi-source data acquisition module to compare it with historical data to identify abnormal trends. The identified abnormal trend data is then marked as high-risk data in the integrated data warehouse.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This AI-powered large-scale language model intelligent interaction system for the power industry acquires multi-dimensional power industry-related data from multiple heterogeneous data sources through a multi-source data acquisition module. It then integrates this data to form an integrated data warehouse, breaking down the "data silos" formed by multi-source heterogeneous data in traditional power information systems. After integration, data of different formats and storage methods can be effectively correlated and shared, providing a unified data foundation for subsequent semantic analysis and knowledge utilization. This allows the system to conduct further processing based on more comprehensive data. The semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse to obtain a semantic feature matrix, which delves into the semantic information contained in the data. Compared with the simple keyword matching of traditional systems, the semantic feature matrix can more accurately capture the inherent logic and semantic relationships of the data. This allows the system's understanding of power industry data to go beyond the surface and touch upon the deeper meaning of the data, providing semantic support for accurately understanding user queries and generating professional responses. The industry knowledge graph construction module builds a power industry knowledge graph based on a semantic feature matrix, presenting professional knowledge of the power industry in a structured form. The knowledge graph covers various relationships among power equipment, power grid structure, operating rules, and fault types, providing the system with rich industry background knowledge. When processing user queries, the system can rely on the related information in the knowledge graph to achieve cross-domain knowledge fusion and retrieval, avoiding the problem of incomplete or inaccurate responses caused by the lack of professional knowledge support in traditional systems. The user query parsing module analyzes user-input queries and extracts query feature vectors, enabling it to more accurately capture users' true needs. Whether users use technical terms or everyday expressions, this module can transform user intent into quantifiable feature vectors through semantic analysis of the query text. This reduces misunderstandings caused by semantic ambiguity or differences in expression, ensuring the system accurately grasps the information users want to obtain. The core semantic filtering module filters core semantic sequences based on semantic feature matrices and query feature vectors, enabling it to focus on key information from complex queries and massive amounts of data. By comparing query features with data semantic features, this module filters out the core semantic content most relevant to user needs, avoiding the impact of information overload on response quality and allowing the system to concentrate resources on analyzing and processing key information during the processing phase. The intelligent response generation module generates intelligent responses based on core semantic sequences and a power industry knowledge graph, achieving a deep integration of semantic understanding and industry knowledge. The generated responses not only accurately address user queries but also supplement and expand upon them with professional knowledge from the knowledge graph, providing users with more comprehensive and professional answers. This response method overcomes the limitations of traditional template-based replies, flexibly generating personalized responses based on different query content and scenarios, thus enhancing the user experience in intelligent interactions within the power industry. Attached Figure Description

[0016] Figure 1 This is a timing diagram of the AI ​​large language model intelligent interaction system for the power industry according to the present invention. Figure 2 A flowchart illustrating the multidimensional data types acquired by the multi-source data acquisition module; Figure 3 A flowchart for data integration from a multi-source data acquisition module; Figure 4 A flowchart illustrating the operation of the core semantic filtering module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides an AI-powered large language model intelligent interaction system for the power industry, the system comprising: The system comprises four modules: a multi-source data acquisition module, a semantic feature extraction module, an industry knowledge graph construction module, a user query parsing module, a core semantic filtering module, and an intelligent response generation module. The multi-source data acquisition module acquires multi-dimensional data related to the power industry from heterogeneous data sources such as power grid operation data, user interaction data, equipment status data, and energy trading data, and integrates this data to form an integrated data warehouse. The semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse, generating a semantic feature matrix. The industry knowledge graph construction module constructs a power industry knowledge graph based on the semantic feature matrix, forming a semantic association network for the power industry. The user query parsing module segments, vectorizes, and performs deep semantic encoding on user-input queries, extracting query feature vectors. The core semantic filtering module combines the semantic feature matrix and query feature vectors to filter out core semantic sequences. The intelligent response generation module matches the core semantic sequences with the power industry knowledge graph to generate an intelligent response, ultimately outputting optimized, formatted response content.

[0019] Example 1: See Figure 2 The multi-source data acquisition module is the core of the system's data input, responsible for acquiring multi-dimensional data from multiple heterogeneous data sources in the power industry and integrating it to form a structured integrated data warehouse. The types of data collected by this module mainly include power grid operation data, user interaction data, equipment status data, and energy trading data, covering the entire process of power system production, transmission, consumption, and market transactions.

[0020] Power grid operation data is core information for real-time monitoring of the power system, including electrical parameters such as voltage, current, frequency, and power, as well as power grid topology, dispatch instructions, historical operation records, and fault event logs. This data typically originates from real-time monitoring equipment such as SCADA systems, EMS systems, and PMU phasor measurement units, as well as operation reports from the power grid dispatch center. Time synchronization must be considered during data acquisition to ensure that data from different monitoring points are aligned in the time dimension, avoiding analysis errors caused by time deviations. For example, voltage monitoring data needs to be matched with load data at the same time to accurately analyze the power grid's operating status.

[0021] User interaction data primarily comes from power companies' customer service systems, mobile applications, and social media channels, including user inquiry records, complaint feedback, and electricity consumption behavior analysis. This type of data is mainly unstructured text, such as electricity-related questions and fault reports submitted by users through online customer service, or electricity experience feedback posted on social media. When collecting data, attention must be paid to semantic diversity; for example, the same problem may appear in different ways, such as "abnormal electricity bill" or "sudden increase in electricity consumption," which may refer to the same type of problem. In addition, user electricity consumption behavior data includes historical electricity consumption, peak and off-peak electricity consumption habits, and smart meter readings, which can be used to analyze user electricity consumption characteristics.

[0022] Equipment status data encompasses the operational health of various devices in the power system, including real-time monitoring parameters, maintenance records, and lifespan prediction data for transformers, circuit breakers, transmission lines, and distribution equipment. Examples include transformer oil temperature, winding temperature, and partial discharge data; circuit breaker opening and closing frequency and mechanical characteristic data; and transmission line current carrying capacity and sag data. This type of data typically originates from equipment monitoring sensors, online monitoring systems, and maintenance reports. When collecting data, it is crucial to pay attention to the unique identifier of the equipment to ensure that historical and real-time data for the same equipment can be correlated, for example, by matching data using the equipment ID or asset number.

[0023] Energy trading data encompasses electricity market transaction records, price fluctuations, and supply-demand balance information, including data from medium- and long-term transactions, spot transactions, and ancillary services markets. Examples include clearing prices, traded volumes, and market supply-demand forecasts for different time periods. This type of data typically originates from publicly available data from electricity trading platforms, market operators, or internal trading systems. When collecting data, it is crucial to pay attention to changes in market rules; for instance, electricity pricing mechanisms may differ across regions, necessitating data categorization and storage according to region.

[0024] The status monitoring module serves as the system's dynamic data input source, responsible for real-time monitoring of the power grid's operational status, generating status monitoring data, and feeding it back to the multi-source data acquisition module to update and integrate the data warehouse. This module collects dynamic data such as power grid load, equipment temperature, and environmental weather conditions in real time through devices deployed at key nodes of the power grid, including sensors, smart meters, and fault recording devices. For example, real-time load data for transmission lines can reflect whether the power grid's transmission capacity is approaching its limit, while transformer oil temperature data can reflect whether the equipment's heat dissipation is normal. The status monitoring module also pays attention to the impact of environmental factors on power grid operation; for example, meteorological data such as temperature, humidity, and wind speed may affect the current carrying capacity of transmission lines or the heat dissipation efficiency of equipment.

[0025] The frequency of status monitoring data collection is dynamically adjusted based on the importance of the data type. For example, temperature data for critical equipment may be collected at a rate of seconds, while meteorological data may be collected at a rate of minutes. After data collection, initial verification is performed to remove obvious outliers, such as sudden changes in data caused by sensor malfunctions. The verified data is then input into a multi-source data acquisition module and compared with historical data to identify abnormal trends. For example, a continuous rise in transformer oil temperature may indicate equipment overload and needs to be marked as high-risk data in the integrated data warehouse.

[0026] The multi-source data acquisition module performs preliminary classification and storage of the collected multi-dimensional data, establishing indexes according to data type, timestamp, data source, and other dimensions to facilitate subsequent semantic analysis and feature extraction. For example, power grid operation data is stored by voltage level and region, user interaction data is stored by question type and time sequence, and equipment status data is stored by equipment type and monitoring parameters. The data storage adopts a hierarchical structure: the raw data layer retains the unprocessed raw records, the intermediate data layer stores cleaned and standardized data, and the semantic data layer stores structured data after semantic mapping.

[0027] During data integration, the multi-source data acquisition module is also responsible for data version management and historical traceability. For example, when equipment status data changes due to maintenance or replacement, the system records the data change history to ensure that the evolution of the data can be traced during analysis. In addition, data integration also needs to consider the time alignment issues of different data sources. For example, the timestamps of user complaint data may differ from the timestamps of power grid fault data, requiring matching through time interpolation or event correlation algorithms.

[0028] The collaborative work between the status monitoring module and the multi-source data acquisition module ensures the dynamic updating of the integrated data warehouse. For example, when the status monitoring module detects a sudden increase in load on a transmission line, the multi-source data acquisition module will correlate the line's historical load data and equipment status data to analyze whether there is a potential overload risk. If the risk is high, the system will mark the line as a key focus in the integrated data warehouse and prioritize the processing of related data in subsequent semantic analysis.

[0029] Through the collaboration of multi-source data acquisition and status monitoring modules, the system can continuously acquire real-time and historical data from the power industry, forming a comprehensive data warehouse covering grid operation, user interaction, equipment status, and energy trading. This data foundation provides comprehensive data support for subsequent semantic feature extraction, knowledge graph construction, and intelligent interaction.

[0030] Example 2: See Figure 3After the multi-source data acquisition module completes data acquisition, it enters the data integration stage. This stage includes three core operations: data cleaning, standardization, and semantic mapping. Data cleaning addresses quality issues in the original multidimensional data, including identifying and deleting duplicate records, such as multiple identical monitoring data points generated by the same device at the same timestamp; filling missing values, such as gaps in power grid load data caused by transmission interruptions, which are filled using linear interpolation of data from adjacent time points; and correcting outliers, such as abrupt changes in current monitoring data caused by sensor malfunctions, which are smoothed using sliding window mean filtering. A copy of the original data must be retained during the data cleaning process to trace the impact of the cleaning operation on the data.

[0031] Standardization focuses on unifying data formats and units. Data sources in the power industry are diverse; for example, voltage data may exist in volts, kilovolts, or per-unit values, and timestamps may use local time zones or UTC format. The standardization phase converts all electrical parameters to the International System of Units (SI): voltage is standardized to kilovolts, current to amperes, and power to megawatts; timestamps are uniformly converted to UTC format and accurate to the millisecond level; text data is uniformly encoded in UTF-8 format; and enumerated data, such as "normal / warning / fault" in equipment status labels, is uniformly mapped to numerical codes 0 / 1 / 2. The standardization process establishes a data conversion rule base, recording the conversion logic of each data field for easy data backtracking.

[0032] Semantic mapping is a crucial step in the integration process, transforming standardized data into semantic entities for the power industry. This process relies on a predefined power semantic ontology library, containing four main categories of semantic tags: equipment, operation, users, and market. Taking equipment status data as an example: transformer oil temperature monitoring values ​​are mapped to the semantic entity "equipment temperature," associated with equipment ID, temperature unit, and threshold range attributes; circuit breaker operation counts are mapped to the semantic entity "mechanical life," associated with equipment type and rated operation count attributes. In user interaction data, text like "electricity bill questions" are mapped to the semantic entity "electricity bill consultation," associated with user ID, electricity consumption month, and question type attributes. Semantic mapping employs a combination of a rule engine and lightweight machine learning. The rule engine handles the direct mapping of structured data, such as field names in a SCADA system directly corresponding to semantic tags; the machine learning model handles unstructured text. Operational semantic tags correspond to the semantic mapping of power grid operation data. For example, 110kV line voltage over-limit data collected by the SCADA system is mapped to the semantic entity "operational parameter over-limit," associated with line ID, over-limit value, duration, and threshold range attributes. Load adjustment command data issued by the dispatch center is mapped to the semantic entity "operational dispatch," associated with command number, execution time, dispatch object, and command content attributes. Line tripping records in fault events are mapped to the semantic entity "operational fault," associated with faulty line ID, tripping time, fault type, and reclosing status attributes. Market semantic tags correspond to the semantic mapping of energy trading data. For example, daily transaction data from the electricity spot trading platform is mapped to the semantic entity "market transaction transaction," associated with transaction number, transaction date, buyer and seller IDs, transaction volume, and transaction price attributes. AGC frequency regulation service data in the ancillary service market is mapped to the semantic entity "market ancillary service," associated with service provider ID, service type, service duration, regulation performance indicators, and settlement amount attributes. The signing data of medium- and long-term contracts is mapped to the semantic entity "market contract," associated with contract number, contracting parties, contract volume, contract price, and performance period attributes.

[0033] After semantic mapping is completed, all dimensions of power semantic data are stored in an entity-attribute structure to form an integrated data warehouse. This warehouse adopts a hybrid architecture of graph database and time-series database: relational data such as equipment status and power grid topology are stored in the graph database to facilitate the expression of the connection relationship between equipment; time-series data such as load curves and voltage fluctuations are stored in the time-series database to support efficient time range queries; the warehouse establishes a global indexing mechanism, for example, by using the equipment ID, its operating data, maintenance records and user repair records can be associated.

[0034] The semantic feature extraction module performs layered processing on the integrated data warehouse. First, it separates the data by dimension: real-time monitoring data and historical operation logs are extracted from the power grid operation dimension; consultation records and electricity consumption behavior data are extracted from the user interaction dimension; sensor monitoring values ​​and maintenance reports are extracted from the equipment status dimension; and market transaction records and price curves are extracted from the energy trading dimension.

[0035] Feature extraction is performed independently for each dimension: The power grid operation dimension focuses on volatility and anomaly patterns. Third-order statistics are calculated from load data to extract daily average volatility and peak-to-valley difference characteristics; voltage data is used to extract the number of over-limit occurrences and duration characteristics; and fault logs are used to extract fault type distribution and spatiotemporal clustering characteristics. Feature extraction employs sliding window analysis, for example, calculating load change gradients with a 15-minute window to identify abrupt change points.

[0036] The user interaction dimension focuses on text semantics and behavioral patterns. NLP technology is used to extract keyword TF-IDF weights and sentiment polarity values ​​from consultation texts; monthly electricity consumption variance and peak-valley electricity consumption ratio features are extracted from electricity consumption behavior data; and regional hotspot issue distribution characteristics are extracted by combining spatiotemporal information, such as generating regional issue tags when a community reports concentrated voltage instability.

[0037] The equipment condition dimension focuses on degradation trends and health indicators. Temperature data is used to extract temperature rise rate and steady-state deviation characteristics; mechanical operation data is used to extract operation frequency deviation characteristics; and maintenance records are combined to extract derived features such as remaining life prediction and failure probability. Feature extraction is incorporated into condition assessment models; for example, transformer insulation condition is analyzed using multivariate analysis of oil chromatography and temperature data to generate a health score.

[0038] The energy trading dimension emphasizes market dynamics and risk indicators. It extracts correlation characteristics of electricity price volatility and trading volume; it extracts balance deviation characteristics from supply and demand data; and it combines external factors to extract cross-dimensional characteristics such as weather impact coefficient and policy sensitivity.

[0039] The semantic feature vectors extracted from each dimension contain original feature values ​​and metadata. The metadata records information such as feature source, calculation time window, and confidence level. For example, the load volatility feature is labeled with data from the SCADA system, calculated over a 24-hour period, and has a 95% confidence interval. All semantic feature vectors from all dimensions are organized into a matrix structure according to dimension numbering. Rows represent time series data, such as one feature vector every 15 minutes, while columns represent feature dimensions, such as load volatility and voltage limit exceedances, forming a spatiotemporally correlated semantic feature matrix. This matrix serves as core intermediate data for the system, supporting subsequent knowledge graph construction and query analysis.

[0040] A version management mechanism is established for the semantic feature matrix. Each data warehouse update triggers incremental calculation of the feature matrix, performing feature extraction only on newly added data, reducing computational overhead. Matrix storage employs columnar compression to optimize performance in high-frequency access scenarios, such as prioritizing the loading of user interaction-related feature columns during real-time queries. The matrix also includes feature correlation annotations, such as the correlation coefficient between load volatility and equipment temperature features, providing preprocessing information for knowledge graph construction.

[0041] The entire implementation process forms a closed loop: the semantic data warehouse output from the data integration phase is used to generate a dynamically updated semantic feature matrix through feature extraction, and the statistical analysis results of the feature matrix can be fed back to the data cleaning rule base. For example, when the temperature characteristics of a certain type of equipment frequently show outliers, the data acquisition module is automatically triggered to perform a verification process on the sensor of that equipment, thereby achieving self-optimization of data quality.

[0042] Example 3: The industry knowledge graph construction module generates a power industry knowledge graph based on a semantic feature matrix. This process is implemented in a hierarchical and progressive manner. First, a target dimension is selected from the semantic feature matrix, such as the power grid operation dimension, and all semantic feature vectors under this dimension are extracted. Each feature vector contains feature values ​​and their spatiotemporal attributes. For example, the load characteristics of a substation include hourly load rate, daily maximum load, and other numerical values, as well as attribute information such as collection time and substation number. The relationship between feature vectors is quantified through feature association. The association degree is calculated using an improved mutual information algorithm, taking into account the temporal characteristics and spatial distribution of power data. For two features... and Its spatiotemporal weighted mutual information The calculation formula is: in: Indicates the total number of time windows. Indicates the total number of spatial regions. It is the time decay factor. It is the spatial weighting coefficient. Representation of features In the time window The probability distribution, It is a joint probability distribution. In this formula... More recent data are given higher weight using an exponential decay method. Determine the spatial weights of key nodes based on the power grid topology.

[0043] After calculating the feature correlation matrix, a community detection algorithm is used to identify feature communities. Those with correlation scores above a threshold are considered... The features are connected by edges to form an initial graph structure. Then, a modularity optimization algorithm is used to adjust the community partitioning, so that features within the same community are closely related, while features between different communities are less related. For example, the power grid operation dimension may form load fluctuation communities, voltage stability communities, and fault mode communities. Each community generates a local semantic association graph, where nodes represent features and edge weights represent association strength. Derived nodes are added to the local graph to represent the overall characteristics of the community. For example, the derived nodes of the load fluctuation community include aggregate indicators such as peak-valley difference rate and fluctuation frequency.

[0044] The user query parsing module employs a multi-stage processing flow. Input queries first undergo electrical terminology identification, matching specialized vocabulary using a pre-built electrical terminology dictionary. Term identification utilizes a maximum matching algorithm, combined with contextual disambiguation; for example, in the phrase "transformer noise is loud," "noise" specifically refers to "equipment noise" in the electrical context. The word segmentation stage retains compound words in the electrical field, such as "reactive power compensation" and "relay protection," as whole segments. Vectorization processing employs a domain-adaptive word embedding model, pre-trained on a general corpus and fine-tuned using electrical industry documents. This model can distinguish the ambiguity of electrical terms, generating initial query vectors that include both word-level and phrase-level embeddings.

[0045] The adaptive word embedding model is the core tool for vectorized processing in the user query parsing module. Its design logic aligns with the semantic association requirements of the power industry knowledge graph and the goal of identifying professional terms in queries. Based on a pre-trained basic word vector model, this model is fine-tuned by incorporating power industry-specific documents (including standards and regulations such as the "Guidelines for the Safety and Stability of Power Systems," power grid operation logs describing SCADA system monitoring data, equipment maintenance reports for transformer / circuit breaker fault analysis, and user interaction history for electricity billing inquiries / fault reporting). Through incremental training, it accurately captures the unique semantics of power industry terms—for example, distinguishing the meaning of "load" in the power context as "power consumption," avoiding ambiguity in general contexts. Simultaneously, it optimizes the vector representation of compound words such as "reactive power compensation" and "relay protection" as overall semantic units, preventing the loss of semantics during word segmentation. During vectorization, the model receives a user query word segmentation sequence (such as "analysis of the cause of 220kV line tripping") and outputs an initial query vector containing power industry semantics. Its vector space structure is adapted to the subsequent semantic feature matrix dimension design, and it can efficiently calculate similarity with feature vectors of equipment, operation, user, and market, providing a "semantic alignment" basis for core semantic filtering. At the same time, this output vector serves as the input of a hierarchical attention mechanism, supporting the first layer of attention to identify power entities such as "220kV" and "line tripping", and the second layer of attention to parse the relationship between entities (such as the voltage level-fault type constraint relationship between "220kV" and "line tripping"), ultimately achieving accurate conversion of user query intent from natural language to structured feature vectors.

[0046] The deep semantic encoding stage employs a hierarchical attention mechanism. The first layer of attention focuses on power entities in the query, such as identifying the voltage level and fault type in "220kV line tripping." The second layer of attention analyzes the relationships between entities, such as determining the constraints between time conditions and electrical parameters in "low voltage during summer peak electricity consumption." The encoder outputs a query feature vector, whose dimensions are consistent with the semantic feature matrix, facilitating subsequent core semantic matching. The encoding process preserves the attention weight distribution to explain the system's grasp of the query's key points.

[0047] Cross-dimensional fusion of knowledge graphs is achieved through hyperedges. When features from different dimensions are related in terms of business, cross-community hyperedge connections are established. For example, the user-dimensional feature of "increased voltage complaints" and the power grid-dimensional feature of "transformer overload" are associated through a hyperedge, with the edge weight determined by the spatial correlation between the geographical distribution of complaints and the transformer load rate. The hyperedge carries dimension transformation rules, such as converting the user-side description of "voltage drop" into the power grid-side technical term "transient voltage dip".

[0048] The knowledge graph's dynamic update mechanism is synchronized with the semantic feature matrix. When new data causes changes in feature distribution, the correlation degree is recalculated and the graph structure is adjusted. For example, if the addition of a new distributed power source causes a change in the power flow direction of a local power grid, the correlation relationships of the corresponding nodes are updated accordingly. The graph version management records every major topology change, supporting historical state backtracking and evolution analysis.

[0049] The context understanding module in query parsing maintains dialogue state tracking. During multi-turn interactions, it records mentioned entities and intents through dialogue state vectors. These state vectors are linked in real-time with entity nodes in the knowledge graph, automatically expanding the query scope and associating with meteorological impact nodes in the graph when the user adds information. The power industry knowledge graph ultimately presents a multi-level structure. The top layer is the inter-dimensional relationship network, the middle layer is the feature communities for each dimension, and the bottom layer is the specific feature nodes and their attributes. This structure supports both macro-level power system correlation analysis and fine-grained queries down to specific equipment or parameters. The graph's visualization interface is customized to meet the needs of different roles; for example, dispatchers focus on the topology of the power grid operation dimension, while customer service personnel focus on the user question classification network. The collaborative work between user query parsing and the knowledge graph is reflected in the semantic mapping process. The parsed query feature vectors establish many-to-many mappings with graph nodes, matching related nodes through vector similarity. The system selects the optimal path based on the context. This mechanism can handle common synonyms and implicit associations in the power field, improving the understanding of complex queries.

[0050] Example 4: See Figure 4The core semantic filtering module filters core semantic sequences based on the semantic feature matrix and query feature vectors. This process is implemented through a multi-level evaluation mechanism. Taking the actual query "Analyze the causes of recent 220kV main transformer load anomalies" from a regional power grid dispatch center as an example, the specific implementation process is illustrated. The system first locates relevant dimensions in the semantic feature matrix; in this example, this involves the power grid operation dimension and the equipment status dimension. The recent semantic feature data stored in the matrix is ​​shown in Table 1.

[0051] Table 1: Examples of semantic features related to 220kV main transformer Semantic dimension entropy calculation considers the volatility of features over time and the spatial dispersion of their distribution. The load factor feature F101 has a high entropy value, reflecting significant recent load fluctuations in the main transformer; the ambient temperature feature F106 has a low entropy value, indicating relatively stable meteorological conditions. In the power industry knowledge graph, these feature nodes have established various relationships: F101 load factor and F104 regional load have a supply correlation of 0.85; F102 oil temperature and F106 ambient temperature have an environmental impact correlation of 0.62; and F103 cooling efficiency and F101 load factor have an equipment performance correlation of 0.78.

[0052] The query feature vector analysis revealed three core elements: voltage level 220kV, equipment type main transformer, and problem nature load anomaly. The load rate feature F101 was found to be directly relevant in the semantic feature matrix. Its core semantic entropy consists of three parts: its own dimensional entropy (0.73), its match degree with the query (0.91), and its average correlation degree in the knowledge graph (0.81). The core semantic entropy was determined using a weighted summation method, calculated as: "Core semantic entropy = Semantic dimensional entropy × 0.4 + Average correlation degree of features in the knowledge graph × 0.6", where the weights are set based on the inherent importance of dimensions in power data and the value of cross-feature associations. Taking the load rate feature of F101 as an example, its semantic dimension entropy is 0.73. In the knowledge graph, its average correlation with the load of F104 region (correlation degree 0.85) and the cooling efficiency of F103 (correlation degree 0.78) is (0.85+0.78) / 2=0.815. Substituting into the formula, we can get the core semantic entropy = 0.73×0.4+0.815×0.6=0.292+0.489=0.781, which is consistent with the core semantic entropy evaluation results mentioned above. Taking the oil temperature feature F102 as an example, its semantic dimension entropy is 0.68. In the knowledge graph, its average correlation with F101 load rate (correlation 0.72) and F106 ambient temperature (correlation 0.62) is (0.72+0.62) / 2=0.67. The calculated core semantic entropy is 0.68×0.4+0.67×0.6=0.272+0.402=0.674. Since this value is higher than the set threshold of 0.65, it is included in the core semantic feature set. The system sets a dynamic threshold to filter out the core semantic feature set {F101,F102,F103,F104}, excluding F105 voltage deviation and F106 ambient temperature, which have low correlation.

[0053] The core semantic sequence generation employs a temporal correlation strategy. Selected features are aligned by timestamps to form a multi-dimensional time series. For example, it was found that when the load rate of F101 reached its peak of 0.93 on July 3rd, the oil temperature of F102 simultaneously rose to 82℃, while the cooling efficiency of F103 dropped to 0.79. This temporal coupling relationship is encoded as the sequence segment "high load → temperature rise → cooling efficiency decrease". Simultaneously, combined with spatial correlation, the load of area F104 shows that the electricity demand in area A-05 increased by 12% during the same period, supplementing the sequence with the prefix "area load increase → main transformer load increase".

[0054] The risk assessment module's involvement is reflected in feature selection. When the F102 oil temperature is detected to exceed the warning threshold, the priority of the oil temperature feature in the sequence is automatically increased. The original sequence "load → temperature → cooling" is adjusted to "temperature → load → cooling" to highlight the risk of equipment overheating. The equipment technical parameters stored in the knowledge graph play a role here. For example, if the rated oil temperature of the main transformer is 85℃, and the current temperature of 82℃ triggers a level-two warning, the core semantic entropy of the temperature feature is increased by a risk correction value of 0.15.

[0055] For complex query scenarios, such as "analysis of multiple line tripping after thunderstorms," ​​multi-dimensional cross-validation is used for core semantic filtering. Trip event features are filtered at the power grid operation dimension, thunderstorm intensity features at the meteorological dimension, and insulation performance features at the equipment dimension. These features are connected via hyperedges in a knowledge graph, forming a composite semantic sequence: "Thunderstorm activity → Insulator flashover → Protection action → Line tripping." The core semantic entropy of each node in the sequence is used to comprehensively calculate its importance within its dimension, the strength of its cross-dimensional association, and its semantic distance from the query.

[0056] The system maintains a historical database of core semantic sequences. When processing a new query, it retrieves similar historical sequences as a reference. It automatically calls upon relevant sequence patterns for matching. This case-based reasoning mechanism can quickly identify the core feature combinations of typical problem scenarios.

[0057] The dynamic updating of feature data in the table affects the filtering results. When the status monitoring module reports a new oil temperature of 85℃, the F102 feature value is updated and the semantic dimension entropy is recalculated. The increase in entropy triggers real-time adjustment of the core semantic sequence, adding a "oil temperature over-limit" warning node, and the sequence becomes "oil temperature over-limit → load rate → cooling efficiency". This dynamic adjustment mechanism ensures that the system response always reflects the latest operating status.

[0058] The core semantic sequence is ultimately represented as a weighted directed graph structure. Nodes are semantic features and their attributes, edges represent the temporal or logical relationships between features, and edge weights reflect the strength of the association. The graph structure retains complete contextual information; for example, the node "regional load growth" is associated with the specific value of 315MW and the time range of July 1-7, supporting the generation of detailed analysis reports. The sequence visualization uses a heatmap, displaying features with high core semantic entropy as red hot zones, allowing users to quickly grasp key factors.

[0059] In power dispatching applications, this implementation method can effectively distinguish between primary and secondary factors of a problem. For example, in a load anomaly analysis, the core sequence selected shows "sudden changes in renewable energy output → grid frequency regulation pressure → main transformer load fluctuations," clearly identifying renewable energy fluctuations as the root cause, rather than directly focusing on the load phenomenon itself. This deep correlation analysis benefits from the grid structure relationships stored in the knowledge graph, correctly associating surface phenomena with system-level influencing factors.

[0060] The adaptability at the user interaction level is reflected in the dynamic interpretation function of semantic sequences. When a user clicks on the "oil temperature" node in the sequence, the system displays a detailed evolution curve of that feature, related equipment parameters, and historical records of similar events. This interactive design aligns with the problem-solving thinking habits of power professionals, starting from core features and gradually expanding the investigation of related factors. Sequence nodes also support manual adjustments. For example, maintenance personnel can manually add "cooling fan failure" to the sequence based on their field experience, and the system will recalculate the core semantic entropy of the relevant features and update the knowledge graph correlation.

[0061] Example 5: The intelligent response generation module constructs the final output based on the core semantic sequence and the power industry knowledge graph. This process begins with the graph matching stage, aligning each feature node in the core semantic sequence with its corresponding node in the knowledge graph. The matching process employs a bidirectional traversal strategy: forward, it searches for influencing factors in the upper-level system from the feature node, such as associating the load rate node with the regional load node; backward, it searches for lower-level consequence nodes, such as associating the oil temperature node with the insulation aging risk node. The matching results form a problem analysis tree, with the root node representing the query topic and branches containing influencing factor chains and consequence chains.

[0062] The initial response draft is generated using a structured narrative framework. The problem analysis tree is converted into a three-part structure: the phenomenon description section directly maps to the feature values ​​of the core sequence, the cause analysis section integrates the chain of influencing factors, and the recommended measures section integrates the consequence chain and industry regulations. The draft retains data traceability markers, and each conclusion corresponds to a specific feature number and knowledge graph node ID, supporting subsequent verification.

[0063] The response optimization phase implements a four-layer filtering mechanism. The first layer, logical verification, detects contradictory statements, triggering a conflict resolution process to re-examine feature relevance. The second layer, terminology standardization, converts technical descriptions into standard power terminology. The third layer, context adaptation, adjusts the wording according to user identity; responses to dispatchers include dispatch instruction codes, and responses to customer service personnel are transformed into user-executable suggestions. The fourth layer, language fluency processing, restructures complex sentences.

[0064] The risk assessment module plays a dual role in response generation. First, it calculates global risk indicators based on a semantic feature matrix, such as the comprehensive load mutation rate, equipment health index, and meteorological disaster probability, generating a risk assessment value from 0 to 100. This value directly affects the response wording: when the risk value exceeds 70, a level-three warning is added to the beginning of the response; when it exceeds 85, emergency operation guidelines are added. Second, risk control logic is integrated into the recommended measures. For example, for abnormal transformer oil temperature scenarios, the system selects different handling solutions based on the risk value: low risk suggests routine inspections, medium risk requires a 72-hour special inspection, and high risk immediately triggers equipment shutdown decision support.

[0065] The response formatting utilizes a dynamic template engine. The template library includes types such as accident analysis reports, operational recommendations, and user notification letters, automatically matching the template based on the query type. Data slots are reserved in the templates, such as labels like [Equipment Number], [Time Range], and [Risk Level], which are populated with specific values ​​by the system. The final output is a multimodal response, with the main text accompanied by data visualization attachments, such as load change curves, equipment temperature heatmaps, and risk indicator radar charts. Visualization elements are linked to key data points in the text; for example, clicking on the text "Oil Temperature 85℃" will expand to display a historical temperature trend chart.

[0066] The system implements a closed-loop verification mechanism for responses. After a user executes the suggested measures in the response, the status monitoring module collects feedback data, such as the equipment temperature after maintenance and the effect of load adjustment. This data returns a semantic feature matrix, triggering an update of the knowledge graph's correlation degree. For example, if the oil temperature drops by 10°C after the "cooling system cleaning" suggestion is implemented, the correlation weight between the cooling efficiency feature at the equipment level and the maintenance measure node increases. This learning mechanism continuously optimizes the quality of subsequent responses. In complex fault diagnosis scenarios, response generation reflects multi-path reasoning capabilities. The system matches multiple causal chains in parallel in the knowledge graph, generating multi-solution responses with probability weights: "65% probability of setpoint anomaly, i.e., feature F201 fluctuation; 30% probability of secondary loop problem, i.e., feature F202 exceeding the standard; 5% probability of electromagnetic interference, i.e., feature F203 existing."

[0067] User interaction experience is enhanced through refined responses. The system records user feedback behaviors to historical responses, such as clicking on details, saving reports, and requesting human assistance, and optimizes subsequent response structures accordingly. For users who prefer technical details, a data analysis appendix is ​​automatically displayed; for operational users, step-by-step guidance is highlighted. In multi-turn conversations, responses automatically inherit previous conclusions; for example, after the initial response analyzes the causes, when the user asks "how to prevent it," the focus shifts directly to the recommended measures.

[0068] The real-time nature of the risk assessment module is particularly crucial in emergency scenarios. When a query for "substation flooding risk" is received during a typhoon warning, the system integrates meteorological data, terrain data, and drainage system status characteristics to calculate a real-time risk value. If the risk value exceeds the threshold, the response integrates elements such as a list of flood control measures, emergency power supply configuration plans, and personnel evacuation route maps, highlighting the warning level with a red border. Simultaneously, a background process is automatically triggered, pushing the high-risk response to relevant responsible personnel.

[0069] A compliance review is conducted before the final response is output. Recommended measures are verified against the power industry regulations database; for example, user power outage operations must comply with Article 73 of the Safety Regulations, and equipment maintenance intervals must meet maintenance procedure standards. If the review fails, the process is reverted to the draft stage and regenerated to ensure all technical recommendations comply with industry norms and safety standards. The response document is automatically generated with a version number and digital signature and incorporated into the enterprise knowledge base as a decision record.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-powered large language model intelligent interaction system for the power industry, characterized in that, include: The multi-source data acquisition module obtains multi-dimensional data related to the power industry from multiple heterogeneous data sources and integrates the multi-dimensional data to obtain an integrated data warehouse. The semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse to obtain a semantic feature matrix. The industry knowledge graph construction module builds a knowledge graph for the power industry based on a semantic feature matrix. The user query parsing module parses the query input by the user and extracts the query feature vector; The core semantic filtering module filters core semantic sequences based on the semantic feature matrix and query feature vector; The intelligent response generation module generates intelligent responses based on core semantic sequences and a knowledge graph of the power industry.

2. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The multi-source data acquisition module obtains multi-dimensional data related to the power industry, including: The power grid operation data includes voltage, current, frequency, power, power grid topology, dispatch instructions, historical operation records, and fault event logs. The user interaction data includes user inquiry records, complaint feedback, and electricity consumption behavior analysis data. The equipment status data includes real-time monitoring parameters, maintenance records, and life prediction data for transformers, circuit breakers, transmission lines, and power distribution equipment. The energy trading data includes medium- and long-term trading data, spot trading data, and ancillary services market data.

3. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The multi-source data acquisition module integrates multi-dimensional data, specifically including: The multidimensional data is preprocessed, including data cleaning, standardization, and semantic mapping. The data cleaning process includes identifying and deleting duplicate records, filling missing values ​​with linear interpolation of data from adjacent time points, correcting outliers by means of a sliding window, and retaining copies of the original data during the data cleaning process. The standardization process includes converting all electrical parameters to the International System of Units (SI), converting timestamps to UTC format with millisecond precision, encoding text data into UTF-8 format, and mapping enumerated data to numerical codes. The standardization process also establishes a data conversion rule base to record the conversion logic of each data field. The semantic mapping includes transforming standardized data into semantic entities for the power industry. Based on a predefined power semantic ontology library containing four categories of semantic tags: equipment, operation, user, and market, semantic mapping is performed on equipment status data, user interaction data, power grid operation data, and energy trading data to obtain power semantic data in various dimensions. An integrated data warehouse is constructed based on power semantic data from all dimensions. The integrated data warehouse adopts a hybrid architecture of graph database and time series database. Equipment status and power grid topology relationship data are stored in the graph database, and load curves and voltage fluctuation time series data are stored in the time series database. The integrated data warehouse establishes a global indexing mechanism to achieve data integration.

4. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The semantic feature extraction module performs semantic analysis and feature extraction on the integrated data warehouse, specifically including: Data in the data warehouse is separated and integrated according to dimensions, namely real-time monitoring data and historical operation logs of power grid operation dimension, consultation records and electricity consumption behavior data of user interaction dimension, sensor monitoring values ​​and maintenance reports of equipment status dimension, and market transaction records and price curves of energy trading dimension. After separating the data for each dimension, feature extraction is performed independently to obtain semantic feature vectors for each dimension; In terms of power grid operation, third-order statistics are calculated for load data to extract features, over-limit correlation features are extracted for voltage data, and fault distribution and clustering features are extracted from fault logs. Sliding window analysis is used for feature extraction. The user interaction dimension extracts consultation text features and sentiment features through NLP technology, extracts electricity fluctuation and proportion features from electricity behavior data, and extracts regional hot issue distribution features by combining spatiotemporal information. The equipment status dimension extracts temperature rise and deviation features from temperature data, and extracts operation frequency deviation features from mechanical operation data. Combined with maintenance records, it extracts lifespan and failure probability derived features. Feature extraction is introduced into the status assessment model. The correlation features between electricity price fluctuations and transaction volume are extracted from the energy trading dimension; the balance deviation features are extracted from supply and demand data; and cross-dimensional features are extracted by combining external factors. All semantic feature vectors of all dimensions are organized into a matrix structure according to the dimension number. The matrix rows represent time series and the columns represent feature dimensions, forming a spatiotemporally related semantic feature matrix, which ultimately enables semantic analysis and feature extraction of the integrated data warehouse.

5. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The industry knowledge graph construction module, based on a semantic feature matrix, constructs a knowledge graph for the power industry, specifically including: Select one dimension from all dimensions of the semantic feature matrix and obtain the semantic feature vector of that dimension. The semantic feature vector contains feature values ​​and their spatiotemporal attributes. The feature correlation degree between each semantic feature in the semantic feature vector of the selected dimension is calculated using an improved mutual information algorithm. The algorithm considers the temporal characteristics and spatial distribution of power data, introduces a time decay factor and a spatial weight coefficient, assigns higher weight to recent data, and determines the spatial weight of key nodes according to the power grid topology. Based on the calculated feature correlation matrix, a community detection algorithm is used to identify feature communities, and feature pairs with correlation higher than a preset threshold θ are connected by edges to form an initial graph structure. The community division is adjusted by a modularity optimization algorithm to make the features within the same community closely related and the features between different communities less related. For each community, a local semantic association graph is generated, where nodes represent features and edge weights represent association strength. Derived nodes are added to the local graph to represent the overall characteristics of the community, ultimately forming a knowledge graph for the power industry.

6. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The user query parsing module parses the user-input query in the following specific ways: The user-input query is segmented and vectorized. The segmentation stage retains compound words in the power field, and the vectorization process adopts a domain-adaptive word embedding model. The model is pre-trained on a general corpus and then fine-tuned using power industry documents to obtain an initial query vector. The initial query vector is subjected to deep semantic encoding, and a hierarchical attention mechanism is adopted. The first layer of attention focuses on the power entities in the query, and the second layer of attention analyzes the relationships between entities, finally obtaining the query feature vector.

7. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The core semantic filtering module filters core semantic sequences based on semantic feature matrices and query feature vectors, specifically including: Calculate the semantic dimension entropy of each semantic feature in the semantic feature matrix. In the knowledge graph of the power industry, all feature correlation degrees corresponding to each semantic feature are extracted, and the correlation degree is calculated by a spatiotemporal weighted mutual information algorithm that introduces time decay factor and spatial weight coefficient. Node connections are established between feature pairs with a correlation degree higher than a preset threshold θ to form a structured correlation relationship between features; The core semantic entropy of each semantic feature is determined by a weighted summation method. The core semantic entropy = semantic dimension entropy × 0.4 + average feature correlation degree in the knowledge graph × 0.6, where the weights are set based on the inherent importance of dimensions and cross-feature correlation value in the power data. Core semantic features are selected based on the core semantic entropy of all semantic features; A temporal correlation strategy is adopted to align the selected core semantic features by timestamps, forming a multi-dimensional time series, namely the core semantic sequence.

8. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, The intelligent response generation module generates intelligent responses based on core semantic sequences and a power industry knowledge graph, specifically including: The core semantic sequence is matched with the knowledge graph of the power industry. A bidirectional traversal strategy is used to align each feature node in the core semantic sequence with the corresponding node in the knowledge graph. The forward search starts from the feature node to find the influencing factors of the upper-level system, and the reverse search finds the lower-level consequence nodes. The matching result forms a problem analysis tree with the query topic as the root node and branches containing the influencing factor chain and the consequence chain. Based on the problem analysis tree, an initial response draft is generated. Using a structured narrative framework, the problem analysis tree is converted into a three-part content: phenomenon description, cause analysis, and suggested measures. The draft retains data traceability markers, and each conclusion corresponds to a specific feature number and knowledge graph node ID. The initial response draft is optimized and formatted. The response optimization implements a four-layer filtering mechanism. The first layer is logical verification to detect contradictory statements and trigger a conflict resolution process to re-examine the relevance of features. The second layer is terminology standardization to convert technical descriptions into standard power terms. The third layer is context adaptation to adjust the expression according to the user's identity. The fourth layer is language fluency processing to reorganize complex sentence structures. The response formatting uses a dynamic template engine, which automatically matches templates from a template library containing accident analysis reports, operation recommendations, and user notification letters based on the query type. The system fills the reserved data slots in the templates with specific values, and the output is a multimodal response with attached data visualizations. The visualization elements are linked with the key data points in the text, resulting in an intelligent response.

9. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, It also includes a risk assessment module; The risk assessment module is used to calculate a global risk index based on a semantic feature matrix. The global risk index integrates load mutation rate, equipment health index, and meteorological disaster probability parameters. The load mutation rate is derived from the feature extraction results of the semantic feature extraction module on the power grid operation dimension data; The equipment health index is derived from the feature extraction and fusion results of the semantic feature extraction module on the equipment status dimension data; The probability of meteorological disasters is derived from real-time data collected by the status monitoring module and processed by the semantic feature extraction module through cross-dimensional feature processing. When the risk assessment module generates a risk assessment value, it first obtains the key parameters mentioned above from the semantic feature extraction module, and then inputs the parameters into a preset risk assessment model that conforms to the risk assessment standards of the power industry. Combining the degree of influence of different parameters in the risk assessment of power grid operation, the module generates a risk assessment value of 0-100 points through the internal calculation logic of the model. When the risk assessment value exceeds 70, add a level 3 warning prompt at the beginning of the smart response; When the risk assessment value exceeds 85, emergency operation guidelines will be added to the intelligent response. Based on the risk assessment value, risk control logic is incorporated into the intelligent response recommendations. Low risk corresponds to daily inspection recommendations, medium risk corresponds to 72-hour special inspection requirements, and high risk corresponds to equipment shutdown decision support.

10. The AI-powered large language model intelligent interaction system for the power industry according to claim 1, characterized in that, It also includes a status monitoring module; The status monitoring module is used to monitor the power grid operation status in real time and generate status monitoring data. Through sensors, smart meters, fault recording devices and other equipment deployed at key nodes of the power grid, dynamic data such as power grid load, equipment temperature and environmental weather are collected in real time. The status monitoring module dynamically adjusts the data acquisition frequency according to the importance of the data type, acquiring temperature data at a second-level frequency and meteorological data at a minute-level frequency. After data collection, the status monitoring module performs preliminary verification of the data to remove obvious outliers, and inputs the verified data into the multi-source data acquisition module to compare it with historical data to identify abnormal trends. The identified abnormal trend data is then marked as high-risk data in the integrated data warehouse.

Images