Dynamic knowledge graph online reasoning method and system for power grid event driving

Abstract

The invention relates to the technical field of power grid data analysis and artificial intelligence, and particularly discloses a dynamic knowledge graph online reasoning method and system for power grid event driving. The method comprises the following steps: semantizing a real-time event stream into a driving event; dynamically constructing a special sub-graph from the power grid knowledge graph; performing multi-path deduction on the sub-graph and calculating a comprehensive confidence coefficient for each path; acquiring actual feedback in a post-event time window as a verification reference; arbitration is carried out based on the verification benchmark and the confidence coefficient, and a target evolution path is screened out; and generating graph evolution operation according to the target path and the competition path information of the target path, and performing incremental adjustment on the knowledge graph. The system correspondingly comprises an event perception module, a dynamic sub-graph construction module, a reasoning evaluation module, a feedback arbitration module, a graph evolution module and the like. According to the method, online reasoning under event driving and adaptive co-evolution of the knowledge graph are realized, and the accuracy and the intelligent level of real-time situation analysis of the power grid are improved.

Description

Online Reasoning Method and System for Power Grid Event-Driven Dynamic Knowledge Graph Technical Field

[0001] This invention relates to the fields of power grid data analysis and artificial intelligence technology, and more specifically, to a dynamic knowledge graph online reasoning method and system for power grid event-driven approaches. Background Technology

[0002] With the advancement of new power system construction, the integration of high proportions of renewable energy and power electronic equipment has led to faster and more complex spatiotemporal changes in the power grid's operating status. Massive amounts of real-time event data, such as equipment status alarms, protection action information, wide-area measurement data, and meteorological environmental information, constitute a key sequence reflecting the dynamic processes of the power grid. This event data not only features multi-source, asynchronous, and massive volumes, but also contains strict physical constraints and operational logic of the power grid. Its evolution process is often accompanied by strict temporal causal relationships (such as the millisecond-level sequence between protection activation, circuit breaker action, and automatic device response) and spatial electrical coupling relationships (such as power flow shifts and voltage fluctuations caused by faults). Real-time perception of the operating status from this highly temporal and strongly coupled event stream, and accurate inference of potential evolutionary trends and risks, is of great significance for the safe and stable operation of the power grid.

[0003] Currently, knowledge graph-based power grid analysis techniques mainly focus on building static correlation models offline using historical data and expert knowledge. These models have significant limitations when facing real-time power grid event flows. First, the static knowledge graph is outdated, failing to incorporate dynamic information such as power grid topology changes, equipment commissioning / decommissioning, or operational mode adjustments in a timely manner. This leads to inferences based on outdated graphs potentially deviating from reality. More fundamentally, existing methods often employ general graph structures and reasoning frameworks, failing to deeply integrate the unique physical laws and operational constraints of the power grid domain. Firstly, the "relationships" in general graphs often lack power grid semantics (such as electrical connectivity, protection coordination logic, and power flow transmission capabilities), resulting in "evolutionary paths" that are electrically infeasible or logically invalid. Secondly, existing methods often respond to events based on isolated judgments of predefined rules, lacking the ability to mine deep correlations from multi-source asynchronous events that conform to the timing logic of power grid protection actions and the causal chain of fault propagation, making it difficult to handle complex cascading or multiple fault scenarios. Third, the reasoning process and knowledge updates are disconnected. The system cannot use the difference between the reasoning results and the actual response of the power grid to automatically correct and enrich its internal electrical correlation model, protection logic knowledge or equipment impact mode. It relies on manual maintenance for a long time and is difficult to adapt to the power grid’s growing demand for real-time, autonomous intelligent analysis.

[0004] Therefore, to address the aforementioned issues, there is an urgent need for a dynamic knowledge graph online reasoning method and system that is deeply adapted to the characteristics of the power grid domain. This method must be able to handle the temporal and asynchronous nature of event flows, and more importantly, deeply integrate the physical constraints of the power grid (such as power flow equations and electrical connectivity), operational logic (such as protection setting coordination and automatic device operation conditions), and expert experience into the entire process of knowledge graph construction, reasoning evaluation, and evolutionary updates. This will enable real-time, accurate, and autonomously evolving situational awareness and intelligent reasoning regarding the power grid's operational status. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an online reasoning method and system for dynamic knowledge graphs driven by power grid events, in order to solve the problems mentioned in the background art.

[0006] To address the problems of existing power grid knowledge graphs being statically lagging, having superficial reasoning, and lacking autonomous evolution, especially the difficulty of general graph reasoning methods in integrating power grid physical constraints, protection action logic, and real-time temporal causal relationships, this invention provides a dynamic knowledge graph online reasoning method and system that is deeply adapted to the characteristics of the power grid domain. This method constructs an online closed loop from power grid event perception to knowledge consolidation. Its core lies in enabling the knowledge graph to evolve synchronously and dynamically in incremental terms, accompanying the reasoning process of real-time events and based on actual power grid feedback, including electrical associations, protection logic relationships, and equipment influence patterns within the graph.

[0007] This method includes the following steps: S1, the power grid event acquisition and semanticization step, which parses and merges the multi-source real-time event stream of the power grid (such as SCADA alarms, protection action signals, switch changes, PMU / fault recording trigger information, etc.) into driving events with clear power grid domain semantics, providing standardized input for subsequent processing.

[0008] The semanticization process needs to integrate the power grid protection logic with its timing characteristics. For example, the successively arriving "overcurrent stage I protection start" signal and "circuit breaker trip" change signal can be integrated into a "line short circuit fault (main protection action)" driving event.

[0009] S2, the dynamic subgraph construction step, responds to driving events and dynamically constructs a dedicated subgraph from the power grid knowledge graph based on event type (such as short circuit fault, equipment overload, protection malfunction) and core entities (such as faulty line, operating switch).

[0010] The scope of this subgraph is not only determined by the general graph association strength, but more importantly by the power grid-specific association strength such as electrical connectivity, protection coordination relationships, and power flow transfer paths, providing a focused context that conforms to the power grid topology and logic for reasoning.

[0011] S3, the multi-path reasoning and evaluation step, performs joint inference on a dedicated subgraph, simulating the possible propagation process of events in the power grid (such as power flow shifts caused by faults, voltage drops, backup protection actions, etc.), generating one or more possible power grid state evolution paths that the event may lead to, and calculating a comprehensive confidence assessment value for each path, thereby quantifying the uncertainty of the reasoning results. The comprehensive confidence assessment deeply integrates the historical success rate of electrical connections traversed by the path, the quality of real-time measurement data (such as the synchronization and integrity of PMU data), and the consistency strength between the path conclusions and the power grid protection logic and safety and stability rules.

[0012] S4, the actual feedback acquisition step of the power grid, acquires the actual power grid measurements (such as fault waveform, steady-state voltage and current recovery values), protection and switching action reports, and the actual switching sequence recorded by the energy management system within a preset time window after the occurrence of the driving event, as a verification benchmark.

[0013] S5, the arbitration and decision-making step, combines the verification benchmark with the comprehensive confidence assessment value of each evolution path to conduct competitive arbitration, selecting the target evolution path that best matches the actual power grid response. The arbitration process must focus on considering the rationality of the protection action timing, the safety of the power flow transfer path, and other power grid-specific constraints.

[0014] S6, the graph incremental evolution step, generates and executes a set of graph evolution operations based on the target evolution path and its competing path information, and performs real-time and local incremental adjustments to the relation weights (such as electrical connection confidence and protection coordination reliability) and connection existence (such as new equipment association and hidden fault links) between corresponding entities in the power grid knowledge graph.

[0015] Furthermore, in step S3, a comprehensive confidence assessment value is calculated for each evolution path. This calculation process integrates multiple sources of power grid confidence.

[0016] The system first obtains the historical verification success rate of the graph relationship edges on which the path depends (such as the "line-protection" action relationship and the "bus-feeder" power supply relationship) (based on the statistics of the power grid historical fault case library), the real-time data quality score of the corresponding equipment of each node in the path (based on the measurement delay, missing rate, and rationality), and derives the logical consistency strength of the path from different power grid business rules (such as the relay protection setting coordination principle, N-1 safety verification rule, and voltage stability criterion).

[0017] Subsequently, a pre-defined fusion calculation model is used, for example, a weighted average: This integrates the heterogeneous information into a unified quantitative confidence level. Weighting coefficients... , , It can be configured differently according to different power grid analysis scenarios (such as focusing on protection action analysis or focusing on power flow security assessment). This upgrades the credibility assessment of power grid event reasoning paths from a single rule matching to a comprehensive metric based on historical statistics, real-time monitoring, and multi-rule logic verification, providing an objective, multi-dimensional decision-making basis that conforms to the characteristics of power grid operation for subsequent arbitration.

[0018] Furthermore, the competitive arbitration in step S5 employs a tiered screening strategy, which incorporates typical power grid decision-making processes.

[0019] The first layer is electrical state matching filtering. It calculates the direct matching degree between the predicted endpoint state (such as the predicted set of switch states, overloaded line sets) and the actual feedback state (such as actual switch positions, SCADA telemetry over-limit information). Paths with a matching degree below the threshold TH1 are quickly eliminated, and erroneous deductions that fundamentally contradict the actual response of the power grid are excluded.

[0020] The second layer involves prioritizing the protection logic. For the remaining paths, their overall confidence scores are used. The priority score is obtained by weighting and adjusting the direct match score. ,For example This allows paths with clearer protection logic, more reliable historical data, and higher data quality to receive higher priority when factual matches are similar.

[0021] The third layer is the complex competition adjudication. When the priority scores of high-scoring paths are close and all exceed the threshold TH2, the topological complexity of the path (such as the number of hops, reflecting the scope of the fault's impact) and its historical adoption frequency are introduced for the final decision. This layer is particularly suitable for multiple assumptions or complex cascading fault scenarios commonly found in power grids, ensuring that when it is difficult to make a direct judgment, the evolutionary path with a simpler structure and more stable historical performance is selected.

[0022] Furthermore, the graph evolution operations generated in step S6 specifically include positive reinforcement, negative suppression, and hypothetical proposals.

[0023] Positive reinforcement operations target the relational edges that constitute the target evolution path and increase their weight according to their criticality. For example, the weight of the edge "specific short-circuit fault characteristics → correct operation of main protection" is increased, so that the power grid correlation that has been proven effective in practice has a greater influence in subsequent reasoning.

[0024] The reverse suppression operation targets competitive paths with high confidence but not selected, and reduces the weight of the key relation edges that cause bifurcation. For example, it suppresses a "alarm-trip" correlation edge that is frequently misinformed due to measurement noise, thereby gently correcting potentially misleading power grid knowledge.

[0025] Hypothetical proposal operations, when the existing graph cannot explain the newly observed grid entity associations (for example, certain weather conditions frequently cause flashovers of specific insulators, but the association is not established in the graph), create a temporary relationship edge with a "hypothesis" label and low initial confidence between the core entity and the new associated entity, providing an entry point for the exploration and verification of new grid knowledge.

[0026] Furthermore, the system implements dynamic lifecycle management for the temporary relationship edges. If a temporary edge is subsequently... If a given reasoning method is successfully used and correctly verified in different power grid events (for example, the newly established "wind speed-line galloping-protection maloperation" temporary edge is confirmed in fault reasoning under multiple strong winds), its confidence level will gradually increase; when it exceeds the conversion threshold... At that time, the "hypothesis" marker is removed and converted to a standard edge, ensuring that only robust associations that have withstood multiple tests of actual power grid operation can be solidified.

[0027] Conversely, if a temporary edge is within the observation period If an element is not effectively utilized, or if its conclusions consistently contradict actual feedback from the power grid, it is automatically removed. This mechanism endows the power grid knowledge graph with the ability to self-verify and self-cleanse, effectively maintaining the long-term simplicity and reliability of the graph while encouraging the exploration of potential new knowledge about the power grid.

[0028] Furthermore, in step S2, the dynamic construction of dedicated subgraphs is achieved through grid strategy combination and semantic pruning. The system matches and loads multiple graph exploration strategies (e.g., adjacency expansion strategy based on electrical topology, logical expansion strategy based on protection setting coordination range, and load transfer path exploration strategy based on power supply reliability) according to the grid semantics of the driving event (such as fault type and voltage level). These strategies are executed in parallel starting from the core entity to obtain subgraph segments, which are then merged into the original fused subgraph.

[0029] Subsequently, the grid semantic relevance of each element in the fused subgraph to the event topic is calculated (e.g., whether nodes belong to the same electrical island, whether edges belong to critical power supply paths under the current operating mode), and elements with relevance below a threshold are removed. This allows for the intelligent construction of a compact subgraph that encompasses the relevant electrical and logical context while eliminating a significant amount of irrelevant noise for each power grid event inference, thereby improving the efficiency and accuracy of subsequent inference.

[0030] The present invention also provides a system for implementing the above method.

[0031] The system includes an event perception and semantic encapsulation module for executing S1; a dynamic subgraph construction module for executing S2, which contains a policy execution and fusion engine; a coupled reasoning and evaluation module for executing S3, which contains a confidence fusion unit; a feedback arbitration and evolutionary decision module for executing S4 and S5, which contains a multi-stage arbitration unit; and an online graph incremental evolution module for executing S6.

[0032] Each module works collaboratively with the data flow through defined interfaces to jointly realize power grid event-driven online reasoning and adaptive evolution of the knowledge graph.

[0033] The technical effects and advantages of this invention are as follows: Compared with the static construction and offline update mode of knowledge graphs in the prior art, this invention establishes a closed-loop mechanism that coordinates event-driven and online graph evolution, enabling the knowledge graph to respond to real-time events for dynamic reasoning and continuously self-optimize based on feedback. After an event is triggered, the method dynamically constructs a focused context subgraph for multi-path inference and evaluation. Subsequently, it uses actual feedback information from the power grid to arbitrate the reasoning paths and, based on the arbitration results, incrementally and locally adjusts the association weights and connections in the graph in real time. This improves the accuracy and timeliness of the knowledge model's representation of the real-time state of the power grid and reduces long-term maintenance costs.

[0034] Compared to existing simple event response or isolated graph query reasoning based on fixed rules, this invention enhances the depth of reasoning and the rationality of decision-making for complex chain reactions by introducing multi-path inference and a competitive arbitration process based on comprehensive confidence. The system simulates the propagation of event impacts on a dedicated subgraph, generating multiple possible state evolution paths and quantifying the credibility of each path by integrating multi-source information such as historical success rate, data quality, and logical consistency. After obtaining real feedback, the system identifies the path that best matches the actual situation from multiple competing paths through hierarchical screening and arbitration strategies. This process enables reasoning to assess the probability of different development possibilities and make more reliable decisions when evidence is sufficient, thereby improving the ability to analyze uncertain power grid events.

[0035] Compared to the existing architecture where the reasoning module and knowledge update module are separated, this invention tightly binds the generation of graph evolution instructions with each event's reasoning process, achieving automated knowledge accumulation through reasoning as learning. Based on the target path and its competing paths generated by arbitration, the system automatically generates refined adjustment instructions, including relation weight strengthening, suppression, and hypothetical relation proposals. Particularly for recurring association patterns that the graph cannot explain, the system can propose creating and progressively verifying temporary relations, ultimately solidifying them into stable knowledge. This mechanism enables the system to continuously learn from successful or failed reasoning cases, transforming practical experience into structured knowledge, achieving autonomous evolution and improvement of the knowledge system, and enhancing its adaptability to novel events. Attached Figure Description

[0036] Figure 1 is a flowchart of the overall process of online reasoning of the power grid event-driven dynamic knowledge graph of the present invention.

[0037] Figure 2 is a flowchart of the multi-path reasoning and arbitration decision-making process of the present invention.

[0038] Figure 3 is a flowchart of the graph evolution operation and path management branch of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] This embodiment details the complete implementation process of the method of the present invention in a power grid, and explains it in conjunction with Figures 1-3. This method aims to solve the problems of inaccurate reasoning and the inability of knowledge to evolve autonomously in real-time event analysis of power grids due to delayed updates and the inability to integrate power grid physical constraints and protection logic.

[0041] S1, Power Grid Event Acquisition and Semanticization.

[0042] The system connects to the real-time data bus of the power grid to acquire multi-source asynchronous event streams from the SCADA system, protection information management system, fault recorder, and synchronous phasor measurement unit (PMU). The event streams include switch position signals with precise timing marks, protection device action / alarm messages (including parameters such as action time, protection section, and fault phase), telemetry over-limit alarms (such as overvoltage, undervoltage, and overload), and fault recording start signals.

[0043] The complexity of power grid events lies in their strict temporal and causal logic. For example, the correct "sequence of actions" for a line fault should be: fault occurrence → abnormal electrical quantity → protection activation → circuit breaker tripping → fault isolation.

[0044] If the system processes these signals in isolation, its underlying logic cannot be understood. To transform low-level signals into high-level semantic events, the system performs event fusion based on a semantic rule base embedded with power grid protection and automation logic. For example, a key rule is: if the same line interval is detected within an extremely short time window... If the "overcurrent stage I protection start" and "circuit breaker trip" signals appear successively within 30 milliseconds, they are merged into a single structured drive event, and its type is identified as "line short circuit fault (main protection action)".

[0045] The standardized format for driving events is defined as follows: .in, The core entity identifier (e.g., faulty line ID), Type (e.g., "short circuit fault," "overload," "protection malfunction"), Timestamp (the event reference time, usually the time of the first critical signal), and Attributes (a set of attributes, which may include initial confidence level (e.g., 0.85), fault phase, operating protection segment, associated substation, etc.) are defined in this step. Through this step, the system transforms the raw signal stream into structured events with clear power grid semantics, providing standardized input for subsequent reasoning and resolving the misjudgment problem caused by the neglect of temporal causal logic in general event processing methods within the power grid.

[0046] S2, Dynamic Subgraph Construction.

[0047] This step aims to dynamically construct a focused reasoning context that conforms to the power grid's operational logic from the vast global power grid knowledge graph for each driving event. (Power Grid Knowledge Graph) It includes not only equipment entities (lines, transformers, circuit breakers, etc.) and their physical connections (topology), but also protection coordination relationships (such as primary protection, near backup, and far backup), control dependency relationships (such as automatic device interlocking), and implicit associations based on historical operating data statistics. Edges are attached with weight attributes to indicate the strength of the association.

[0048] General graph traversal methods typically consider only a single type of association (such as topological connectivity), which can lead to the construction of subgraphs that either omit crucial logical associations (such as ignoring backup protection) or introduce a large number of irrelevant topological nodes in a power grid. Therefore, this invention employs a power grid strategy combination and semantic pruning mechanism.

[0049] Specifically: First, based on the type (Type) of the driving event, explore the strategy library from the power grid graph. Multiple grid-specific strategies can be invoked in parallel. These strategies may include: topology extension strategies based on electrical wiring diagrams. (Locating electrically connected devices), protection association extension strategy based on protection setting coordination logic (Identifying potentially activated adjacent or upstream backup protections), and power supply impact analysis strategies based on the N-1 safety criterion and power flow calculation model. (After identifying the faulty equipment, it is possible that it may be at risk of overload due to power flow shift).

[0050] by Starting with a node, execute each selected strategy in parallel. Each strategy traverses the graph independently. And return a sub-image segment Perform a union operation on the nodes and edges of all returned sub-image segments to obtain an initial merged graph. .

[0051] Subsequently, semantic relevance of the power grid is pruned. System calculations. Each element (node) or edge ) and the current driving event Semantic relevance of power grid or The correlation calculation comprehensively considers: electrical distance (topology hop count), functional relevance (whether they belong to the same power supply area or protection zone), and the probability of co-occurrence of historical events. For example, in a line fault event, bus nodes that are electrically distant and not within the range of any backup protection will have a very low correlation.

[0052] Those with a correlation coefficient below a preset threshold Remove nodes and edges to generate the final dedicated subgraph. Threshold The value typically ranges from 0.1 to 0.5 and can be adjusted according to different event types (e.g., 0.3 for local faults and 0.1 for system oscillations). Through this step, the system constructs a compact subgraph for each inference, which encompasses the relevant electrical and logical context while eliminating a large amount of irrelevant noise, thus providing a foundation for subsequent accurate and efficient inference.

[0053] S3, Multi-path reasoning and comprehensive confidence assessment.

[0054] This step is in the dedicated subgraph. The above will be used for deduction and calculation. As a source of disturbance, based on pre-built power grid impact propagation models in the knowledge graph (e.g., fault current propagation model, protection action logic model, power flow transfer calculation model after switch tripping), its effects are simulated. Propagation in the network, generation Possible state evolution paths .

[0055] Each evolutionary path It is represented as a sequence of nodes and edges, pointing to one or more deduced conclusion states. For example, path The possible path is: "Line L1 fault → its main protection P1 operates → circuit breaker CB1 trips → fault isolation successful"; The possible manifestations are: "Line L1 fault → main protection P1 fails to operate → backup protection P2 operates with a delay → CB1 trips → may cause adjacent line L2 to overload."

[0056] At the same time, for each evolutionary path Calculate a comprehensive confidence score This calculation process deeply integrates multiple sources of credibility for the power grid: 1. Historical verification success rate: the graph relationship edge set traversed by the calculation path. Each edge Historical verification success rate The success rate is based on the number of times the edge has been arbitrated as valid in historical reasoning. Rather than the total number of times it is used The ratio calculation, i.e. This reflects the long-term statistical reliability of the power grid correlations in the graph.

[0057] 2. Real-time data quality scoring: For the set of nodes in the path... Each node The corresponding physical devices are evaluated to assess the quality score of their current real-time data. The rating is based on the delay in data reporting. Missing rate Sum of numerical rationality indicators A comprehensive judgment is made through a scoring function. calculate: The function output is normalized to the interval [0, 1]. This introduces the instantaneous reliability of the current power grid operating environment.

[0058] 3. Strength of logical consistency of power grid rules: This will be used to extrapolate the conclusions. With an independent power grid business rule base middle Related rules Output results By comparing the results, the consistency strength can be obtained. .

[0059] The rules include relay protection setting coordination principles, N-1 safety and stability verification rules based on the current operating mode, and voltage stability criteria. For example, calculating the average Jaccard similarity: This ensures that the reasoning path does not violate the basic principles of power grid operation.

[0060] Subsequently, the above three types of information are input into a preset fusion computing model. Output path Overall confidence level assessment value .

[0061] In one embodiment, the model takes a linear weighted form: .

[0062] in, The average historical validation success rate for all edges on the path; The average real-time data quality score for all nodes along the path; As defined above. , , The preset weighting coefficients satisfy... and .

[0063] The coefficient configuration can emphasize different analysis objectives. For example, when focusing on the analysis of the correctness of protection actions, the following settings can be configured: , , This approach relies more on historical experience and protection rules; when focusing on operational risk assessment, it can set... , , To focus more on real-time data quality and stability rules.

[0064] Through this step, the system not only generates multiple possible evolution paths, but also assigns each path a comprehensive credibility metric that integrates historical statistics, real-time monitoring, and multi-rule logical verification, providing an objective and multi-dimensional decision-making basis for subsequent arbitration.

[0065] S4. Feedback Acquisition Steps.

[0066] The system sets a configurable time window. Its length is set according to the type of power grid event. For transient faults, It could be from a few seconds to tens of seconds (e.g.) For complex faults that could trigger a chain reaction, It may take several minutes (e.g.) ).

[0067] exist Time window after Timestamp Inside, the system continuously monitors and collects actual response data from the power grid. This data includes: final steady-state measurements (voltage, current) of relevant equipment, complete sequence of actions (SOE) reports of protection and automation devices, final position signal sequences of switches, and waveform files and analysis conclusions (such as fault type and distance measurement) generated by the fault recorder.

[0068] The collected data was organized into a verification benchmark for this event. It is a record of the actual state of relevant entities within a time window. A set of. It is the actual state of the physical entity of the power grid after the event, serving as the objective core standard for judging the correctness of reasoning.

[0069] S5. Arbitration and Decision-Making Steps.

[0070] This step utilizes verification benchmarks. For multiple evolutionary paths and its confidence level The selection process was conducted to determine the most suitable objective evolution path based on the actual situation. The screening process employs a tiered arbitration mechanism.

[0071] First layer: Grid state matching degree filtering. Calculate each evolution path. The deduction conclusion of the state set With verification benchmark The actual set of states in direct matching degree The matching degree is calculated using the Jaccard similarity coefficient: .

[0072] Set a first threshold TH1 to eliminate all [players / entities]. The path. A typical value for TH1 is 0.6. This step quickly eliminates erroneous extrapolations that fundamentally contradict the actual response of the power grid.

[0073] Second layer: Priority scoring for enhanced protection logic. This applies to the set of paths filtered by the first layer. Using its comprehensive confidence assessment value The direct match score is adjusted to obtain the priority score. The calculation formula is: ,in This is an adjustment factor, typically ranging from 0.1 to 0.3, used to control the degree of influence of confidence level on the score. For example, it can be taken as... This means that, when factual matches are similar, paths with clearer protection logic, more reliable historical data, and higher data quality will receive higher priority.

[0074] Third layer: Complex competition adjudication. This occurs when at least two paths exist. The priority score PS is higher than the second threshold TH2, and the difference between their scores is less than 1. When, that is, satisfied , ,and If so, then the decision will proceed to this level. Typical values ​​for TH2 are higher than TH1, for example... ; The typical value is 0.05.

[0075] In this case, the system calculates a final decision score for each candidate path. : .

[0076] in, For path The number of jumps (sides) reflects the extent of the electrical impact of the fault. This reflects considerations of topological complexity, favoring simpler paths; This path has been historically adopted under similar power grid operating conditions. The frequency of is normalized to the [0, 1] interval; These are the weighting coefficients, and (For example, it is advisable) ).choose The highest path is used as the target evolution path. This tiered mechanism ensures that arbitration results respect objective facts while also taking into account the intrinsic quality of the reasoning process, enabling stable decisions to be made in complex competitive situations by leveraging structural and historical knowledge.

[0077] S6, Incremental Evolution Steps of the Spectrum.

[0078] This step, based on the arbitration result, examines the power grid knowledge graph. Make adjustments online. The adjustments are based on the target evolution path. and the competing paths that are activated but not selected in S3 in the dedicated subgraph The adjustment is achieved by performing a set of evolutionary operations. Implementation, specifically including: positive reinforcement operations : Target evolution path Each relation edge on Increase its weight value according to its importance in this reasoning process. The weight update formula is: .in, For the edge The original weights, This is the learning rate parameter, with a value ranging from 0.01 to 0.2 (e.g., ...). ), The increment calculated based on the importance of the edge in the path (which can be set to a constant of 1, or based on the importance of the edge in the path). (Calculation of grid centrality index in the data). This increases the probability that grid correlations proven effective in practice will be selected in subsequent reasoning.

[0079] Inverse suppression operation For those with a high overall confidence level ( However, the competing path was not selected. Location leads to and exist The key relationship edge that generates bifurcation The weight of that side is attenuated. The attenuation formula is: .in, This is the original weight of the edge. This is the attenuation factor, with a value ranging from 0.9 to 0.99 (e.g., ...). ).level It is typically set to 0.5 to 0.7 (e.g.) This gentle correction, rather than outright deletion, of potentially misleading grid correlations reflects a gradual adjustment to uncertain knowledge.

[0080] Hypothetical proposal operation When verifying the benchmark Indicates two entities and There is a correlation between them (for example, under specific weather conditions of "persistent dense fog," a certain type of "outdoor insulator" at substation A will trigger a slight discharge alarm each time), and the graph... When the corresponding relationship is missing in the data, the system can... and Create a new relationship edge The edge is marked as an attribute. And assign a low initial confidence value. , The value range is from 0.1 to 0.3 (e.g. This operation allows the map to explore and record potential new correlations in the power grid.

[0081] For those marked as Relationship The system performs continuous lifecycle management: Promotion mechanism: If In the subsequent series It is used in reasoning about different events (i.e., it is included in the corresponding (in the middle), and all paths it occupies are arbitrated as correct (i.e., become) (a portion of the system), then the system will proceed according to each... The magnitude of the increase in its confidence level . For example, an integer greater than 1. ; Typical values ​​are 0.1 to 0.2 (e.g.) ).when The confidence level was raised to exceed the preset conversion threshold. At that time, that is The system removes its hypothetical label and converts it to a standard edge. The value range is from 0.6 to 0.8 (e.g. ).

[0082] Elimination mechanism: If During the set observation period (For example If an edge is not used in any reasoning process within a given day, or if its path is arbitrated as incorrect every time it is used, the system will remove the edge from the graph. Delete it.

[0083] The above evolution mechanism endows the power grid knowledge graph with the ability to self-verify and self-purify. While encouraging the exploration of potential new knowledge about the power grid, it effectively maintains the long-term simplicity and reliability of the graph, prevents the accumulation of invalid or erroneous knowledge, and realizes the autonomous evolution of the knowledge system.

[0084] Example 2 (Example of a typical fault reasoning scenario in a power grid): To more specifically and intuitively demonstrate the working process of the present invention in a power grid, a typical scenario is illustrated using a 220kV line L1 experiencing a transient fault, with the main protection operating correctly and the reclosing successfully.

[0085] S1 event semantics: The system receives signals with closely related timestamps in sequence: "Line L1 overcurrent I-stage protection starts" at time t1, "Circuit breaker CB1 opens" at time t2 (t2-t1=25ms<30ms threshold) at time t3, "Reclosing device starts" at time t4, "Circuit breaker CB1 closes" at time t5, "Line L1 current and voltage return to normal" at time t5.

[0086] Based on the semantic rule base, the system merges the t1 and t2 signals into a driving event. The core entity is line L1, the type is "line L1 transient short circuit fault (main protection action, reclosing successful)", the timestamp is t1, and the initial confidence level is set to 0.85.

[0087] S2 Dynamic Subgraph Construction: Starting with L1 as the initial entity, the topology strategy (obtaining adjacent buses B1 and B2, transformer T1), protection strategy (obtaining L1's main protection P1, its backup protection P2, and failure protection P3) and power supply impact strategy based on N-1 (obtaining line L2 that is electrically coupled to L1 and may be affected by power flow transfer) are executed in parallel.

[0088] After merging the preliminary results, the semantic relevance of each element to the topic of "line short circuit fault" is calculated.

[0089] Assume a feeder with an electrical distance of 4 stops and no protection coordination with L1 has an correlation of 0.1 (below the threshold). If the value is 0, it will be clipped. This results in a dedicated subgraph. It includes entities such as L1, CB1, B1, B2, T1, P1, P2, P3, and L2, as well as their relationships.

[0090] S3 Multi-Path Reasoning and Evaluation: In The fault propagation simulation was performed, generating two main paths: Path (Target path): L1 fault → main protection P1 operates correctly → CB1 trips → fault arc extinguishes → reclosing operation → CB1 closes → system recovers.

[0091] path (Competition path): L1 fault → main protection P1 fails to operate → backup protection P2 delays operation → CB1 trips → fault isolation, but may result in load loss.

[0092] Calculate the overall confidence level: For Its reliance on edges such as "L1-P1 action" has a high historical success rate. The average historical verification success rate is calculated based on all edges on the path. The relevant node data is of good quality, with an average data quality score. Consistency of rules (complies with the main protection's fast-acting and reclosing logic) .

[0093] Pick ,but .

[0094] for Its key edge "P1 refusal to move" has a historically extremely low success rate. Data quality Rule consistency (complies with backup protection logic but violates the high reliability of primary protection) ,but .

[0095] S4 Feedback Acquisition: In the post-event time window Internal, actual feedback collected This indicates that CB1 tripped at time 2 and closed at time 4, the electrical quantities of line L1 returned to normal, and no other protection actions were taken. This reflects the actual state. {CB1 trips, CB1 closes, L1 recovers}.

[0096] S5 Arbitration Decision: First-level Matching Filtering: {CB1 trips, CB1 closes, L1 recovers} ; {CB1 tripped (delayed)}, (Assumption), let , Eliminated.

[0097] Second-level priority scoring: of .

[0098] The third layer was not triggered (only one path passed the filter).

[0099] therefore, Determined as the target path .

[0100] S6 map incremental evolution: positive reinforcement: path The critical edge "L1 fault → main protection P1 action" has the original weight. According to the formula Strengthen it.

[0101] Reverse inhibition: Overall confidence level Therefore, no suppression is performed.

[0102] Hypothetical proposal: No new associations were found in this case that cannot be explained by the graph, so it was not triggered.

[0103] This embodiment demonstrates that the present invention closely aligns with the physical characteristics, protection logic, and real-time data of the power grid, achieving accurate online reasoning. Through this event processing, the weight of the core association "line fault → main protection action" in the knowledge graph increased from 0.82 to 0.92. When the system encounters similar faults subsequently, the confidence level for deduceing the correct path will be higher, reflecting the positive accumulation and consolidation of knowledge. Simultaneously, since no hypothetical proposal was triggered, it indicates that the current event did not expose any unknown defects in the knowledge graph; the system is in a knowledge consolidation phase rather than an exploration phase, demonstrating the system's adaptive operating mode under different scenarios.

[0104] A system embodiment provides an online reasoning system for dynamic knowledge graphs driven by power grid events, implementing the above method. The system includes the following modules connected via a data interface: an event perception and semantic encapsulation module, used to access real-time power grid data streams and execute step S1 as described in Embodiment 1, transforming the original event sequence into driving events with power grid domain semantics. .

[0105] The dynamic subgraph construction module connects the event perception and semantic encapsulation module and the power grid knowledge graph. This is used to perform step S2 as described in Embodiment 1, constructing a dedicated subgraph based on the driving event. Internally, it includes a strategy execution and fusion engine, which dynamically selects and combines power grid traversal strategies based on the semantic features of driving events, and performs pruning on the preliminary exploration results based on the semantic relevance of the power grid.

[0106] The coupled reasoning and evaluation module, connected to the dynamic subgraph construction module, is used to execute step S3 as described in Embodiment 1, performing multi-path inference on the dedicated subgraph and generating an evaluation value with comprehensive confidence. The reasoning results are presented. Internally, a confidence fusion unit is included to aggregate the inherent confidence of the relationships upon which the evolution path depends, real-time data quality reliability indicators, and the consistency of multi-rule reasoning conclusions. The comprehensive confidence assessment value is then output through a fusion function.

[0107] The feedback arbitration and evolutionary decision-making module, connected to the real-time feedback data stream from the power grid and linked to the coupled reasoning and evaluation module, is used to execute steps S4 and S5 as described in Embodiment 1, and to derive the target evolution path based on feedback arbitration. It generates graph evolution instructions. Internally, it contains a multi-stage arbitration unit for performing coarse-grained screening based on fact matching, fine-grained ranking based on confidence weighting, and invoking decision routines based on topology and historical patterns when competition occurs.

[0108] The online graph incremental evolution module, connected to the feedback arbitration and evolution decision module and the power grid knowledge graph, is used to execute step S6 as described in Embodiment 1, executing the graph evolution instructions, and... The relation weights and connections are adjusted incrementally.

[0109] Each module is deployed as a software service, and exchanges data through message middleware or service call interfaces. Together, they complete a closed-loop process from power grid event input to knowledge graph update, and jointly realize the adaptive and collaborative evolution of online reasoning and knowledge graph driven by power grid events.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0111] Finally, several points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation," "connection," and "linkage" should be interpreted broadly, and can refer to mechanical or electrical connections, or internal connections between two components, or direct connections. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may change. Second, the accompanying drawings of the embodiments disclosed in this invention only involve structures relevant to the embodiments disclosed in this invention; other structures can refer to common designs. Where there is no conflict, the same embodiment and different embodiments of this invention can be combined with each other. Finally, the above descriptions are merely preferred embodiments of this invention and are not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A dynamic knowledge graph online reasoning method driven by power grid events, characterized in that, The method includes the following steps: S1. Acquire real-time power grid data stream, parse and fuse the multi-source event sequences into driving events with power grid domain semantics; S2. Based on the type and core entity of the driving event, dynamically construct a dedicated subgraph from the power grid knowledge graph, the boundary of which is determined according to the event semantics and the strength of the graph association path; S3. Perform joint deduction on the dedicated subgraph to generate at least one power grid state evolution path caused by the driving event, and calculate a comprehensive confidence assessment value for each evolution path; S4. Acquire the actual power grid measurement and action feedback information within a time window T after the occurrence of the driving event as a verification benchmark, where T is a preset duration; S5. Arbitrate based on the verification benchmark and the comprehensive confidence assessment value of each evolution path to select a target evolution path; S6. Based on the information of the target evolution path and its competing paths, generate a set of graph evolution operations to incrementally adjust the relation weights and connections in the power grid knowledge graph.

2. The online reasoning method for dynamic knowledge graphs driven by power grid events according to claim 1, characterized in that, Step S3, which calculates a comprehensive confidence assessment value for each evolution path, includes: S31, obtaining at least two types of information from the historical verification success rate of the graph relationship edges on which the evolution path depends, the real-time data quality score of the nodes in the path, and the logical consistency strength of the path derived from different business rules; S32, inputting the obtained information into a preset fusion calculation model and outputting the quantitative comprehensive confidence assessment value of the path.

3. The online reasoning method for dynamic knowledge graphs driven by power grid events according to claim 2, characterized in that, The arbitration in step S5 adopts a tiered screening strategy: S51, calculate the direct matching degree between the deduced endpoint state of each evolution path and the verification benchmark, and eliminate paths with a matching degree lower than the first preset threshold TH1; S52, for the remaining paths, use their respective comprehensive confidence evaluation values ​​to weight and correct the direct matching degree to obtain a priority score; S53, if there are at least two paths with priority scores higher than the second preset threshold TH2 and the difference between them is less than Δ, then the topological complexity and historical adoption frequency of each path are introduced for final adjudication to determine the target evolution path.

4. The online reasoning method for dynamic knowledge graphs driven by power grid events according to claim 3, characterized in that, The graph evolution operations generated in step S6 include: S61, increasing the weight attribute value of the relation edges constituting the target evolution path according to their criticality; S62, for competing paths that are activated but not selected in inference and whose overall confidence evaluation value is higher than level L, reducing the weight of the critical relation edges that cause bifurcation in the dedicated subgraph; S63, when the verification benchmark indicates the existence of unexplained associations in the graph, creating a path with a "hypothesis" label and initial confidence between the core entity of the driving event and the new association entity reflected by the verification benchmark. Temporary relationship.

5. The online reasoning method for dynamic knowledge graphs driven by power grid events according to claim 4, characterized in that, It also includes the management of the temporary relationship edge: if the temporary relationship edge is successfully used and verified correctly in subsequent N consecutive event reasoning, its confidence level is increased, and when it exceeds the conversion threshold... Remove the "hypothesis" tag and it becomes a standard relation edge, where N>1; if the temporary relation edge is within the observation period If an element is not effectively utilized or its utilization leads to incorrect conclusions, it will be removed from the graph.

6. The online reasoning method for dynamic knowledge graphs driven by power grid events according to claim 1, characterized in that, Step S2, which involves dynamically constructing a dedicated subgraph, includes: S21, parsing the semantics of the driving event, matching and loading at least one basic exploration strategy from a pre-configured strategy library; S22, starting from the core entity, executing each loaded strategy in parallel to generate preliminary subgraph exploration results; S23, performing a union operation on nodes and edges of all preliminary results to form an original fused subgraph; S24, calculating the correlation between each node and edge in the original fused subgraph and the current event topic, removing the parts with correlation below the threshold θ, and outputting the dedicated subgraph.

7. A dynamic knowledge graph online reasoning system driven by power grid events, characterized in that, include: The event awareness and semantic encapsulation module is used to transform the original event sequence into driving events; A dynamic subgraph construction module, connected to the event perception and semantic encapsulation module and the power grid knowledge graph, is used to construct a dedicated subgraph based on the driving event; The coupled reasoning and evaluation module, connected to the dynamic subgraph construction module, is used to perform inferences on the dedicated subgraph and generate multi-path reasoning results with comprehensive confidence evaluation values. The feedback arbitration and evolution decision module is connected to the real-time feedback data stream of the power grid and the coupled reasoning and evaluation module. It is used to derive the target evolution path based on feedback arbitration and generate graph evolution instructions. The online graph incremental evolution module is connected to the feedback arbitration and evolution decision module and the power grid knowledge graph. It is used to execute the graph evolution instructions.

8. The online reasoning system for dynamic knowledge graphs driven by power grid events according to claim 7, characterized in that, The coupled reasoning and evaluation module includes a confidence fusion unit, which aggregates the inherent confidence of the relationships on which the evolution path depends, the real-time data quality confidence index, and the consistency of multi-rule reasoning conclusions, and outputs the comprehensive confidence evaluation value through a fusion function.

9. The online reasoning system for dynamic knowledge graphs driven by power grid events according to claim 7, characterized in that, The feedback arbitration and evolutionary decision-making module includes a multi-stage arbitration unit for performing coarse-grained screening based on fact matching, fine-grained sorting based on confidence weighting, and invoking decision routines based on topology and historical patterns when competition occurs.

10. The online reasoning system for dynamic knowledge graphs driven by power grid events according to claim 7, characterized in that, The dynamic subgraph construction module includes a strategy execution and fusion engine, which is used to dynamically select and combine graph traversal strategies based on the semantic features of the driving events, and to perform semantic relevance-based pruning on the preliminary exploration results.

Images