Knowledge graph-based power dispatch operation risk linkage modeling method and system

Abstract

The application discloses a power dispatching operation risk linkage modeling method and system based on a knowledge graph, and belongs to the technical field of power system automation, which comprises the following steps: constructing an electric power business semantic field and a physical topology graph; generating a business semantic vector facing topology anchoring by using a neural network; identifying a risk event and encapsulating it as a risk logic cluster containing an influence chain and disposal knowledge; linking the risk logic to the topology graph by a bidirectional coupling engine to generate a risk linkage knowledge graph; and capturing artificial intervention decision differences based on visual deduction to realize double-space co-evolution closed loop. The application adopts semantic recognition and topology model deep fusion technology, can break through the digital linkage channel of risks and equipment and users, realizes dynamic risk perception and penetrating analysis, and improves the timeliness and accuracy of power dispatching in response to real-time operation risks.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and in particular to a power dispatching operation risk linkage modeling method and system based on a knowledge graph. Background Art

[0002] Power dispatching operation is the core link to ensure the safe, stable and economic operation of the power grid, and its risk management is crucial for preventing large-scale power outages. During the operation of the power system, a large amount of multi-source heterogeneous data will be generated, mainly including structured data represented by the power grid topology structure, equipment inventory, and real-time measurement data, as well as unstructured text data represented by dispatching logs, alarm information, defect records, and operation tickets. How to effectively utilize this data for accurate risk assessment and linkage analysis is an important topic in the field of power dispatching.

[0003] In the prior art, power dispatching risk analysis usually relies on multiple independent systems and manual processes. Dispatchers mainly obtain real-time power grid operation parameters and alarm information through monitoring automation systems such as energy management systems. When abnormalities occur, they need to conduct manual research and judgment by combining text information such as dispatching logs and historical defect records, and analyze potential risks through offline simulation tools such as power flow calculations. These systems and data sources are separated from each other, and the identification and assessment of risks largely depend on the personal professional knowledge and work experience of dispatchers.

[0004] The prior art solutions have several technical defects. First, the structured power grid topology data and the unstructured risk text data are in a separated state, lacking effective technical means to automatically associate fuzzy text descriptions with precise physical topology entities, resulting in insufficient utilization of risk information. Second, the impact analysis of risks is mostly static and local, and it is difficult to dynamically and completely reveal the conduction path and linkage effect of risks in a complex power grid. Finally, the analysis rules or models固化 in the system are updated laggingly, unable to effectively precipitate and reuse the disposal experience of dispatch experts, and lacking the ability of self-learning and continuous optimization. Summary of the Invention

[0005] To solve the above problems, the present invention provides a power dispatching operation risk linkage modeling method and system based on a knowledge graph, which can achieve in-depth fusion modeling and intelligent evolution of power dispatching risks by constructing a dual space of physical topology and business semantics and establishing a human-machine feedback closed loop, and improve the scientificity of risk situation awareness and disposal decision-making.

[0006] The above objectives can be achieved through the following solutions:

[0007] A knowledge graph-based method for modeling power dispatching and operation risks includes: acquiring structured topology data and unstructured risk text data of the power system, and constructing a power business semantic field based on the business context and historical handling cases in the unstructured risk text data; constructing a power grid physical topology graph containing equipment entity nodes and connection relationships based on the structured topology data, wherein the equipment entity nodes in the power grid physical topology graph are configured with dynamic attribute slots and logical association interfaces; using a neural network model that integrates the rules of the power business semantic field to vectorize the unstructured risk text data, generating business semantic vectors anchored to the physical topology; and identifying risk events and generating corresponding risk logic clusters based on the business semantic vectors using a preset risk event extraction model. The risk logic cluster includes a risk subject, an impact chain, and a disposal knowledge package. The risk logic cluster is linked to the power grid physical topology map via entity alignment. Through a bidirectional coupling engine, the impact chain logic of the risk logic cluster is injected into the logical association interface, and dynamic risk attribute values ​​are written into the corresponding dynamic attribute slots to generate a risk linkage knowledge graph. Based on the risk linkage knowledge graph, risk visualization simulation is performed to obtain the dispatcher's manual intervention actions based on the simulation results, and the decision differences between the manual intervention actions and the system simulation results are captured. These decision differences are simultaneously used as a first feedback sample and a second feedback sample. The first feedback sample is used to optimize the parameter update of the neural network model and the generation of business semantic vectors, while the second feedback sample is used to optimize the derivation rules of the impact chain in the risk logic cluster.

[0008] Optionally, the acquisition of structured topology data and unstructured risk text data of the power system, and the acquisition of business context and historical handling cases in the unstructured risk text data, includes: acquiring a set of power monitoring alarm terms, a set of standard equipment names, a set of historical defect records, and a set of load anomaly patterns as basic corpus; analyzing the basic corpus to establish a triplet association rule base with power entities and risk events as lexical units, the contextual environment description of the lexical units as the first association, and the typical business consequences caused by the lexical units as the second association; and fusing the triplet association rule base with a general semantic model to generate a power business semantic field.

[0009] Optionally, vectorizing the unstructured risk text data using a neural network model that integrates the rules of the power business semantic field includes: inputting the unstructured risk text data into the neural network model for word segmentation and contextual analysis to obtain text units; labeling the text units with business context based on the power business semantic field to generate labeled text sequences; and training the neural network model with the vector representation of the set of equipment nodes affected by the labeled text sequence in the power grid physical topology map as the approximation target, so that the distance between its output vector representation and the target vector representation in the vector space is minimized, thereby generating business semantic vectors.

[0010] Optionally, identifying risk events and generating corresponding risk logic clusters based on the business semantic vector using a preset risk event extraction model includes: inputting the business semantic vector into the risk event extraction model to determine the risk type and the risk subject equipment; calling the logical association interface connected to the risk subject equipment in the power grid physical topology map to automatically deduce the affected upstream and downstream equipment and user paths, forming an impact chain; using the risk type and the risk subject equipment as search conditions, matching and extracting relevant handling step texts from the unstructured risk text data or contingency plan knowledge base to generate a handling knowledge package; and encapsulating the risk subject, the impact chain, and the handling knowledge package to generate a risk logic cluster.

[0011] Optionally, the step of injecting the influence chain logic of the risk logic cluster into the logical association interface and writing the dynamic risk attribute value into the corresponding dynamic attribute slot through the bidirectional coupling engine includes: the bidirectional coupling engine monitoring structural change events of the power grid physical topology map; when a structural change event is detected, triggering a recalculation of the influence chain of all attached risk logic clusters and updating the risk attribute value in the dynamic attribute slot; when a new risk logic cluster attachment event is detected, triggering attribute value injection and state update of the dynamic attribute slot of the relevant node in the power grid physical topology map along the influence chain; based on the updated dynamic attribute slot, rendering a risk heat map in real time on the visualization interface of the power grid physical topology map to form a risk linkage knowledge graph.

[0012] Optionally, when a new risk logic cluster connection event is detected, triggering attribute value injection and state update of the dynamic attribute slots of relevant nodes in the power grid physical topology map along the influence chain includes: obtaining an initial risk attribute value from a preset risk-attribute mapping table according to the risk type of the risk logic cluster; performing attenuation calculation on the initial risk attribute value according to the electrical distance and impedance parameters of the equipment entity nodes traversed by the influence chain to obtain the local risk attribute value of each node; writing the local risk attribute value into the dynamic attribute slot of the corresponding node; and aggregating and calculating the network-wide or regional risk indicators based on the values ​​of the dynamic attribute slots of all nodes.

[0013] Optionally, optimizing the generation of the business semantic vector for the neural network model by using the first feedback sample includes: converting the decision difference into a text description and forming a comparison sample pair with the original unstructured risk text data; using the comparison sample pair to perform comparative learning training on the neural network model, adjusting the model parameters so that the business semantic vector generated by the model for input containing the decision difference text is closer to the vector representation of the set of equipment nodes updated in the power grid physical topology map as pointed to by the manual intervention operation.

[0014] Optionally, optimizing the derivation rules of the influence chain in the risk logic cluster using the second feedback sample includes: parsing the decision differences and identifying the equipment entities that are additionally included or excluded by the manual intervention operation; analyzing the electrical connection relationship or protection logic relationship between the additionally included or excluded equipment entities and the risk subject equipment to obtain potential association logic; and based on the potential association logic, correcting or supplementing the association rules in the logical association interface, or adjusting the weight parameters of different connection types in the influence chain derivation process.

[0015] Optionally, risk visualization simulation based on the risk linkage knowledge graph includes: receiving simulated handling instructions for specific risk logic clusters in the risk linkage knowledge graph; simulating the operation of relevant equipment entity node states in the power grid physical topology graph in an independent sandbox environment according to the simulated handling instructions and the handling knowledge package, and performing power flow calculations to obtain the simulated state; calculating the propagation results of the simulated state along the influence chain through the bidirectional coupling engine, updating the full graph risk heat rendering, and outputting a risk comparison view and key indicator changes before and after the simulation operation.

[0016] Based on the same inventive concept, this invention also provides a knowledge graph-based power dispatching operation risk linkage modeling system. The system includes: a data acquisition and semantic field construction module, used to acquire structured topology data and unstructured risk text data of the power system, and construct a power business semantic field based on the business context and historical handling cases in the unstructured risk text data; a physical topology construction module, used to construct a power grid physical topology graph containing equipment entity nodes and connection relationships based on the structured topology data, wherein the equipment entity nodes in the power grid physical topology graph are configured with dynamic attribute slots and logical association interfaces; a semantic vector generation module, used to vectorize the unstructured risk text data using a neural network model that integrates the rules of the power business semantic field, generating business semantic vectors anchored to the physical topology; and a risk event extraction and encapsulation module, used to identify risk events and generate corresponding risk events based on the business semantic vectors using a preset risk event extraction model. The system comprises a risk logic cluster, wherein the risk logic cluster includes a risk subject, an impact chain, and a disposal knowledge package; a bidirectional coupling and linkage knowledge graph generation module, used to attach the risk logic cluster to the power grid physical topology graph through entity alignment, and inject the impact chain logic of the risk logic cluster into the logic association interface through a bidirectional coupling engine, and write the dynamic risk attribute values ​​into the corresponding dynamic attribute slots to generate a risk linkage knowledge graph; a visualization deduction and difference capture module, used to perform risk visualization deduction based on the risk linkage knowledge graph, obtain the dispatcher's manual intervention operation on the deduction result, and capture the decision difference between the manual intervention operation and the system deduction result; and a semantic optimization and rule optimization module, used to use the decision difference as both a first feedback sample and a second feedback sample, using the first feedback sample to optimize the parameter update of the neural network model to generate the business semantic vector, and using the second feedback sample to optimize the deduction rules of the impact chain in the risk logic cluster.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] This invention achieves deep integration and unified representation of structured topology data and unstructured risk text data by constructing a dual spatial model of power grid physical topology map and power business semantic field. It breaks down the barriers between different data sources and forms a holistic view that comprehensively reflects the risks of power grid operation, thereby improving the completeness and systematic nature of risk analysis.

[0019] This invention proposes an automated generation mechanism from semantic vectors to risk logic clusters, which can proactively deduce the impact chain of risk events and match disposal knowledge. Combined with risk visualization and simulation functions, it enables dispatchers to pre-assess the consequences of different disposal strategies, transforming risk management from post-event response to pre-event prediction and proactive intervention, thereby enhancing the scientific nature and foresight of dispatching decisions.

[0020] This invention designs a dual-space co-evolutionary closed loop of semantic space and logical rule space. By capturing and utilizing the differences between dispatcher expert decisions and system inference, it simultaneously optimizes the semantic understanding ability of the neural network and the inference rules of the influence chain, enabling the system to have the ability to continuously learn and iteratively optimize itself. It can continuously adapt to changes in the power grid and accumulate expert experience, thereby maintaining and improving the accuracy and intelligence level of risk modeling in the long term.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the knowledge graph-based power dispatching operation risk linkage modeling method according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the knowledge graph-based power dispatching operation risk linkage modeling system according to an embodiment of the present invention.

[0025] Figure 3 This is a risk decay curve diagram on the influence chain of an embodiment of the present invention.

[0026] Figure 4 This is a diagram showing the evolution of the business semantic vector space distribution according to an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Reference Figure 1One embodiment of the present invention proposes a knowledge graph-based method for linking power dispatching operation risks. By constructing a dual space of physical topology and business semantics and establishing a human-machine feedback loop, it can achieve deep integration modeling and intelligent evolution of power dispatching risks, thereby improving the scientific nature of risk situation perception and handling decisions.

[0029] The method described in this embodiment specifically includes:

[0030] Obtain structured topology data and unstructured risk text data of the power system, and construct a power business semantic field based on the business context and historical handling cases in the unstructured risk text data;

[0031] Based on the structured topology data, a power grid physical topology map containing device entity nodes and connection relationships is constructed, wherein the device entity nodes in the power grid physical topology map are configured with dynamic attribute slots and logical association interfaces.

[0032] The unstructured risk text data is vectorized using a neural network model that integrates the semantic field rules of the power business to generate a business semantic vector anchored to the physical topology.

[0033] Based on the business semantic vector, risk events are identified and corresponding risk logic clusters are generated through a preset risk event extraction model. The risk logic clusters include risk subjects, impact chains, and disposal knowledge packages.

[0034] The risk logic cluster is attached to the power grid physical topology graph through entity alignment, and the influence chain logic of the risk logic cluster is injected into the logic association interface through the bidirectional coupling engine. The dynamic risk attribute value is written into the corresponding dynamic attribute slot to generate a risk linkage knowledge graph.

[0035] Based on the risk linkage knowledge graph, risk visualization simulation is performed to obtain the dispatcher's manual intervention operation based on the simulation results, and the decision difference between the manual intervention operation and the system simulation results is captured.

[0036] The decision difference is used as both the first and second feedback samples. The first feedback sample is used to optimize the generation of the business semantic vector by updating the parameters of the neural network model, and the second feedback sample is used to optimize the derivation rules of the influence chain in the risk logic cluster.

[0037] Specifically, a dual-space model integrating power physical topology and business semantics was constructed, and the co-evolution of the two spaces was achieved through a closed-loop feedback mechanism. By acquiring structured topology data and unstructured risk text, a power grid physical topology map representing the physical entities of the power grid and a power business semantic field representing domain knowledge were constructed, respectively. Deep learning technology was used to map the unstructured text containing risk information into business semantic vectors anchored to the physical topology, and further decoded these vectors into structured risk logic clusters containing risk subjects, impact chains, and disposal knowledge. The core lies in dynamically attaching and injecting these logical-level risk clusters into the physical topology map through a bidirectional coupling engine, generating a knowledge graph that can reflect the risk linkage relationship in real time. A human-computer interactive deduction stage was introduced to capture the difference between system deduction and dispatcher expert decisions, and this difference was used as a dual feedback signal to simultaneously optimize the front-end semantic understanding model and the back-end logical deduction rules, thereby forming an intelligent closed-loop system capable of self-learning and self-improvement.

[0038] Optionally, the acquisition of structured topology data and unstructured risk text data of the power system, and the processing based on the business context and historical handling cases in the unstructured risk text data, includes:

[0039] The set of power monitoring alarm terms, the set of standard equipment names, the set of historical defect records, and the set of load anomaly patterns are used as the basic corpus.

[0040] Based on the analysis of the basic corpus, a triplet association rule base is established, which uses power entities and risk events as lexical units, the contextual environment of the lexical units as the first association, and the typical business consequences caused by the lexical units as the second association.

[0041] The triple association rule base is fused with a general semantic model for training to generate a power business semantic field.

[0042] Specifically, the dispersed, multi-source, and heterogeneous text data in the power dispatching field is transformed into a computable semantic space with professional domain knowledge, laying the foundation for subsequent accurate risk text vectorization processing. The basic corpus is acquired by collecting and integrating data from multiple business systems through data interfaces. The collected power monitoring alarm terminology set mainly comes from SCADA or EMS systems, typically containing alarm logs from the past 3 to 5 years, with data volumes reaching TB levels, such as standard alarm information like "35 kV bus voltage exceeds upper limit"; the equipment standard naming set comes from the equipment asset management system, ensuring entity uniqueness, such as "Chengnan Substation 220 kV No. 1 main transformer"; the historical defect record set comes from the operation and maintenance records in the production management system (PMS), which is unstructured text rich in contextual information, such as "Discharge phenomenon found in phase A insulator string of tower 32 on 110 kV Chengshang line during inspection"; and the load anomaly pattern set comes from the load forecasting or measurement system, containing descriptive text of specific patterns, such as "The evening peak load of the industrial park disappeared during holidays". After data collection, the construction of the triple association rule base was initiated. The implicit knowledge in the corpus was made explicit and structured. Using natural language processing technology, entity and event identification was first performed on the basic corpus, defining core concepts such as "220 kV line" and "transformer overload" as lexical units. A lexical unit is the smallest indivisible semantic unit in the power dispatching field, representing a specific equipment entity or a defined operating state. Next, through dependency parsing and co-occurrence statistics, the associations between lexical units were mined. The first association was described by the context of the lexical unit; for example, the first association of the lexical unit "line tripping" might be "thunderstorm weather." The second association was described by the typical business consequences caused by the lexical unit; for example, the second association of "line tripping" might be "downstream user power loss." This resulted in triple association rules in the form of <thunderstorm weather, impact, line tripping> and <line tripping, causing, user power loss>, which were stored in the rule base. Finally, fusion training was performed to generate the final power business semantic field. By deeply integrating general language knowledge with knowledge from the power industry, a professional vector space model is formed. In engineering, a power business semantic field is a specially trained neural network language model capable of mapping any power-related text fragment to specific coordinates in a high-dimensional vector space, ensuring that semantically similar texts have similar Euclidean distances within this space. A pre-trained general semantic model such as BERT is selected as the basic architecture, and a triple association rule base is used as knowledge constraints for fine-tuning the model. The training objective function consists of two parts, mathematically expressed as:

[0043] ;

[0044] in, This is the total loss function, and the goal of training is to minimize it; It is the standard language model loss of the general semantic model on the basic corpus, used to maintain the model's generalized language understanding ability. It is calculated through self-supervised tasks such as masked language modeling on the corpus. It is a knowledge graph embedding loss built on a triple association rule base. It forces the model to learn the professional associations defined in the rule base, such as making the vector representation of "line tripping" closer to the vector representation of "user power loss". This loss is usually obtained by calculating the distance or score function between the vectors of the head entity, relation and tail entity in the triple. This is a weight hyperparameter used to balance the contributions of the two parts of the loss. Its value typically ranges from 0.1 to 1.0 and is determined through experimental tuning. The goal is to minimize the total loss function. Through iterative training, the final model obtained is the semantic field of power business.

[0045] Optionally, vectorizing the unstructured risk text data using a neural network model that integrates the semantic field rules of the power business includes:

[0046] The unstructured risk text data is input into the neural network model for word segmentation and contextual analysis to obtain text units;

[0047] Based on the aforementioned power business semantic field, the text units are annotated with business context to generate an annotated text sequence;

[0048] Using the vector representation of the set of device nodes affected by the labeled text sequence in the physical topology map of the power grid as the approximation target, the neural network model is trained to minimize the distance between its output vector representation and the target vector representation in the vector space, thereby generating a business semantic vector.

[0049] Specifically, a neural network model is used to vectorize unstructured risk text data, generating a business semantic vector that accurately maps to the physical topology of the power grid, achieving a direct association from textual description to the scope of influence of physical entities. First, text processing begins by inputting a single piece of unstructured risk text data, such as "abnormal secondary voltage of the PT on the 220kV I-section busbar of Chengbei Substation," into a neural network model that integrates rules from the power business semantic field. This model incorporates a word segmenter optimized for power industry terminology, dividing the input text into a sequence of text units with independent business meanings, such as "Chengbei Substation," "220kV," "I-section busbar," "PT," "secondary voltage," and "abnormal." Next, the business context annotation stage begins. The core of this stage is to utilize the constructed power business semantic field to perform deep semantic understanding and entity linking of the text units. The power business semantic field model identifies "Chengbei Substation" as a substation entity, "220kV I-section busbar" as a specific piece of equipment, and "abnormal secondary voltage of the PT" as a risk event type, thereby generating an annotated text sequence with structured tags. Finally, physical topology-oriented approximation training is performed to generate the final business semantic vector. Essentially, this aligns the semantic representation of the text with its physical consequences in the vector space. Historical risk data is used as the training set; each data point contains the risk text and the set of device nodes actually affected by that risk in the power grid physical topology map. For each training sample, a target vector is computed, which is an aggregation of the vector representations of the affected device node set in the map, typically calculated using the mean or weighted average. The goal of the training process is to minimize the spatial distance between the vector representation of the risk text output by the neural network model and the target vector. This optimization objective can be achieved using the following loss function. express:

[0050] ;

[0051] in, This represents the loss value that needs to be minimized. It is a distance metric function in vector space, such as the negative of Euclidean distance or cosine distance, used to measure the similarity between two vectors. It is a neural network model for input text The processed output business semantic vector. This represents the total number of device nodes actually affected by this risk event in the power grid physical topology map. It is the first A vector representation of each affected device node, which is itself pre-trained on the power grid physical topology graph by a graph embedding algorithm, typically with dimensions of 256 or 512. Summation symbol. Indicates all The vectors of each affected device node are summed element-wise, and then divided by . The aggregated target vector is obtained. The neural network is then continuously adjusted using optimization algorithms such as gradient descent. Internal parameters to reduce The business semantic vector output by the model can naturally be located near the vector center of the group of devices that the risk it describes may affect in the vector space, thus achieving a deep anchoring of semantics and physical topology.

[0052] Optionally, based on the business semantic vector, identifying risk events and generating corresponding risk logic clusters through a preset risk event extraction model includes:

[0053] The business semantic vector is input into the risk event extraction model to determine the risk type and the risk subject device;

[0054] By calling the logical association interface in the power grid physical topology map that is connected to the risk subject equipment, the affected upstream and downstream equipment and user paths are automatically deduced to form an impact chain;

[0055] Using the risk type and the risk subject equipment as search criteria, relevant handling step texts are matched and extracted from the unstructured risk text data or contingency plan knowledge base to generate a handling knowledge package;

[0056] The risk subject, the impact chain, and the disposal knowledge package are encapsulated to generate a risk logic cluster.

[0057] Specifically, the abstract vector anchored to the physical topology is decoded and visualized into a structured business object containing complete risk elements, such as the subject, impact, and handling, enabling it to be directly invoked and executed. Risk subject identification is initiated by inputting a business semantic vector into a pre-defined risk event extraction model. This model is typically a multi-task neural network composed of a classifier and a sequence labeling module. The classifier is responsible for parsing the risk type information in the vector and outputting predefined risk labels such as "equipment overload" and "voltage limit exceeded"; the sequence labeling module is responsible for identifying and outputting the core risk subject equipment corresponding to the vector, for example, by performing a nearest neighbor search in the equipment entity database of the power grid physical topology map to locate a unique equipment ID, such as "Chengnan Substation 220 kV No. 1 Main Transformer". The derivation of the impact chain is automatically triggered. Taking the identified risk subject equipment as the starting node, the logical association interface configured for that node in the power grid physical topology map is invoked. This interface, in engineering terms, is a set of rules defining electrical connection relationships and protection coordination logic. Based on these rules, a customized graph traversal algorithm, such as breadth-first search with electrical constraints, is executed to automatically deduce the affected upstream and downstream equipment and key user paths. For example, starting from the overloaded main transformer node, the traversal follows its connected buses, outgoing switches, and lines, all the way to downstream substations or important user terminals. The traversal depth is typically limited to 5 to 7 electrical levels to ensure computational efficiency and tightness of association, ultimately forming an impact chain containing node sequences and connection relationships. A disposal knowledge package is then generated. The identified risk type "equipment overload" and the risk subject equipment type "main transformer" are used as composite search conditions to initiate a query against unstructured risk text data in the background, such as historical defect reports, operation tickets, and a structured contingency plan knowledge base. This query employs vector retrieval technology, encoding the search conditions into query vectors, and matching and extracting the most semantically similar nodes in the pre-vectorized knowledge base. The relevant handling steps are documented in text, such as "verify the overload multiple and duration," "attempt to transfer part of the load through switching operations," and "contact dispatch to request an adjustment to the operating mode." An encapsulation operation is performed, combining the risk subject equipment ID as an influence chain in an ordered graph structure and the handling knowledge package as a text list, into a standardized data object, namely, a risk logic cluster.

[0058] Optionally, the step of injecting the influence chain logic of the risk logic cluster into the logical association interface through a bidirectional coupling engine and writing the dynamic risk attribute value into the corresponding dynamic attribute slot includes:

[0059] The bidirectional coupling engine monitors structural change events in the power grid physical topology map;

[0060] When the structural change event is detected, the impact chain of all attached risk logic clusters is recalculated, and the risk attribute values ​​in the dynamic attribute slots are updated.

[0061] When a new risk logic cluster connection event is detected, attribute value injection and state update are triggered along the influence chain to the dynamic attribute slots of the relevant nodes in the power grid physical topology map;

[0062] Based on the updated dynamic attribute slots, a risk heatmap is rendered in real time on the visualization interface of the power grid physical topology map, forming a risk linkage knowledge graph.

[0063] Specifically, a risk-linked knowledge graph is generated through a bidirectional coupling engine, establishing a real-time synchronization mechanism capable of responding to changes in the power grid's physical structure and new risk events. This ensures that the risk profile presented in the knowledge graph remains dynamically consistent with the actual state of the power grid. This bidirectional coupling engine is a persistent background monitoring service that continuously monitors two core event sources. The first is the monitoring of structural change events in the power grid's physical topology graph. When a structural change event is detected, such as a circuit breaker changing from closed to open, or a line being put into operation or decommissioned, the engine is triggered. After triggering, the engine traverses all risk logic clusters already attached to the graph and, for each risk logic cluster, starts with its risk subject and re-invokes the influence chain derivation algorithm based on the changed topology. This recalculation process ensures that the original risk propagation path can adapt to the new power grid operation mode. Subsequently, it updates the risk attribute values ​​in the dynamic attribute slots of relevant equipment nodes according to the new influence chain. The second is the monitoring of new risk logic cluster attachment events. When an upstream entity successfully generates a new risk logic cluster and requests attachment, the engine is triggered again. At this point, the engine parses the impact chain within the risk logic cluster and performs attribute injection operations along the device node paths defined by the impact chain in the power grid physical topology graph. It sequentially accesses the dynamic attribute slots of each node in the chain, writes the calculated risk attribute values, and updates the node status, such as marking the operating status as "threatened." Dynamic attribute slots are pre-defined read-write areas in the data structure of each device entity node, specifically used to store non-topology, dynamically changing business attributes, such as risk values ​​and alarm levels. After completing any triggering process, the engine ultimately renders the graph in real-time on the visualization interface based on the updated dynamic attribute slots of all nodes. The rendering engine marks nodes or connections with different colors or glow effects according to different risk attribute value ranges; for example, low risk is blue, and high risk is red, thus forming a dynamic and intuitive risk heatmap. This visualization view, composed of the physical topology, the connected risk logic, and the dynamically rendered risk attributes, is the final generated risk linkage knowledge graph. Its data update latency is typically controlled within seconds to meet the real-time requirements of scheduling operations.

[0064] Optionally, when a new risk logic cluster connection event is detected, triggering attribute value injection and state update of the dynamic attribute slots of relevant nodes in the power grid physical topology map along the influence chain includes:

[0065] Based on the risk type of the risk logic cluster, the initial risk attribute value is obtained from the preset risk-attribute mapping table;

[0066] Based on the electrical distance and impedance parameters of the device entity nodes through which the influence chain passes, the initial risk attribute value is attenuated to obtain the local risk attribute value of each node.

[0067] Write the local risk attribute value into the dynamic attribute slot of the corresponding node;

[0068] Based on the values ​​of the dynamic attribute slots of all nodes, aggregate and calculate the risk indicators for the entire network or a region.

[0069] Specifically, the process of injecting attribute values ​​and updating states transforms the abstract risk impact chain into a measurable and computable quantitative risk distribution on the power grid physical topology map, enabling refined modeling of the degree of risk impact. When a new risk logic cluster connection event is triggered, the system first accesses a preset risk-attribute mapping table based on the risk type within that cluster, such as "single-phase grounding of a line." This table is a key-value database that stores dimensionless initial risk attribute values ​​corresponding to different risk types, typically normalized to between 0 and 100. The initial risk attribute value is retrieved based on the risk type. Subsequently, risk attenuation calculations are performed on each device entity node along the impact chain of the risk logic cluster to obtain the local risk attribute value for each node. This calculation process aims to simulate the propagation and attenuation effects of risk in the electrical system. The core calculation formula is as follows:

[0070] ;

[0071] in, Is it affecting the first on-chain The local risk attribute value of each node, i.e. the value that needs to be calculated and written. It is the initial risk attribute value obtained from the risk-attribute mapping table. This represents the natural exponential function, used to simulate the nonlinear decay characteristics of risk. It is the key risk attenuation coefficient, a dimensionless hyperparameter, with a value generally ranging from 0.01 to 0.5. Its specific value is obtained through regression analysis of historical data or expert experience based on the physical propagation characteristics of different risk types. Represents the process from the risk subject equipment to the current node. The electrical distance is obtained by accumulating the per-unit impedance parameters of all series-connected devices along the chain path. These impedance parameters are directly read from the static attributes of the corresponding devices in the power grid physical topology map. After calculation, the resulting local risk attribute value is... Write to the corresponding node The dynamic attribute slots are then used to calculate a comprehensive regional or network-wide risk indicator. This indicator can be used for high-level risk alerts and trend assessments.

[0072] Optionally, optimizing the parameters of the neural network model and updating the generation of the business semantic vector using the first feedback sample includes:

[0073] The decision differences are transformed into text descriptions and compared with the original unstructured risk text data to form a comparative sample pair;

[0074] The neural network model is trained using the comparison samples to adjust its parameters, so that the business semantic vector generated by the model in response to input containing the decision difference text is closer to the vector representation of the set of device nodes updated in the power grid physical topology map to which the manual intervention operation is directed.

[0075] Specifically, the neural network model is optimized using the first feedback sample. By introducing expert intervention knowledge from dispatchers, the vectorization process of unstructured risk text is refined and calibrated to improve the accuracy of predicting the impact range of physical topology by business semantic vectors. First, the decision discrepancy between the dispatcher's manual intervention and the system's deduction results is captured. For example, the deduction might suggest that a certain risk text only affects devices A and B, but the dispatcher's action indicates that device C is also affected. This discrepancy is transformed into a structured text description, such as "This risk should also be associated with device C," and concatenated or associated with the original unstructured risk text data to form a positive-negative sample pair. A comparison sample pair contains three parts: the original risk text anchor, the negative sample text description corresponding to the system's erroneous deduction result, and the positive sample text description containing the dispatcher's decision discrepancy. Next, these constructed comparison sample pairs are used to train the neural network model, which incorporates the rules of the power business semantic field, through comparative learning. Comparative learning is a self-supervised learning paradigm whose core idea is to teach the model how to bring similar samples closer together in the vector space and push dissimilar samples further apart. In this scenario, the goal of training is to adjust the parameters of the neural network model so that the distance between the processed anchor sample's business semantic vector and the positive sample's business semantic vector is minimized, while the distance with the negative sample's business semantic vector is maximized. This optimization process is typically achieved using a triple loss function, mathematically expressed as:

[0076] ;

[0077] in, It is a triple loss value, and the training objective is to minimize it. These are the business semantic vectors generated by the neural network model from the anchor point, positive sample, and negative sample, respectively. It is a distance function in vector space, such as Euclidean distance. This is a marginal hyperparameter representing the minimum expected distance between positive and negative sample pairs, typically set to a small positive number, such as 0.2. After multiple rounds of iterative training by minimizing this loss function, the parameters of the neural network model are updated. When processing similar risky texts, the updated model's output business semantic vector will be more likely to approximate the aggregated vector representation of the correct set of device nodes updated in the power grid physical topology map, as indicated by the dispatcher's manual intervention, thus achieving closed-loop optimization of semantic understanding capabilities.

[0078] Optionally, optimizing the derivation rules of the influence chain in the risk logic cluster using the second feedback sample includes:

[0079] Analyze the decision differences to identify the additional or excluded equipment entities in the manual intervention operation;

[0080] Analyze the electrical connection or protection logic relationship between the additional included or excluded equipment entities and the risk subject equipment to obtain potential association logic;

[0081] Based on the potential association logic, the association rules in the logical association interface are modified or supplemented, or the weight parameters of different connection types in the influence chain derivation process are adjusted.

[0082] Specifically, the influence chain derivation rules are optimized using second feedback samples to correct and enhance the inherent logical model of risk propagation in the power grid physical topology map, enabling the automatic derivation capability of the risk impact range to learn from expert experience and continuously evolve. The system compares dispatcher manual intervention operations with the results of automatic derivation in a programmed manner to accurately identify the list of equipment entities that have been additionally included in the handling scope or excluded from the influence chain by the dispatcher. For example, if the derivation shows that a line fault only affects one downstream substation, but the dispatcher also opened the tie switch of an adjacent line, then that tie switch is identified as an additional equipment entity. Next, the mining of potential correlation logic is initiated. For each identified additionally included or excluded equipment entity, a deep correlation path analysis is performed on the power grid physical topology map, using the entity and the original risk subject equipment as endpoints. This analysis not only checks direct electrical connections, such as whether they are on the same power supply path, but more importantly, it queries the background protection logic database to analyze whether there are non-direct topology protection logic relationships such as automatic transfer switch logic, main and backup protection coordination relationships, or remote trip signals. The analysis results in one or more verified potential correlation logics, such as discovering that an additional tie switch is one of the backup protection exits of the original faulty line. Finally, based on the discovered potential correlation logics, the rule base is modified or parameters are adjusted. Two optimization paths are provided. The first is to modify or supplement the correlation rules in the logic correlation interface. If the discovered potential correlation logic is a hard rule not defined in the system, such as a specific interlocking condition, it will be automatically formatted and added to the risk derivation rule base to ensure direct application in future similar risk derivations. The second is to adjust the weight parameters of different connection types during the influence chain derivation process. If the potential correlation logic is a probabilistic or empirical correlation, it will be used as training data to update the weight model in the derivation algorithm. For example, when traversing the graph, the influence chain derivation algorithm assigns different propagation weights to different types of connections, such as trunk lines, tie lines, and protection signals. The initial weights might be... Main line: 0.9, tie line: 0.6, protection signal: 0.7 When it is discovered that the scheduler frequently considers specific types of connection lines, the weight parameter of the "connection line" connection type is fine-tuned from 0.6 to 0.65 using methods such as gradient ascent. This makes the derivation model more inclined to extend the influence chain along this type of path in future calculations. Through these two methods, the implicit knowledge of experts is solidified into explicit model rules or parameters, completing the closed-loop optimization of the influence chain derivation logic.

[0083] Optionally, risk visualization and deduction based on the aforementioned risk linkage knowledge graph includes:

[0084] Receive simulated handling instructions for specific risk logic clusters in the risk linkage knowledge graph;

[0085] Based on the simulated disposal instructions and the disposal knowledge package, the state of relevant equipment entity nodes in the power grid physical topology map is simulated in an independent sandbox environment, and power flow calculation is performed to obtain the simulated state.

[0086] The bidirectional coupling engine calculates the propagation results of the simulated state along the influence chain, updates the full-map risk heat rendering, and outputs a risk comparison view and key indicator changes before and after the simulation operation.

[0087] Specifically, risk visualization simulations are performed based on a risk-linked knowledge graph, providing dispatchers with an interactive and risk-free decision-making pre-simulation environment. This allows them to quantitatively assess and compare the effectiveness and potential secondary risks of different handling schemes before executing actual operations. This process is triggered by a user-initiated simulation handling command. The dispatcher selects a specific risk logic cluster on the visualization interface of the risk-linked knowledge graph and issues a simulation handling command, such as "Execute the load transfer plan for substation XX". Upon receiving the command, the simulation is first executed in an isolated, independent sandbox environment. The sandbox environment is a complete memory snapshot of the current power grid physical topology, ensuring that any simulation operation does not affect the real-time production system. Based on the command content and referring to the handling knowledge package encapsulated in the risk logic cluster, the command is decomposed into a series of modification operations on the state of equipment entity nodes, such as changing the state of a specific circuit breaker in the sandbox environment from "closed" to "open". After completing the state simulation change, the background power flow calculation engine is immediately invoked to perform full or incremental power flow calculations on the power grid model in the sandbox environment. The calculation method typically uses the Newton-Raphson method to ensure accuracy. This calculation process generates a simulated post-state containing the voltage and branch power distribution of all nodes. Subsequently, a bidirectional coupling engine is activated, using this simulated post-state as input to reassess the propagation of risks in the power grid. The engine updates the risk attribute values ​​of each node along the original influence chain based on the new power flow distribution and checks for any new grid constraint violations caused by the simulation operation, such as line overloads or voltage drops. If any are found, a new risk logic cluster is dynamically generated. Based on the updated dynamic attribute slots of all nodes in the full graph, the risk heatmap rendering of the sandbox environment is updated in real time. Finally, the original risk view before the simulation and the risk view after the simulation are presented side-by-side on the user interface, forming an intuitive risk comparison view. Simultaneously, changes in a series of key indicators are calculated and displayed, such as changes in the highest line load rate, the lowest bus voltage, and the rise and fall of the comprehensive risk indicator score, thus providing dispatchers with quantitative basis for decision-making.

[0088] Based on the same inventive concept, such as Figure 2 As shown, the present invention also provides a knowledge graph-based power dispatching operation risk linkage modeling system, the system comprising:

[0089] The data acquisition and semantic field construction module is used to acquire structured topology data and unstructured risk text data of the power system, and construct the power business semantic field based on the business context and historical handling cases in the unstructured risk text data.

[0090] The physical topology construction module is used to construct a power grid physical topology map containing device entity nodes and connection relationships based on the structured topology data, wherein the device entity nodes in the power grid physical topology map are configured with dynamic attribute slots and logical association interfaces.

[0091] The semantic vector generation module is used to vectorize the unstructured risk text data using a neural network model that integrates the semantic field rules of the power business, and generate a business semantic vector anchored to the physical topology.

[0092] The risk event extraction and encapsulation module is used to identify risk events and generate corresponding risk logic clusters based on the business semantic vector and a preset risk event extraction model. The risk logic cluster includes a risk subject, an impact chain, and a disposal knowledge package.

[0093] The bidirectional coupling and linkage knowledge graph generation module is used to attach the risk logic cluster to the power grid physical topology graph through entity alignment, and inject the influence chain logic of the risk logic cluster into the logic association interface through the bidirectional coupling engine, and write the dynamic risk attribute value into the corresponding dynamic attribute slot to generate a risk linkage knowledge graph.

[0094] The visualization simulation and difference capture module is used to perform risk visualization simulation based on the risk linkage knowledge graph, obtain the dispatcher's manual intervention operation on the simulation result, and capture the decision difference between the manual intervention operation and the system simulation result.

[0095] The semantic optimization and rule optimization module is used to simultaneously use the decision difference as a first feedback sample and a second feedback sample, use the first feedback sample to optimize the generation of the business semantic vector by updating the parameters of the neural network model, and use the second feedback sample to optimize the derivation rules of the influence chain in the risk logic cluster.

[0096] Example 1:

[0097] To verify the feasibility of this invention in practice, it was applied to a municipal power grid dispatch center to perform real-time modeling and intelligent analysis of the operational risks of its power grid. This power grid includes multiple 220 kV and 110 kV substations. Traditional risk management relies on dispatchers' personal experience to comprehensively judge scattered alarm information and defect reports, lacking the ability to systematically analyze and dynamically extrapolate the interconnectedness of risks. This embodiment aims to construct a risk interconnection knowledge graph that deeply integrates textual risk information with the physical topology of the power grid, and to achieve continuous model evolution through human-machine collaboration.

[0098] The testing period for this embodiment was 6 months, focusing on the actual operational data of the municipal power grid. Structured topology data: CIM model data of the municipal power grid was acquired, including 5 220 kV substations, 22 110 kV substations, and a total of 5,800 related line and equipment nodes, constructing a power grid physical topology map containing 5,800 equipment entity nodes and 8,200 connection relationships. Dynamic attribute slots for storing risk values ​​and logical association interfaces for defining electrical associations were configured for each node. Unstructured risk text data: Historical defect records, operation tickets, scheduling logs, and other text data from the past 5 years were collected, totaling approximately [amount missing]. Construction of the semantic field for power business: Based on the above text data, a triplet association rule base containing 3,000 power monitoring alarm terms, 15,000 standard equipment names, and 80,000 historical defect records was extracted. BERT was selected as the basic general semantic model, and the rule base was used as a knowledge constraint for fusion training. The total loss function of the training objective was set as... ,in Set to 0.4, and after 200 rounds of iterative training, a semantic field for electricity business was generated that can map electricity text to a 512-dimensional vector space.

[0099] A real-world event during the summer peak electricity consumption period was used as a test case. An unstructured risk text was received from the PMS system: "Chengnan Substation 220kV No. 1 main transformer temperature overheating alarm, load rate reached 95%, overload risk." Semantic vector generation: This text was input into a neural network model that incorporates the semantic field of power business. The model, through physical topology-guided approximation training, converted the text into a 512-dimensional business semantic vector. The objective loss function of this training process is... This ensures that the generated vector is spatially proximate to the vector center of the set of devices affected by similar historical events. Risk logic cluster generation: The business semantic vector is input into the risk event extraction model for decoding; the specific process and results are recorded in the table below.

[0100] Table 1 Example of Risk Logic Cluster Generation Process

[0101] step Output content Parameters / Description Risk type identification Equipment overload The output is based on the business semantic vector by the classifier module. Risk subject identification Chengnan Substation 220kV No. 1 Main Transformer (Equipment ID: T-CN-220-01) Location is determined by nearest neighbor search in the entity database. Influence chain derivation T-CN-220-01→B-CN-220-1 (Busbar)→L-CNX-220-1 (Line)→S-XN-220-01 (West Suburb Substation) Call the logical association interface of T-CN-220-01 to execute the graph traversal algorithm, with the traversal depth limited to 5 electrical levels. Knowledge package generation 1. Verify the overload factor and duration. 2. Attempt to transfer part of the load through switching operations. 3. Contact dispatch to request an adjustment to the operating mode. Using "equipment overload" and "main transformer" as search criteria, the top 3 relevant texts were extracted from the historical case database.

[0102] Bidirectional Coupling and Attribute Injection: The generated risk logic cluster is attached to the T-CN-220-01 node via entity alignment. The bidirectional coupling engine is triggered, initiating dynamic risk attribute value injection along the influence chain. This process employs the risk decay calculation formula. The specific parameter settings are as follows: initial risk attribute value ( According to the risk-attribute mapping table, the initial value of the "equipment overload" risk is set to 90. Risk attenuation coefficient ( Based on expert experience, the overload risk propagation attenuation coefficient for the main power grid is set at 0.15. Electrical distance ( ): The cumulative per-unit impedance between devices, read from the physical topology map. The calculation process and results are shown in the table below.

[0103] Table 2 Example of calculating the decay of risk attribute values ​​affecting the chain

[0104] Affecting chain nodes Device ID Electrical distance to the risk subject ( ) Calculation process Local risk attribute value ( ) Chengnan Substation No. 1 Main Transformer T-CN-220-01 0.00 90.0 Chengnan Bian busbar B-CN-220-1 0.01 89.9 South City-West Suburbs Line L-CNX-220-1 0.05 89.3 West Suburbs Substation S-XN-220-01 0.06 89.2

[0105] like Figure 3 As shown, the nonlinear decay trend of the risk value with increasing electrical distance is illustrated, where the scatter points represent the various physical device nodes on the influence chain.

[0106] The calculated local risk attribute values ​​are written to the corresponding dynamic attribute slots of each node in real time. A red risk heat map, centered on "Chengnan Substation" and spreading towards "Xijiao Substation," is then rendered on the visualization interface, forming a risk linkage knowledge graph. During the risk simulation phase, the on-duty dispatcher reviewed the impact chain generated by the system. Based on historical experience, the dispatcher pointed out that the Xijiao Substation (S-XN-220-01) is also connected to the Yuanjiao Substation (S-YJ-110-01) via a connecting line (L-XYL-110-1) with a backup automatic transfer logic. Under heavy load conditions, voltage fluctuations at the Xijiao Substation may trigger this logic, thus affecting the Yuanjiao Substation. Therefore, the dispatcher additionally included the Yuanjiao Substation in the monitoring scope during the simulation. Decision discrepancy capture: The decision discrepancy between the dispatcher's manual intervention and the system simulation results was captured, with the manual intervention involving the additional inclusion of the "Yuanjiao Substation."

[0107] The second feedback sample optimization (rule space) analyzes this difference, automatically identifies the correlation between "West Suburbs Substation" and "Remote Suburbs Substation," and discovers a "backup automatic transfer" relationship in the protection logic database. Based on this potential correlation logic, the influence chain derivation rules are optimized. Specifically, the weight parameters of the connection type are adjusted.

[0108] Table 3 Comparison of Influence Chain Derivation Rules Before and After Optimization

[0109] Parameters Before optimization After optimization Optimization basis Connection type weight: tie line 0.60 0.65 By capturing the discrepancies in dispatcher decisions and uncovering the "backup self-connection" logic, the importance of this type of connection in the derivation was enhanced. Derivation results (re-simulation) "Suburban substations" are not included. It includes "suburban substations". With the increased weight, the risk propagation model will prioritize connecting paths when traversing the graph, thus incorporating suburban areas into the impact chain.

[0110] First feedback sample optimization (semantic space): Simultaneously, this decision difference is transformed into the text description "the risk should also be associated with suburban substations," and compared with the original risk text to form a contrasting sample pair. Using this sample pair, a triple loss function is applied... Among them, marginal hyperparameters A value of 0.2 was used to perform a round of comparative learning training on the neural network model. After optimization, when processing risk texts related to "western suburbs change" in the future, the model's output business semantic vector will be closer to the vector representation of the correct set of devices, including "far suburbs change," improving semantic understanding accuracy by approximately 8%. Figure 4 As shown, this illustrates the process by which the risk text anchor vector moves closer to the positive sample cluster center and further away from the negative sample cluster center after optimization using a triple loss function.

[0111] As can be seen from the above embodiments, this invention can effectively transform unstructured textual risks into a measurable and inferable interconnected risk model based on physical topology. As shown in Table 2, this method can calculate the propagation and attenuation of risks in the power grid in a refined manner based on electrical parameters, reducing the risk value from an initial 90 to 89.2 at the downstream substation, thus quantifying the degree of impact. More importantly, as shown in Table 3, by capturing the differences between human and machine decision-making, the system can learn and optimize the influence chain derivation rules, increasing the weight parameter of the "tethering line" connection type from 0.60 to 0.65, thereby solidifying the dispatcher's implicit experience into explicit model rules, realizing the closed-loop evolution of the knowledge graph, and improving the intelligence level and accuracy of power dispatching operation risk analysis.

[0112] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0113] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A knowledge graph-based method for linking power dispatching and operation risks, characterized in that, The method includes: Obtain structured topology data and unstructured risk text data of the power system, and construct a power business semantic field based on the business context and historical handling cases in the unstructured risk text data; Based on the structured topology data, a power grid physical topology map containing device entity nodes and connection relationships is constructed, wherein the device entity nodes in the power grid physical topology map are configured with dynamic attribute slots and logical association interfaces. The unstructured risk text data is vectorized using a neural network model that integrates the semantic field rules of the power business to generate a business semantic vector anchored to the physical topology. Based on the business semantic vector, risk events are identified and corresponding risk logic clusters are generated through a preset risk event extraction model. The risk logic clusters include risk subjects, impact chains, and disposal knowledge packages. The risk logic cluster is attached to the power grid physical topology graph through entity alignment, and the influence chain logic of the risk logic cluster is injected into the logic association interface through the bidirectional coupling engine. The dynamic risk attribute value is written into the corresponding dynamic attribute slot to generate a risk linkage knowledge graph. Based on the risk linkage knowledge graph, risk visualization simulation is performed to obtain the dispatcher's manual intervention operation based on the simulation results, and the decision difference between the manual intervention operation and the system simulation results is captured. The decision difference is used as both the first and second feedback samples. The first feedback sample is used to optimize the generation of the business semantic vector by updating the parameters of the neural network model, and the second feedback sample is used to optimize the derivation rules of the influence chain in the risk logic cluster.

2. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, The acquisition of structured topology data and unstructured risk text data of the power system, and the application of business context and historical handling cases in the unstructured risk text data, include: The set of power monitoring alarm terms, the set of standard equipment names, the set of historical defect records, and the set of load anomaly patterns are used as the basic corpus. Based on the analysis of the basic corpus, a triplet association rule base is established, which uses power entities and risk events as lexical units, the contextual environment of the lexical units as the first association, and the typical business consequences caused by the lexical units as the second association. The triple association rule base is fused with a general semantic model for training to generate a power business semantic field.

3. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, Vectorization of the unstructured risk text data using a neural network model that integrates the semantic field rules of the power business includes: The unstructured risk text data is input into the neural network model for word segmentation and contextual analysis to obtain text units; Based on the aforementioned power business semantic field, the text units are annotated with business context to generate an annotated text sequence; Using the vector representation of the set of device nodes affected by the labeled text sequence in the physical topology map of the power grid as the approximation target, the neural network model is trained to minimize the distance between its output vector representation and the target vector representation in the vector space, thereby generating a business semantic vector.

4. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, Based on the business semantic vector, risk events are identified and corresponding risk logic clusters are generated through a preset risk event extraction model, including: The business semantic vector is input into the risk event extraction model to determine the risk type and the risk subject device; By calling the logical association interface connected to the risk subject equipment in the power grid physical topology map, the affected upstream and downstream equipment and user paths are automatically deduced to form an impact chain; Using the risk type and the risk subject equipment as search criteria, relevant handling step texts are matched and extracted from the unstructured risk text data or contingency plan knowledge base to generate a handling knowledge package; The risk subject, the impact chain, and the disposal knowledge package are encapsulated to generate a risk logic cluster.

5. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, The step of injecting the influence chain logic of the risk logic cluster into the logical association interface through a bidirectional coupling engine and writing the dynamic risk attribute value into the corresponding dynamic attribute slot includes: The bidirectional coupling engine monitors structural change events in the power grid physical topology map; When the structural change event is detected, the impact chain of all attached risk logic clusters is recalculated, and the risk attribute values ​​in the dynamic attribute slots are updated. When a new risk logic cluster connection event is detected, attribute value injection and state update are triggered along the influence chain to the dynamic attribute slots of the relevant nodes in the power grid physical topology map; Based on the updated dynamic attribute slots, a risk heatmap is rendered in real time on the visualization interface of the power grid physical topology map, forming a risk linkage knowledge graph.

6. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 5, characterized in that, When a new risk logic cluster connection event is detected, triggering attribute value injection and state update of the dynamic attribute slots of relevant nodes in the power grid physical topology map along the influence chain includes: Based on the risk type of the risk logic cluster, the initial risk attribute value is obtained from the preset risk-attribute mapping table; Based on the electrical distance and impedance parameters of the device entity nodes through which the influence chain passes, the initial risk attribute value is attenuated to obtain the local risk attribute value of each node. Write the local risk attribute value into the dynamic attribute slot of the corresponding node; Based on the values ​​of the dynamic attribute slots of all nodes, aggregate and calculate the risk indicators for the entire network or a region.

7. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, The generation of the business semantic vector, which optimizes the parameters of the neural network model using the first feedback sample, includes: The decision differences are transformed into text descriptions and compared with the original unstructured risk text data to form a comparative sample pair; The neural network model is trained using the comparison samples to adjust its parameters, so that the business semantic vector generated by the model in response to input containing the decision difference text is closer to the vector representation of the set of device nodes updated in the power grid physical topology map to which the manual intervention operation is directed.

8. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, Optimizing the derivation rules of the influence chain in the risk logic cluster using the second feedback sample includes: Analyze the decision differences to identify the additional or excluded equipment entities in the manual intervention operation; Analyze the electrical connection or protection logic relationship between the additional included or excluded equipment entities and the risk subject equipment to obtain potential association logic; Based on the potential association logic, the association rules in the logical association interface are modified or supplemented, or the weight parameters of different connection types in the influence chain derivation process are adjusted.

9. The knowledge graph-based power dispatching operation risk linkage modeling method according to claim 1, characterized in that, Risk visualization and deduction based on the aforementioned risk linkage knowledge graph includes: Receive simulated handling instructions for specific risk logic clusters in the risk linkage knowledge graph; Based on the simulated disposal instructions and the disposal knowledge package, the state of relevant equipment entity nodes in the power grid physical topology map is simulated in an independent sandbox environment, and power flow calculation is performed to obtain the simulated state. The bidirectional coupling engine calculates the propagation results of the simulated state along the influence chain, updates the full-map risk heat rendering, and outputs a risk comparison view and key indicator changes before and after the simulation operation.

10. A knowledge graph-based power dispatching and operation risk linkage modeling system, characterized in that, The system is used in the knowledge graph-based power dispatching operation risk linkage modeling method as described in any one of claims 1-9, and the system comprises: The data acquisition and semantic field construction module is used to acquire structured topology data and unstructured risk text data of the power system, and construct the power business semantic field based on the business context and historical handling cases in the unstructured risk text data. The physical topology construction module is used to construct a power grid physical topology map containing device entity nodes and connection relationships based on the structured topology data, wherein the device entity nodes in the power grid physical topology map are configured with dynamic attribute slots and logical association interfaces. The semantic vector generation module is used to vectorize the unstructured risk text data using a neural network model that integrates the semantic field rules of the power business, and generate a business semantic vector anchored to the physical topology. The risk event extraction and encapsulation module is used to identify risk events and generate corresponding risk logic clusters based on the business semantic vector and a preset risk event extraction model. The risk logic cluster includes a risk subject, an impact chain, and a disposal knowledge package. The bidirectional coupling and linkage knowledge graph generation module is used to attach the risk logic cluster to the power grid physical topology graph through entity alignment, and inject the influence chain logic of the risk logic cluster into the logic association interface through the bidirectional coupling engine, and write the dynamic risk attribute value into the corresponding dynamic attribute slot to generate a risk linkage knowledge graph. The visualization simulation and difference capture module is used to perform risk visualization simulation based on the risk linkage knowledge graph, obtain the dispatcher's manual intervention operation on the simulation result, and capture the decision difference between the manual intervention operation and the system simulation result. The semantic optimization and rule optimization module is used to simultaneously use the decision difference as a first feedback sample and a second feedback sample, use the first feedback sample to optimize the generation of the business semantic vector by updating the parameters of the neural network model, and use the second feedback sample to optimize the derivation rules of the influence chain in the risk logic cluster.

Images