Dynamic knowledge graph construction and application method, system, equipment and medium oriented to energy policy influence conduction analysis

Abstract

The invention discloses a dynamic knowledge graph construction and application method, system and device for energy policy influence conduction analysis and a medium, and the method comprises the steps: carrying out the multi-dimensional knowledge extraction of an energy policy text through employing a first adaptive extraction model, and outputting a structured knowledge triple; constructing a dynamic multi-subject knowledge graph; performing deduction search on the potential association path through a first influence search algorithm, and identifying a key policy influence conduction chain; an event-driven knowledge graph closed-loop self-evolution mechanism is established, policy change events are captured in real time through automatic monitoring, and incremental updating and structure correction of the dynamic multi-subject knowledge graph are triggered; and constructing a scene-oriented personalized application service interface according to the dynamic multi-subject knowledge graph, converting a user natural language query into a graph query instruction, and generating a scene response in combination with a retrieval result. According to the invention, the cost of manual arrangement and maintenance is obviously reduced, and the working efficiency and expandability of the system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atlas construction and application, and particularly relates to a dynamic knowledge graph construction and application method, system, device and medium for energy policy influence conduction analysis. BACKGROUND

[0002] With the rapid development of the energy industry and the promotion of the national "double carbon" goal, the formulation and execution of energy policies are extremely frequent and complex. In recent years, digitalization and intelligentization of policy governance have become the only way to build a new energy system and promote the healthy development of the industry. However, existing energy policy analysis and knowledge management means are mostly limited to traditional manual induction, simple policy full-text retrieval or static knowledge graph construction, and are difficult to support policy deep logical reasoning, interest relationship mining and multi-scenario intelligent services.

[0003] On the one hand, traditional rule-based text analysis methods often face the problems of structural fragmentation, low extraction accuracy, lack of semantic consistency and the like when facing highly diversified and frequently changing policy texts. Even if some intelligent word segmentation and entity recognition tools can achieve a certain degree of structurization, the final results are mostly isolated fragments, which are difficult to reflect the overall picture of the policy and the internal semantic hierarchy and system relationship. On the other hand, existing policy knowledge management platforms usually only mark the subject name and attributes, and cannot detail the status change, dynamic response and strategy linkage of the subject in the policy chain, and cannot support scientific decision-making or flexible response to industrial changes.

[0004] In addition, the execution and impact results of energy policies are highly dependent on complex logical chains. For example, adjustment of a certain electricity price mechanism may cause changes in the cost structure of power generation enterprises, affect the enthusiasm of terminal electricity users, and then produce a chain economic effect. The existing power policy knowledge graph construction technology usually only builds a static relationship network, and is difficult to reveal the influence propagation and upstream and downstream linkage among multiple subjects. The lack of intelligent reasoning mechanism limits the application of the knowledge system in real scenarios, and users cannot obtain chain explanation and effective early warning, and it is even more difficult to accurately match policy content according to specific scenario requirements. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] Therefore, the present application provides a dynamic knowledge graph construction and application method, system, device and medium for energy policy influence conduction analysis, which solves the problems of knowledge extraction fragmentation, relationship network static, lack of deep logical reasoning and update lag in the existing energy policy knowledge management technology.

[0007] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides a dynamic knowledge graph construction and application method for energy policy impact transmission analysis, comprising: adopting a first adaptive extraction model to perform multi-dimensional knowledge extraction on energy policy texts and output structured knowledge triples; constructing a dynamic multi-agent knowledge graph based on the structured knowledge triples; based on the dynamic multi-agent knowledge graph, performing deducing search on potential correlation paths through a first impact search algorithm to identify key policy impact transmission chains; establishing an event-driven knowledge graph closed-loop self-evolution mechanism to automatically capture policy change events in real time through automatic monitoring, triggering incremental update and structure correction of the dynamic multi-agent knowledge graph; constructing a scenario-oriented personalized application service interface based on the dynamic multi-agent knowledge graph, converting user natural language queries into graph query instructions, and generating scenario-based responses in combination with search results.

[0008] As a preferred scheme of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis, wherein: the multi-dimensional knowledge extraction on energy policy texts by the first adaptive extraction model comprises: constructing a first adaptive extraction model based on a pre-trained language encoder, the first adaptive extraction model comprising a shared encoding layer and an entity extraction head, a semantic relationship extraction head and a logical relationship extraction head connected to the shared encoding layer, respectively; identifying policy subjects, policy objects and policy tools in policy texts through the entity extraction head; judging the semantic relationship types between entities through the semantic relationship extraction head; identifying deep logical relationships between entity pairs through the logical relationship extraction head and outputting corresponding logical relationship confidence.

[0009] As a preferred scheme of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis, wherein: the construction of the dynamic multi-agent knowledge graph comprises: extending each triple in the structured knowledge triples into a five-tuple form containing timestamp information, and constructing a knowledge graph data model containing a time dimension based on the extended five-tuples; establishing a time-series attribute graph model for the subject nodes of the knowledge graph, storing the static and dynamic attributes of the nodes, persistently storing the static attributes as the basic attributes of the nodes, and attaching the dynamic attributes as attribute lists with timestamps to the corresponding nodes; record the state changes and behavior feedback of the subject nodes at different time points through the timestamp information; Based on the time sequence attribute graph model, a dynamic multi-agent knowledge graph capable of reflecting a policy life cycle and evolution of subject behaviors is constructed.

[0010] As a preferred scheme of the dynamic knowledge graph construction and application method for energy policy influence conduction analysis, the method comprises the following steps: Based on the dynamic multi-agent knowledge graph, an influence conduction analysis query is received, and a starting node and a target node of path search are determined; All possible paths connecting the starting node and the target node are searched by using a first search strategy; An influence conduction score is calculated for each searched path, and the influence conduction score is aggregated based on single-step influence strengths of edges on the path; The all possible paths are sorted based on the influence conduction scores, and several paths with the highest influence conduction scores are selected as key policy influence conduction chains and output.

[0011] The preferred technical scheme has the beneficial effect that how a policy is conducted through a complex chain with multiple hops and multiple levels can be clearly revealed, and unprecedented decision-making insight and risk warning capability are provided for users.

[0012] As a preferred scheme of the dynamic knowledge graph construction and application method for energy policy influence conduction analysis, the method comprises the following steps: A policy information source list is configured; Policy information sources are monitored in real time through automatic monitoring, and policy addition, revision or abolition events are detected, and a polling detection and change comparison combination is used to identify an update state of a policy document; A detected policy change is encapsulated as a standardized event object, and a corresponding knowledge graph update operation is triggered according to an event type; For an added policy event, a knowledge extraction module is called for analysis and injection into the graph; for a revised policy event, a change content is located based on semantic comparison and updated affected nodes; and for an abolished policy event, an effective time attribute of a corresponding relationship is updated and marked as an invalid state.

[0013] The preferred technical scheme has the beneficial effect that costs of manual sorting and maintenance are significantly reduced, work efficiency and scalability of the system are greatly improved while ensuring real-time content.

[0014] As a preferred scheme of the dynamic knowledge graph construction and application method for energy policy influence conduction analysis, the method comprises the following steps: Construct a contextualized response framework based on search enhancement generation; The scenario-based response framework converts the natural language query input by the user into a structured graph query instruction, executes the graph query instruction on the dynamic multi-agent knowledge graph, retrieves relevant fact nodes, relation edges and logical relation paths, and forms a contextual knowledge set; The user's original query and the contextual knowledge set are organized into structured prompt information, and the structured prompt information is input into the generation engine model to obtain the final contextualized response content.

[0015] As a preferred embodiment of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis described in this invention, wherein: the first adaptive extraction model is trained and optimized end-to-end using a joint loss function, wherein the joint loss function... The formula is: in, The cross-entropy loss for entity extraction and the log-likelihood loss function of CRF are used. The cross-entropy loss function for relation classification. The cross-entropy loss function is a binary classification function for logical relationships. , , This represents the hyperparameters used to balance the importance of different tasks; The policy text is parsed based on the first adaptive extraction model that has been trained, and the output is a structured knowledge triple containing entities, semantic relations and logical relations.

[0016] Secondly, this invention provides a dynamic knowledge graph construction and application system for analyzing the transmission of the impact of energy policies, including: The multidimensional knowledge extraction module is used to extract multidimensional knowledge from energy policy texts using the first adaptive extraction model and output structured knowledge triples. The knowledge graph construction module is used to construct dynamic multi-agent knowledge graphs based on structured knowledge triples. The transmission path deduction module is used to deduce and search for potential related paths based on the dynamic multi-agent knowledge graph and through the first influence search algorithm to identify the transmission chain of key policy influences; The self-evolution update module is used to establish an event-driven closed-loop self-evolution mechanism for the knowledge graph. It captures policy change events in real time through automated monitoring, triggering incremental updates and structural corrections of the dynamic multi-agent knowledge graph. The scenario service module is used to construct scenario-oriented personalized application service interfaces based on the dynamic multi-agent knowledge graph, convert user natural language queries into graph query instructions, and generate scenario-based responses based on the search results.

[0017] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of a dynamic knowledge graph construction and application method for energy policy impact transmission analysis.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of a method for constructing and applying a dynamic knowledge graph for energy policy impact transmission analysis.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention, through a domain-adaptive multi-task joint extraction method, can systematically analyze the hierarchical and logical relationships between policies, regulations, detailed rules, and measures, overcoming the shortcomings of traditional methods that can only obtain fragmented knowledge pieces, and constructing a semantically consistent knowledge system. This invention introduces a time dimension to construct a knowledge graph, which can accurately depict the complete lifecycle of policies, including their release, revision, and repeal, as well as the dynamic behavioral changes of various entities under the influence of policies, solving the problem of static relationship networks in existing technologies that cannot reflect the evolution of the real industrial ecosystem. The innovative dynamic impact path quantification and deduction algorithm can clearly reveal how policies are transmitted through complex multi-hop, multi-level chains, providing users with unprecedented decision-making insights and risk warning capabilities, filling the gap in existing technologies that lack deep logical reasoning mechanisms. Based on a retrieval-enhanced generation application framework, this invention can accurately understand users' natural language query intentions in specific business scenarios and provide intelligent answers anchored to facts and traceable sources. Furthermore, the event-driven closed-loop self-evolution mechanism provided by this invention automates the entire process of large-scale policy data collection, processing, and knowledge updating, significantly reducing the cost of manual organization and maintenance, and greatly improving the system's efficiency and scalability while ensuring the real-time nature of the content. Attached Figure Description

[0020] 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.

[0021] Figure 1This is a schematic diagram illustrating the overall process logic of a dynamic knowledge graph construction and application method for energy policy impact transmission analysis, provided as an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for constructing and applying dynamic knowledge graphs for energy policy impact transmission analysis is provided, such as... Figure 1 The specific steps shown are as follows: S100: The first adaptive extraction model is used to extract multidimensional knowledge from energy policy texts and output structured knowledge triples. S200: Constructing a dynamic multi-agent knowledge graph based on structured knowledge triples; S300: Based on a dynamic multi-agent knowledge graph, it uses the first influence search algorithm to infer and search for potential related paths, and identify the transmission chain of key policy influences; S400: Establish an event-driven closed-loop self-evolution mechanism for knowledge graphs, and capture policy change events in real time through automated monitoring to trigger incremental updates and structural corrections of dynamic multi-agent knowledge graphs; S500: Based on a dynamic multi-agent knowledge graph, it constructs a scenario-oriented personalized application service interface, converts user natural language queries into graph query commands, and generates scenario-based responses based on the search results.

[0024] It should be noted that, to address the problems of fragmented knowledge extraction, static relationship networks, lack of deep logical reasoning, and lagging updates in existing energy policy knowledge graph application management technologies, steps S100-S500 above utilize a domain-adaptive multi-task joint extraction method. This method systematically analyzes the hierarchical and logical relationships between policies, regulations, detailed rules, and measures, overcoming the shortcomings of traditional methods that can only acquire scattered knowledge fragments and constructing a semantically consistent knowledge system. This invention introduces a time dimension to construct the knowledge graph, accurately depicting the complete lifecycle of policies, including their release, revision, and repeal, as well as the dynamic behavioral changes of various entities under the influence of policies. This solves the problems of static relationship networks and the inability to reflect the real evolution of the industrial ecosystem in existing technologies. The innovative dynamic impact path quantification and deduction algorithm clearly reveals how policies are transmitted through complex multi-hop, multi-level chains, providing users with unprecedented decision-making insights and risk warning capabilities, filling the gap in existing technologies that lack deep logical reasoning mechanisms. Based on a retrieval-enhanced generation application framework, this invention can accurately understand users' natural language query intentions in specific business scenarios and provide fact-based, traceable, and intelligent answers. Furthermore, the event-driven closed-loop self-evolution mechanism provided by this invention automates the entire process of large-scale policy data collection, processing, and knowledge updating, significantly reducing the cost of manual organization and maintenance, and greatly improving the system's efficiency and scalability while ensuring the real-time nature of the content.

[0025] In this embodiment of the invention, step S100 above uses a first adaptive extraction model to extract multidimensional knowledge from energy policy text and outputs structured knowledge triples, including: Specifically, approximately 50,000 policy documents published on the official websites of national and local energy bureaus and national development and reform commissions since 2010 were crawled. After automated text cleaning, deduplication, and format standardization, a domain-specific corpus was constructed. ; Specifically, BERT was chosen as the base encoder, and in Domain-adaptive fine-tuning is performed to obtain a first adaptive extraction model capable of deeply understanding the semantics of energy policies. This first adaptive extraction model is then used to extract multi-dimensional knowledge from energy policy texts. Detailed steps include: The first adaptive extraction model is constructed based on a pre-trained language encoder. The first adaptive extraction model includes a shared encoding layer and an entity extraction head, a semantic relation extraction head, and a logical relation extraction head, which are respectively connected to the shared encoding layer. Identify policy subjects, policy objects, and policy instruments in policy texts using entity extraction heads; The semantic relationship type between entities is determined by extracting the semantic relationship header. The logical relationship extraction head identifies the deep logical relationships between entity pairs and outputs the corresponding logical relationship confidence level.

[0026] It should be noted that the entity extraction head uses a Conditional Random Field (CRF) layer for sequence labeling to identify predefined entity types, including policy subjects (such as government agencies and enterprises), policy objects (such as industries and projects), and policy tools (such as subsidies and taxes); the semantic relationship extraction head uses a relationship classifier, based on entity pair representation vectors, and uses a linear layer and a Softmax function to determine semantic relationships; the logical relationship extraction head uses an attention-based relationship classifier to specifically identify deep logical relationships between entity pairs and output the corresponding logical relationship confidence.

[0027] Specifically, the first adaptive extraction model is trained and optimized end-to-end using a joint loss function. The formula is: in, The cross-entropy loss for entity extraction and the log-likelihood loss function of CRF are used. The cross-entropy loss function for relation classification. The cross-entropy loss function is a binary classification function for logical relationships. , , This represents the hyperparameters used to balance the importance of different tasks; , , The value is determined by using a grid search on the reserved validation set.

[0028] Specifically, the policy text is parsed based on the first adaptive extraction model that has been trained, and the output is a structured knowledge triple containing entities, semantic relations and logical relations.

[0029] In an optional embodiment, the first adaptive extraction model can also be a text-to-text generative extraction framework based on a pre-trained language model. By unifying entity recognition, relation extraction, and logical relation judgment into a sequence generation task, and leveraging the powerful encoding-decoding capabilities of the T5 model, it directly inputs policy text paragraphs and generates structured triple sequences, thereby avoiding the error accumulation problem of traditional pipeline models and achieving end-to-end knowledge extraction.

[0030] In another alternative embodiment, the first adaptive extraction model can also be a joint extraction architecture based on the fusion of graph neural networks and pre-trained models. The pre-trained model is used to obtain the contextual representation of the text, then a document-level word graph is constructed and a graph attention network is applied to capture the global dependencies between word nodes. Entities and relationships are decoded synchronously through a graph reasoning mechanism.

[0031] It should be noted that step S100 above systematically analyzes the hierarchical and logical relationships between policies, regulations, rules and measures through a domain-adaptive multi-task joint extraction method. This overcomes the shortcomings of traditional methods that can only obtain fragmented knowledge pieces and significantly improves the accuracy and completeness of extracting entities, semantic relationships and deep logical relationships in policy texts.

[0032] In this embodiment of the invention, step S200, which involves constructing a dynamic multi-agent knowledge graph based on structured knowledge triples, includes: Each triple in the structured knowledge triple is expanded into a quintuple containing timestamp information, and a knowledge graph data model containing the time dimension is constructed based on the expanded quintuples. Establish a temporal attribute graph model for the main nodes of the knowledge graph, store the static and dynamic attributes of the nodes, persist the static attributes as the basic attributes of the nodes, and attach the dynamic attributes as a list of attributes with timestamps to the corresponding nodes. The system uses timestamp information to record the status changes and behavioral feedback of the main node at different points in time. A dynamic multi-agent knowledge graph that reflects the policy life cycle and the evolution of agent behavior can be constructed based on a time-series attribute graph model.

[0033] Specifically, a knowledge graph is defined as a tuple containing a time dimension. Where N is the set of nodes, E is the set of edges, and T is the time axis. Each extracted knowledge triple (head entity, relation, tail entity) is expanded into a tuple containing a timestamp when injected into the graph database, formalized as: in, and Don't use the beginning and end nodes. For relation types, and These represent the effective and expiration dates of the relationship, automatically extracted from the policy text (e.g., publication date, effective clauses). If the expiration date is not explicitly specified, the default value is used. .

[0034] Specifically, for the main nodes in the graph The node's static attributes are represented using a temporal property graph model. (Such as enterprise size and type) are stored as its basic attributes. Its dynamic attributes... (For example, behavioral status at different points in time, policy feedback), these are directly appended to the node as a list of attributes with timestamps.

[0035] Specifically, when the dynamic changes of a node are recorded as a series of events or state changes, for example: {attribute: 'behavior state', value: 'project A', timestamp: '2023-04-04'}, when it is necessary to obtain the node state at a specific time t, it is determined by directly querying the valid attribute values ​​at that time point.

[0036] In an optional embodiment, the step of constructing a dynamic multi-agent knowledge graph can also be based on the dynamic graph construction method of event-series networks. By modeling key nodes such as policy release, revision and subject behavior as event nodes, causal and temporal dependency edges between events are established to form a policy evolution network centered on events, thereby intuitively showing the transmission process and key turning points of policy impact on the timeline.

[0037] In another optional embodiment, the step of constructing a dynamic multi-agent knowledge graph can also adopt a graph management mode that combines versioned snapshots and incremental updates. Time slice snapshots of the knowledge graph are generated periodically, while the differences between adjacent versions are recorded through change streams. This preserves the historical picture of policy evolution and supports quick backtracking of the graph state at any point in time.

[0038] It should be noted that the above step S200 can accurately depict the complete life cycle of the policy and the behavioral evolution of each subject under the influence of the policy, solve the problem of the static nature of traditional knowledge graph relationship networks, and form a dynamic multi-subject knowledge network that can trace history and reflect real-time changes.

[0039] In this embodiment of the invention, step S300, based on a dynamic multi-agent knowledge graph, uses a first influence search algorithm to perform deductive search on potential related paths to identify key policy influence transmission chains, including: Based on dynamic multi-agent knowledge graph reception influence transmission analysis query, the starting node and target node of path search are determined; The first search strategy is used to search for all possible paths connecting the starting node and the target node; For each path found, an influence propagation score is calculated, which is obtained by aggregating the single-step influence strength of each edge on the path; Based on the impact transmission score, all possible paths are sorted, and the paths with the highest impact transmission scores are selected as the key policy impact transmission chain outputs.

[0040] In an optional embodiment, the first influence search algorithm can also be a probabilistic exploration algorithm based on Monte Carlo tree search, which simulates the random walk process of policy influence in the knowledge graph, counts the access frequency and reward value of each path, and gradually focuses on the discovery of high-value transmission paths.

[0041] In another optional embodiment, the first influence search algorithm can also adopt a path search framework that combines multi-objective optimization and genetic algorithm, which transforms multiple evaluation indicators such as path length, relationship strength, and timeliness into optimization objectives, and generates and filters the optimal transmission path set through population evolution mechanism, which can simultaneously output diverse influence transmission schemes that meet different decision preferences.

[0042] Specifically, when the system receives a deduction query, for example from a starting node... (e.g., a policy) to the target node When performing impact analysis on a specific type of enterprise, the algorithm is activated. The algorithm employs a beam search strategy to find all possible paths connecting the starting and ending points in the graph.

[0043] In an optional embodiment, the first search strategy can also be a hybrid search strategy based on bidirectional breadth-first search and heuristic evaluation. By simultaneously expanding the search from the starting node and the target node, using a heuristic function constructed from policy domain knowledge to evaluate the importance of nodes, and performing path fusion when they meet, the search efficiency and solution quality of long chain transmission relationships can be significantly improved.

[0044] In another alternative embodiment, the first search strategy may also employ a graph mining algorithm that combines random walks with personalized PageRank. By simulating multiple rounds of random walks with restarts, the probability of influence between nodes is calculated. Combined with a topic-sensitive weight adjustment mechanism, the most representative policy influence transmission patterns are automatically identified, which is particularly suitable for discovering implicit cross-domain transmission paths.

[0045] Specifically, to identify the most critical transmission chains from a massive number of paths, the algorithm calculates an Impact Propagation Score (IPS) for each path P. The IPS is calculated using the geometric mean, which is calculated from all edges along the path. Single-step influence strength It is formed by aggregation, that is: Where k represents the path length, i.e. the number of relation edges contained in the path, used to normalize the total influence strength of the path and prevent the score of long paths from being excessively weakened.

[0046] Specifically, the intensity of single-step influence It can then be calculated using a weighted formula: in, Representing the strength of the logical relationship, this score is directly output by the relation classifier of the logical relationship extraction head in step one, and the value range is [0,1]. It quantifies the certainty of the model's judgment on the association. The authority of the policy source that defines the relationship is quantified based on the level of the issuing agency (central = 1.0, ministry = 0.9, province = 0.7, city = 0.5, association = 0.4). This represents the temporal density of adjacent events on the chain. This value is dynamically calculated at each step of the path search, measuring the current edge to be expanded. With an edge on the path The temporal continuity between them is derived from a decay function: in, The time difference (in days) between the start of the next event and the end of the previous event. It is a time decay constant. =30, ensuring that the path edges are arranged in chronological order; and , , These are preset weight hyperparameters, whose values ​​are optimized using methods such as grid search on a small-scale expert-annotated validation set to ensure the reasonableness of the scores.

[0047] Specifically, the algorithm returns the top few paths with the highest IPS scores and uses them as the most likely policy impact transmission chains to provide a basis for decision-making.

[0048] It should be noted that step S300 above systematically explores the multi-hop and multi-level transmission paths of policies through cluster search strategies and impact transmission score quantification mechanisms. It can clearly reveal the chain effects of policy adjustments on upstream and downstream entities, provide interpretable penetrating reasoning analysis, and provide users with in-depth decision-making insights and forward-looking risk warnings, thus filling the gap in the lack of logical reasoning capabilities in existing technologies.

[0049] In this embodiment of the invention, step S400 establishes an event-driven closed-loop self-evolution mechanism for the knowledge graph, which triggers incremental updates and structural corrections of the dynamic multi-agent knowledge graph by automatically monitoring and capturing policy change events in real time. Configure a list of policy information sources; The policy information source is monitored in real time through automated monitoring to detect policy additions, revisions or repeals, and the update status of policy documents is identified by a combination of polling detection and change comparison. The detected policy changes are encapsulated as standardized event objects, and the corresponding knowledge graph update operation is triggered according to the event type. For new policy events, the knowledge extraction module is invoked to parse them and inject them into the graph; for revised policy events, the changed content is located based on semantic comparison and the affected nodes are updated; for repealed policy events, the effective time attribute of the corresponding relationship is updated and marked as invalid.

[0050] In an optional embodiment, the knowledge graph update step can also adopt a continuous update architecture based on streaming processing and real-time computing engines. By connecting the data stream of the policy release platform to a distributed message queue, the stream processing framework is used to parse policy document changes in real time, and the knowledge graph topology is dynamically updated in combination with incremental graph computing algorithms, so as to achieve real-time knowledge synchronization and evolution with sub-second latency.

[0051] In another optional embodiment, the knowledge graph update step can also adopt an intelligent update strategy that combines collaborative filtering and proactive discovery. By analyzing user query patterns and policy relevance, potential knowledge update needs can be predicted. At the same time, web crawlers can be used to proactively discover sources of policies that have not been included in the database. Combined with a confidence assessment mechanism, high-value update content can be automatically screened, forming a dual-cycle update system of demand-driven and proactive mining.

[0052] Specifically, a list of policy information sources is pre-configured during deployment, including the official websites and policy disclosure platforms of the National Energy Administration, the National Development and Reform Commission, the Ministry of Ecology and Environment, and provincial energy management departments. Monitoring operates using a combination of polling detection and change comparison, including: ① The system is set to automatically access the policy directory page at fixed time intervals to obtain a structured summary of the page (title, publication date, document number, link, etc.); ② Compare with the latest entry in the local cache index. If a newly added or modified document number is detected, it is determined to be a new event. ③ The system can also receive policy updates in real time through the RSS subscription interface or web push API and automatically download new documents.

[0053] Specifically, the event structure is defined as: Event = (event_type, source, doc_id, timestamp, content_vector, affected_nodes), where event_type represents the event type (new, revised, or deprecated), content_vector is the semantic vector representation of the document, and affected_nodes is the set of nodes that may be affected.

[0054] Specifically, when the event type is "new," the system automatically calls the knowledge extraction module to parse the new policy and inject it into the graph. When the event type is "revision," the system performs semantic comparison based on historical document versions to locate the changed content and only updates the affected nodes. When the event type is "repealed," the system does not delete the relevant nodes but updates the effective time attribute of the corresponding relationship to the event timestamp and marks it as invalid to preserve the policy's historical trajectory. The semantic comparison algorithm uses a combination of Sentence-BERT semantic vector matching and tree edit distance. The system first performs similarity filtering at the sentence level and then identifies the location of clause changes through syntactic tree structure alignment, thereby achieving minimal and accurate updates. The monitoring module and the event-driven mechanism are connected through a unified message queue interface to ensure full automation and real-time performance from monitoring to triggering updates.

[0055] It should be noted that step S400 above realizes the fully automated updating and maintenance of the knowledge graph, which greatly reduces the cost of manual intervention, ensures the timeliness, accuracy and traceability of the knowledge system, and significantly improves the system's practicality and sustainable development capabilities.

[0056] In this embodiment of the invention, step S500, which involves constructing a scenario-oriented personalized application service interface based on a dynamic multi-agent knowledge graph, converting user natural language queries into graph query instructions, and generating scenario-based responses based on the search results, includes: Construct a contextualized response framework based on search enhancement generation; The contextualized response framework converts the natural language query input by the user into a structured graph query instruction, executes the graph query instruction on a dynamic multi-agent knowledge graph, retrieves relevant fact nodes, relation edges and logical relation paths, and forms a contextual knowledge set; The user's original query and the set of contextual knowledge are organized into structured prompts, and the structured prompts are input into the generation engine model to obtain the final contextualized response content.

[0057] Specifically, when a user inputs a natural language question At this stage, the contextualized response framework based on retrieval enhancement first enters the retrieval phase. An intent parser is achieved using a T5 (Text-to-Text Transfer Transformer) model fine-tuned for the Text-to-GraphQuery task. , will Convert to structured graph query commands .

[0058] Specifically, a semi-automated parallel corpus generation strategy is used to fine-tune the model. First, a series of graph query templates are defined; then, a large language model is used to inversely generate diverse natural language questions based on the templates; finally, the generated data is manually filtered and corrected to form a training set containing thousands of high-quality samples. Next, after the command is executed on the knowledge graph, it retrieves the fact nodes, relation edges, and logical relationship paths deduced from step three that are most relevant to the question, forming a contextual knowledge set. .

[0059] Specifically, the user's original question With retrieved contextual knowledge Together they form a structured prompt, which is then fed into the generation engine. In Google Gemini or other alternative large-scale generative models (or equivalent on-premises models). The model leverages its powerful natural language understanding and content generation capabilities to ultimately answer... This is generated from, that is: in, This refers to the structured prompts input to the generative model, including user questions, contextual knowledge retrieved from the graph, and task constraints, which guide the model to generate factually based answers.

[0060] It should be noted that the above step S500 is based on a retrieval enhancement generation framework, which accurately parses natural language queries into graph instructions and integrates retrieval knowledge to generate intelligent answers, thereby upgrading the service model and providing users with decision support that is anchored to facts, traceable, and tailored to business scenarios. This greatly improves the intelligence level of policy knowledge services and the user experience.

[0061] Example 2, based on the previous example, provides an application example of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis, to verify and illustrate the technical effects of the method.

[0062] This embodiment selects 12 renewable energy electricity price subsidy policies issued by national and local energy authorities between 2015 and 2024 as experimental data, including the "Notice on Improving Wind Power On-Grid Tariff Policies" and the "Notice on Leveraging Price Mechanisms to Promote the Healthy Development of the Photovoltaic Industry Chain." These policy documents total more than 500,000 words, covering sub-sectors such as wind power, photovoltaics, and biomass power generation, and the policy tools encompass various types, including fixed electricity prices, competitive bidding subsidies, and tax reductions.

[0063] After parsing the aforementioned policies using the domain-adaptive multi-task joint extraction algorithm of this invention, a total of 2843 entities were identified, including policy subjects, policy objects, and policy tools; 3517 semantic relations were extracted, and 892 deep logical relations were identified. To quantify the extraction effect, a comparative experiment was conducted on a manually labeled test set: the existing technology (rule + NER) achieved F1 scores of 76.3% and 68.9% for entity recognition and relation extraction, respectively, and could not identify logical relations; while this invention achieved F1 scores of 93.5%, 90.2%, and 87.8% for these three tasks, respectively, significantly improving the completeness and accuracy of knowledge extraction.

[0064] Based on the extraction results, a temporal knowledge graph containing 4215 nodes and 5602 edges was constructed. Each relation edge is accompanied by an effective and ineffective timestamp, and the dynamic attributes of the main nodes are continuously updated over time, forming a dynamic network that reflects the policy lifecycle and changes in the behavior of the main entities. In the impact transmission analysis, this embodiment uses "how the adjustment of wind power subsidy policy affects the end-user electricity cost" as a query case. The system successfully identified multiple transmission paths through dynamic impact path search and quantitative scoring algorithms. The most representative path clearly shows the complete transmission chain from the subsidy policy to wind power company investment, installed capacity growth, and a decrease in grid-connected electricity price, ultimately affecting the electricity cost of industrial users. The IPS score of this path is 0.82, and the system also returns a quantitative assessment and visualization of the impact strength. This result shows that the present invention can reveal multi-hop, cross-level transmission mechanisms that traditional methods cannot present.

[0065] Finally, through the RAG-based intelligent question-and-answer interface, users can obtain answers that combine accurate facts and deep logical relationships. For example, when a wind power equipment manufacturing company inquired about the impact of policies on R&D investment, the system not only provided a quantitative prediction of "an average annual increase of 8%-12% in R&D investment," but also clarified the transmission path of "subsidies → project investment → technology upgrades," realizing a service upgrade from "people seeking policies" to "policies finding people."

[0066] Example 3: This example provides a dynamic knowledge graph construction and application system for analyzing the transmission of energy policy impacts, including: The multidimensional knowledge extraction module is used to extract multidimensional knowledge from energy policy texts using the first adaptive extraction model and output structured knowledge triples. The knowledge graph construction module is used to construct dynamic multi-agent knowledge graphs based on structured knowledge triples. The transmission path inference module is used to infer and search for potential related paths based on a dynamic multi-agent knowledge graph and the first influence search algorithm to identify the transmission chain of key policy influences. The self-evolution update module is used to establish an event-driven closed-loop self-evolution mechanism for knowledge graphs. It captures policy change events in real time through automated monitoring, triggering incremental updates and structural corrections of dynamic multi-agent knowledge graphs. The scenario service module is used to build scenario-oriented personalized application service interfaces based on dynamic multi-agent knowledge graphs, convert user natural language queries into graph query instructions, and generate scenario-based responses based on search results.

[0067] It should be noted that the technical solution of the dynamic knowledge graph construction and application system for energy policy impact transmission analysis is based on the same concept as the technical solution of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis described above. For details not described in detail in the technical solution of the dynamic knowledge graph construction and application system for energy policy impact transmission analysis in this embodiment, please refer to the description of the technical solution of the dynamic knowledge graph construction and application method for energy policy impact transmission analysis described above.

[0068] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0069] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for constructing and applying a dynamic knowledge graph for analyzing the transmission of energy policy impacts. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0070] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0071] The storage medium proposed in this embodiment belongs to the same inventive concept as the method 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.

[0072] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0073] 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 constructing and applying a dynamic knowledge graph for energy policy impact transmission analysis, characterized in that, The method comprises the following steps: adopting a first adaptive extraction model to perform multi-dimensional knowledge extraction on energy policy texts, and outputting structured knowledge triples; constructing a dynamic multi-agent knowledge graph based on the structured knowledge triples; based on the dynamic multi-agent knowledge graph, performing deductive search on potential associated paths through a first influence search algorithm, and identifying key policy influence transmission chains; establishing an event-driven knowledge graph closed-loop self-evolution mechanism, capturing policy change events in real time through automatic monitoring, triggering incremental update and structure correction of the dynamic multi-agent knowledge graph; constructing scenario-oriented personalized application service interfaces based on the dynamic multi-agent knowledge graph, converting user natural language queries into graph query instructions, and generating scenario-based responses combined with search results. 2.The method of claim 1, wherein, The method comprises the following steps: based on a pre-trained language encoder, a first adaptive extraction model is constructed, which comprises a shared encoding layer and entity extraction heads, semantic relationship extraction heads and logical relationship extraction heads connected to the shared encoding layer respectively; through the entity extraction heads, policy subjects, policy objects and policy tools in the policy texts are identified; through the semantic relationship extraction heads, the semantic relationship types between entities are determined; through the logical relationship extraction heads, deep logical relationships between entity pairs are identified, and corresponding logical relationship confidence is outputted. 3.The method of claim 2, wherein, The method comprises the following steps: each triple in the structured knowledge triples is expanded into a five-tuple form containing timestamp information, and a knowledge graph data model containing a time dimension is constructed based on the expanded five-tuples; a time-series attribute graph model is established for the subject nodes of the knowledge graph, and the static attributes and dynamic attributes of the nodes are stored, the static attributes are used as the basic attributes of the nodes for persistent storage, and the dynamic attributes are attached to the corresponding nodes as attribute lists with timestamps; through the timestamp information, the state changes and behavior feedbacks of the subject nodes at different time points are recorded; based on the time-series attribute graph model, a dynamic multi-agent knowledge graph reflecting the policy life cycle and the evolution of subject behavior is constructed. 4.The method of claim 3, wherein, The method comprises the following steps: based on the dynamic multi-agent knowledge graph, an influence transmission analysis query is received, and the starting node and the target node of the path search are determined; a first search strategy is adopted to search all possible paths connecting the starting node and the target node; for each path searched, an influence transmission score is calculated, which is aggregated based on the single-step influence strength of each edge in the path; based on the influence transmission scores, the all possible paths are sorted, and the several paths with the highest influence transmission scores are selected as the key policy influence transmission chains and outputted. 5.The method of claim 4, wherein, The method comprises the following steps: a policy information source list is configured; through automatic monitoring, the policy information sources are monitored in real time, policy addition, revision or abolition events are detected, and a combination of polling detection and change comparison is adopted to identify the update status of the policy documents; The detected policy change is packaged into a standardized event object, and a corresponding knowledge graph update operation is triggered according to the event type; For a new policy event, the knowledge extraction module is called for analysis and injection into the graph; for a revised policy event, the changed content is located based on semantic comparison and the affected nodes are updated; for a repealed policy event, the validity time attribute of the corresponding relationship is updated and marked as invalid. 6.The method of claim 5, wherein, The construction of the scenario-oriented personalized application service interface includes: Constructing a scenario-based response framework generated based on retrieval enhancement; The scenario-based response framework converts the natural language query input by the user into a structured graph query instruction, executes the graph query instruction on the dynamic multi-agent knowledge graph, retrieves related fact nodes, relationship edges and logical relationship paths to form a context knowledge set; The user's original query and the context knowledge set are jointly organized into structured prompt information, and the structured prompt information is input into the generation engine model to obtain the final scenario-based response content.

7. The dynamic knowledge graph construction and application method for energy policy impact conduction analysis according to claim 2, wherein, The first adaptive extraction model is trained and optimized end-to-end using a joint loss function, the joint loss function is given by the formula: wherein, is the cross-entropy loss function for entity extraction, is the cross-entropy loss function for relation classification, is the binary cross-entropy loss function for logical relation, , , denotes a hyper-parameter for balancing the importance of different tasks; Based on the first adaptive extraction model trained, the policy text is parsed to output structured knowledge triples containing entities, semantic relationships and logical relationships.

8. The dynamic knowledge graph construction and application system for energy policy impact transmission analysis, applying the dynamic knowledge graph construction and application method for energy policy impact transmission analysis according to any one of claims 1-7, characterized in that, It includes: A multi-dimensional knowledge extraction module is used to extract multi-dimensional knowledge from energy policy text using a first adaptive extraction model and output structured knowledge triples; A knowledge graph construction module is used to construct a dynamic multi-agent knowledge graph based on structured knowledge triples; A conduction path deduction module is used to search for potential associated paths based on the dynamic multi-agent knowledge graph through a first influence search algorithm to identify key policy influence transmission chains; A self-evolution update module is used to establish an event-driven knowledge graph closed-loop self-evolution mechanism to automatically capture policy change events in real time and trigger incremental updates and structural corrections of the dynamic multi-agent knowledge graph; A scenario service module is used to construct a scenario-oriented personalized application service interface based on the dynamic multi-agent knowledge graph, convert natural language queries from users into graph query instructions, and generate scenario-based responses in combination with retrieval results. 9.An electronic device comprising a memory and a processor, the electronic device characterized by: The memory is used to store computer executable instructions, and the processor executes the computer executable instructions to realize the steps of the dynamic knowledge graph construction and application method for energy policy influence transmission analysis according to any one of claims 1-7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: The computer executable instructions are executed by the processor to realize the steps of the dynamic knowledge graph construction and application method for energy policy influence transmission analysis according to any one of claims 1-7.

Images