Power grid knowledge graph fusion intention recognition and dynamic routing decision method, system and device, and storage medium

Abstract

The invention discloses a power grid knowledge graph fusion intention recognition and dynamic routing decision method, system and device, and a storage medium, and relates to the technical field of power grid intelligent question answering and decision support, and the method comprises the steps: carrying out the power field specialized intention understanding, recognizing a query intention, and linking a related entity based on a power grid related query inputted by a user; the method comprises the following steps: constructing and fusing a three-layer knowledge graph covering equipment, topology and state on the basis of multi-source power grid service data to form a structured domain knowledge base; selecting a processing flow based on an intention type obtained by identifying the query intention, a query complexity score determined through a complexity evaluation model and an emergency degree score determined through an emergency degree evaluation model; according to the selected processing flow, generating a response result and outputting the response result to the user; according to the method, the recognition precision of the professional query intention of the power grid can be remarkably improved, complex reasoning is supported, differentiated response is realized, emergency faults can be quickly processed, and the method has continuous self-optimization capability.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering and decision support technology for power grids, and in particular to a method, system, device, and storage medium for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion. Background Technology

[0002] With the continuous improvement of power grid intelligence, the demand for professional, efficient, and accurate intelligent question-answering and decision support systems is becoming increasingly urgent in scenarios such as power grid operation and maintenance management, fault diagnosis and handling, and power dispatching and command. However, existing power grid intelligent question-answering systems often struggle to accurately understand the user's true intent when faced with queries containing a large number of technical terms, complex topological relationships, and implicit domain semantics. They are unable to provide differentiated response strategies, resulting in insufficient response efficiency and professionalism, and failing to meet the high standards required for real-time power grid operation and maintenance and emergency command.

[0003] However, existing technologies have the following shortcomings in power grid professional application scenarios: General intent recognition frameworks struggle to accurately understand power grid terminology and implicit semantics. When processing queries involving professional concepts such as line loss, power flow, and relay protection, intent recognition accuracy drops significantly, failing to accurately distinguish between different intent categories such as power grid operation and maintenance queries, fault diagnosis requests, and dispatch instructions. This leads to subsequent information retrieval and answer generation deviating from actual user needs. Knowledge graph construction does not fully integrate characteristics of power grid topology, equipment relationships, and operating status. It lacks in-depth modeling of electrical connections, spatial relationships, and functional dependencies between power grid equipment. When processing queries involving multi-level voltage transformations, cross-regional power grid coordination, and complex fault propagation chains, it cannot provide accurate relationship reasoning support. Furthermore, it lacks a mechanism for dynamic routing decisions based on intent type, problem complexity, and urgency. A unified processing flow is used for different types of requests, such as emergency power grid fault handling, routine operation and maintenance queries, and historical data analysis, failing to achieve differentiated responses. This results in delayed emergency fault handling responses, impacting the safe and stable operation of the power grid. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a method, system, device and storage medium for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion.

[0005] Therefore, the technical problem solved by this invention is: how to accurately understand the professional intent of the power grid, deeply integrate multi-dimensional knowledge of the power grid, and dynamically allocate processing resources based on query characteristics for intent identification and routing decisions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion, comprising: Based on user-inputted power grid-related queries, perform power sector-specific intent understanding, identify query intent, and link relevant entities; Based on multi-source power grid business data, a three-layer knowledge graph covering equipment, topology, and status is constructed and integrated to form a structured domain knowledge base; Based on the intent type obtained from identifying the query intent, the query complexity score determined by the complexity evaluation model and the urgency score determined by the urgency evaluation model, the processing flow is selected. Based on the selected processing flow, generate the response result and output it to the user.

[0007] The beneficial effects of this technical solution are as follows: by constructing a complete technical closed loop of query intent understanding, knowledge graph fusion, dynamic routing decision and differentiated response generation, it realizes end-to-end intelligent processing of complex queries in the power grid field, overcomes the limitations of existing technologies in handling power grid professional queries such as inaccurate intent recognition, insufficient knowledge support and single response strategy, and significantly improves the accuracy, timeliness and professionalism of the power grid intelligent question answering system in core scenarios such as operation and maintenance, diagnosis and scheduling.

[0008] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The process of understanding power sector-specific intents based on user-input power grid-related queries, identifying query intents, and linking relevant entities includes: The basic categories of query intent are extracted through semantic analysis units: the query is segmented to obtain a word sequence; word graphs and their dependencies are constructed based on dependency parsing; the word graph is processed using a graph attention network, and the deep semantic representation of word is learned by aggregating the neighbor node information of each word, and the overall semantic representation of the query is generated through pooling operations; the query intent category is determined based on the overall semantic representation.

[0009] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The process of understanding power sector-specific intents based on user-input power grid-related queries, identifying query intents, and linking relevant entities also includes: The entity linking unit links the technical terms and entity descriptions in the query to the corresponding nodes in the power grid knowledge graph: identifying entity mentions in the query; retrieving candidate entities corresponding to the entity mentions from the power grid knowledge graph; comprehensively considering the semantic matching degree between the candidate entities and the entity mentions, as well as the topological distance constraints of the candidate entities in the power grid knowledge graph, determining the final entity linking result through weighted calculation; and verifying the entity linking result.

[0010] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The aforementioned construction and fusion of a three-layer knowledge graph covering equipment, topology, and status, based on multi-source power grid business data, to form a structured domain knowledge base includes: Based on multi-source power grid business data, a three-layer power grid knowledge graph is constructed and maintained through a combination of automated and manual verification. The three-layer architecture includes: The device layer is used to store static attribute information of various power equipment in the form of nodes; The topology layer is used to store the connection relationships and spatial location relationships between the power devices; The status layer is used to store dynamic operating status data and historical fault records of the power equipment. Establish a dynamic update mechanism for the power grid knowledge graph to maintain consistency with the actual state of the power grid; based on the constructed power grid knowledge graph, provide knowledge query and reasoning services, including inferring the scope of fault impact based on connection relationships and predicting fault risks based on historical fault records.

[0011] The beneficial effects of this preferred technical solution are as follows: by constructing a knowledge graph that integrates equipment, topology, and status, static attributes, dynamic operation, and historical fault data are organically combined to form a panoramic and in-depth characterization of the power grid entity and its relationships, providing comprehensive, accurate, and real-time updated domain knowledge support for subsequent intent understanding, fault reasoning, and impact assessment.

[0012] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The process of selecting the processing flow based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model includes: Based on intent type, complexity assessment score, and urgency assessment score, a processing flow is selected for the query according to preset rules. The processing flow includes a fast response flow, a standard processing flow, and a deep analysis flow. The preset rules are as follows: for fault diagnosis queries, the flow selection is prioritized based on the urgency score; for operation and maintenance queries, the flow selection is prioritized based on the complexity score; and for data analysis queries, the deep analysis flow is initiated.

[0013] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The selection process, based on the intent type obtained from identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model, further includes: An optimization mechanism based on reinforcement learning is adopted, with intent type, complexity evaluation score, urgency evaluation score and current system resource status as state inputs, and the selected processing flow as action output. Reward signals are calculated based on user satisfaction, response time and resource consumption after query processing to continuously update and optimize the routing decision parameters in the preset rules.

[0014] As a preferred scheme for intent recognition and dynamic routing decision-making through power grid knowledge graph fusion, wherein: The step of generating a response result and outputting it to the user according to the selected processing flow includes: Based on the selected processing flow, the corresponding knowledge resources and computing resources are invoked to generate response results. The processing flows include a rapid response flow, a standard processing flow, and a deep analysis flow. When the processing flow is a rapid response flow, based on the fault phenomenon and faulty equipment information in the query, historical fault cases are retrieved and matched through case reasoning to quickly identify the fault type; based on the topological connection relationship of the faulty equipment in the power grid knowledge graph, the potential scope of the fault's impact is assessed; based on the fault type and impact scope assessment results, emergency handling recommendations containing immediate actions, diagnostic steps, and recovery plans are retrieved from the emergency handling plan library and generated; the fault type identification results, impact scope assessment results, and emergency handling recommendations are integrated into a key information summary output; When the processing flow is the standard processing flow, relevant information is retrieved from the multi-source knowledge base according to the query intent, the search results are deduplicated, sorted and integrated to form a complete answer, and the integrated information is filled into a predefined standardized answer template to generate a structured detailed answer; When the processing flow is a deep analysis flow, historical data is extracted from the time series database for statistical analysis, correlation analysis and cluster analysis, and a time series forecasting model is used for trend forecasting analysis. The analysis results are then used to generate a comprehensive report in the form of a combination of charts and text. The generated response is then output to the user after quality checks, which include fact-checking, logical consistency checks, and language fluency checks.

[0015] Secondly, the present invention provides an intent recognition and dynamic routing decision system based on power grid knowledge graph fusion, comprising: The query receiving and intent recognition module is used to perform power sector-specific intent understanding based on user-input power grid-related queries, identify query intent, and link relevant entities; The knowledge graph fusion module is used to construct and fuse a three-layer knowledge graph covering equipment, topology and status based on multi-source power grid business data to form a structured domain knowledge base; The dynamic routing decision module is used to select the processing flow based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model. The response generation module is used to generate response results and output them to the user based on the selected processing flow.

[0016] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the intent recognition and dynamic routing decision method for power grid knowledge graph fusion are implemented.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement steps of an intent recognition and dynamic routing decision method for power grid knowledge graph fusion.

[0018] The beneficial effects of this invention are as follows: By constructing a specialized intent understanding framework for the power sector, combined with semantic analysis and entity linking technologies, this invention can accurately identify complex query intents in professional scenarios such as power grid operation and maintenance, fault handling, and power dispatching. The intent recognition accuracy is improved by 15% to 22% compared to general methods, significantly enhancing the system's ability to understand power grid terminology and implicit semantics. Through deep integration of a multi-level power grid knowledge graph covering power equipment, power grid topology, operating parameters, and fault types, accurate modeling of the relationships and operating states of power grid equipment is achieved. This effectively supports query processing involving multi-level voltage transformation, cross-regional power grid coordination, and complex fault propagation chains, and improves the accuracy of relational reasoning. Improvements of over 20%; Through the innovative design of a dynamic routing decision mechanism, a complexity assessment model and an urgency assessment model are introduced to quantitatively evaluate queries. The processing flow is selected based on intent type, complexity score, and urgency score, enabling differentiated responses to different types of queries. Emergency fault handling response time is shortened to less than 5 seconds, and the average response time for regular queries is controlled at around 20 seconds, resulting in an overall system response efficiency improvement of over 40%. Through a routing strategy optimization mechanism based on reinforcement learning, the system can continuously learn and improve the routing decision effect, leading to continuous improvement in system performance and user satisfaction over long-term operation. It has good adaptive capabilities and sustainable optimization characteristics. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of an intent recognition and dynamic routing decision method based on power grid knowledge graph fusion provided by the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides an intent recognition and dynamic routing decision method based on power grid knowledge graph fusion, comprising: S1: Based on user-inputted power grid-related queries, perform power sector-specific intent understanding, identify query intent, and link relevant entities; S2: Based on multi-source power grid business data, construct and integrate a three-layer knowledge graph covering equipment, topology and status to form a structured domain knowledge base; S3: Based on the intent type obtained from identifying the query intent, the query complexity score determined by the complexity evaluation model and the urgency score determined by the urgency evaluation model, select the processing flow; S4: Generate a response result and output it to the user based on the selected processing flow.

[0023] It should be noted that through steps S1-S4, a complete technical closed loop from professional query understanding to differentiated decision response was constructed. This achieved deep integration of complex semantics and multimodal knowledge in the power grid field, and could adaptively select the optimal processing path based on the real-time context and inherent characteristics of the query. This not only significantly improved the accuracy and timeliness of intelligent question answering in core scenarios such as operation and maintenance, diagnosis, and scheduling, but also provided intelligent decision support with continuous optimization capabilities for power grid operation and management, ultimately achieving a systematic leap in response efficiency, professional depth, and user experience.

[0024] Example 2, refer to Figure 1As one embodiment of the present invention, based on the previous embodiment, a method for intent recognition and dynamic routing decision-making by fusion of power grid knowledge graph is provided, including: In this embodiment, step S1 above, based on the user-inputted power grid-related query, performs power-domain-specific intent understanding, identifies the query intent, and links relevant entities, including: The system receives and preprocesses power grid-related queries input by users through a query receiving module; these queries involve scenarios related to power grid operation and maintenance management, fault diagnosis and handling, or power dispatch and command. User queries can come from various input methods, including text input, voice input, and graphical interface operations. For text input, the query receiving module directly receives the text content. For voice input, the query receiving module calls a speech recognition engine to convert speech into text. For graphical interface operations, the query receiving module generates the corresponding query text based on the user's selections and clicks. Query preprocessing includes operations such as removing redundant characters, correcting spelling errors, and standardizing terminology. The query receiving module also records the contextual information of the query, including query time, user identity, user role, and historical query records. This contextual information plays an important role in subsequent intent recognition and routing decisions.

[0025] The intent recognition module performs power-domain-specific intent understanding on queries. The intent recognition module includes a semantic analysis unit and an entity linking unit. The semantic analysis unit is used to extract the basic categories of query intent, and the entity linking unit is used to link the technical terms and entity descriptions in the query to the corresponding nodes in the power grid knowledge graph.

[0026] The semantic analysis unit's workflow includes four steps: query segmentation, word graph construction, graph attention network processing, and intent classification. Query segmentation utilizes a segmentation tool optimized for the power industry, accurately identifying power-related technical terms. Word graph construction, based on dependency parsing, constructs a graph structure from the query's tokens, with tokens as nodes and dependency relations as edges. The graph attention network employs a multi-layered architecture, aggregating the neighbor node information of each token through an attention mechanism to learn the deep semantic representation of the tokens.

[0027] Specifically, for the first in the query The word element, in its th word element, The layer is represented as: in, For the first The word element in the first Layer representation vector, For the first The set of neighbor nodes of each word element. Let be the attention weight, representing the th The word pair The importance of each word element For the first The weight matrix of the layer, For the first The word element in the first Layer representation vector, For activation functions. Attention weights. Calculated using the attention mechanism: in, The attention score is calculated using the attention function: in, For attention parameter vectors, This represents a vector concatenation operation. The activation function is defined as ELU. After processing by a multi-layer graph attention network, an overall semantic representation of the query is generated through average pooling or max pooling. The intent classifier maps this semantic representation to predefined intent categories and outputs the probability distribution of each intent category. In this embodiment, the preferred graph attention network uses a 3-layer architecture with 128 hidden layer dimensions, 8 attention heads, and an ELU activation function. Intent categories include fault diagnosis, operation and maintenance query, data analysis, operation instruction, and information consultation, etc.

[0028] The entity linking unit's workflow includes four steps: named entity recognition, candidate entity generation, entity disambiguation, and link verification. Named entity recognition uses a BiLSTM-CRF-based sequence labeling model to identify entities such as equipment, location, personnel, and parameters in the query. Candidate entity generation retrieves candidate entities from the knowledge graph that match or are semantically similar to the entity mention string. Entity disambiguation employs a topological distance-based strategy, considering both semantic matching and topological distance for entity selection. Semantic matching is calculated by comparing the context of the entity mention with the candidate entity's description information, using cosine similarity based on word vectors. Topological distance is calculated using path lengths in the knowledge graph; for the power grid domain, topological distance prioritizes voltage level and geographical region affiliation, with entities at the same voltage level and in the same region having a closer topological distance. The entity disambiguation score is calculated using the following formula: in, Candidate entities The disambiguation score, For entity reference, For semantic matching degree, The topological distance score is used to determine the score. The weighting coefficient is 0.6, which is preferred in this embodiment, indicating that the semantic matching degree has a slightly higher weight than the topological distance. The topological distance score is calculated based on the path length normalization; the shorter the path, the higher the score. Link verification ensures the accuracy of the link results through entity consistency checks and context matching verification.

[0029] For example, in the intent recognition module, suppose a user inputs the query "The current in phase B of the high-voltage side of the No. 1 main transformer at the 220kV Xijiao Substation is too high, please analyze the reason." The query receiving module preprocesses the query and then passes it to the intent recognition module. The semantic analysis unit first segments the query into words, obtaining a word sequence {220kV, Xijiao, substation, No. 1, main transformer, high-voltage side, phase B, current, too high, please, analyze, reason}. Then, a word graph is constructed. According to dependency parsing, 220kV modifies Xijiao Substation, No. 1 modifies main transformer, high-voltage side modifies phase B, current is an attribute of phase B, too high modifies current, analysis is the predicate, and reason is the object of analysis. An adjacency matrix is ​​constructed based on these dependency relationships. Next, a three-layer graph attention network is used to process the word graph, learning the deep semantic representation of each word, and average pooling is used to obtain the overall semantic representation of the query. The semantic representation is input into the intent classifier to obtain the probability distribution of each intent category. The fault diagnosis category has the highest probability (0.92), therefore the query is classified as a fault diagnosis query. Simultaneously, the entity linking unit performs named entity recognition on the query, identifying 220kV Xijiao Substation as a substation entity mention, No. 1 main transformer as an equipment entity mention, high-voltage side B phase as an equipment location mention, and current as a monitoring quantity mention. For 220kV Xijiao Substation, candidate entities retrieved from the knowledge graph include 220kV Xijiao Substation and 110kV Xijiao Substation. After calculating semantic matching degree and topological distance, 220kV Xijiao Substation has the highest score and is selected as the link result. Similarly, other entity mentions are also accurately linked to their corresponding nodes in the knowledge graph. Finally, the intent recognition module outputs the query intent category "fault diagnosis" and a list of entities linked to the knowledge graph.

[0030] In another possible implementation, when processing word graphs, graph attention networks can also introduce heterogeneous information networks. In this case, the word graph not only includes word nodes and dependency edges, but also part-of-speech nodes, entity type nodes, etc., forming a heterogeneous graph containing multiple types of nodes and relations. The graph attention mechanism is then extended to a model for processing heterogeneous graph relations, enabling it to distinguish the importance of different relation types (such as subject-verb, verb-object, entity modifier) ​​when aggregating information, thereby learning a richer and more accurate hybrid syntactic and semantic representation.

[0031] In another possible implementation, when processing word graphs, graph attention networks can also employ dynamic graph neural networks. Instead of relying on static dependency parsing results, the system dynamically calculates the association weights between lexical units based on their initial embeddings to construct an initial graph structure, or dynamically learns and adjusts the edge weights of a fixed graph built on dependency relationships. This approach can adaptively strengthen connections between semantically highly related but syntactically distant lexical units (e.g., "current" and "large"), thus more flexibly capturing the semantic core of the query.

[0032] In another possible implementation, topology distance constraints can be achieved by introducing an electrical distance and impedance model. When performing entity disambiguation, in addition to considering the number of hops of the device in the connection graph, the electrical distance of the candidate entity in the power grid electrical topology is also calculated (e.g., the equivalent electrical length calculated based on line impedance). For scenarios involving queries of electrical quantities such as power flow and voltage, entities with closer electrical distances (tighter electrical connections) are preferred, making the link results more consistent with the physical characteristics and actual operation of the power grid.

[0033] In another possible implementation, topological distance constraints can be implemented by combining functional area division, functionally dividing the power grid according to dispatch jurisdiction, power supply zones, or voltage level zones. During entity disambiguation, if a mentioned entity exists in multiple physical locations, the system prioritizes linking to entities within the same functional zone as the query context (such as other mentioned devices or zone names) or the current user's permission zone. For example, when a dispatcher queries "Main Transformer No. 1," the system prioritizes linking to the corresponding main transformer within its managed power grid zone, rather than devices with the same name in other zones.

[0034] It should be noted that the intent recognition module can accurately understand specialized queries in the power grid field, identify the query intent, and link relevant entities, providing high-quality input for subsequent dynamic routing decisions and response generation. Compared with existing technologies, the intent recognition module provided by this invention achieves an intent classification accuracy of 93.5% for specialized queries in the power grid field, an improvement of 18% compared to general intent recognition methods. The entity linking accuracy reaches 89.7%, accurately handling complex entities and technical terms in the power grid field.

[0035] In another possible implementation, a pre-trained language model for the power sector can be introduced to understand power-specific intents. This pre-trained language model is incrementally pre-trained on massive amounts of power grid regulations, equipment manuals, and historical work order records, enabling its word vectors and contextual representations to deeply capture the domain semantics of professional terms such as "overload" and "relay protection action." When a user enters a query, the query text is directly encoded using this domain model, and the encoded representation vector is matched with a predefined intent vector to complete intent classification.

[0036] In another possible implementation, a power domain intent rule and template library can be constructed to assist in understanding power domain-specific intents. This library systematically summarizes typical sentence templates and keyword rules for different scenarios (such as fault alarms, status queries, and operation guidance). (For example, words containing "trip" or "abnormal" are usually associated with fault diagnosis intent.) The system quickly matches and weights user queries against these rule templates, serving as an important supplement or initial judgment beyond the deep learning model. This approach is suitable for processing well-structured dispatch instructions or standardized queries with standardized sentence structures.

[0037] In this embodiment, step S2 above involves constructing and fusing a three-layer knowledge graph covering equipment, topology, and status based on multi-source power grid business data to form a structured domain knowledge base, including: A multi-level power grid knowledge graph covering power equipment, power grid topology, operating parameters, and fault types is constructed through a knowledge graph fusion module. The power grid knowledge graph includes equipment nodes, topological relationships, attribute information, and status data.

[0038] The knowledge graph fusion module is responsible for building and maintaining a multi-layered power grid knowledge graph, providing rich domain knowledge support for intent recognition and dynamic routing decisions. The multi-layered power grid knowledge graph adopts a three-layer architecture design, including a device layer, a topology layer, and a state layer.

[0039] The equipment layer contains various power equipment nodes and their static attribute information. Power equipment nodes include substations, transformers, circuit breakers, disconnectors, busbars, transmission lines, capacitors, reactors, etc. Each equipment node is associated with a series of static attributes, such as equipment number, equipment name, equipment model, rated capacity, rated voltage, rated current, manufacturer, manufacturing date, commissioning date, region, and voltage level. This static attribute information is obtained from the power company's Production Management System (PMS) to ensure data accuracy and authority. The nodes and attributes of the equipment layer are stored using a graph database for efficient querying and relationship traversal.

[0040] The topology layer characterizes the connectivity and spatial relationships between electrical equipment. There are various types of relationships between electrical equipment, primarily including electrical connections, inclusion relationships, and positional relationships. Electrical connections reflect the circuit connections between equipment, such as the high-voltage side of a transformer connecting to the high-voltage bus and the low-voltage side connecting to the low-voltage bus, and transmission lines connecting to the busbars of two substations at each end. Electrical connections are the core of the power grid topology, determining the propagation paths of current and voltage within the grid. Inclusion relationships reflect the hierarchical affiliation of equipment, such as a substation containing multiple transformers, a transformer containing multiple windings, and a busbar containing multiple segments. Positional relationships reflect the spatial distribution of equipment, such as equipment within the same substation being geographically adjacent, and different substations being spatially connected via transmission lines. The relationship information in the topology layer is obtained from the Supervisory Control and Data Acquisition (SCADA) system and the Energy Management System (EMS). Power grid topology analysis algorithms identify the connectivity relationships between equipment, and Geographic Information Systems (GIS) are used to obtain the spatial location information of the equipment.

[0041] The state layer contains dynamic operating status data and historical fault records of power equipment. Dynamic operating status data includes real-time monitored and calculated quantities, such as transformer voltage, current, active power, reactive power, power factor, temperature, and oil level; transmission line current, active power, and reactive power; and bus voltage and frequency. This dynamic data is acquired in real-time from the SCADA system, with a sampling period typically between 2 and 5 seconds, reflecting the real-time operating status of the power grid. Historical fault records contain information on faults that have occurred in the equipment's history, such as fault occurrence time, fault type, fault cause, impact range, handling measures, and recovery time. These historical fault records are obtained from the fault information management system, providing important references for fault diagnosis and preventative maintenance. The state layer data is stored using a time-series database, supporting efficient time-range queries and aggregation calculations.

[0042] In another possible implementation, the three-layer architecture of the knowledge graph can be extended to a four-layer architecture including a mechanism model layer. Based on the device layer, topology layer, and state layer, a mechanism model layer is added. This layer encapsulates the dynamic behavior mechanisms of key power equipment (such as transformers and generators), power grid flow calculation models, and typical fault evolution models in the form of computable models (such as difference equations and algebraic equations). This enables the knowledge graph to not only answer factual questions such as "what," "where," and "what state," but also to support simulations and causal analyses of "what if..."

[0043] In another possible implementation, the three-layer architecture of the knowledge graph can be enhanced by introducing a decision-making plan layer. Above the state layer, a dedicated layer for storing structured decision knowledge is added to associate specific equipment, specific fault modes, or specific operating states with a series of standardized operating plans, processing procedures, maintenance strategies, or scheduling instructions. When the system recognizes a fault diagnosis intent or conducts an impact assessment, it can directly trigger and organize relevant emergency plans from this layer, greatly shortening the time from analysis to decision support.

[0044] The construction of the multi-layered power grid knowledge graph adopts a combination of automation and manual verification. First, data is automatically collected from various business systems (PMS, SCADA, EMS, GIS, fault information management systems, etc.) using data integration tools, and then cleaned and formatted to ensure data quality. Next, knowledge extraction algorithms extract entities, attributes, and relationships from the collected data to construct an initial version of the knowledge graph. Then, domain experts manually verify the knowledge graph, correcting errors and filling in gaps to ensure its accuracy and completeness. Finally, a dynamic update mechanism for the knowledge graph is established, periodically synchronizing the latest data from business systems to maintain consistency between the knowledge graph and the actual power grid status. For static information at the equipment and topology layers, updates are performed daily; for dynamic data at the state layer, updates are performed in real-time or near real-time.

[0045] The knowledge graph fusion module also provides knowledge graph query and reasoning services for other modules to call. The query service supports flexible queries based on entities, relationships, and attributes, such as querying all attributes of a specific device, querying all devices directly connected to a specific device, or querying all devices within a specific region. The reasoning service supports logical and rule-based reasoning based on the knowledge graph, such as inferring the potential impact range of a fault based on device connectivity, or predicting the fault risk of a device based on its historical fault records. These query and reasoning services provide robust knowledge support for intent recognition, dynamic routing decisions, and response generation.

[0046] For example, in the knowledge graph fusion module, assume there is a 220kV substation, Xijiao Substation, in the power grid. This substation contains two main transformers, No. 1 and No. 2. At the equipment layer, a substation node is created for Xijiao Substation, containing attributes such as: equipment number ST-220-001, equipment name Xijiao Substation, voltage level 220kV, region Xicheng District, and commissioning date 2015-06-01. A transformer node is created for No. 1 main transformer, containing attributes such as: equipment number TR-220-001-01, equipment name No. 1 main transformer, equipment model SFPSZ9-120000 / 220, rated capacity 120MVA, rated voltage 220kV / 110kV / 10.5kV, manufacturer [Company Name], and commissioning date 2015-06-01. At the topology layer, an inclusion relationship is created between Xijiao Substation and No. 1 main transformer. An electrical connection is established between the high-voltage side of main transformer No. 1 and the 220kV busbar section I, and between the low-voltage side and the 110kV busbar section I. At the state level, real-time monitoring data of main transformer No. 1 is stored, such as the high-voltage side A-phase current (648A), B-phase current (652A), and C-phase current (6A), high-voltage side voltage (225.3kV), active power (78.5MW), reactive power (12.3Mvar), and oil temperature (65.2℃). Historical fault records of main transformer No. 1 are also stored, such as a cooling system fault occurring on March 15, 2020 at 14:23:17, affecting the transformer's operating temperature by increasing; the corrective action was replacing the cooling fan; and the recovery time was March 15, 2020 at 16:00:00. Through this multi-layered knowledge graph, the system can comprehensively grasp the static attributes, topological relationships, and dynamic states of power grid equipment, providing rich knowledge support for intelligent question answering.

[0047] It should be noted that by constructing a multi-layered power grid knowledge graph, the system possesses a comprehensive and in-depth understanding of the power grid, enabling it to support complex query and reasoning tasks. The knowledge graph contains over 100,000 device nodes, over 500,000 relationships, and over 2 million attributes, covering the main equipment and connections of the power grid. Based on this knowledge graph, the system can quickly answer various queries such as "What is the rated capacity of the No. 1 main transformer at the 220kV Xijiao Substation?", "What are all the lines directly connected to the 220kV Xijiao Substation?", and "How many 220kV substations are there in Xicheng District?", with a query response time of less than 100ms and an accuracy rate exceeding 99%.

[0048] In this embodiment, the process selected in step S3 above, based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model, includes: The processing flow includes a rapid response process, a standard processing process, and an in-depth analysis process.

[0049] The decision-making process is carried out through a dynamic routing decision-making module, which includes four steps: intent type judgment, complexity assessment, urgency assessment, and routing strategy selection.

[0050] Based on the intent type output by the intent recognition module, queries are categorized into three main types: fault diagnosis, operation and maintenance, and data analysis. Fault diagnosis queries involve the diagnosis and analysis of power grid faults and emergency handling, such as "Power outage in a certain area, please analyze the cause" or "A certain device's protection has activated, please determine the fault type." These queries are related to power grid safety and power supply reliability and require rapid response. Operation and maintenance query type queries involve information queries and technical consultations in daily operation and maintenance work, such as the technical parameters of a certain equipment, the requirements of a certain operating procedure, the maintenance cycle of a certain equipment, etc. These types of queries are the daily needs of operation and maintenance personnel and require detailed and accurate answers. Data analysis queries involve statistical analysis and trend prediction of historical data, such as the average load of a region over the past year, the failure rate trend of a certain type of equipment, and the load forecast for the next week. These queries are used for management decision-making and planning and require in-depth data mining and analysis.

[0051] The complexity assessment module quantitatively evaluates the complexity of queries, primarily measuring it from the following dimensions: First, the number of entities involved in the query; the more entities involved, the more complex the query. Second, the types and levels of relationships involved; queries requiring multi-hop reasoning within the knowledge graph have higher complexity. Third, the complexity of the query's language expression; queries with complex sentence structures, clauses, or multiple conditions have higher complexity. Fourth, the amount of data processing required; queries requiring the analysis of large amounts of historical data or complex calculations have higher complexity. The complexity assessment employs a combination of rule-based and machine learning methods. Simple queries (such as single-entity attribute queries) can be quickly determined using rules, while complex queries are predicted using a trained complexity assessment model. The complexity assessment model takes the query's language features, intent category, number of involved entities, and relationship types as input features, and outputs a complexity score ranging from 0 to 1, where 0 represents the simplest and 1 represents the most complex.

[0052] The urgency assessment module determines the urgency of a query, primarily considering the following factors: First, whether the query involves a power grid fault or anomaly; queries involving faults have a high urgency level. Second, the severity and scope of the fault or anomaly involved; faults with a large impact, potentially leading to widespread power outages, have an extremely high urgency level. Third, time-related terms in the query, such as "urgent," "immediately," and "right away," indicate the urgency of the query. Fourth, the role and permissions of the query submitter; queries submitted by dispatchers or leaders typically have a higher urgency level. The urgency assessment also employs a combination of rule-based and machine learning methods. The urgency assessment model outputs an urgency score for the query, ranging from 0 to 1, where 0 indicates no urgency and 1 indicates the most urgent.

[0053] The routing strategy selection module comprehensively considers intent type, complexity score, and urgency score to select the most appropriate processing flow. The routing decision rules are as follows: For fault diagnosis queries, if the urgency score is greater than 0.7, a rapid response process is triggered. This process aims to minimize response time, ideally providing preliminary fault analysis and emergency suggestions within 5 seconds. If the urgency score is between 0.3 and 0.7, a standard processing flow is executed. This flow balances response time and answer quality, ideally generating a detailed fault diagnosis report within 10 to 30 seconds. If the urgency score is less than 0.3, a deep analysis process is initiated. This process aims to optimize the comprehensiveness and depth of the answer, allowing for a longer processing time (ideally 2 to 5 minutes) to retrieve historical fault cases and compare multiple solutions.

[0054] For operation and maintenance queries, if the complexity score is less than 0.3, a fast response process is executed. For simple information queries such as device parameter queries, the answer can be directly retrieved from the knowledge graph. If the complexity score is between 0.3 and 0.7, a standard processing procedure is executed, which requires knowledge base retrieval, multi-source information integration, and standardized answer generation. If the complexity score is greater than 0.7, a deep analysis process is initiated, which requires integrating multiple knowledge sources, performing complex logical reasoning, and matching expert rules.

[0055] For data analysis queries, which typically require extensive data processing and statistical analysis, a deep analysis workflow is usually initiated. However, for very simple statistical queries (such as a specific monitoring value at a given moment), a standard processing workflow can be executed to quickly return the results.

[0056] The routing strategy selection employs a reinforcement learning-based optimization mechanism. By learning from the processing effects of historical queries and user feedback, it dynamically adjusts routing decision parameters to maximize the overall system response quality. The reinforcement learning model takes the query intent type, complexity score, urgency score, and current system resource status as input, and outputs an action that selects between a fast response process, a standard processing process, or a deep analysis process. Reward signals are obtained based on the final user satisfaction and resource consumption of the query, and the model parameters are updated using the Q-learning algorithm to optimize the routing strategy. The state space of the reinforcement learning is defined as follows: in, For a set of intent types, These correspond to complexity scores and urgency scores, respectively. This refers to the status of system resources (such as current load, available computing resources, etc.).

[0057] Action space is defined as Rapid response, standard processing, in-depth analysis .

[0058] The reward function is defined as: in, Indicates the reward value. User satisfaction is measured using explicit user feedback (such as ratings and likes) or implicit feedback (such as whether an answer is accepted or whether further questions are asked). The normalized response time is defined as the ratio of the actual response time to the expected response time. This represents the normalized resource consumption, defined as the ratio of the computational resources consumed in this query to the average consumption. Let be the weighting coefficient, satisfying This embodiment is preferred. , , This indicates that user satisfaction is the most important optimization goal, followed by response time, and lastly resource consumption.

[0059] This reward function encourages the system to minimize response time and reduce resource consumption while ensuring user satisfaction, thus achieving a balanced optimization of response quality, response speed, and resource efficiency.

[0060] The Q-value update formula for the Q-learning algorithm is: in, Representing state Take action below The Q value, i.e., the long-term expected reward, This represents the learning rate, which controls the magnitude of Q-value updates; in this invention, it is set to 0.1. For instant rewards, This is a discount factor that controls the weight of future rewards. In this invention, it is set to 0.9, indicating a greater emphasis on long-term returns. Indicates the execution of an action The new state afterward Indicates a new state The following actions can be taken. Indicates a new state The maximum Q-value among all possible actions represents the optimal expected reward in the future. The entire formula implements Q-value updates based on temporal difference, learning the optimal routing strategy through interaction with the environment.

[0061] Through reinforcement learning, the dynamic routing decision module can continuously optimize routing strategies, thereby improving response quality and user satisfaction over long-term operation.

[0062] For example, suppose the query intent type output by the intent recognition module is fault diagnosis, the complexity assessment model outputs a complexity score of 0.5, and the urgency assessment model outputs an urgency score of 0.85. According to the routing decision rules, this query is a fault diagnosis type and the urgency score is greater than 0.7, thus triggering the rapid response process. The system immediately calls the fault diagnosis expert knowledge base and emergency response plan library, quickly retrieves similar historical fault cases based on the faulty equipment and fault phenomena linked in the query, analyzes possible fault causes, and provides preliminary emergency response suggestions. The entire rapid response process is completed within 3.2 seconds, returning a concise fault analysis summary to the user: Based on the abnormally high B-phase current on the high-voltage side, it is preliminarily determined that it may be caused by an inter-turn short circuit in the B-phase winding or flashover of the B-phase bushing. It is recommended to immediately shut down and inspect the No. 1 main transformer, focusing on checking the insulation of the B-phase winding and the cleanliness of the B-phase bushing. This answer meets the requirements of rapid response and provides timely support for dispatchers' emergency decision-making. Meanwhile, the system records the processing results and user feedback for this query (assuming the user gave positive feedback), and uses this sample to train the reinforcement learning model to optimize the routing strategy.

[0063] It should be noted that, through the dynamic routing decision-making mechanism, the system can select the most appropriate processing flow based on different types and characteristics of queries, achieving differentiated responses. For emergency fault queries, the average response time of the fast response process is 4.1 seconds, which is 75% shorter than the unified processing flow, significantly improving emergency response capabilities. For routine maintenance queries, the average response time of the standard processing flow is 18.3 seconds, with an accuracy rate of 94.2%, meeting the daily needs of maintenance personnel. For data analysis queries, the deep analysis process can provide comprehensive and in-depth analysis reports, with a user satisfaction rate of 91.5%. Through routing strategy optimization based on reinforcement learning, after 6 months of operation, the overall user satisfaction increased by 12% compared to the initial state, and the response time was reduced by 8%, fully demonstrating the effectiveness and adaptability of the dynamic routing decision-making mechanism.

[0064] In this embodiment, step S4 above, which generates a response result and outputs it to the user according to the selected processing flow, includes: The response generation module generates response results based on the processing flow selected by the dynamic routing decision module and outputs them to the user.

[0065] The rapid response process is specifically designed for urgent fault diagnosis queries, with the optimization goal of providing preliminary fault analysis and emergency handling suggestions in the shortest possible time.

[0066] The rapid response process includes three steps: rapid identification of the fault type, preliminary assessment of the scope of impact, and generation of emergency response recommendations.

[0067] Specifically, based on the description of the fault phenomenon and the linked faulty equipment in the query, the fault type is quickly identified. Fault type identification employs a case-based reasoning method, retrieving the most similar historical cases from the fault case database, and using the fault type of these historical cases as a preliminary judgment of the current fault. The fault case database contains various fault cases that have occurred throughout the power grid's history, with each case recording information such as the faulty equipment, fault phenomenon, fault type, fault cause, and handling measures. Case similarity calculation comprehensively considers the matching degree of the faulty equipment type, the textual similarity of the fault phenomenon description, and the similarity of the operating state at the time of the fault. The similarity calculation formula is: in, Indicates the current fault case Compared with historical failure cases The overall similarity between them, with values ​​ranging from [0, 1], This indicates the matching degree of the faulty equipment type. A value of 1 is assigned if the two cases involve completely identical equipment types, 0.7 if they belong to the same major category (e.g., both are transformers), and 0 if they are different categories. The text similarity representing the fault phenomenon description is calculated using cosine similarity based on word vectors. This indicates the similarity of operating conditions at the time of the fault, calculated by comparing factors such as load level, voltage level, and weather conditions at the time of the fault. Let be the weighting coefficient, satisfying This embodiment is preferred. , , .

[0068] Retrieve the top 5 historical cases with the highest similarity, analyze the distribution of fault types in these cases, and select the fault type with the highest frequency as the initial judgment of the current fault.

[0069] Based on the topological location and electrical connections of the faulty equipment within the power grid, a preliminary assessment of the potential impact range of the fault is conducted. The impact range assessment employs a knowledge graph-based graph search algorithm, starting from the faulty equipment node and expanding outwards along the electrical connections to identify potentially affected downstream equipment and users. The propagation rules for the impact range differ depending on the type of fault. For circuit breaker tripping faults, the impact range includes all equipment and users downstream of the circuit breaker. For line faults, the impact range includes equipment at both ends of the line and users downstream of the line. For transformer faults, the impact range includes all equipment and users on the low-voltage side of the transformer. The impact range assessment also considers the power grid's operating mode and protection configuration, such as the existence of backup lines and whether the impact can be mitigated through load transfer. The results of the impact range assessment are quantified in terms of the number of affected users and the affected load capacity, providing a basis for emergency decision-making.

[0070] Based on the fault type identification and impact scope assessment results, matching emergency response plans are retrieved from the emergency response plan database to generate emergency response recommendations. The database contains standardized emergency response procedures and operational guidelines for various fault types, with each plan specifying the operational steps, precautions, and safety measures for emergency response. Emergency response recommendations consist of three parts: immediate actions, further diagnostic steps, and a power restoration plan, presented in concise and clear language for easy understanding and execution by dispatchers.

[0071] All computations in the rapid response process are executed in parallel on high-performance servers. Case retrieval is accelerated using vectorized retrieval, and knowledge graph search is optimized using indexing, ensuring that the entire process is preferably completed within 5 seconds.

[0072] For example, suppose a user queries "Circuit breaker 20 of Beijiao Line at 220kV Beijiao Substation tripped, please analyze the cause and provide handling suggestions." This query is identified as a fault diagnosis type with high urgency, triggering a rapid response process. First, the fault type rapid identification module extracts the fault symptom of circuit breaker tripping and the faulty equipment, "Circuit breaker 20 of Beijiao Line at 220kV Beijiao Substation," from the query. A search of the fault case database finds the five most similar historical cases: a 220kV line circuit breaker tripped due to a line short circuit, a 220kV line circuit breaker tripped due to overcurrent protection operation, a 220kV line circuit breaker tripped due to differential protection operation, a 220kV line circuit breaker tripped due to distance protection operation, and a 220kV line circuit breaker tripped due to switch failure. Statistics on the fault types in these five cases show that line short circuit occurred three times; therefore, the initial judgment is that the fault type is a line short circuit. Then, the preliminary impact assessment module queries the knowledge graph for downstream equipment of the Beijiao Line 20 circuit breaker and finds that it connects to the Xijiao Substation, supplying power to 5,000 users with a load capacity of 80MW. Therefore, the impact assessment result is that 5,000 users and 80MW of load are affected. Finally, the emergency response suggestion generation module retrieves the emergency response plan for line short circuit faults from the contingency plan library and generates the following emergency response suggestion: Immediate Action: Notify the Xijiao Substation to prepare for load transfer. Diagnostic Steps: Dispatch patrol personnel to conduct a full-line inspection of the Beijiao Line, focusing on checking towers, conductors, and insulators to locate the short circuit point. Recovery Plan: After the short circuit point is located and isolated, attempt to reclose the Beijiao Line 20 circuit breaker to restore power supply. The entire rapid response process is completed within 4.5 seconds, providing timely decision support for dispatchers.

[0073] It should be noted that, through the rapid response process, the system can provide preliminary fault analysis and emergency handling suggestions for urgent fault queries in a very short time, significantly improving the power grid's emergency response capabilities. The average response time of the rapid response process is 4.1 seconds, the fault type identification accuracy rate is 87.3%, the average error of impact range assessment is within 10%, and the adoption rate of emergency handling suggestions reaches 92.5%, fully meeting the real-time requirements of power grid emergency command.

[0074] Furthermore, the standard processing flow is applicable to routine operation and maintenance queries, including knowledge base retrieval, multi-source information integration, and standardized answer generation.

[0075] The knowledge base retrieval system searches and queries relevant information from multiple knowledge sources, including the power grid knowledge graph, regulations and specifications database, and technical manual database. The retrieval strategy determines the search scope and keywords based on the query intent and entity link results. For equipment parameter queries, static attributes are directly retrieved from the equipment layer of the knowledge graph. For operating procedure queries, relevant clauses are retrieved from the regulations and specifications database. For technical consultation queries, technical instructions and operating guidelines are retrieved from the technical manual database. Search results are sorted by relevance score, prioritizing information with high relevance.

[0076] Multi-source information integration involves deduplicating, sorting, and integrating multiple retrieved information entries to ensure the completeness and consistency of the answer. The deduplication process identifies information entries with identical or highly similar content, retaining only one. The sorting process ranks information entries based on their authority, timeliness, and relevance, placing the most important information first. The integration process organizes scattered information fragments into logically coherent content, forming a complete answer.

[0077] Standardized answer generation selects an appropriate answer template based on the query type, fills the integrated information into the template, and generates a structured and detailed answer. Answer templates include those for parameter queries, procedures, and technical consultations. Answer formats include text descriptions, tables, and charts, with the most suitable presentation method selected based on the information type and user needs.

[0078] The response time of the standard processing procedure is preferably controlled between 10 and 30 seconds, and the answers are detailed and comprehensive, meeting the daily information needs of operation and maintenance personnel.

[0079] Furthermore, the in-depth analysis process is suitable for data analysis queries, including historical data mining, trend prediction analysis, and comprehensive report generation.

[0080] Historical data mining extracts large amounts of historical data from time-series databases, performing statistical analysis, correlation analysis, and cluster analysis to discover patterns and regularities within the data. Statistical analysis calculates statistical indicators such as the mean, variance, maximum, and minimum values ​​of the data. Correlation analysis identifies the relationships between different variables, such as the correlation between load and temperature, or the correlation between equipment failure rate and years of operation. Cluster analysis groups the data and identifies different data patterns, such as classifying load curves into weekday patterns, weekend patterns, and holiday patterns.

[0081] Trend forecasting analysis uses time series forecasting models (such as ARIMA and LSTM) to predict future trends. These models learn from the periodicity, trend, and randomness of historical data to predict data changes over a future period. The prediction results include point forecasts and confidence intervals, providing a reference for decision-making.

[0082] The comprehensive report generator presents the analysis results in a combination of charts and text, producing a professional analytical report. The report includes sections on analysis objectives, data sources, analysis methods, analysis results, and conclusions / recommendations. Charts include various types such as line charts, bar charts, pie charts, and scatter plots, allowing you to choose the most suitable visualization method based on the characteristics of your data. The report format is standardized, facilitating reading and decision-making for managers.

[0083] The optimal processing time for in-depth analysis is 2 to 5 minutes, providing strong data support for management decisions and planning.

[0084] The response generation module, based on the processing flow selected by the dynamic routing decision module, invokes relevant knowledge and computing resources to generate response results and outputs them to the user in an appropriate format. For rapid response flows, it generates a concise summary of key information, highlighting the most important fault analysis and emergency recommendations, with a word count limited to 200 words. For standard processing flows, it generates a structured, detailed answer, including a problem description, information sources, detailed solutions, and references, typically between 500 and 1000 words. For in-depth analysis flows, it generates a comprehensive report containing charts and data analysis, typically between 2000 and 5000 words, including multiple charts and detailed data analysis. The response generation module is also responsible for quality checks on the response results, ensuring the accuracy, completeness, and readability of the answers. Quality checks include fact-checking, logical consistency checks, and language fluency checks. Fact-checking verifies the accuracy of technical data and parameters in the answer. Logical consistency checks ensure the reasoning process of the answer is reasonable and the conclusions are consistent with the arguments. Language fluency checks ensure the answer is clearly expressed and easy to understand.

[0085] Example 3 illustrates a schematic scheme for an intent recognition and dynamic routing decision method based on power grid knowledge graph fusion according to this embodiment. It should be noted that the technical solution of the intent recognition and dynamic routing decision system based on power grid knowledge graph fusion and the technical solution of the aforementioned intent recognition and dynamic routing decision method based on power grid knowledge graph fusion belong to the same concept. Details not described in detail in the technical solution of the intent recognition and dynamic routing decision system based on power grid knowledge graph fusion in this embodiment can be found in the description of the aforementioned intent recognition and dynamic routing decision method based on power grid knowledge graph fusion.

[0086] This embodiment also provides an intent recognition and dynamic routing decision system based on power grid knowledge graph fusion, including: The query receiving and intent recognition module is used to perform power sector-specific intent understanding based on user-input power grid-related queries, identify query intent, and link relevant entities; The knowledge graph fusion module is used to construct and fuse a three-layer knowledge graph covering equipment, topology and status based on multi-source power grid business data to form a structured domain knowledge base; The dynamic routing decision module is used to select the processing flow based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model. The response generation module is used to generate response results and output them to the user based on the selected processing flow.

[0087] This embodiment also provides an electronic device applicable to a power grid knowledge graph fusion-based intent recognition and dynamic routing decision method, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement an intent recognition and dynamic routing decision method for power grid knowledge graph fusion as proposed in the above embodiments.

[0088] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements an intent recognition and dynamic routing decision method for power grid knowledge graph fusion as proposed in the above embodiments.

[0089] The storage medium proposed in this embodiment belongs to the same inventive concept as the intent recognition and dynamic routing decision method for power grid knowledge graph fusion proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0090] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion, characterized in that, include: Based on user-inputted power grid-related queries, perform power sector-specific intent understanding, identify query intent, and link relevant entities; Based on multi-source power grid business data, a three-layer knowledge graph covering equipment, topology, and status is constructed and integrated to form a structured domain knowledge base; Based on the intent type obtained from identifying the query intent, the query complexity score determined by the complexity evaluation model and the urgency score determined by the urgency evaluation model, the processing flow is selected. Based on the selected processing flow, generate the response result and output it to the user.

2. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 1, characterized in that, The process of understanding power sector-specific intents based on user-input power grid-related queries, identifying query intents, and linking relevant entities includes: The basic categories of query intent are extracted through semantic analysis units: the query is segmented to obtain a word sequence; word graphs and their dependencies are constructed based on dependency parsing; the word graph is processed using a graph attention network, and the deep semantic representation of word is learned by aggregating the neighbor node information of each word, and the overall semantic representation of the query is generated through pooling operations; the query intent category is determined based on the overall semantic representation.

3. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 2, characterized in that, The process of understanding power sector-specific intents based on user-input power grid-related queries, identifying query intents, and linking relevant entities also includes: The entity linking unit links the technical terms and entity descriptions in the query to the corresponding nodes in the power grid knowledge graph: identifying entity mentions in the query; retrieving candidate entities corresponding to the entity mentions from the power grid knowledge graph; comprehensively considering the semantic matching degree between the candidate entities and the entity mentions, as well as the topological distance constraints of the candidate entities in the power grid knowledge graph, determining the final entity linking result through weighted calculation; and verifying the entity linking result.

4. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 3, characterized in that, The aforementioned construction and fusion of a three-layer knowledge graph covering equipment, topology, and status, based on multi-source power grid business data, to form a structured domain knowledge base includes: Based on multi-source power grid business data, a three-layer power grid knowledge graph is constructed and maintained through a combination of automated and manual verification. The three-layer architecture includes: The device layer is used to store static attribute information of various power equipment in the form of nodes; The topology layer is used to store the connection relationships and spatial location relationships between the power devices; The status layer is used to store dynamic operating status data and historical fault records of the power equipment. Establish a dynamic update mechanism for the power grid knowledge graph to maintain consistency with the actual state of the power grid; based on the constructed power grid knowledge graph, provide knowledge query and reasoning services, including inferring the scope of fault impact based on connection relationships and predicting fault risks based on historical fault records.

5. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 4, characterized in that, The process of selecting the processing flow based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model includes: Based on intent type, complexity assessment score, and urgency assessment score, a processing flow is selected for the query according to preset rules. The processing flow includes a fast response flow, a standard processing flow, and a deep analysis flow. The preset rules are as follows: for fault diagnosis queries, the flow selection is prioritized based on the urgency score; for operation and maintenance queries, the flow selection is prioritized based on the complexity score; and for data analysis queries, the deep analysis flow is initiated.

6. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 5, characterized in that, The selection process, based on the intent type obtained from identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model, further includes: An optimization mechanism based on reinforcement learning is adopted, with intent type, complexity evaluation score, urgency evaluation score and current system resource status as state inputs, and the selected processing flow as action output. Reward signals are calculated based on user satisfaction, response time and resource consumption after query processing to continuously update and optimize the routing decision parameters in the preset rules.

7. The method for intent recognition and dynamic routing decision-making based on power grid knowledge graph fusion as described in claim 6, characterized in that, The step of generating a response result and outputting it to the user according to the selected processing flow includes: Based on the selected processing flow, the corresponding knowledge resources and computing resources are invoked to generate response results. The processing flows include a rapid response flow, a standard processing flow, and a deep analysis flow. When the processing flow is a rapid response flow, based on the fault phenomenon and faulty equipment information in the query, historical fault cases are retrieved and matched through case reasoning to quickly identify the fault type; based on the topological connection relationship of the faulty equipment in the power grid knowledge graph, the potential scope of the fault's impact is assessed; based on the fault type and impact scope assessment results, emergency handling recommendations containing immediate actions, diagnostic steps, and recovery plans are retrieved from the emergency handling plan library and generated; the fault type identification results, impact scope assessment results, and emergency handling recommendations are integrated into a key information summary output; When the processing flow is the standard processing flow, relevant information is retrieved from the multi-source knowledge base according to the query intent, the search results are deduplicated, sorted and integrated to form a complete answer, and the integrated information is filled into a predefined standardized answer template to generate a structured detailed answer; When the processing flow is a deep analysis flow, historical data is extracted from the time series database for statistical analysis, correlation analysis and cluster analysis, and a time series forecasting model is used for trend forecasting analysis. The analysis results are then used to generate a comprehensive report in the form of a combination of charts and text. The generated response is then output to the user after quality checks, which include fact-checking, logical consistency checks, and language fluency checks.

8. A power grid knowledge graph fusion-based intent recognition and dynamic routing decision system, employing the method described in any one of claims 1 to 7, characterized in that, include: The query receiving and intent recognition module is used to perform power sector-specific intent understanding based on user-input power grid-related queries, identify query intent, and link relevant entities; The knowledge graph fusion module is used to construct and fuse a three-layer knowledge graph covering equipment, topology and status based on multi-source power grid business data to form a structured domain knowledge base; The dynamic routing decision module is used to select the processing flow based on the intent type obtained by identifying the query intent, the query complexity score determined by the complexity evaluation model, and the urgency score determined by the urgency evaluation model. The response generation module is used to generate response results and output them to the user based on the selected processing flow.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images