Construction method and system of power distribution network knowledge graph

Abstract

The invention relates to the technical field of artificial intelligence, and particularly discloses a construction method and system of a power distribution network knowledge graph, and the method comprises the steps: obtaining and analyzing power distribution network multi-modal data from a heterogeneous data source; generating a mixed feature vector for a to-be-aligned entity by adopting a mixed representation model enhanced by a graph neural network, and constructing a static knowledge layer of the power distribution network; analyzing event elements in the dynamic text data stream, and generating an event atomic stream carrying a standardized semantic tag; performing time sequence association on events in the event atomic flow and entities in the static knowledge layer to construct a directed event evolution chain, and effectively solving the problems of alignment and conflict resolution of multi-source heterogeneous data through a mixed representation model enhanced by a graph neural network and a fusion strategy based on confidence; and in combination with time sequence association and evolution chain construction, it is ensured that the knowledge graph can accurately reflect the dynamic change of the power grid.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically a method and system for constructing a knowledge graph for a power distribution network. Background Technology

[0002] With the increasing scale and automation level of power distribution networks, massive amounts of multimodal heterogeneous data are generated during their operation. These include equipment ledgers from production management systems (PMS), real-time measurements and alarms from SCADA systems, and operation tickets and fault reports filled out by maintenance personnel. Efficiently integrating, managing, and deeply mining the knowledge contained in this data is crucial for improving the fault diagnosis, condition assessment, and decision support capabilities of power distribution networks, and is a core element in achieving intelligent operation and maintenance of the power grid.

[0003] However, existing technologies still face numerous challenges in constructing and applying power distribution network knowledge graphs. First, the integration of multi-source heterogeneous data is difficult, with "data silos" existing between different business systems. Inconsistent device naming and attribute descriptions make it difficult to construct accurate and complete power grid topology knowledge. Second, existing knowledge graphs are mostly static models, struggling to effectively analyze and correlate the massive, real-time dynamic event streams in the power grid, failing to fully reflect the evolution of the power grid's operational status. Furthermore, existing analysis methods often remain at the level of apparent event relationships, lacking the ability to uncover the deep causal logic between events, limiting the value of knowledge graphs in core applications such as fault root cause tracing. Simultaneously, existing knowledge graphs lack effective adaptive optimization mechanisms, unable to iteratively improve themselves based on newly discovered knowledge.

[0004] Therefore, there is an urgent need to propose a new technical solution to address the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for constructing a knowledge graph of a power distribution network, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a knowledge graph for a power distribution network, the method comprising: The system acquires and parses multimodal data of the distribution network from heterogeneous data sources, performs entity extraction and standardization; uses a hybrid representation model enhanced by graph neural network to generate hybrid feature vectors for entities to be aligned, performs cross-source entity alignment based on the hybrid feature vectors, and uses a weighted fusion strategy based on data source confidence assessment to resolve conflicting attributes and construct a static knowledge layer of the distribution network. A joint extraction model based on prompt word engineering and template matching adaptive routing is used to parse event elements in dynamic text data streams and generate event atomic streams carrying standardized semantic tags. The events in the event atomic flow are temporally associated with the entities in the static knowledge layer to construct a directed event evolution chain. A two-stage hybrid-driven causal discovery algorithm is used to mine the potential causal relationships between events and construct a temporal causal knowledge graph. Establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

[0007] As a further embodiment of the present invention, the entity extraction and standardization process specifically includes: A multimodal entity extractor is built based on rule templates, domain dictionaries, and named entity recognition models; Construct standardized data models for power distribution network equipment; Generate topology-consistent signatures for different source entities based on electrical topology attributes and perform pre-alignment.

[0008] As a further embodiment of the present invention, the step of generating the hybrid feature vector specifically includes: A large language model is used to encode entity text information to generate initial semantic feature vectors; Construct a heterogeneous topology network that includes electrical connections and logical attribution relationships; Generate structural topological feature vectors based on state-aware heterogeneous graph neural networks; The initial semantic feature vector and the structural topological feature vector are concatenated to generate a hybrid feature vector.

[0009] As a further aspect of the present invention, the steps of performing cross-source entity alignment based on the hybrid feature vector and resolving conflicting attributes using a weighted fusion strategy based on data source confidence assessment specifically include: Aligned entity pairs are determined by cosine similarity and topological consistency verification based on hybrid feature vectors. Assign differentiated confidence scores to each data source based on attribute type; A weighted fusion strategy is used to generate a unique fusion entity.

[0010] As a further embodiment of the present invention, the joint extraction model parsing process includes: For the input dynamic text data stream, the adaptive routing module adaptively routes the text to the template matching path or the large model parsing path based on the degree of text structuring; Template matching paths use regular expression templates to extract event elements from structured text; The large model parsing path utilizes enhanced cue words from the static knowledge layer to parse event elements from unstructured text.

[0011] As a further embodiment of the present invention, the construction of the event evolution chain includes: Based on electrical topology relationships, related event atoms are aggregated around triggering events to form a candidate event set, wherein the triggering event is the event identified as the root cause in the event atom stream; Construct protection-priority and upstream-priority sorting rules, and perform causal logic sorting of events based on the sorting rules to generate a directed event evolution chain; Annotate the directed event evolution chain with structured attributes.

[0012] As a further embodiment of the present invention, the two-stage hybrid-driven causal discovery algorithm includes: Causal candidate pruning based on domain knowledge base: Define causal pruning rules based on static knowledge layer to filter candidate causal sets from directed event evolution chains; Asymmetric relationship verification based on environmental perception: Collect environmental state data, calculate the environmental perception condition transition probability of candidate causal sets, and confirm the causal relationship by comparing the positive and negative condition transition probabilities.

[0013] As a further aspect of the present invention, the step of establishing a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversing the correction of the confidence level of entity attributes and the association relationship between entities in the static knowledge layer, specifically includes: The confidence level of data source attributes is iteratively adjusted based on the number of causal relationship conflicts in the time-series causal knowledge graph. The implicit association strength is calculated based on the transition probability of environmental perception conditions, and the implicit association relationships in the static knowledge layer are supplemented. Incrementally update the causal relationship probabilities in the temporal causal knowledge graph.

[0014] As a further embodiment of the present invention, the causal pruning rules include protection-enforcement rules, upstream-downstream influence rules, and fault-symptom rules.

[0015] This invention also provides a system for constructing a distribution network knowledge graph, used to implement the aforementioned method for constructing a distribution network knowledge graph, the system comprising: The data access and preprocessing module is used to acquire and parse multimodal data of the power distribution network from heterogeneous data sources, and to perform entity extraction and standardization processing. The static knowledge fusion module is used to generate hybrid feature vectors for entities to be aligned using a hybrid representation model enhanced by graph neural networks, perform cross-source entity alignment based on the hybrid feature vectors, and resolve conflicting attributes using a weighted fusion strategy based on data source confidence assessment to construct a static knowledge layer for the distribution network. The dynamic event parsing module is used to parse event elements in dynamic text data streams using a joint extraction model based on prompt word engineering and template matching adaptive routing, and generate event atomic streams carrying standardized semantic tags. The temporal graph construction module is used to temporally associate events in the event atomic flow with entities in the static knowledge layer to construct a directed event evolution chain. It uses a two-stage hybrid-driven causal discovery algorithm to mine potential causal relationships between events and construct a temporal causal knowledge graph. The knowledge closed-loop feedback module is used to establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and to reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

[0016] Compared with the prior art, the beneficial effects of the present invention are: by using a hybrid representation model enhanced by graph neural networks and a fusion strategy based on confidence, the problem of alignment and conflict resolution of multi-source heterogeneous data is effectively solved, and the constructed static power grid topology knowledge has high reliability; The designed adaptive joint extraction model can parse various dynamic event data streams in real time and efficiently. Combined with temporal correlation and evolution chain construction, it ensures that the knowledge graph can accurately reflect the dynamic changes of the power grid. The original two-stage hybrid-driven causal discovery algorithm achieves a leap from correlation analysis to causal inference, and can uncover the deep causal chain hidden behind events that conforms to electrical logic, providing information support for fault tracing and risk warning; The innovative closed-loop feedback mechanism enables knowledge graphs to break free from the limitations of static databases, allowing them to self-correct and iteratively enhance themselves based on dynamically discovered knowledge. This achieves continuous accumulation and improvement of knowledge, giving them high long-term application value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0018] Figure 1 A flowchart illustrating the overall steps of a method for constructing a power distribution network knowledge graph provided by this invention; Figure 2 A flowchart illustrating the construction process of the static knowledge layer in a method for constructing a power distribution network knowledge graph provided by this invention. Figure 3 A flowchart illustrating the construction process of the dynamic event layer in a method for constructing a knowledge graph of a power distribution network provided by this invention. Figure 4 A flowchart illustrating the construction process of a time-series causal knowledge graph, which is part of a method for constructing a knowledge graph for a power distribution network provided by this invention. Figure 5 The flowchart illustrates the closed-loop feedback and adaptive optimization of a method for constructing a distribution network knowledge graph provided by this invention. Detailed Implementation

[0019] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0020] Figure 1 This is a flowchart illustrating a method for constructing a distribution network knowledge graph. In this embodiment of the invention, a method for constructing a distribution network knowledge graph includes: The system acquires and parses multimodal data of the distribution network from heterogeneous data sources, performs entity extraction and standardization; uses a hybrid representation model enhanced by graph neural network to generate hybrid feature vectors for entities to be aligned, performs cross-source entity alignment based on the hybrid feature vectors, and uses a weighted fusion strategy based on data source confidence assessment to resolve conflicting attributes and construct a static knowledge layer of the distribution network. A joint extraction model based on prompt word engineering and template matching adaptive routing is used to parse event elements in dynamic text data streams and generate event atomic streams carrying standardized semantic tags. The events in the event atomic flow are temporally associated with the entities in the static knowledge layer to construct a directed event evolution chain. A two-stage hybrid-driven causal discovery algorithm is used to mine the potential causal relationships between events and construct a temporal causal knowledge graph. Establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

[0021] In this embodiment, step one, static knowledge layer construction, involves acquiring and parsing data from multiple heterogeneous data sources in the distribution network, such as Production Management System (PMS), SCADA, and line ledgers, and extracting and standardizing the equipment entities contained therein. Subsequently, a hybrid representation model enhanced by graph neural networks is used to generate hybrid feature vectors containing both semantic and structural information for entities representing the same physical equipment but originating from different data systems. Based on these vectors, cross-source entity alignment is performed, and a weighted fusion strategy based on data source confidence assessment is used to resolve attribute conflicts between aligned entities, ultimately constructing a highly reliable static knowledge layer for the distribution network. Step 2: Dynamic Event Layer Construction: For dynamic text data streams such as fault reports, operation tickets, and alarm logs generated during the operation of the distribution network, a joint extraction model based on prompt word engineering and template matching adaptive routing is adopted to parse event elements in real time and generate event atomic streams carrying standardized semantic tags. The model can efficiently process structured high-frequency events and unstructured low-frequency and novel events. Step 3: Generation of Temporal Causal Knowledge Graph: Above the static knowledge layer, events in the event atomic flow are temporally associated with one or more uniquely identified entities in the static knowledge layer to construct an event evolution chain; then, a two-stage hybrid-driven causal discovery algorithm is used to mine the potential causal relationships between events, thereby constructing a dynamically evolving temporal causal knowledge graph. Step 4, Closed-loop feedback and adaptive optimization: Based on the causal relationships or frequent event patterns with high confidence mined above, a closed-loop feedback loop is established from the temporal causal knowledge graph to the static knowledge layer. This loop corrects or enhances the confidence of entity attributes and the relationship between entities in the static knowledge layer, thereby achieving self-improvement and accuracy enhancement of the constructed knowledge graph. Step 5: Provide the optimized knowledge graph to downstream applications such as intelligent operation and maintenance analysis of the power distribution network.

[0022] like Figure 2 As shown, in a preferred embodiment of the present invention, the entity extraction and standardization process specifically includes: A multimodal entity extractor is built based on rule templates, domain dictionaries, and named entity recognition models; Construct standardized data models for power distribution network equipment; Generate topology-consistent signatures for different source entities based on electrical topology attributes and perform pre-alignment.

[0023] In this embodiment, power grid entity identification based on multimodal feature fusion is implemented as follows: For structured data in PMS, time-series measurement data in SCADA system, and semi-structured text in line ledgers, a multimodal entity extractor based on rule templates, domain dictionary, and named entity recognition (NER) model is constructed and integrated; the rule template extracts entities according to equipment naming specifications, the domain dictionary includes power grid equipment models, manufacturers, and terminology, and the NER model is used to identify equipment entities and their attributes from the descriptive text of the ledgers; Standardized model construction for topology relationships: Based on the "GB / T 33601-2017 General Model Data Naming Specification for Power Grid Equipment", a standardized data model for distribution network equipment with topological connection relationships and electrical attributes is defined and extended. This model enforces constraints on key fields such as equipment ID, voltage level, upstream and downstream connection nodes, and phase, which are used to uniformly map and clean the heterogeneous entity information extracted in the previous steps to form structured standard equipment objects. Entity pre-alignment based on electrical topology consistency: During the standardization process, topology consistency signatures are generated for entities from different sources based on strongly correlated electrical topology attributes such as feeder, switch status, and voltage level. By comparing these signatures, pre-alignment labels are applied to multiple standardized device objects that may belong to the same physical device, providing a high-confidence candidate set for subsequent entity alignment.

[0024] like Figure 2 As shown, in a preferred embodiment of the present invention, the step of generating the hybrid feature vector specifically includes: A large language model is used to encode entity text information to generate initial semantic feature vectors; Construct a heterogeneous topology network that includes electrical connections and logical attribution relationships; Generate structural topological feature vectors based on state-aware heterogeneous graph neural networks; The initial semantic feature vector and the structural topological feature vector are concatenated to generate a hybrid feature vector.

[0025] In this embodiment, a hybrid representation model enhanced by graph neural networks is used to generate hybrid feature vectors. Furthermore, a pre-trained large-scale power language model is introduced to inject prior knowledge of power expertise, specifically including: Domain semantic feature extraction: A large language model, incrementally trained and fine-tuned on a corpus specific to the power industry, is used to encode the textual information such as the name, type, and description of the entity i to be aligned from each data source, generating an initial semantic feature vector representing the prior knowledge of the power industry. ; Construction of heterogeneous topology networks for distribution networks: Abstracting the topology of distribution networks into a heterogeneous information network containing multiple node types and edge types. Among them, the node set V includes physical nodes of equipment such as transformers, circuit breakers, and busbars; the edge set E contains two main types of relationships: one is the electrical connection relationship that represents the physical wiring between equipment (at this time, the attributes of the edge carry electrical parameters such as line impedance, length, and rated current), and the other is the logical affiliation relationship of equipment in space or management (such as switching equipment belonging to a certain switching station, or a certain piece of equipment being installed on a specific pole). State-aware directed heterogeneous graph aggregation: On the heterogeneous information network, a state-aware heterogeneous graph neural network is applied to encode structural information to obtain structural topological features. When aggregating the characteristics of neighboring nodes, the network can distinguish between upstream (power supply side) and downstream (load side) and perform differentiated processing based on different relationship types such as electrical connection and logical affiliation. At the same time, it introduces SCADA real-time switch status as a dynamic adjustment factor. When the switch is open, it blocks or attenuates the information transmission on the corresponding path to ensure that the encoded structural information is consistent with the actual operating state of the power grid. Hybrid feature vector generation: This involves generating the initial semantic feature vector. The runtime-aware structural topology feature vector aggregated with S1-2-3 The features are concatenated and merged to generate a final hybrid feature vector used for entity alignment. : ; in, It is a learnable fusion weight matrix, where CAT represents the feature concatenation operation. It is a bias vector. It is a non-linear activation function.

[0026] The constructed heterogeneous information network can be formally represented as: .

[0027] Where V is the set of device entity nodes, and E is the set of edges. It is a collection of node types (such as transformers, circuit breakers, busbars, etc.). It is a collection of edge types (such as electrical connections, logical affiliations, etc.). Each node They all have a type Each edge They all have a type .

[0028] In practice, the step of encoding structural information to obtain structural topological features using the state-aware heterogeneous graph neural network is specifically implemented by processing any entity node in the network. The feature vectors are updated layer by layer, and the feature vectors of the last layer are concatenated to obtain the structural topology feature vectors. The update process for each layer is as follows: Define the information transmission coefficient for state awareness: For a connection path between entity node i and any of its neighboring nodes j, calculate the information transmission coefficient based on the real-time status of all switches on that path provided by the SCADA system. When the path is completely closed, The value approaches 1 when any switch in the path is open. The value approaches 0 or is a preset attenuation value, thereby achieving dynamic adjustment of information transmission; Aggregate directional and heterogeneous neighborhood information: Divide the neighboring nodes of entity node i into upstream neighbor sets. and downstream neighbor sets For the feature vectors of neighbor node j in these two sets respectively Aggregation is performed; before aggregation, the corresponding learnable weight matrix is ​​selected based on the relationship type r between node i and node j (including but not limited to electrical connection, logical affiliation, etc.). right Perform the transformation and multiply by the information transfer coefficient defined in step (1). ; Update the generated structural topology features: Concatenate the aggregated information of upstream neighbors with the aggregated information of downstream neighbors, and combine them with the features of node i itself, using a nonlinear activation function. The layer is updated to ultimately generate the encoded structural topological feature vector. The complete update formula is as follows: ; ; ; Among them, AGG up and AGG down These are the aggregation results for upstream and downstream neighbors, respectively. and It is the upstream and downstream feature transformation matrix related to relation r, and AVG represents the calculation of the feature mean.

[0029] like Figure 2 As shown, in a preferred embodiment of the present invention, the steps of performing cross-source entity alignment based on the hybrid feature vector and resolving conflicting attributes using a weighted fusion strategy based on data source confidence assessment specifically include: Aligned entity pairs are determined by cosine similarity and topological consistency verification based on hybrid feature vectors. Assign differentiated confidence scores to each data source based on attribute type; A weighted fusion strategy is used to generate a unique fusion entity.

[0030] In this embodiment, cross-source entity alignment is performed based on hybrid feature vectors, and a weighted fusion strategy is used to construct a highly reliable static knowledge layer, specifically including: Entity alignment based on hybrid similarity and topological consistency: First, calculate the hybrid feature vector (H) of all entity pairs (i,j) to be aligned from different data sources. i H jThe cosine similarity between the entities is used to filter out entity pairs with similarity higher than a preset threshold, forming a candidate matching set. Then, the consistency of their local topology in their respective source graphs is verified. The strong constraint characteristics of the power grid topology are used to eliminate false matches with similar features but incorrect electrical connection positions, thereby determining the final aligned entity pairs. Attribute-based dynamic data source confidence assessment: Construct an attribute-aware confidence assessment model to assign differentiated confidence scores to each data source (PMS, SCADA, ledgers, etc.) based on different attribute categories. Where s represents the data source, To represent attribute categories, this model follows these principles: For equipment ledger attributes (such as equipment model, commissioning date, etc.), assign the highest confidence level to PMS; For real-time status and measurement attributes (such as switch status, voltage and current values), assign the highest confidence level to the SCADA system; For topology connection and naming convention attributes, assign a high confidence level to the data source of the line ledger.

[0031] Weighted fusion and conflict resolution for electrical knowledge: For multiple aligned data source instances representing the same physical entity, all their attributes are traversed, and weighted fusion is performed to generate a unique, highly reliable fused entity. Resolution rules include, but are not limited to: Category attribute fusion: For conflicting category or text attributes (such as device model), the value provided by the data source with the highest confidence among the aforementioned dynamic data sources is directly adopted; Numerical attribute fusion: For conflicting numerical attributes (such as line resistance and rated capacity), a weighted average method with the confidence scores of the aforementioned dynamic data source as weights is used for calculation; Status attribute verification: The switch status provided by the SCADA source is forced to be the final fusion value. Based on this, the connectivity analysis of the fused static topology is performed to ensure that the final knowledge layer accurately reflects the current main wiring diagram of the power grid, thereby completing the construction of a high-reliability distribution network static knowledge layer.

[0032] like Figure 3 As shown, in a preferred embodiment of the present invention, the joint extraction model parsing process includes: For the input dynamic text data stream, the adaptive routing module adaptively routes the text to the template matching path or the large model parsing path based on the degree of text structuring; Template matching paths use regular expression templates to extract event elements from structured text; The large model parsing path utilizes enhanced cue words from the static knowledge layer to parse event elements from unstructured text.

[0033] In this embodiment, dynamic text data streams are collected and input, including but not limited to fault reports, SCADA alarms, and other data; Establish an adaptive routing module for event types: For the input dynamic text data stream, pre-determine its text structuring level and event type, and based on the determination result, adaptively route the text to an efficient template matching path or a knowledge-enhanced large model parsing path; Efficient template matching path: Apply pre-designed regular expression templates to quickly match and extract key event elements from structured, high-frequency text (such as SCADA standard alarms); The large-scale model parsing path of knowledge enhancement: The prompt words of the large language model are dynamically enhanced by using the static knowledge layer constructed above. First, the device entities and their topological context knowledge related to the text content are vectorized and retrieved in the static knowledge layer. Then, these knowledge and the original text are used to construct an enhanced prompt word to guide the large language model to deeply parse event elements from unstructured, low-frequency text (such as fault analysis reports). The parsed event elements are generated into an event atomic stream carrying standardized semantic tags, thereby completing the construction of the dynamic event layer.

[0034] Specifically, the parsed event elements are generated into an event atomic stream carrying standardized semantic tags, including: Standardization of unified event elements: The extracted event elements are uniformly mapped to a predefined standardized event atomic structure that includes fields such as unique event ID, timestamp, event type label, involved entities, and attribute description; Forced linking of event entities to static knowledge: When generating event atoms, the device entities involved are forcibly linked to the unique identifiers of the corresponding entities in the static knowledge layer, ensuring that each dynamic event can be accurately anchored to the corresponding node in the static power grid topology; Constructing a dynamic knowledge layer that evolves over time: A series of standardized and entity-linked event atoms are organized in time stamp order to form a dynamic event knowledge layer that reflects the temporal evolution of the power grid state and is closely related to the physical topology.

[0035] like Figure 4 As shown, in a preferred embodiment of the present invention, the construction of the event evolution chain includes: Based on electrical topology relationships, related event atoms are aggregated around triggering events to form a candidate event set, wherein the triggering event is the event identified as the root cause in the event atom stream; Construct protection-priority and upstream-priority sorting rules, and perform causal logic sorting of events based on the sorting rules to generate a directed event evolution chain; Annotate the directed event evolution chain with structured attributes.

[0036] In this embodiment, the process of generating the temporal causal knowledge graph includes: The events in the event atomic flow are temporally associated with one or more uniquely identified entities in the static knowledge layer to construct an event evolution chain; A two-stage hybrid-driven causal discovery algorithm is used to mine potential causal relationships between events. The first stage is causal candidate pruning based on the domain knowledge base to form a high-confidence candidate causal set. A two-stage hybrid-driven causal discovery algorithm is used to mine potential causal relationships between events. The second stage is based on environment-aware asymmetric relationship verification, thereby constructing a dynamically evolving temporal causal knowledge graph.

[0037] Specifically, the events in the event atomic flow are temporally associated with one or more uniquely identified entities in the static knowledge layer to construct an event evolution chain, including: Event aggregation and grouping based on electrical topology: For each triggering event (such as fault alarm or protection action) identified as the root cause in the event atom flow, a dynamic spatiotemporal topology window is defined using the precise electrical topology relationships in the static knowledge layer. This window aggregates related event atoms with the triggering event as the center, forming a candidate event set that is highly correlated with a specific power grid disturbance. Event sequencing and chaining based on electrical timing logic: Within each candidate event set, a sequencing rule that integrates protection priority and upstream priority is applied to causally sequence the events and connect them into a directed event evolution chain with the root cause event as the starting node, time as the axis, and related entities as the carriers. The sequencing rule includes, but is not limited to: Protection priority principle: Force protection logic events such as "protection device action" and "overcurrent / overvoltage alarm" to be prioritized before equipment execution events such as "circuit breaker / switch trip". Upstream priority principle: When timestamps are very close, events that occur closer to the power source (upstream) in the electrical topology are forced to be prioritized before downstream events.

[0038] Attribute labeling and structuring of event evolution chains: For each generated event evolution chain, structured attributes are calculated and labeled to characterize its overall features. These structured event evolution chains are then used as input for subsequent causal discovery algorithms. These attributes include, but are not limited to: Root cause event type: Indicates the type of the starting event in this chain (e.g., lightning strike, short circuit, overload). Topology impact range: By aggregating all related entities on the chain, the specific lines, substations and customer ranges affected by the event chain are calculated and marked in the static knowledge layer; Evolutionary characteristics: Calculate and label the total evolution time and fault clearing time of the chain from the first event to the last event.

[0039] As a preferred embodiment of the present invention, the two-stage hybrid-driven causal discovery algorithm includes: Causal candidate pruning based on domain knowledge base: Define causal pruning rules based on static knowledge layer to filter candidate causal sets from directed event evolution chains; Asymmetric relationship verification based on environmental perception: Collect environmental state data, calculate the environmental perception condition transition probability of candidate causal sets, and confirm the causal relationship by comparing the positive and negative condition transition probabilities.

[0040] In this embodiment, causal pruning rules based on a static knowledge layer are defined: relying on a highly reliable static knowledge layer and combined with electrical principles, a set of deterministic causal pruning rules are defined, including but not limited to protection-execution rules, upstream-downstream influence rules, and fault-symptom rules. Protection-Execution Rules: Verify whether there is a clear protection and protected relationship between the entities associated with event A (e.g., overcurrent protection activation) and event B (e.g., circuit breaker tripping) in the static knowledge layer; Upstream-downstream influence rule: Query the electrical topology location of the entities associated with event A and event B from the aforementioned static knowledge layer, verify whether the entity of event A is indeed upstream (power supply side) of the entity of event B, and thus determine whether the causal constraint of power flow direction is satisfied; Fault-Symptom Rules: Verify whether the entities associated with event A (e.g., single-phase ground fault) and event B (e.g., zero-sequence current surge) are the same device or adjacent electrical nodes, and confirm that their device types conform to the fault mechanism.

[0041] Filtering candidate causal pairs: Perform causal pruning on the aforementioned event evolution chain, and verify any event pair in the chain. Only when their relationship does not violate any of the above rules are they retained, thus forming a high-confidence candidate causal set.

[0042] Verification based on environment-aware asymmetric relationships includes: Collect concurrent environmental state data for events: For the aforementioned high-confidence candidate causal set, synchronously collect the power grid environmental state vector at the time of its occurrence, including but not limited to weather information (such as lightning, strong wind, icing), load level (such as heavy load and light load recorded by SCADA) and power grid operation mode (such as planned maintenance information recorded by PMS). Calculate the conditional transition probability of environmental perception: Based on the aforementioned historical event evolution chain dataset as a statistical sample, and using the collected environmental state vector as a condition, calculate the conditional transition probability of environmental perception for candidate causal pairs (assuming A is the cause and B is the effect). This probability represents the probability under a specific environmental state. Given that event A has occurred, the probability of event B occurring. Asymmetric causality confirmation: The asymmetry of causality is verified by comparing the positive and negative conditional transition probabilities. A causal relationship between a pair of events (A, B) can be confirmed if and only if the following conditions are met: ; That is, under specific circumstances, the occurrence of A significantly increases the probability of the occurrence of B, and this increasing effect is far greater than the reverse relationship; all the confirmed causal relationships together constitute a dynamically evolving temporal causal knowledge graph.

[0043] like Figure 5 As shown, in a preferred embodiment of the present invention, the step of establishing a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversing the confidence level of entity attributes and the association relationship between entities in the static knowledge layer, specifically includes: The confidence level of data source attributes is iteratively adjusted based on the number of causal relationship conflicts in the time-series causal knowledge graph. The implicit association strength is calculated based on the transition probability of environmental perception conditions, and the implicit association relationships in the static knowledge layer are supplemented. Incrementally update the causal relationship probabilities in the temporal causal knowledge graph.

[0044] In this embodiment, the iterative correction of the data source attribute confidence is as follows: based on the high-confidence causal relationships in the dynamic temporal causal knowledge graph, the number of times they conflict with the entity attributes of the static knowledge layer is counted. Confidence of data source attributes Perform iterative decay: ; in, and These represent the confidence scores before and after the update, respectively. This is the preset penalty coefficient; Quantitative supplementation of implicit entity relationships: For entity pairs that exhibit strong correlations in dynamic causal relationships but have no explicit connections in static topology. Aggregate all environment-aware conditional transition probabilities associated with the entity pair and calculate the implicit association strength. : ; in, and They are respectively related to entity ei and e j The relevant l-th and k-th events, This refers to the environmental conditions at the time. When the correlation strength... When the threshold is exceeded, a new implicit relationship is added to the entity pair in the static knowledge layer, and its strength is used as the confidence level of the relationship. Online updates of the temporal causal knowledge graph: As new event evolution chains are continuously generated, the probabilities of causal relationships in the temporal causal knowledge graph are also continuously and incrementally updated. For any causal pair (A→B) in a specific environment... The conditional transition probability is updated as follows: ; Among them, C t Represents the observation count at time t. This represents the count of newly observed events.

[0045] This invention also provides a system for constructing a distribution network knowledge graph, used to implement the method for constructing the distribution network knowledge graph, the system comprising: The data access and preprocessing module is used to acquire and parse multimodal data of the power distribution network from heterogeneous data sources, and to perform entity extraction and standardization processing. The static knowledge fusion module is used to generate hybrid feature vectors for entities to be aligned using a hybrid representation model enhanced by graph neural networks, perform cross-source entity alignment based on the hybrid feature vectors, and resolve conflicting attributes using a weighted fusion strategy based on data source confidence assessment to construct a static knowledge layer for the distribution network. The dynamic event parsing module is used to parse event elements in dynamic text data streams using a joint extraction model based on prompt word engineering and template matching adaptive routing, and generate event atomic streams carrying standardized semantic tags. The temporal graph construction module is used to temporally associate events in the event atomic flow with entities in the static knowledge layer to construct a directed event evolution chain. It uses a two-stage hybrid-driven causal discovery algorithm to mine potential causal relationships between events and construct a temporal causal knowledge graph. The knowledge closed-loop feedback module is used to establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and to reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a knowledge graph for a power distribution network, characterized in that, The method includes: Acquire and parse multimodal data of power distribution network from heterogeneous data sources, and perform entity extraction and standardization processing; A hybrid representation model enhanced by graph neural networks is used to generate hybrid feature vectors for entities to be aligned. Cross-source entity alignment is performed based on the hybrid feature vectors, and a weighted fusion strategy based on data source confidence assessment is used to resolve conflicting attributes and construct a static knowledge layer for the distribution network. A joint extraction model based on prompt word engineering and template matching adaptive routing is used to parse event elements in dynamic text data streams and generate event atomic streams carrying standardized semantic tags. The events in the event atomic flow are temporally associated with the entities in the static knowledge layer to construct a directed event evolution chain. A two-stage hybrid-driven causal discovery algorithm is used to mine the potential causal relationships between events and construct a temporal causal knowledge graph. Establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

2. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The entity extraction and standardization process specifically includes: A multimodal entity extractor is built based on rule templates, domain dictionaries, and named entity recognition models; Construct standardized data models for power distribution network equipment; Generate topology-consistent signatures for different source entities based on electrical topology attributes and perform pre-alignment.

3. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The specific steps for generating the hybrid feature vector include: A large language model is used to encode entity text information to generate initial semantic feature vectors; Construct a heterogeneous topology network that includes electrical connections and logical attribution relationships; Generate structural topological feature vectors based on state-aware heterogeneous graph neural networks; The initial semantic feature vector and the structural topological feature vector are concatenated to generate a hybrid feature vector.

4. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The steps of performing cross-source entity alignment based on the hybrid feature vectors and resolving conflicting attributes using a weighted fusion strategy based on data source confidence assessment specifically include: Aligned entity pairs are determined by cosine similarity and topological consistency verification based on hybrid feature vectors. Assign differentiated confidence scores to each data source based on attribute type; A weighted fusion strategy is used to generate a unique fusion entity.

5. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The joint extraction model parsing process includes: For the input dynamic text data stream, the adaptive routing module adaptively routes the text to the template matching path or the large model parsing path based on the degree of text structuring; Template matching paths use regular expression templates to extract event elements from structured text; The large model parsing path utilizes enhanced cue words from the static knowledge layer to parse event elements from unstructured text.

6. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The construction of the event evolution chain includes: Based on electrical topology relationships, related event atoms are aggregated around triggering events to form a candidate event set, wherein the triggering event is the event identified as the root cause in the event atom stream; Construct protection-priority and upstream-priority sorting rules, and perform causal logic sorting of events based on the sorting rules to generate a directed event evolution chain; Annotate the directed event evolution chain with structured attributes.

7. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The two-stage hybrid-driven causal discovery algorithm includes: Causal candidate pruning based on domain knowledge base: Define causal pruning rules based on static knowledge layer to filter candidate causal sets from directed event evolution chains; Asymmetric relationship verification based on environmental perception: Collect environmental state data, calculate the environmental perception condition transition probability of candidate causal sets, and confirm the causal relationship by comparing the positive and negative condition transition probabilities.

8. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The step of establishing a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and reversing the confidence level of entity attributes and the relationship between entities in the static knowledge layer, specifically includes: The confidence level of data source attributes is iteratively adjusted based on the number of causal relationship conflicts in the time-series causal knowledge graph. The implicit association strength is calculated based on the transition probability of environmental perception conditions, and the implicit association relationships in the static knowledge layer are supplemented. Incrementally update the causal relationship probabilities in the temporal causal knowledge graph.

9. The method for constructing a distribution network knowledge graph according to claim 1, characterized in that, The causal pruning rules include protection-enforcement rules, upstream-downstream influence rules, and fault-symptom rules.

10. A system for constructing a distribution network knowledge graph, used to implement the method for constructing a distribution network knowledge graph according to any one of claims 1-9, characterized in that, The system includes: The data access and preprocessing module is used to acquire and parse multimodal data of the power distribution network from heterogeneous data sources, and to perform entity extraction and standardization processing. The static knowledge fusion module is used to generate hybrid feature vectors for entities to be aligned using a hybrid representation model enhanced by graph neural networks, perform cross-source entity alignment based on the hybrid feature vectors, and resolve conflicting attributes using a weighted fusion strategy based on data source confidence assessment to construct a static knowledge layer for the distribution network. The dynamic event parsing module is used to parse event elements in dynamic text data streams using a joint extraction model based on prompt word engineering and template matching adaptive routing, and generate event atomic streams carrying standardized semantic tags. The temporal graph construction module is used to temporally associate events in the event atomic flow with entities in the static knowledge layer to construct a directed event evolution chain. It uses a two-stage hybrid-driven causal discovery algorithm to mine potential causal relationships between events and construct a temporal causal knowledge graph. The knowledge closed-loop feedback module is used to establish a feedback loop from the temporal causal knowledge graph to the static knowledge layer, and to reversely correct the confidence of entity attributes and the relationship between entities in the static knowledge layer.

Images