Power system graph relationship completion method and system based on multi-source evidence and causal reasoning

Abstract

The application belongs to the technical field of power system artificial intelligence and knowledge graph construction, and discloses a power system graph relationship completion method and system based on multi-source evidence and causal reasoning, which comprises the following steps: obtaining multi-source heterogeneous operation data to construct a multi-source heterogeneous evidence graph; extracting the static semantics and dynamic operation characteristics of entities based on a double-flow encoder, and obtaining the joint representation of the entities through cross-modal attention mechanism fusion; calculating the Granger causality index to generate a causal prior mask matrix, and injecting a graph attention network to aggregate the causal constraints and reason the dynamic relationship probability; using D-S evidence theory to synthesize the rules and fuse the multi-source prediction probability, and outputting the confidence interval to filter the noise relationship; and updating the original knowledge graph by using the high-confidence triplets. The application can mine the implicit linkage relationship between devices, block the information propagation that violates the causality law, quantify the uncertainty of the completion result, and improve the ability and reliability of the graph completion under dynamic conditions.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and knowledge graph construction technology for power systems, specifically involving a method and system for completing power system graph relationships based on multi-source evidence and causal reasoning. Background Technology

[0002] In recent years, knowledge graphs, with their powerful knowledge representation and reasoning capabilities, have played a crucial role in artificial intelligence applications such as intelligent question answering, recommendation systems, and fault diagnosis. In the power system field, knowledge graphs are widely used in scenarios such as power grid topology analysis, dispatching support decision-making, and intelligent operation and maintenance. Although existing power system knowledge graphs contain a large number of entities and static relationships, they remain incomplete in actual operational scenarios, particularly lacking dynamic relationships between devices arising from changes in operating states, such as "cascading tripping," "fault propagation," and "operational linkage." Based on this, the task of knowledge graph relationship completion has been proposed to fill in the missing relationship facts in the graph.

[0003] With the development of deep learning technology, many research works on knowledge graph relation completion mainly revolve around the related technologies of knowledge graph representation learning. Existing relation completion methods mainly include the following categories: (1) Methods based on translation models, such as translation embedding model (TransE) and rotation embedding model (RotatE), which regard relations as translation or rotation operations of head and tail entities in a low-dimensional vector space. These methods are mostly based on the assumption of a closed world and regard relations as static translation vectors. They cannot perceive the temporal changes of power grid operation and are difficult to characterize the dynamic causal directionality caused by fault propagation between power equipment; (2) Methods based on graph neural networks Methods based on graph neural networks (GNNs), such as graph convolutional networks (GCNs) and graph attention networks (GATs), predict relationships by aggregating the features of neighboring nodes. However, such methods only transmit information based on static topology and lack explicit modeling of temporal causal logic. They are very easy to fill in false associations that violate the laws of electrical physics, such as predicting devices without electrical connections as linked relationships. (3) Methods based on pre-trained language models extract entity description features by leveraging the powerful text encoding capabilities of models such as BERT. However, in the field of power systems, simple text description information is relatively limited and cannot capture the temporal dynamic features of equipment operating status.

[0004] Furthermore, power systems exhibit multi-source heterogeneous data characteristics. Real-time telemetry and teleindication data from Supervisory Control and Data Acquisition (SCADA) systems, equipment ledger data from Power Production Management (PMS) systems, and inspection texts reflect the relationships between equipment from different dimensions. Existing zero-sample or small-sample completion methods typically rely on a single data source (such as relying solely on PMS ledger data or solely on inspection texts), failing to effectively integrate complementary evidence from multi-source heterogeneous data. Moreover, they lack a measure of the uncertainty of the completion results, leading to low confidence levels in long-tail equipment and complex operating conditions, making them difficult to directly apply to power grid control and protection operations with high reliability requirements.

[0005] However, the dynamic relationships between devices in a power system often exhibit strict causal transitivity and temporal dependence. We hope to introduce external multi-source information with richer semantics and temporal characteristics to improve the ability and reliability of power system knowledge graph relationship completion under complex dynamic conditions. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method and system for completing power system knowledge graph relationships using multi-source evidence and causal reasoning. By constructing a multi-source heterogeneous evidence graph that integrates multi-dimensional information and introducing a temporal causal reasoning mechanism and DS evidence theory, it better solves the problems of completing dynamic implicit relationships and calibrating confidence in power system knowledge graphs.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A method for completing the power system graph relationship using multi-source evidence and causal reasoning includes the following steps: Step S1: Obtain multi-source heterogeneous data related to the power system knowledge graph and construct a multi-source heterogeneous evidence graph. The multi-source heterogeneous data includes SCADA telemetry and telecontrol data as time-series data, PMS ledger data as structural data, and inspection text data as text data. Two devices with a relationship to be completed (such as a circuit breaker and a disconnector) are selected as candidate entity pairs. Cross-source data alignment is performed using the candidate entity pairs as anchor points. Time-series waveform segments, discrete ledger attributes, and text semantic segments are respectively used as heterogeneous evidence of the existence of the relationship to construct a multi-source heterogeneous evidence graph.

[0008] Step S2: Utilize a dual-stream encoder to jointly represent entities in the multi-source heterogeneous evidence graph. Design a text-attribute stream encoder and a temporal stream encoder to extract static semantic features and dynamic operational features of entities from the multi-source heterogeneous evidence graph, respectively, to obtain dual-stream features, which serve as independent feature representations for each data source. Further, fuse the dual-stream features through a cross-modal attention mechanism to obtain a joint entity representation containing multi-source evidence information.

[0009] Step S3: Introduce a temporal causal mechanism to discover implicit relationships by combining entity joint representations. Calculate the Granger causality index between dynamic features of entities to generate a causal prior mask matrix; inject the causal prior mask matrix into a graph attention network to suppress invalid information flow in non-causal directions, infer the probability of dynamic relationships between candidate entity pairs, and output the joint feature probability of candidate entity pairs.

[0010] Step S4: Multi-source fusion and confidence calculation based on DS evidence theory. Based on the joint feature probability and the independent feature probabilities calculated through the independent feature representation of each data source, the multi-source probabilities are fused using the synthesis rules of DS evidence theory to output the final probability and confidence interval of the dynamic relationship. Low-confidence noisy relationships are filtered out according to the set threshold to obtain high-confidence relationship triples.

[0011] Step S5: Update the original knowledge graph using high-confidence dynamic relation triples. Relationships with confidence levels meeting a preset threshold are written into the graph to complete the power system graph relationships, forming a dynamic power system knowledge graph that supports time-series queries and causal backtracking.

[0012] Furthermore, step S2 is detailed as follows: The text-attribute stream encoder uses a pre-trained language model to encode textual semantic fragments in a multi-source heterogeneous evidence graph, and combines this with embedded representations of discrete ledger attributes to extract static semantic representations of entities through average pooling. The temporal stream encoder uses a temporal convolutional network to extract temporal waveform segments from multi-source heterogeneous evidence graphs, and inputs them into a temporal Transformer to capture long-distance state dependencies and extract the dynamic operational representation of entities. Subsequently, using static semantic representations as queries and dynamic runtime representations as keys and values, cross-modal attention scores are calculated, and joint entity representations are fused and output.

[0013] Furthermore, step S3 is as follows: Calculate the Granger causality scores between candidate entity pairs and their temporal features, and construct a causal prior mask matrix. M Construct a graph attention network and correlate its attention coefficients with a causal prior mask matrix. M Multiply, normalize, and then perform weighted fusion to obtain the updated joint entity representation. Finally, the joint feature probability of the candidate entity pair is calculated using the Sigmoid function.

[0014] Furthermore, step S4 is detailed as follows: The joint feature probabilities are transformed into mass functions; graph convolutional networks are constructed for structural data sources, BERT models are constructed for text data sources, and LSTM models are constructed for time-series data sources, serving as their respective prediction models; independent feature representations from each data source are input into their respective prediction models to obtain independent feature probabilities, which are then transformed into independent mass functions; the Dempster synthesis rule is used to calculate the fused mass function; and a confidence function for the validity of the relationship is calculated based on the fused mass function. Bel With likelihood function Pl ,Will[ Bel , Pl As a confidence interval for a completion relation, it is only when... Bel When the value exceeds the set threshold, the relationship completion is deemed valid.

[0015] A power system graph relationship completion system based on multi-source evidence and causal reasoning, used to implement the above method, includes: Multi-source evidence alignment and construction module: Deployed on the data access server, with a built-in SCADA parser and ledger data interface, responsible for aligning multi-source heterogeneous data and constructing local evidence subgraphs with entity pairs as the core; Dual-stream joint representation module: used to construct the underlying feature extraction base of the system; it includes text-attribute stream coding unit and temporal stream coding unit, and outputs entity joint representation that fuses multi-source information through cross-modal attention layer; Temporal causal reasoning module: It has a built-in Granger causal calculation unit and a causal prior mask matrix generation unit. It injects the causal prior mask matrix during graph attention network aggregation and outputs dynamic relationship predictions that conform to physical logic. Evidence fusion and calibration module: Receives prediction results from various data source branches, performs probabilistic fusion based on the DS evidence theory synthesis rules, and outputs the confidence interval for the validity of the relationship. Bel , Pl [and filter low-confidence noise;] Dynamic graph update module: Deployed on the graph database server, it writes dynamic relation triples with timestamps and confidence intervals into the graph database, supporting querying and visualization of time series graphs.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Unlike existing completion methods that rely on a single graph structure or text information, this invention integrates SCADA time series data, ledger structure data and text data based on multi-source heterogeneous evidence graphs, and introduces the real-time operating status of the power system into the graph completion process. It can uncover the implicit linkage relationship caused by changes in the operating mode, and make the graph have dynamic evolution capabilities.

[0017] (2) Unlike existing graph reasoning methods that lack modeling of physical causal logic, this invention innovatively introduces a priori mask based on Granger causality in graph attention networks, which blocks the propagation of information that violates the law of electrical causality, fundamentally reduces the error completion caused by “false co-occurrence”, and enhances the interpretability and physical compliance of relation prediction.

[0018] (3) Unlike existing methods that output deterministic 0 / 1 predictions, this invention uses DS evidence theory to fuse and quantify the uncertainty of prediction results of heterogeneous data. This not only improves the robustness of long-tail device relationship prediction, but also provides a confidence interval for high reliability applications in power grid systems, supporting decision control based on risk thresholds. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the power system graph relationship completion method based on multi-source evidence and causal reasoning provided for an embodiment; Figure 2 This is a schematic diagram illustrating the construction of a multi-source heterogeneous evidence graph provided in the embodiments. Figure 3 A diagram of a dual-stream encoder and cross-modal fusion architecture is provided for this embodiment. Figure 4 A schematic diagram of the temporal causal reasoning network structure provided for the embodiment; Figure 5 This is a schematic diagram illustrating the confidence level calculation based on the DS evidence theory, provided for an example. Detailed Implementation

[0020] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0021] This invention presents a method for completing power system knowledge graph relationships using multi-source evidence and causal reasoning. It aims to address the problems in power system knowledge graphs where static topology cannot reflect dynamic linkages, evidence from a single data source is insufficient, and graph reasoning violates physical causality. The implementation process of this method is as follows: Figure 1 As shown, it mainly includes the following core steps: Step S1: Obtain multi-source heterogeneous data related to the power system knowledge graph and construct a multi-source heterogeneous evidence graph. Relationships in the power system are often corroborated by multiple sources of data. For a specific candidate entity... (That is, two entities awaiting determination of their relationship) The acquired multi-source heterogeneous data is evidence that can reflect the dynamic relationship between devices. Figure 2 The construction logic of the multi-source heterogeneous evidence graph is demonstrated, specifically including: Step S1.1: Using the state change or limit violation event of the candidate entity pair in the SCADA telemetry and information database as the trigger point, extract the characteristic event window before and after the event occurs. Timing waveform segments within the time frame serve as timing evidence. ; Step S1.2: Perform a correlation query on the candidate entity pair in the PMS ledger database to extract the attributes of the two devices (such as the plant or substation to which they belong, voltage level, etc.), and align them to the same dimension, using discrete ledger attributes as structural evidence. ; Step S1.3: Scan and extract text semantic fragments that appear together in the same sentence or paragraph as text evidence from the inspection text database. ; Step S1.4: Take the entities in the candidate entity pair as device nodes and the three types of evidence as evidence nodes. Construct a local evidence subgraph attached to the candidate entity pair through the relationship between devices and the relationship between devices and evidence, thereby obtaining a multi-source heterogeneous evidence graph.

[0022] Step S2: Jointly represent entities in the multi-source heterogeneous evidence graph using a two-stream encoder. To enable entity representations to simultaneously include static semantics and dynamic operational features, this invention designs a two-stream encoder architecture. For example... Figure 3 The diagram illustrates the specific process of dual-stream encoder and cross-modal fusion: Step S2.1, Static Stream Encoding. The pre-trained language model BERT is used to encode the textual evidence. Combined with structural evidence The embedding representation is then used to obtain the static semantic representation through average pooling. : in, This represents the encoding operation of the BERT model. This represents the vectorized embedding operation of ledger attributes. This represents a vector concatenation operation. This indicates that the average pooling operation is performed on the vector sequence.

[0023] Step S2.2, Dynamic Stream Coding. For temporal evidence... A temporal convolutional network (TCN) is used to extract local waveform features, which are then input into a temporal transformer to capture long-range state dependencies, resulting in a dynamic representation of the operation. : in, This represents the feature extraction operation of a temporal convolutional network. This represents the feature extraction operation of the temporal Transformer encoder.

[0024] Step S2.3, Cross-modal fusion. Using static semantic representation as the query and dynamic runtime representation as the key and value, calculate the cross-modal attention score and fuse the output entity joint representation. Z : in, Q For querying the matrix, K The key matrix, V For value matrices, d k is the dimension of the key matrix, used to scale the dot product to prevent gradient vanishing. For normalized exponential functions, Represents vector concatenation operations. Joint entity representation. Z It includes both the textual specification of the entity and the temporal commonality of its operational state.

[0025] Step S3: Implicit Relation Discovery Based on Temporal Causality Mechanism. Traditional graph attention networks rely solely on topological distance aggregation, which can easily lead to incorrect connections of irrelevant devices. This invention introduces causal priors, such as... Figure 4 As shown, the structure and mask generation logic of the temporal causal mechanism inference network are illustrated: Step S3.1: Granger causality index calculation. Dynamic operational representation based on entities. Calculate candidate entity pairs middle right Granger causality score .like The historical time sequence characteristics significantly reduced the influence on The prediction error of the current state indicates the existence of causal explanatory power, as shown in the following formula: in, Representing entities At any moment t The true time-series feature values, Representing entities At any moment t Previous historical time series feature values, Representing entities At any moment t Previous historical chronological characteristics, This is the autoregressive prediction function. The numerator represents only the function using... The prediction error at that time, the denominator represents the joint error. The larger the score, the greater the prediction error. right The stronger the causal explanatory power, the better.

[0026] Step S3.2: Construct the causal prior mask matrix M If Granger causality index Below the set threshold Then the matrix M The corresponding element is 0 if it is 0, otherwise it is 1, that is: in, Representing entities right The mask value.

[0027] Step S3.3: Causal Constraint Graph Attention Aggregation. In the graph attention network, the attention coefficients are combined with the causal prior mask matrix. M Multiplication enables neighborhood aggregation under causal constraints: in, W To share the learnable linear transformation weight matrix, , , for , , Joint representation of entities This represents a vector concatenation operation. For the learnable weight vector in the attention mechanism, For activation function, For entities The set of neighboring device nodes, for right Attention coefficient It is a non-linear activation function. Updated after aggregation with causal constraints Joint representation of entities.

[0028] Step S3.4: Joint Feature Probability Calculation. Based on the updated joint entity representation. calculate right The joint feature probability is: ,in, Representing entities The updated joint entity representation, Representing entities The updated entity joint representation.

[0029] Step S4: Multi-source fusion and confidence calculation based on DS evidence theory. Since a single data source may contain noise (such as remote signaling jitter or textual ambiguity), this step quantifies the uncertainty of the probability prediction results. For example... Figure 5 The diagram illustrates the confidence calibration process based on the DS evidence theory: Step S4.1: Construct independent mass functions. For structured data sources, construct a graph convolutional network; for text data sources, construct a BERT model; and for time-series data sources, construct an LSTM model. The output is denoted as... , indicating that the candidate entity is paired with In the data source k Independent feature probabilities under, k It can be a structural, textual, or temporal data source. Let's define the recognition framework. ,in, This indicates that the relationship is established. If the relationship is not established, then for the data source... k Its independent feature probability p k The corresponding mass function is as follows: , in, Indicates data source k The probability that the support relationship is true. Indicates data source k The probability that the support relationship is not valid.

[0030] Step S4.2: Dempster synthesis rules fuse the mass function. For joint feature probabilities... Then the mass function under the joint feature representation is The conflict coefficient is calculated by fusing the mass functions from three data sources and the mass function from the joint feature representation using the orthogonal summation rule. K : in, A , B , C , D The recognition frameworks represent data sources from structural, textual, temporal, and joint feature representations, respectively. a subset of The subset includes , 、 ; K Let be the conflict coefficient between pieces of evidence, used to measure the degree of conflict between predictions from different data sources. Then the fused mass function is: in, This is a normalization factor to ensure that the sum of the mass functions after fusion is 1; E For identification framework A subset of.

[0031] Step S4.3: Calculate the confidence interval. Calculate the confidence function. With likelihood function : in, Let be the trust function, representing the probability that the relationship is explicitly supported. Let be the likelihood function, representing the probability that the relation does not oppose the validity of the relation (including the uncertain part). The confidence interval that constitutes the complement relation. If ( If a threshold is set (e.g., 0.85), then the dynamic relationship completion is deemed valid, and... This serves as a reliability attribute for the relationship.

[0032] Step S5: Update the original knowledge graph using high-confidence dynamic relation triples. The completed dynamic relation triples, i.e. (device) Dynamic relationships, equipment The original knowledge graph stored in the graph database (which was pre-built using knowledge graph technologies such as data collection, entity recognition, and static relationship extraction) is updated by adding a valid timestamp and confidence interval, forming a dynamic power knowledge graph that supports time-series querying and causal backtracking.

[0033] The aforementioned power system knowledge graph relationship completion method based on multi-source evidence and causal reasoning, employing a dual-stream coding and evidence fusion model, retains static specification knowledge while introducing dynamic temporal features. Furthermore, the designed causal prior mask introduces power system-specific causal constraints into the graph reasoning network, thereby providing a more reliable reasoning basis for the dynamic relationship completion task of the power system knowledge graph, enhancing the ability and confidence of knowledge graph completion under dynamic conditions. To better illustrate the effectiveness of the method provided in the above embodiments, this embodiment uses the knowledge graph of a regional power grid system's dispatch logs, SCADA telemetry database, and PMS ledger database as an example.

[0034] First, aligned data is extracted from multi-source systems to construct an evidence graph. For example, for circuit breaker CB1 and disconnector DS3, the temporal evidence consists of the remote signaling change sequences of both before and after operation (length 10, sampling frequency 1Hz), the textual evidence is the description of the "interlocking relationship between the circuit breaker and the disconnector" in the regulations (converted to 128-dimensional word vectors after jieba word segmentation and stop word removal), and the structural evidence is the substation affiliation attribute of both (e.g., "substation S1", converted to 32-dimensional one-hot encoding). Specifically, 12,356 operation records from January to June 2023 are selected from the dispatch log, 1,200,000 remote signaling change data from the SCADA remote signaling and telemetry database are collected during the same period, and 8,764 equipment ledger data are extracted from the PMS ledger database. These three are aligned using a time window of ±5 seconds to ensure the accuracy of the temporal evidence.

[0035] Subsequently, joint representations were extracted based on the two-stream encoder: the static stream encoding used a pre-trained language model BERT (12 layers, 768 hidden layers, 12 multi-head attention heads) to encode textual evidence. Combined with structural evidence The 32-dimensional vector embedding is used to obtain the static semantic representation through average pooling (taking the global average of the BERT output sequence). Dynamic stream coding for time-series evidence A temporal convolutional network (TCN, kernel size 3, number of layers 2, hidden layer size 64) is used to extract local waveform features, which are then input into a temporal transformer (number of layers 2, hidden layer size 128, multi-head attention heads 8) to capture long-range state dependencies and obtain dynamic runtime representations. Subsequently, attention mechanisms will be used to... and Fusion yields a joint representation of entities. Z In the temporal causal reasoning stage, SCADA data shows that the action of CB1 strictly precedes that of DS3 in time (CB1 closes at 10:00:00, and DS3 closes at 10:00:05). Using CB1 as the dependent variable and DS3 as the independent variable, a Granger causality index with a lag order of 2 (significance level) was calculated. The exponent value is 0.87 (higher than the threshold of 0.8), therefore the causal prior mask matrix... M set up (Allow information flow) (Suppress information flow) to prevent reverse erroneous reasoning.

[0036] Finally, through the DS evidence theory fusion: time-series evidence is allocated based on a 99.5% accuracy rate in remote signaling displacement. , Textual evidence is allocated authoritatively according to procedures. , Structural evidence is allocated based on the frequency of ledger updates. , and , Based on this, the conflict coefficient is calculated. K =0.05, the synthesized trust function Bel =0.88 (exceeding the threshold γ=0.85), likelihood function Pl =0.92. Finally, the dynamic relationship (CB1, operation linkage, DS3) is robustly completed and the confidence interval is assigned to [0.88, 0.92].

[0037] This invention also provides a power system graph relationship completion system based on multi-source evidence and causal reasoning, used to implement the method described in Embodiment 1, including: Multi-source evidence alignment and construction module: Deployed on the data access server, with a built-in SCADA parser and ledger data interface, responsible for aligning multi-source heterogeneous data and constructing local evidence subgraphs with entity pairs as the core.

[0038] Dual-stream Joint Representation Module: Used to construct the underlying feature extraction foundation of the system. This module includes a text-attribute stream coding unit and a temporal stream coding unit, which outputs a joint entity representation that fuses multi-source information through a cross-modal attention layer.

[0039] Temporal causal reasoning module: The core computing module of the system, which has a built-in Granger causal computing unit and a causal prior mask matrix generation unit. It injects the causal prior mask matrix during graph attention network aggregation and outputs dynamic relationship predictions that conform to physical logic.

[0040] Evidence Fusion and Calibration Module: This module receives the prediction results from each data source branch, performs probabilistic fusion based on the DS evidence theory synthesis rules, and outputs the confidence interval for the validity of the relationship. Bel , Pl [, and filter low-confidence noise.]

[0041] Dynamic graph update module: Deployed on the graph database server, it writes dynamic relation triples with timestamps and confidence intervals into the graph database, supporting querying and visualization of time series graphs.

Claims

1. A method for completing the power system graph relationship based on multi-source evidence and causal reasoning, characterized in that, Includes the following steps: Step S1: Obtain multi-source heterogeneous data related to the power system knowledge graph and construct a multi-source heterogeneous evidence graph; the multi-source heterogeneous data includes SCADA telemetry and telecontrol data as time-series data, PMS ledger data as structural data, and inspection text data as text data; the two devices to be completed are taken as candidate entity pairs, and cross-source data alignment is performed with the candidate entity pairs as anchor points; time-series waveform segments, discrete ledger attributes, and text semantic segments are respectively taken as heterogeneous evidence of the existence of the relationship, and a multi-source heterogeneous evidence graph is constructed. Step S2: Use a dual-stream encoder to jointly represent entities in the multi-source heterogeneous evidence graph; design a text-attribute stream encoder and a time-series stream encoder to extract static semantic features and dynamic operation features of entities from the multi-source heterogeneous evidence graph, respectively, to obtain dual-stream features as independent feature representations of each data source. Furthermore, by fusing dual-stream features through a cross-modal attention mechanism, a joint entity representation containing multi-source evidence information is obtained; Step S3: Introduce a temporal causal mechanism by combining entity joint representation to discover implicit relationships; calculate the Granger causality index between dynamic features of entities and generate a causal prior mask matrix. Injecting a causal prior mask matrix into a graph attention network suppresses the flow of invalid information in non-causal directions, infers the probability of dynamic relationships between candidate entity pairs, and outputs the joint feature probability of candidate entity pairs. Step S4: Multi-source fusion and confidence calculation based on DS evidence theory; Based on the joint feature probability and the independent feature probability calculated by the independent feature representation of each data source, the multi-source probability is fused using the synthesis rules of DS evidence theory to output the final probability and confidence interval of the dynamic relationship. Low-confidence noise relationships are filtered out according to the set threshold to obtain high-confidence relationship triples. Step S5: Update the original knowledge graph using high-confidence dynamic relation triples; Relationships with confidence levels meeting preset thresholds are written into the graph to complete the power system graph relationships, forming a dynamic power system knowledge graph that supports time-series queries and causal backtracking.

2. The method for completing the power system graph relationship based on multi-source evidence and causal reasoning according to claim 1, characterized in that, In step S2, the text-attribute stream encoder uses a pre-trained language model to encode text semantic fragments in the multi-source heterogeneous evidence graph, and combines the embedded representations of discrete ledger attributes to extract the static semantic representation of entities through average pooling. The temporal stream encoder uses a temporal convolutional network to extract temporal waveform segments from multi-source heterogeneous evidence graphs, and inputs them into a temporal Transformer to capture long-distance state dependencies and extract the dynamic operational representation of entities. ; Subsequently, using static semantic representations as queries and dynamic runtime representations as keys and values, cross-modal attention scores are calculated, and joint entity representations are fused and output. In step S3, Granger causality scores between temporal features of candidate entity pairs are calculated, and a causal prior mask matrix is ​​constructed. M Construct a graph attention network and correlate its attention coefficients with a causal prior mask matrix. M Multiply, normalize, and then perform weighted fusion to obtain the updated joint entity representation. Finally, the joint feature probability of the candidate entity pair is calculated using the Sigmoid function. In step S4, the joint feature probabilities are transformed into mass functions; a graph convolutional network is constructed for the structural data source, a BERT model is constructed for the text data source, and an LSTM model is constructed for the time-series data source, serving as their respective prediction models; the independent feature representations of each data source are input into their respective prediction models to obtain independent feature probabilities, which are then transformed into independent mass functions; the Dempster synthesis rule is used to calculate the fused mass function; and the confidence function for the validity of the relationship is calculated based on the fused mass function. Bel With likelihood function Pl ,Will[ Bel , Pl As a confidence interval for a completion relation, it is only when Bel When the value exceeds the set threshold, the relationship completion is deemed valid.

3. A method for completing power system graph relationships based on multi-source evidence and causal reasoning according to claim 1 or 2, characterized in that, In step S1, for a certain candidate entity pair The acquired multi-source heterogeneous data serves as evidence reflecting the dynamic relationships between devices, specifically including: Step S1.1: Using the state change or limit violation event of the candidate entity pair in the SCADA telemetry and information database as the trigger point, extract the characteristic event window before and after the event occurs. Timing waveform segments within the time frame serve as timing evidence. ; Step S1.2: Perform a correlation query on the candidate entity pair in the PMS ledger database, extract the attributes of the two devices, align them to the same dimension, and use the discrete ledger attributes as structural evidence. ; Step S1.3: Scan and extract text semantic fragments that appear together in the same sentence or paragraph as text evidence from the inspection text database. ; Step S1.4: Take the entities in the candidate entity pair as device nodes and the three types of evidence as evidence nodes. Construct a local evidence subgraph attached to the candidate entity pair through the relationship between devices and the relationship between devices and evidence, thereby obtaining a multi-source heterogeneous evidence graph.

4. A method for completing power system graph relationships based on multi-source evidence and causal reasoning according to claim 1 or 2, characterized in that, Step S2 is as follows: Step S2.1: Static stream encoding; using the pre-trained language model BERT to encode textual evidence. Combined with structural evidence The embedding representation is then used to obtain the static semantic representation through average pooling. : in, This represents the encoding operation of the BERT model. This represents the vectorized embedding operation of ledger attributes. This represents a vector concatenation operation. This indicates that average pooling is performed on the vector sequence. Step S2.2, Dynamic Stream Coding; for temporal evidence A temporal convolutional network (TCN) is used to extract local waveform features, which are then input into a temporal transformer to capture long-range state dependencies, resulting in a dynamic runtime representation. : in, This represents the feature extraction operation of a temporal convolutional network. This represents the feature extraction operation of a temporal Transformer encoder. Step S2.3, Cross-modal fusion: Using static semantic representation as the query and dynamic runtime representation as the key and value, calculate the cross-modal attention score and fuse the output entity joint representation. Z : in, Q For querying the matrix, K The key matrix, V For value matrices, d k Let be the dimension of the key matrix. For normalized exponential functions, This indicates a vector concatenation operation.

5. A method for completing power system graph relationships based on multi-source evidence and causal reasoning according to claim 1 or 2, characterized in that, Step S3 is as follows: Step S3.1: Granger causality index calculation; entity-based dynamic operational representation. Calculate candidate entity pairs middle right Granger causality score ;like The historical time sequence characteristics significantly reduced the influence on The prediction error of the current state indicates the existence of causal explanatory power, as shown in the following formula: in, Representing entities At any moment t The true time-series feature values, Representing entities At any moment t Previous historical time series feature values, Representing entities At any moment t Previous historical chronological characteristics, It is an autoregressive prediction function; Step S3.2: Construct the causal prior mask matrix M If Granger causality index Below the set threshold Then the matrix M The corresponding element is 0 if it is 0, otherwise it is 1, that is: in, Representing entities right The mask value; Step S3.3, Graph Attention Aggregation with Causal Constraints; In the graph attention network, the attention coefficients are combined with the causal prior mask matrix. M Multiplication enables neighborhood aggregation under causal constraints: in, W To share the learnable linear transformation weight matrix, , , for , , Joint representation of entities This represents a vector concatenation operation. For the learnable weight vector in the attention mechanism, For activation function, For entities The set of neighboring device nodes, for right Attention coefficient It is a non-linear activation function. Updated after aggregation with causal constraints Joint representation of entities; Step S3.4: Calculate joint feature probability; based on the updated joint entity representation. calculate right The joint feature probability is: ,in, Representing entities The updated joint entity representation, Representing entities The updated entity joint representation.

6. A method for completing power system graph relationships based on multi-source evidence and causal reasoning according to claim 1 or 2, characterized in that, Step S4 is as follows: Step S4.1: Construct independent mass functions; for structured data sources, construct a graph convolutional network; for text data sources, construct a BERT model; for time-series data sources, construct an LSTM model. The output is denoted as... , indicating that the candidate entity is paired with In the data source k Independent feature probabilities under, k It is a structural, textual, or temporal data source; let the recognition framework be defined. ,in, This indicates that the relationship is established. If the relationship is not established, then for the data source... k Its independent feature probability p k The corresponding mass function is as follows: ， in, Indicates data source k The probability that the support relationship is true. Indicates data source k The probability that the support relationship is not valid; Step S4.2: Dempster synthesis rules fuse the mass function; for joint feature probabilities Then the mass function under the joint feature representation is The conflict coefficient is calculated by fusing the mass functions from three data sources and the mass function from the joint feature representation using the orthogonal summation rule. K : in, A , B , C , D The recognition frameworks represent data sources from structural, textual, temporal, and joint feature representations, respectively. a subset of The subset includes , 、 The resulting mass function is: in, This is a normalization factor to ensure that the sum of the mass functions after fusion is 1; E For identification framework A subset of; Step S4.3: Calculate the confidence interval; calculate the confidence function. With likelihood function : in, Let be the trust function, representing the probability that the relationship is explicitly supported. Let be the likelihood function, representing the probability that the relation does not oppose the validity of the relation. The confidence interval that constitutes the complement relation; if , If a threshold is set, the dynamic relationship completion is determined to be valid, and then... This serves as a reliability attribute for the relationship.

7. A method for completing power system graph relationships based on multi-source evidence and causal reasoning according to claim 1 or 2, characterized in that, Step S5 is as follows: Update the original knowledge graph using high-confidence dynamic relation triples; add the completed dynamic relation triples, i.e. (devices) Dynamic relationships, equipment The original knowledge graph stored in the graph database is updated by adding a valid timestamp and confidence interval, forming a dynamic power knowledge graph that supports time-series queries and causal backtracking.

8. A power system graph relationship completion system based on multi-source evidence and causal reasoning, used to implement the method described in any one of claims 1-7, characterized in that, include: Multi-source evidence alignment and construction module: Deployed on the data access server, with a built-in SCADA parser and ledger data interface, responsible for aligning multi-source heterogeneous data and constructing local evidence subgraphs with entity pairs as the core; Dual-stream joint representation module: used to construct the underlying feature extraction base of the system; it includes text-attribute stream coding unit and temporal stream coding unit, and outputs entity joint representation that fuses multi-source information through cross-modal attention layer; Temporal causal reasoning module: It has a built-in Granger causal calculation unit and a causal prior mask matrix generation unit. It injects the causal prior mask matrix during graph attention network aggregation and outputs dynamic relationship predictions that conform to physical logic. Evidence fusion and calibration module: Receives prediction results from various data source branches, performs probabilistic fusion based on the DS evidence theory synthesis rules, and outputs the confidence interval for the validity of the relationship. Bel , Pl [and filter low-confidence noise;] Dynamic graph update module: Deployed on the graph database server, it writes dynamic relation triples with timestamps and confidence intervals into the graph database, supporting querying and visualization of time series graphs.

Images