Graph-model complementary driven power grid fault intelligent auxiliary analysis method and system

Abstract

The invention relates to a graph-model complementary driven power grid fault intelligent auxiliary analysis method and system, and the method comprises the analysis steps: collecting fault event related multi-source business data in a power grid dispatching control system, carrying out the unstructured information modeling of the data, and constructing a power grid fault affair knowledge graph; performing monitoring system element-level fault diagnosis based on related multi-source service data of the fault event, and performing alarm correlation analysis in combination with a power grid fault affair knowledge graph to obtain interlocking fault event information and interlocking alarm event information of the fault event; based on interlocking fault event information and interlocking alarm event information of the fault event, a pre-training language model and a graph neural network model assist a fault affair knowledge graph to carry out knowledge reasoning, so that whole-process information of fault disposal is obtained, and a power grid fault disposal navigation graph is generated; and fusing the large-scale language model and the power grid fault affair knowledge graph, performing power grid fault disposal decision reasoning, and generating a power grid fault disposal scheme.

Description

Technical Field

[0001] The present invention relates to the field of intelligent auxiliary analysis of power grid faults, and in particular to an intelligent auxiliary analysis method and system for power grid faults driven by complementary graph and model. Background Art

[0002] The main processing process of the application of intelligent auxiliary analysis of power grid faults can be divided into three stages: intelligent alarm of power grid, risk analysis, and auxiliary disposal. Through the step-by-step connection of functions such as intelligent alarm of power grid, risk analysis of perception and judgment, accident plan matching, and auxiliary decision-making for risk elimination, auxiliary disposal of power grid faults is realized. Among them, the comprehensive intelligent alarm of power grid includes functions such as component-level fault alarm and panoramic information display, event-level alarm and its chain fault analysis. In the prior art, the event-level alarm analysis of power grid often relies on the direct relationship between time association and topological association, or is realized through the chain fault scenario rules configured by the dispatcher, which has the problems of poor interpretability and scalability and low accuracy. In addition, there is a lack of effective knowledge guidance means and effective information that dispatchers are concerned about in the auxiliary fault disposal stage, such as lack of panoramic fault information such as equipment defects, maintenance, and power generation plans. Therefore, there is an urgent need for a power grid fault auxiliary analysis solution that integrates comprehensive intelligent alarm of power grid and intelligent navigation of fault disposal to further improve the intelligent level of power grid fault handling. Summary of the invention

[0003] In view of the problems in the prior art of poor interpretability and low accuracy of intelligent power grid alarms, and lack of effective knowledge guidance in fault handling scenarios, the present invention provides a method and system for intelligent auxiliary analysis of power grid faults driven by graph-model complementarity. By constructing a knowledge graph of power grid faults, power grid fault-related events are analyzed and the judgment process and mechanism of comprehensive intelligent alarms are inferred. Based on the large control model, various types of high-dimensional and scattered information are summarized, and an intelligent auxiliary analysis system for power grid faults is established in combination with the knowledge graph to quickly generate fault analysis briefs and fault handling navigation maps, thereby realizing intelligent auxiliary analysis of power grid faults.

[0004] The present invention solves the above technical problems as follows: a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, comprising the following steps:

[0005] Collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a knowledge graph of power grid fault events;

[0006] Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, so as to obtain the interlocking fault event information and interlocking alarm event information of the fault event;

[0007] Based on the interlocking fault event information and interlocking alarm event information of fault events, the pre-trained language model and graph neural network model are used to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of fault handling, and generate a power grid fault handling navigation map;

[0008] Integrate large-scale language models with knowledge graphs of power grid faults to conduct decision reasoning on power grid fault handling and generate power grid fault handling plans.

[0009] Preferably, the collecting of multi-source business data related to fault events in the power grid dispatching and control system, performing unstructured information modeling on the multi-source business data related to the fault events, and constructing a knowledge graph of power grid fault events include:

[0010] Collect multi-source business data related to fault events in the power grid dispatching and control system, and pre-process the multi-source business data related to fault events. The steps are as follows:

[0011] The Z-score method is used to remove outliers for numerical data, as follows:

[0012] Assuming that the data follows a Gaussian distribution, the outliers are far away from the mean of all data and are at the tail of the distribution. All data points are normalized using the following formula:

[0013]

[0014] In the formula, xi is a data point; μ is the mean of all data; σ is the standard deviation of all data; zi is the normalized data point. A threshold zthr (usually set to 2.5, 3.0 or 3.5) is set for all normalized points, and the definition satisfies |z i |>z thr The data points are outliers and are removed.

[0015] Gaussian filter and histogram equalization method are used to perform image denoising and enhancement on image type data;

[0016] For text type data, text cleaning, text segmentation and stop word removal methods are used to finally remove irrelevant characters and unify the text format.

[0017] The preprocessed multi-source business data related to the fault event is labeled. For text type data, the Doccano tool is used to manually label the entities, relationships and attributes related to the alarm to obtain the labeled text data. For image type data, the LabelImg tool is used to manually label the alarm related objects in the image, and its label is used as the prefix information of the image, corresponding to the entity in the knowledge graph, which can be used for subsequent multimodal knowledge fusion and alignment.

[0018] The format conversion and knowledge extraction of the annotated multi-source business data are performed. Based on the mechanism and judgment rules of power grid fault alarm, the fault entities and relationships are organized in the form of "entity-relationship-entity" and stored in the Neo4j graph database to build a knowledge graph of power grid fault events. The steps are as follows:

[0019] The annotated power grid fault text data is processed by the BERT model, BiLSTM model, and RelEx relation classifier to extract fault knowledge;

[0020] An image description generation network is constructed based on the Vision Transformer model and the LLaMA model to extract fault knowledge from image type data.

[0021] The fault event statements and image description texts after fault knowledge extraction are respectively input into the UIE framework and the general information extraction framework, and the fault entities and relationships therein are extracted by the structured extraction language and the structural pattern director to complete the fault knowledge extraction of multi-source business data.

[0022] The similarity between entities is calculated by combining the edit distance similarity, overlap coefficient similarity and cosine similarity. If the similarity between entities is higher than the set threshold, the two refer to the same object category, and the entity name is modified to be unified to achieve fault knowledge fusion and alignment. Finally, the fault entities and relationships are organized into the key points of power grid fault alarm judgment in the form of "entity-relationship-entity" by using the mechanism and judgment rules (expert reasoning rules) of power grid fault alarm, and these key points in triple format are stored in the Neo4j graph database to construct a knowledge graph of power grid fault facts.

[0023] Preferably, the monitoring system component-level fault diagnosis is performed based on the relevant multi-source business data of the fault event, and the power grid event-level alarm correlation analysis is performed in combination with the power grid fault event knowledge graph to obtain the interlocking fault event information and interlocking alarm event information of the fault event, including:

[0024] Based on the Deep Walk node embedding algorithm, the entities in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text and image multimodal information in the multi-source business data related to the fault event are entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules;

[0025] Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database; the fault diagnosis mechanism and the conditions satisfied by the export results are visualized in the form of a graph, thereby assisting dispatchers to conduct intuitive and rapid autonomous analysis;

[0026] The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model. The simple, intuitive and highly scalable TransE model is used to convert the triples in the knowledge graph into head entity vectors, relationship vectors and tail entity vectors.

[0027] A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space; among them, the policy neural network is constructed using a fully connected network, and the state vector is mapped to the probability values ​​and rewards corresponding to all possible next actions by updating and optimizing the parameters of the policy function.

[0028] By training the single-step optimal strategy network and the cascade relationship reasoning strategy network of the alarm events in the fault event knowledge graph, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

[0029] Preferably, the single-step optimization strategy network and the cascade relationship reasoning strategy network in the training strategy neural network specifically include:

[0030] For each triple in the knowledge graph, add (e h ,r,e t ) adds a reverse relationship to form a new triple (e t ,r rev ,e h ), all triples form a set;

[0031] Take out the first two parts of each triple, merge the same entities or relations into bigrams, and construct a training sample set Btrain for subsequent training of the single-step optimization strategy network;

[0032] Each entity embedding of the training sample set Btrain is converted into a state vector and input into the policy neural network to select the best next action. The reward value is updated according to whether the selected action matches the binary in the training sample set Btrain until the gradient change of the policy network parameters tends to zero and the training is completed; the details are as follows:

[0033] Take each entity in the training sample set as the starting state est, embed this state into the state vector st of the agent, input it into the policy neural network, obtain the relations corresponding to all possible actions, and select a relation rst as the next action;

[0034] The starting state est and the relationship rst form a new tuple, and determine whether this tuple belongs to the training sample set Btrain. If it does, the reward value γ obtained by selecting this action is increased by 1; if it does not, the reward value does not change. The agent updates the policy network according to the reward value to maximize the reward. The present invention adopts a reward update method based on policy gradient, and the expression of this process is as follows:

[0035] ▽ ω logπ(a=r st ,s t ;ω)γ(e t ,r st )

[0036] Where: ω is the parameter of the policy network; π(a=r st ,s t ; ω) is the probability of selecting the relation rst under the state vector st; γ is the global reward value obtained by selecting this relation; e t is the target state.

[0037] Extract all triples containing a specific cascade relationship r' from the knowledge graph, construct a dataset for training the cascade relationship reasoning strategy network, and divide it into a training set and a test set;

[0038] Initialize the reasoning environment of the reinforcement learning agent and load the updated single-step optimization policy network parameters. Use the head entity and tail entity of the triple in the training set as the starting state est and target state et of the agent's reasoning. Starting from the starting state, the agent learns through the policy neural network to obtain other cascade relationships that can reach the target state in addition to the known r'. Finally, the agent continuously interacts with the knowledge graph reasoning environment, uses a variety of reward functions to train the cascade relationship reasoning policy network, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

[0039] Preferably, the interlocking fault event information and interlocking alarm event information based on the fault event are assisted by the pre-trained language model and the graph neural network model to perform knowledge reasoning on the fault event knowledge graph, obtain the full process information of the power grid fault handling, generate the power grid fault handling navigation map, and provide effective knowledge guidance means for the dispatching personnel, including:

[0040] The BERT pre-trained language model creates an initial semantic representation for the entities and relations in the "fault-relationship-treatment measures" triple (h, r, t) of the power grid fault event knowledge graph, and maps it into a low-dimensional embedding vector, which is represented as:

[0041] h0=Embbedding(h), t0=Embbeding(t), r0=Embbeding(r)

[0042] Where h, t, r are the head entity vector, tail entity vector and relationship vector respectively; h0, t0, r0 are the low-dimensional embedding vectors corresponding to the head entity, tail entity and relationship respectively.

[0043] The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures plan in the fault event knowledge graph, which are recorded as v1 and v2 respectively; the cosine similarity between the two vectors is calculated to measure the similarity between the fault event and each fault handling measure in the knowledge graph, and semantic search is performed based on this to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event. The principle of cosine similarity calculation is shown in the following formula.

[0044]

[0045] All fault handling nodes and relationships similar to the fault event are input into the GNN model, and higher weights are given to entities and relationships with high similarity;

[0046] The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event. The principle of node information aggregation is shown in the following formula.

[0047]

[0048] Wherein, v1 and v2 are the power grid fault event node and the fault handling measure node respectively; r is the relationship between the two nodes; (l-1) is the (l-1)th layer of the GNN model; l is the lth layer of the GNN model after information aggregation; N(v1) is the set of neighbor nodes of the power grid fault event node; f is the activation function, which adopts a simple feedforward neural network; m is the message transmitted from the fault handling measure node to the fault event node.

[0049] While aggregating neighbor node information, the embedded representation of each fault event node is gradually updated in combination with semantic search results. The definition of node update is shown in the following formula.

[0050]

[0051] Where W is the weight matrix; b is the bias value; σ is the ReLU activation function.

[0052] The GNN model performs multi-layer forward propagation, integrates the node's neighbor information, uses the breadth-first search algorithm to track high-attention weight paths, and infers the full process information of power grid fault handling; specifically, it integrates the deeper neighbor node information around each node. The output of the last layer of the model is the final representation of each fault node. The breadth-first search algorithm is used to track the high-attention weight transmission path "from the fault event node to all possible next-step fault handling measure nodes", and finally returns the fault handling measures for each step in turn, so that the full process information of power grid fault handling is inferred.

[0053] Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault disposal is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault disposal operation is converted into a power grid fault disposal navigation map.

[0054] Preferably, a large language model is integrated with a knowledge graph of power grid fault events to perform power grid fault handling decision reasoning and generate a power grid fault handling plan, including:

[0055] Based on the KG-GPT framework, the large-scale language model in the field of power grid control and the knowledge graph of power grid faults are integrated to form a KG-GPT model.

[0056] The entities, relationships and attribute information related to the fault events are extracted through the knowledge graph of power grid faults and input into the knowledge layer of the KG-GPT model, as follows:

[0057] Named entity recognition: The BiLSTM-CRF model is used to identify fault entities in the query sentence. First, the context representation of each entity word is generated through the BiLSTM layer, and the corresponding hidden state vector is output; then the transition probability between entity labels is modeled through the conditional random field model, and the Viterbi algorithm is used to find the global optimal label sequence.

[0058] Knowledge graph query: Use the Cypher query language to search for information related to the fault event entity in the query statement in the power grid fault event knowledge graph.

[0059] Power grid fault event information extraction: Extract the attribute and relationship information of the fault event entity from the search results in the previous step and return it to the knowledge layer of the KG-GPT model.

[0060] Through the knowledge layer, the entities, relationships and attribute information related to the fault event are connected to the fault alarm sentence in the alarm event information to form a sentence tree structure, as follows;

[0061] Entity information query: All entities contained in the fault alarm statement are defined as set A: {e1,...,e s}, where s is the total number of entities in the statement. Use Cypher query language to query the triples related to each entity in A in the knowledge graph module, including "fault-relationship-plan information", "plan information-relationship-risk information", "risk information-relationship-risk control strategy", "risk control strategy-relationship-key factors for power restoration decision-making", etc., which can be uniformly recorded as the following formula:

[0062] E=E query (A,K)

[0063] Where E is the set of all queried triples; K is the knowledge graph of power grid fault events.

[0064] Construct a sentence tree: Take each triple in E as a branch, connect them and insert them into the fault alarm sentence to construct a sentence tree structure. This process is shown as follows:

[0065] ST=C input (A,E)

[0066] In the formula, C input is the connection operation between the input entity information and the queried triples; ST is the constructed sentence tree.

[0067] The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model, as follows;

[0068] The sentence tree structure is converted into a vector sequence through the embedding layer as the input of the subsequent decoder module. This representation vector consists of a token embedding vector, a position encoding vector, and an entity information embedding vector. Among them, token embedding is implemented by a learnable embedding matrix. The dimension of the matrix depends on the size of the fault handling information vocabulary, and each token has a unique vector representation in the embedding matrix; the position encoding vector is used to add position information to each token in the input sentence as the basis for the input data structure conversion; the fault entity in the sentence tree is encoded as an entity embedding vector, and added to the entity position information to obtain the entity position encoding vector, which is calculated as follows:

[0069] p j =pos i +s i

[0070] In the formula, si is the embedding vector representation of the i-th entity; pos i is the location information corresponding to the i-th entity; p j is the position encoding vector corresponding to the i-th entity, 1≤j≤n.

[0071] Assume that the input sequence is {s1,s2,...,s n}, where the i-th position is the fault entity and the embedding vector is s i , according to the entity alignment relationship p j =f(s i ), the entity embedding vector is replaced by the corresponding entity position encoding vector, and the input vector sequence is converted to {s1,...,p j ,...,s n}, as input to the decoder layer of the model.

[0072] The Transformer decoder layer associates the vector sequence according to the process sequence of power grid fault handling, completes fault information processing, and generates fault analysis briefings and power grid fault handling plans, as follows:

[0073] Starting from the first decoder, the input vector sequence is processed sequentially through the self-attention mechanism and the feedforward neural network, and the result is used as the input of the next decoder.

[0074] Each decoder uses a multi-head self-attention mechanism that combines entity information. A matrix R that reflects the relationship between entities is added to the traditional self-attention mechanism, which is defined as follows:

[0075]

[0076] In the formula, s i ⊙s j Represents entity s i and j Located on the same branch of the sentence tree; Represents entity s i and j Located on different branches of the sentence tree.

[0077] Calculate the attention matrix of each attention head, as defined below:

[0078]

[0079] Where Q, K, and V are the query vector, key vector, and value vector of the fault alarm entity respectively; K is the dimension of the key vector; the attention weights are normalized using the softmax function.

[0080] The outputs of each attention head are concatenated in series to obtain the multi-head attention weight matrix, which is defined as follows:

[0081] a i =Attention(Q,K,V,R)

[0082] Mh_Attention(Q,K,V)=Concat(a1,a2,...,a i )W l

[0083] In the formula, a i is the output of the i-th attention head; W l is the weight matrix of linear transformation; Mh_Attention is the weight matrix of multi-head attention.

[0084] After updating the entity information embedding, each decoder fuses the entity token sequence and the non-entity token sequence and calculates the output of each entity token and non-entity token, as defined in the following formula. After N layers of decoder structure, the vector representation of entity tokens and non-entity tokens is continuously updated, and the output of the last layer is used as the output of the model decoder layer.

[0085]

[0086] In the formula, s p is the intermediate hidden vector that combines entity and non-entity tokens; σ is the nonlinear activation function; and are the conversion weight matrices corresponding to non-entity tokens and entity tokens respectively; and are the pth non-entity token of the ith decoder and the qth entity token of the ith decoder respectively; and are the bias vectors of non-entity tokens and entity tokens respectively; b i is the overall bias vector.

[0087] Make full use of the fault entity information and the output of the decoder layer to infer and generate relevant information on plan matching, risk analysis, risk treatment, and forced power transmission condition judgment process, and quickly obtain power grid fault analysis briefings and effective fault treatment plans.

[0088] The present invention also provides a power grid fault intelligent auxiliary analysis system driven by complementary graphics and models, comprising:

[0089] A power grid fault event knowledge graph construction module is used to collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a power grid fault event knowledge graph;

[0090] The power grid integrated intelligent alarm module is used to perform component-level fault diagnosis of the monitoring system based on the relevant multi-source business data of the fault event, and to perform power grid event-level alarm correlation analysis in combination with the knowledge graph of power grid fault events to obtain the interlocking fault event information and interlocking alarm event information of the fault event;

[0091] The power grid fault handling navigation map generation module is used to generate interlocking fault event information and interlocking alarm event information based on fault events. It uses the pre-trained language model and graph neural network model to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of fault handling, and generate a power grid fault handling navigation map.

[0092] The power grid fault handling plan output module is used to integrate the large language model with the power grid fault event knowledge graph, perform power grid fault handling decision reasoning, and generate a power grid fault handling plan.

[0093] Preferably, the collecting of multi-source business data related to fault events in the power grid dispatching and control system, performing unstructured information modeling on the multi-source business data related to the fault events, and constructing a knowledge graph of power grid fault events include:

[0094] Collect multi-source business data related to fault events in the power grid dispatching and control system, and pre-process the multi-source business data related to fault events.

[0095] Label the pre-processed multi-source business data related to the fault event;

[0096] The format conversion and knowledge extraction are performed on the labeled multi-source business data. Based on the mechanism and judgment rules of power grid fault alarm, the fault entities and relationships are organized in the form of "entity-relationship-entity" and stored in the Neo4j graph database to construct a knowledge graph of power grid fault events.

[0097] Preferably, based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is performed, and the power grid event-level alarm correlation analysis is performed in combination with the power grid fault event knowledge graph to obtain the interlocking fault event information and interlocking alarm event information of the fault event, including:

[0098] Based on the Deep Walk node embedding algorithm, the entities in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text and image multimodal information in the multi-source business data related to the fault event are entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules;

[0099] Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database;

[0100] The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model;

[0101] A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space;

[0102] By training the single-step optimal strategy network and the cascade relationship reasoning strategy network of the alarm events in the fault event knowledge graph, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

[0103] Preferably, the single-step optimization strategy network and the cascade relationship reasoning strategy network in the training strategy neural network specifically include:

[0104] Add a reverse relationship to each triple in the knowledge graph, extract the head entity and relationship from the triple to form a bigram, and construct a sample set Btrain for training the single-step optimization strategy network;

[0105] Each entity embedding of the training sample set Btrain is converted into a state vector and input into the policy neural network to select the best next action. The reward value is updated according to whether the selected action matches the binary in the training sample set Btrain, until the gradient change of the policy network parameters tends to zero, and the training is completed;

[0106] Extract all triples containing specific cascade relations from the knowledge graph, construct a dataset for training the cascade relation reasoning strategy network, and divide it into training set and test set;

[0107] Initialize the agent's reasoning environment and load the updated single-step optimal strategy network parameters. Use the head entity and tail entity of the triplet in the training set as the starting state and target state of the agent's reasoning, respectively. The agent learns through the policy neural network, expands the cascade relationship, trains the cascade relationship reasoning strategy network through the reward function, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

[0108] Preferably, the interlocking fault event information and interlocking alarm event information based on the fault event are assisted by a pre-trained language model and a graph neural network model to perform knowledge reasoning on the fault event knowledge graph, obtain the full process information of the power grid fault handling, and generate a power grid fault handling navigation map, including:

[0109] The BERT pre-trained language model creates an initial semantic representation for the entities and relations in the "fault-relation-treatment measure" triple (h, r, t) of the power grid fault event knowledge graph, and maps it into a low-dimensional embedding vector;

[0110] The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures plan in the fault event knowledge graph; the cosine similarity between the text feature encodings is calculated for semantic search to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event;

[0111] All fault handling nodes and relationships similar to the fault event are input into the GNN model, and higher weights are given to entities and relationships with high similarity;

[0112] The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event.

[0113] The GNN model performs multi-layer forward propagation, integrates node neighbor information, uses a breadth-first search algorithm to track high attention weight paths, and infers the full process information of power grid fault handling;

[0114] Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault disposal is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault disposal operation is converted into a power grid fault disposal navigation map.

[0115] Preferably, a large language model is integrated with a knowledge graph of power grid fault events to perform power grid fault handling decision reasoning and generate a power grid fault handling plan, including:

[0116] Based on the KG-GPT framework, the KG-GPT model is formed by integrating the large language model and the knowledge graph of power grid faults.

[0117] Through the knowledge graph of power grid fault events, the entities, relations and attribute information related to the fault events are extracted and input into the knowledge layer of the KG-GPT model;

[0118] Through the knowledge layer, the entities, relations and attribute information related to the fault event are connected to the fault alarm sentences in the alarm event information to form a sentence tree structure;

[0119] The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model;

[0120] The Transformer decoder layer associates the vector sequence according to the process sequence of power grid fault handling, completes fault information processing, and generates fault analysis briefings and power grid fault handling plans.

[0121] The present invention also provides a computer storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models as described above is implemented.

[0122] The present invention also provides an electronic device, comprising a memory and a processor: the memory is used to store computer executable instructions, the processor is used to execute the computer executable instructions, and when the computer executable instructions are executed by the processor, the intelligent auxiliary analysis method for power grid faults driven by complementary graphics and models as described above is implemented.

[0123] The beneficial effects of the present invention are as follows: the present invention provides an intelligent auxiliary analysis method and system for power grid faults that integrates a large-scale control model and a knowledge graph. Through the knowledge graph of power grid faults, the judgment process and mechanism of comprehensive intelligent alarms are inferred, the fault diagnosis mechanism is visualized, and the knowledge graph is used to represent the reasoning rules, and the correlation and causal relationship between events are analyzed, which solves the problem that dispatchers cannot independently and quickly analyze when power grid alarms occur, and the alarm analysis has poor explanatory power and low accuracy. An interactive decision-making reasoning engine that integrates large-scale language model scene intention recognition, knowledge graph reasoning, and computational analysis services is constructed to perceive power grid fault events, trigger the entity nodes of the knowledge graph of power grid faults to form a fault handling navigation map, and based on the large model, the whole process of intelligent execution of steps such as plan matching, risk analysis, risk handling, and judgment of forced power transmission conditions is driven, which solves the problem of lack of effective knowledge guidance means in fault handling scenarios, and thus provides an effective handling solution for power grid fault handling, improving handling efficiency and accuracy.

[0124] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. The specific implementation of the present invention is given in detail by the following embodiments and their accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0125] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0126] Figure 1 A flow chart of a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models provided in Example 1;

[0127] Figure 2 A diagram showing the line fault diagnosis mechanism and the conditions satisfied by the export results in Example 1 in the form of a graph;

[0128] Figure 3 A module diagram of an intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models provided in Example 2. DETAILED DESCRIPTION

[0129] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0130] like Figure 1 As shown, the present invention provides a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, comprising the following steps:

[0131] 1. Collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a knowledge graph of power grid fault events. The specific steps are as follows:

[0132] 1.1 In this embodiment, taking AC fault as an example, multi-source business data is collected, including steady-state monitoring, dynamic monitoring (WMAS), protection and recording data, real-time indicators, operation tickets, AGC, AVC, equipment defects, maintenance, power generation plans, etc. These data are multimodal data, among which steady-state monitoring, dynamic monitoring (WMAS), real-time indicators, AGC, and AVC are numerical type data; protection and recording data are image type data; operation tickets, equipment defects, maintenance, and power generation plans are text type data;

[0133] 1.2 Preprocess the collected multi-source business data according to the data type:

[0134] Numerical data: Some data in steady-state monitoring, dynamic monitoring (WMAS), real-time indicators, AGC, and AVC are often saved in table form, and outliers need to be removed. The Z-score method is used to remove outliers. The principle is as follows:

[0135] Assuming that the data follows a Gaussian distribution, the outliers are far away from the mean of all data and are at the tail of the distribution. All data points are normalized using the following formula:

[0136]

[0137] In the formula, x i is a data point; μ is the mean of all data; σ is the standard deviation of all data; z i is the normalized data point. Set a threshold z for all normalized points thr(usually set to 2.5, 3.0 or 3.5), defined to satisfy |z i |>z thr The data points are outliers and are removed.

[0138] Image type data: Gaussian filter and histogram equalization method are used for image denoising and enhancement.

[0139] Text type data: Use text cleaning, text segmentation and stop word removal methods to finally remove irrelevant characters and unify the text format. The specific methods are: standardize the full names and abbreviations of power grid professional terms, variable units, etc., unify the text case; use the N-gram model for Chinese word segmentation; build a stop word list in the field of power grid regulation and remove stop words.

[0140] 1.3 Label the pre-processed multi-source business data related to fault events; for text type data, use the Doccano tool to manually label the entities, relationships and attributes related to the alarm to obtain labeled text data; for image type data, use the LabelImg tool to manually label the alarm related objects in the image, and use its label as the prefix information of the image, corresponding to the entity in the knowledge graph, which can be used for subsequent multimodal knowledge fusion and alignment.

[0141] The entity tags in the annotation system used in this embodiment include: plant, unit, line, substation, voltage level, fault event, handling operation, power grid operation index, local power grid name and location, equipment defect, maintenance action, etc.;

[0142] The relationship labels between entities include: contains, does not contain, belongs to, subject-predicate relationship, conditional relationship, etc.

[0143] The label of image type data is determined by the information contained in the business scenario and alarm interface, such as: monitoring time, equipment name, telesignaling status, telemetry value, fault event, fault phase, protection action status, reclosing action status, etc.

[0144] 1.4 Perform format conversion and knowledge extraction on the annotated multi-source business data. Based on the mechanism and judgment rules of power grid fault alarm, organize the fault entities and relationships into the form of "entity-relationship-entity" and store them in the Neo4j graph database to build a knowledge graph of power grid fault events. The steps are as follows:

[0145] The preprocessed and annotated power grid fault text data is input into the BERT model, and pre-trained on the input text through a self-supervised learning method to capture the contextual semantic information of entity words and generate word vector representations;

[0146] The context of each generated token is embedded into the BiLSTM model to continue capturing the forward and backward dependencies between entity words. The BiLSTM model further encodes and processes the token sequence through a bidirectional LSTM structure, integrating forward and backward information to obtain the hidden state of the sequence, which can be represented as a three-dimensional tensor. Finally, the three-dimensional tensor finally output by the BiLSTM model is transmitted to the conditional random field (CRF) model, and the Viterbi algorithm selects the fault entity label sequence with the largest conditional probability as the model's prediction output to obtain the fault entity recognition result. Based on the identified entities, the RelEx relationship classifier is used or the method of expanding the label type in the text annotation process is adopted to extract the fault relationship.

[0147] Based on the Vision Transformer model and the LLaMA model, an image description generation network is constructed to extract fault knowledge from image type data. The specific process is as follows:

[0148] First, the visual prefix of the image data is extracted through the encoder structure of the Vision Transformer model and converted into a semantically embedded title. Secondly, the semantically embedded title is input into the LLaMA model as a precondition to predict the title of the next frame. Finally, the semantically embedded title is trained by the LLaMA model to generate a complete image description text, thereby converting the image label data into text label data.

[0149] After the fault knowledge extraction, the fault event statements and image description texts are respectively input into the universal information extraction (UIE) framework, and the fault entities and relations therein are extracted by the structured extraction language and the structural pattern director to complete the fault knowledge extraction of the multi-source data of the power grid. The similarity between the fault entities is further calculated based on the knowledge extraction results, and it is determined whether different entities point to the same object according to the set threshold. The principle of the similarity calculation method between entities is shown in the following formula.

[0150]

[0151]

[0152] In the formula, S ed ,S oc ,S cosare edit distance similarity, overlap coefficient similarity and cosine similarity respectively; S is entity similarity; p, q are two strings in the entity; ED is edit distance; length(p), length(q) are the lengths of strings p, q respectively; I is the number of identical characters in the two strings; set(p+q) is the set formed by removing the identical characters from the two strings; V p ,V q is the representation of two strings in the vector space; ||V p ||,||V q || represents the length of two vectors.

[0153] The similarity between entities is calculated by combining the edit distance similarity, overlap coefficient similarity and cosine similarity. If the similarity between entities is higher than the set threshold, the two refer to the same object category, and the entity name is modified to be unified to achieve fault knowledge fusion and alignment. Finally, the fault entities and relationships are organized into the key points of power grid fault alarm judgment in the form of "entity-relationship-entity" by using the mechanism and judgment rules (expert reasoning rules) of power grid fault alarm, and these key points in triple format are stored in the Neo4j graph database to construct a knowledge graph of power grid fault facts.

[0154] 2. Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, and the interlocking fault event information and interlocking alarm event information of the fault event are obtained. The specific steps are as follows:

[0155] 2.1 This embodiment takes a line fault as an example to perform component-level fault diagnosis of the monitoring system, as follows:

[0156] Based on the Deep Walk node embedding algorithm, the entities corresponding to different modal data such as steady-state monitoring, dynamic monitoring (WMAS), protection and recording data, real-time indicators, equipment defects, maintenance, power generation plan, etc. in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text information in the multi-source business data related to the fault event is entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules;

[0157] Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database; the fault diagnosis mechanism and the conditions satisfied by the export results are visualized in the form of a graph, thereby assisting dispatchers to conduct intuitive and rapid autonomous analysis, such as Figure 2 As shown;

[0158] 2.2 Based on the idea of ​​fault diagnosis at the component level of the monitoring system, combined with the complex cascade relationship between chain failure scenarios and alarm events, on the basis of the knowledge graph of power grid faults, a knowledge reasoning method integrating distributed knowledge representation and deep reinforcement learning is adopted to conduct power grid event-level alarm correlation analysis and sequentially reason out chain failure events, as follows:

[0159] The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model. The simple, intuitive and highly scalable TransE model is used to convert the triples in the knowledge graph into head entity vectors, relationship vectors and tail entity vectors.

[0160] A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space; among them, the policy neural network is constructed using a fully connected network, and the state vector is mapped to the probability values ​​and rewards corresponding to all possible next actions by updating and optimizing the parameters of the policy function.

[0161] By training the single-step optimal strategy network and the cascade relationship reasoning strategy network in the alarm events in the fault event knowledge graph, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

[0162] Among them, the single-step optimization strategy network and the cascade relationship reasoning strategy network in the training strategy neural network specifically include:

[0163] For each triple in the knowledge graph, add (e h ,r,e t ) adds a reverse relationship to form a new triple (e t ,r rev ,e h ), all triples form a set;

[0164] Take the first two parts of each triple, merge the same entities or relations into bigrams, and construct the training sample set B train , used for subsequent training of single-step optimal strategy network;

[0165] The training sample set B train Each entity embedding of is converted into a state vector and input into the policy neural network to select the best next action. train The reward value is updated by checking whether the binary pairs in match until the gradient change of the policy network parameters approaches zero and the training is completed; the details are as follows:

[0166] Each entity in the training sample set is taken as the starting state e st, the entity embedding of this state is converted into the state vector s of the agent t , input into the policy neural network, obtain the relations corresponding to all possible actions, and select a relation r st As the next step;

[0167] The starting state e st With the relationship st Construct a new bigram and determine whether this bigram belongs to the training sample set B train If it is, the reward value γ obtained by selecting this action is increased by 1; if it is not, the reward value does not change. The agent updates the policy network according to the reward value to maximize the reward. This embodiment adopts a reward update method based on policy gradient. The expression of this process is as follows:

[0168] ▽ ω logπ(a=r st ,s t ;ω)γ(e t ,r st )

[0169] Where: ω is the parameter of the policy network; π(a=r st ,s t ; ω) is the state vector s t Next select relation r st The probability of choosing this relationship; γ is the global reward value obtained by choosing this relationship; e t is the target state.

[0170] Extract all triples containing a specific cascade relationship r' from the knowledge graph, construct a dataset for training the cascade relationship reasoning strategy network, and divide it into a training set and a test set;

[0171] Initialize the reasoning environment of the reinforcement learning agent and load the updated single-step optimization strategy network parameters. Use the head entity and tail entity of the triple in the training set as the starting state of the agent reasoning. st and the target state e t The agent starts from the initial state and learns through the policy neural network to obtain other cascade relationships that can reach the target state in addition to the known r'. Finally, the agent continuously interacts with the knowledge graph reasoning environment, uses a variety of reward functions to train the cascade relationship reasoning policy network, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

[0172] 3. Based on the interlocking fault event information and interlocking alarm event information of the fault event, the pre-trained language model and graph neural network model are used to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of the fault handling, and generate the power grid fault handling navigation map. The specific steps are as follows:

[0173] 3.1 The BERT pre-trained language model creates an initial semantic representation for the entities and relations in the "fault-relationship-treatment measures" triple (h, r, t) of the power grid fault event knowledge graph, and maps it into a low-dimensional embedding vector, which is represented as:

[0174] h0=Embbedding(h), t0=Embbeding(t), r0=Embbeding(r)

[0175] Where h, t, r are the head entity vector, tail entity vector and relationship vector respectively; h0, t0, r0 are the low-dimensional embedding vectors corresponding to the head entity, tail entity and relationship respectively.

[0176] 3.2 The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures plan in the fault event knowledge graph, which are recorded as v1 and v2 respectively; the cosine similarity between the two vectors is calculated to measure the similarity between the fault event and each fault handling measure in the knowledge graph, and semantic search is performed based on this to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event. The principle of cosine similarity calculation is shown in the following formula.

[0177]

[0178] 3.3 Input the fault handling nodes and relationships with similar fault events into the GNN model, and assign higher weights to entities and relationships with high similarity;

[0179] 3.4 The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event; the principle of node information aggregation is shown in the following formula.

[0180]

[0181] Wherein, v1 and v2 are the power grid fault event node and the fault handling measure node respectively; r is the relationship between the two nodes; (l-1) is the (l-1)th layer of the GNN model; l is the lth layer of the GNN model after information aggregation; N(v1) is the set of neighbor nodes of the power grid fault event node; f is the activation function, which adopts a simple feedforward neural network; m is the message transmitted from the fault handling measure node to the fault event node.

[0182] While aggregating neighbor node information, the embedded representation of each fault event node is gradually updated in combination with semantic search results. The definition of node update is shown in the following formula.

[0183]

[0184] Where W is the weight matrix; b is the bias value; σ is the ReLU activation function.

[0185] The 3.5GNN model performs multi-layer forward propagation, integrates the node's neighbor information, uses the breadth-first search algorithm to track high-attention weight paths, and infers the full process information of power grid fault handling; specifically, it integrates the deeper neighbor node information around each node. The output of the last layer of the model is the final representation of each fault node. The breadth-first search algorithm is used to track the high-attention weight transmission path "from the fault event node to all possible next-step fault handling measure nodes", and finally returns the fault handling measures for each step in turn, so that the full process information of power grid fault handling is inferred.

[0186] 3.6 Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault handling is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault handling operation is converted into a power grid fault handling navigation map.

[0187] 4. Integrate the large language model with the knowledge graph of power grid faults to make decision-making inferences on power grid fault handling, generate fault analysis briefs and power grid fault handling plans. The specific steps are as follows:

[0188] 4.1 Based on the KG-GPT framework, the large-scale language model in the field of power grid regulation and the knowledge graph of power grid faults are integrated to form the KG-GPT model

[0189] 4.2 Extract the entities, relationships and attribute information related to the fault event through the knowledge graph of power grid fault events and input them into the knowledge layer of the KG-GPT model, as follows:

[0190] Named entity recognition: The BiLSTM-CRF model is used to identify fault entities in the query sentence. First, the context representation of each entity word is generated through the BiLSTM layer, and the corresponding hidden state vector is output; then the transition probability between entity labels is modeled through the conditional random field model, and the Viterbi algorithm is used to find the global optimal label sequence.

[0191] Knowledge graph query: Use the Cypher query language to search for information related to the fault event entity in the query statement in the power grid fault event knowledge graph.

[0192] Power grid fault event information extraction: Extract the attributes and relationship information of the fault event from the search results in the second step and return it to the knowledge layer of the KG-GPT model.

[0193] 4.3 Through the knowledge layer, the entities, relationships and attribute information related to the fault event are connected to the fault alarm sentences in the alarm event information to form a sentence tree structure, as follows;

[0194] Entity information query: All entities contained in the fault alarm statement are defined as set A: {e1,...,e s}, where s is the total number of entities in the statement. Use Cypher query language to query the triples related to each entity in A in the knowledge graph module, including "fault-relationship-plan information", "plan information-relationship-risk information", "risk information-relationship-risk control strategy", "risk control strategy-relationship-key factors for power restoration decision-making", etc., which can be uniformly recorded as the following formula:

[0195] E=E query (A,K)

[0196] Where E is the set of all queried triples; K is the knowledge graph of power grid fault events.

[0197] Construct a sentence tree: Take each triple in E as a branch, connect them and insert them into the fault alarm sentence to construct a sentence tree structure. This process is shown as follows:

[0198] ST=C input (A,E)

[0199] In the formula, C input is the connection operation between the input entity information and the queried triples; ST is the constructed sentence tree.

[0200] 4.4 The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model, as follows;

[0201] The sentence tree structure is converted into a vector sequence through the embedding layer as the input of the subsequent decoder module. This representation vector consists of a token embedding vector, a position encoding vector, and an entity information embedding vector. Among them, token embedding is implemented by a learnable embedding matrix. The dimension of the matrix depends on the size of the fault handling information vocabulary, and each token has a unique vector representation in the embedding matrix; the position encoding vector is used to add position information to each token in the input sentence as the basis for the input data structure conversion; the fault entity in the sentence tree is encoded as an entity embedding vector, and added to the entity position information to obtain the entity position encoding vector, which is calculated as follows:

[0202] p j =pos i +s i

[0203] In the formula, s i is the embedding vector representation of the i-th entity; pos i is the location information corresponding to the i-th entity; p j is the position encoding vector corresponding to the i-th entity, 1≤j≤n.

[0204] Assume that the input sequence is {s1,s2,...,s n}, where the i-th position is the fault entity and the embedding vector is s i , according to the entity alignment relationship p j =f(s i ), the entity embedding vector is replaced by the corresponding entity position encoding vector, and the input vector sequence is converted to {s1,...,p j ,...,s n}, as input to the decoder layer of the model.

[0205] 4.5 Transformer decoder layer associates the relevant vector sequence of fault handling plan according to the vector sequence corresponding to the power grid fault information, and obtains the fault equipment ledger, three spans, switch action, protection action, plant and station weather, maintenance plan, equipment defects, accident plan, PMU and other information. According to the above method, sequential association is performed to obtain risk information (including power grid flow over-limit, voltage over-limit, weak link, power-off load, power-off user, etc.), risk control strategy information (including unit adjustment, mode adjustment, zone load transfer, power restriction, etc.), recommended power restoration factors (including N-1 over-limit, weak operation mode, heavy load around the fault, over-limit around the fault, over-limit around the fault in the future, over-limit in the section in the future, etc.) and non-power restoration factors (including cable lines, three-phase short circuit, station fault, cross-over line, equipment failure, incomplete signal, maintenance plan, cross-over railway, family defect, severe meteorological disaster, live operation plan, cross-over highway, etc.), so as to assist in the analysis of key factors of power restoration decision. Finally, the conclusion of power restoration is given based on the on-site station and line inspection conditions (for example: the primary / secondary equipment in the station is in good condition, power supply to the faulty tripped line will not pose a safety threat to the station personnel, line inspection and confirmation will not affect normal operation, no line tower collapse, no wire damage, no personnel climbing, etc.). The above method completes the information processing of the entire fault process, generates a fault analysis briefing and a power grid fault handling plan, and presents a recommendation to the dispatcher. The specific principles are as follows:

[0206] Starting from the first decoder, the input vector sequence is processed sequentially through the self-attention mechanism and the feedforward neural network, and the result is used as the input of the next decoder.

[0207] Each decoder uses a multi-head self-attention mechanism that combines entity information. A matrix R that reflects the relationship between entities is added to the traditional self-attention mechanism, which is defined as follows:

[0208]

[0209] In the formula, s i ⊙s j Represents entity s i and j Located on the same branch of the sentence tree; Represents entity s i and j Located on different branches of the sentence tree.

[0210] Calculate the attention matrix of each attention head, as defined below:

[0211]

[0212] Where Q, K, and V are the query vector, key vector, and value vector of the fault alarm entity respectively; K is the dimension of the key vector; the attention weights are normalized using the softmax function.

[0213] The outputs of each attention head are concatenated in series to obtain the multi-head attention weight matrix, which is defined as follows:

[0214] a i =Attention(Q,K,V,R)

[0215] Mh_Attention(Q,K,V)=Concat(a1,a2,...,a i )W l

[0216] In the formula, a i is the output of the i-th attention head; W l is the weight matrix of linear transformation; Mh_Attention is the weight matrix of multi-head attention.

[0217] After updating the entity information embedding, each decoder fuses the entity token sequence and the non-entity token sequence and calculates the output of each entity token and non-entity token, as defined in the following formula. After N layers of decoder structure, the vector representation of entity tokens and non-entity tokens is continuously updated, and the output of the last layer is used as the output of the model decoder layer.

[0218]

[0219] In the formula, s pis the intermediate hidden vector that combines entity and non-entity tokens; σ is the nonlinear activation function; and are the conversion weight matrices corresponding to non-entity tokens and entity tokens respectively; and are the pth non-entity token of the ith decoder and the qth entity token of the ith decoder respectively; and are the bias vectors of non-entity tokens and entity tokens respectively; b i is the overall bias vector.

[0220] Make full use of the fault entity information and the output of the decoder layer to infer and generate relevant information on plan matching, risk analysis, risk treatment, and forced power transmission condition judgment process, and quickly obtain power grid fault analysis briefings and effective fault treatment plans.

[0221] Example 2

[0222] like Figure 3 As shown, this embodiment provides a power grid fault intelligent auxiliary analysis system driven by complementary graphics and models, including:

[0223] A power grid fault event knowledge graph construction module is used to collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a power grid fault event knowledge graph;

[0224] The power grid integrated intelligent alarm module is used to perform component-level fault diagnosis of the monitoring system based on the relevant multi-source business data of the fault event, and to perform power grid event-level alarm correlation analysis in combination with the knowledge graph of power grid fault events to obtain the interlocking fault event information and interlocking alarm event information of the fault event;

[0225] The power grid fault handling navigation map generation module is used to generate interlocking fault event information and interlocking alarm event information based on fault events. It uses the pre-trained language model and graph neural network model to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of fault handling, and generate a power grid fault handling navigation map.

[0226] The power grid fault handling plan output module is used to integrate the large language model with the power grid fault event knowledge graph, perform power grid fault handling decision reasoning, and generate fault analysis briefings and power grid fault handling plans.

[0227] The method collects multi-source business data related to fault events in the power grid dispatching and control system, performs unstructured information modeling on the multi-source business data related to the fault events, and constructs a knowledge graph of power grid fault events, including:

[0228] Collect multi-source business data related to fault events in the power grid dispatching and control system, and pre-process the multi-source business data related to fault events.

[0229] Label the pre-processed multi-source business data related to the fault event;

[0230] The format conversion and knowledge extraction are performed on the labeled multi-source business data. Based on the mechanism and judgment rules of power grid fault alarm, the fault entities and relationships are organized into the "entity-relationship-entity" format and stored in the Neo4j graph database to build a knowledge graph of power grid fault events.

[0231] Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, and the interlocking fault event information and interlocking alarm event information of the fault event are obtained, including:

[0232] Based on the Deep Walk node embedding algorithm, the entities in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text and image multimodal information in the multi-source business data related to the fault event are entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules;

[0233] Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database;

[0234] The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model;

[0235] A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space;

[0236] By training the single-step optimal strategy network and the cascade relationship reasoning strategy network of the alarm events in the fault event knowledge graph, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

[0237] The single-step optimization strategy network and the cascade relationship reasoning strategy network in the training strategy neural network specifically include:

[0238] Add a reverse relationship to each triple in the knowledge graph, extract the head entity and relationship from the triple to form a bigram, and construct a sample set B for training the single-step optimization strategy network train ;

[0239] The training sample set B train Each entity embedding of is converted into a state vector and input into the policy neural network to select the best next action. train Whether the binary pairs in match, the reward value is updated until the gradient change of the policy network parameters tends to zero, and the training is completed;

[0240] Extract all triples containing specific cascade relations from the knowledge graph, construct a dataset for training the cascade relation reasoning strategy network, and divide it into training set and test set;

[0241] Initialize the agent's reasoning environment and load the updated single-step optimal strategy network parameters. Use the head entity and tail entity of the triplet in the training set as the starting state and target state of the agent's reasoning, respectively. The agent learns through the policy neural network, expands the cascade relationship, trains the cascade relationship reasoning strategy network through the reward function, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

[0242] The interlocking fault event information and interlocking alarm event information based on the fault event are used to perform knowledge reasoning on the fault event knowledge graph assisted by a pre-trained language model and a graph neural network model, so as to obtain the full process information of the power grid fault handling and generate a power grid fault handling navigation map, including:

[0243] The BERT pre-trained language model creates an initial semantic representation for the entities and relations in the "fault-relation-treatment measure" triple (h, r, t) of the power grid fault event knowledge graph, and maps it into a low-dimensional embedding vector;

[0244] The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures in the fault event knowledge graph; the cosine similarity between the text feature encodings is calculated for semantic search to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event;

[0245] All fault handling nodes and relationships similar to the fault event are input into the GNN model, and higher weights are given to entities and relationships with high similarity;

[0246] The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event.

[0247] The GNN model performs multi-layer forward propagation, integrates node neighbor information, uses a breadth-first search algorithm to track high attention weight paths, and infers the full process information of power grid fault handling;

[0248] Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault disposal is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault disposal operation is converted into a power grid fault disposal navigation map.

[0249] Integrate large-scale language models with knowledge graphs of power grid faults to conduct power grid fault handling decision reasoning, generate fault analysis briefs and power grid fault handling plans, including:

[0250] Based on the KG-GPT framework, the large-scale language model in the field of power grid control and the knowledge graph of power grid faults are integrated to form a KG-GPT model.

[0251] Through the knowledge graph of power grid fault events, the entities, relations and attribute information related to the fault events are extracted and input into the knowledge layer of the KG-GPT model;

[0252] Through the knowledge layer, the entities, relations and attribute information related to the fault event are connected to the fault alarm sentences in the alarm event information to form a sentence tree structure;

[0253] The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model;

[0254] The Transformer decoder layer associates the vector sequence according to the process sequence of power grid fault handling, completes fault information processing, and generates fault analysis briefings and power grid fault handling plans.

[0255] Example 3

[0256] This embodiment provides a computer storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models as described in Embodiment 1 is implemented.

[0257] Example 4

[0258] This embodiment provides an electronic device, including a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the intelligent auxiliary analysis method for power grid faults driven by complementary graphics and models as described in Example 1 is implemented.

[0259] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0260] The present disclosure is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0261] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0262] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0263] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any ordinary technician in the industry can smoothly implement the present invention as shown in the drawings and described above. However, any equivalent changes, modifications and evolutions made by technicians familiar with the profession without departing from the scope of the technical solution of the present invention using the technical content disclosed above are all equivalent embodiments of the present invention. At the same time, any equivalent changes, modifications and evolutions made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, characterized in that: The following steps are involved: Collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a knowledge graph of power grid fault events; Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, so as to obtain the interlocking fault event information and interlocking alarm event information of the fault event; Based on the interlocking fault event information and interlocking alarm event information of fault events, the pre-trained language model and graph neural network model are used to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of fault handling, and generate a power grid fault handling navigation map; Integrate large-scale language models with knowledge graphs of power grid faults to conduct decision reasoning on power grid fault handling and generate power grid fault handling plans.

2. According to claim 1, a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, characterized in that: The method collects multi-source business data related to fault events in the power grid dispatching and control system, models the unstructured information of the multi-source business data related to the fault events, and constructs a knowledge graph of power grid fault events, including: Collect multi-source business data related to fault events in the power grid dispatching and control system, and pre-process the multi-source business data related to fault events. Label the pre-processed multi-source business data related to the fault event; The format conversion and knowledge extraction are performed on the labeled multi-source business data. Based on the mechanism and judgment rules of power grid fault alarm, the fault entities and relationships are organized in the form of entity, relationship, and entity triples and stored in the Neo4j graph database to construct a power grid fault knowledge graph.

3. According to claim 1, a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, characterized in that: Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, and the interlocking fault event information and interlocking alarm event information of the fault event are obtained, including: Based on the Deep Walk node embedding algorithm, the entities in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text and image multimodal information in the multi-source business data related to the fault event are entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules; Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database; The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model; A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space; By training the single-step optimal strategy network and the cascade relationship reasoning strategy network of alarm events in the knowledge graph of power grid fault events, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

4. According to claim 3, a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, characterized in that: The training of the single-step optimization strategy network and the cascade relationship reasoning strategy network for alarm events in the power grid fault event knowledge graph specifically includes: Add a reverse relationship to each triple in the knowledge graph, extract the head entity and relationship from the triple to form a bigram, and construct a sample set B for training the single-step optimization strategy network. train ; The training sample set B train Each entity embedding of is converted into a state vector and input into the policy neural network to select the best next action. train Whether the binary pairs in match, the reward value is updated until the gradient change of the policy network parameters tends to zero, and the training is completed; Extract all triples containing specific cascade relations from the knowledge graph, construct a dataset for training the cascade relation reasoning strategy network, and divide it into training set and test set; Initialize the agent's reasoning environment and load the updated single-step optimal strategy network parameters. Use the head entity and tail entity of the triplet in the training set as the starting state and target state of the agent's reasoning, respectively. The agent learns through the policy neural network, expands the cascade relationship, trains the cascade relationship reasoning strategy network through the reward function, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

5. According to claim 3, a method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models, characterized in that: The interlocking fault event information and interlocking alarm event information based on the fault event are used to perform knowledge reasoning on the fault event knowledge graph assisted by a pre-trained language model and a graph neural network model, so as to obtain the full process information of the power grid fault handling and generate a power grid fault handling navigation map, including: The BERT pre-trained language model creates initial semantic representations for entities and relations in the triples (h, r, t) of faults, relations, and treatment measures in the knowledge graph of power grid faults, and maps them into low-dimensional embedding vectors. The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures in the fault event knowledge graph; the cosine similarity between the text feature encodings is calculated for semantic search to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event; All fault handling nodes and relationships similar to the fault event are input into the GNN model, and higher weights are given to entities and relationships with high similarity; The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event. The GNN model performs multi-layer forward propagation, integrates node neighbor information, uses a breadth-first search algorithm to track high attention weight paths, and infers the full process information of power grid fault handling; Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault disposal is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault disposal operation is converted into a power grid fault disposal navigation map.

6. The method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models according to claim 1 is characterized in that: Integrate large-scale language models with the knowledge graph of power grid faults to conduct power grid fault handling decision reasoning and generate power grid fault handling solutions, including: Based on the KG-GPT framework, the large-scale language model in the field of power grid control and the knowledge graph of power grid faults are integrated to form a KG-GPT model. Through the knowledge graph of power grid fault events, the entities, relationships and attribute information related to the fault events are extracted and input into the knowledge layer of the KG-GPT model; Through the knowledge layer, the entities, relations and attribute information related to the fault event are connected to the fault alarm sentences in the alarm event information to form a sentence tree structure; The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model; The Transformer decoder layer associates the vector sequence according to the process sequence of power grid fault handling, completes fault information processing, and generates fault analysis briefings and power grid fault handling plans.

7. An intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models, characterized in that: include: A power grid fault event knowledge graph construction module is used to collect multi-source business data related to fault events in the power grid dispatching and control system, perform unstructured information modeling on the multi-source business data related to the fault events, and construct a power grid fault event knowledge graph; The power grid integrated intelligent alarm module is used to perform component-level fault diagnosis of the monitoring system based on the relevant multi-source business data of the fault event, and to perform power grid event-level alarm correlation analysis in combination with the knowledge graph of power grid fault events to obtain the interlocking fault event information and interlocking alarm event information of the fault event; The power grid fault handling navigation map generation module is used to generate interlocking fault event information and interlocking alarm event information based on fault events. It uses the pre-trained language model and graph neural network model to assist the fault event knowledge graph to perform knowledge reasoning, obtain the full process information of fault handling, and generate a power grid fault handling navigation map. The power grid fault handling plan output module is used to integrate the large language model with the power grid fault event knowledge graph, perform power grid fault handling decision reasoning, and generate a power grid fault handling plan.

8. The intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models according to claim 7 is characterized in that: The method collects multi-source business data related to fault events in the power grid dispatching and control system, models the unstructured information of the multi-source business data related to the fault events, and constructs a knowledge graph of power grid fault events, including: Collect multi-source business data related to fault events in the power grid dispatching and control system, and pre-process the multi-source business data related to fault events. Label the pre-processed multi-source business data related to the fault event; The format conversion and knowledge extraction are performed on the labeled multi-source business data. Based on the mechanism and judgment rules of power grid fault alarm, the fault entities and relationships are organized in the form of entity, relationship, and entity triples and stored in the Neo4j graph database to construct a power grid fault knowledge graph.

9. The intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models according to claim 7 is characterized in that: Based on the relevant multi-source business data of the fault event, the monitoring system component-level fault diagnosis is carried out, and the grid event-level alarm correlation analysis is carried out in combination with the knowledge graph of the power grid fault, and the interlocking fault event information and interlocking alarm event information of the fault event are obtained, including: Based on the Deep Walk node embedding algorithm, the entities in the multi-source business data related to the fault event are aligned; based on the logical relationship between the fault event alarm information and the basic elements of the fault event, the text and image multimodal information in the multi-source business data related to the fault event are entity linked, and based on the aligned entities and the configuration information of the entity link, the characteristic events of the fault alarm are generated, and the causal relationship between the characteristic events and the fault category is used to generate the fault analysis rules; Based on the fault analysis rules and the knowledge graph of power grid faults, a triple structure of the fault alarm analysis process is constructed and stored in the Neo4j graph database; The entities and relations in the knowledge graph are mapped to a low-dimensional vector space through the TransE model; A deep reinforcement learning model consisting of state space S, action space A, state transition P, reward function β and policy neural network is established in the vector space; By training the single-step optimal strategy network and the cascade relationship reasoning strategy network of the alarm events in the fault event knowledge graph, the network parameters of the deep reinforcement learning model are updated to obtain the interlocking fault event information and interlocking alarm event information of the fault events.

10. The intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models according to claim 7, characterized in that: The single-step optimization strategy network and the cascade relationship reasoning strategy network in the training strategy neural network specifically include: Add a reverse relationship to each triple in the knowledge graph, extract the head entity and relationship from the triple to form a bigram, and construct a sample set B for training the single-step optimization strategy network. train ; The training sample set B train Each entity embedding of is converted into a state vector and input into the policy neural network to select the best next action. train Whether the binary pairs in match, the reward value is updated until the gradient change of the policy network parameters tends to zero, and the training is completed; Extract all triples containing specific cascade relations from the knowledge graph, construct a dataset for training the cascade relation reasoning strategy network, and divide it into training set and test set; Initialize the agent's reasoning environment and load the updated single-step optimal strategy network parameters. Use the head entity and tail entity of the triplet in the training set as the starting state and target state of the agent's reasoning, respectively. The agent learns through the policy neural network, expands the cascade relationship, trains the cascade relationship reasoning strategy network through the reward function, and updates the network parameters until the number of training steps reaches the set maximum number of reasoning steps.

11. The intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models according to claim 7, characterized in that: The interlocking fault event information and interlocking alarm event information based on the fault event are used to perform knowledge reasoning on the fault event knowledge graph assisted by a pre-trained language model and a graph neural network model, so as to obtain the full process information of the power grid fault handling and generate a power grid fault handling navigation map, including: The BERT pre-trained language model creates initial semantic representations for entities and relations in the triples (h, r, t) of faults, relations, and treatment measures in the knowledge graph of power grid faults, and maps them into low-dimensional embedding vectors. The BERT model is used to extract the interlocking fault event information and interlocking alarm event information of the fault event and the text feature encoding corresponding to the fault handling measures in the fault event knowledge graph; the cosine similarity between the text feature encodings is calculated for semantic search to obtain the fault handling measure nodes and relationships with high semantic similarity to the queried fault event; All fault handling nodes and relationships similar to the fault event are input into the GNN model, and higher weights are given to entities and relationships with high similarity; The entities in the knowledge graph are used as nodes of the GNN model, and the relationships are used as edges. The information of different neighbor nodes is fused through the message passing mechanism, and the embedded representation of each fault node is gradually updated. The embedded representation of each fault event node is gradually updated by combining the most relevant fault handling measure nodes and relationships of the fault event. The GNN model performs multi-layer forward propagation, integrates node neighbor information, uses a breadth-first search algorithm to track high attention weight paths, and infers the full process information of power grid fault handling; Through the Neo4j graph database and Cypher query operation, the full process information of power grid fault disposal is visualized and analyzed in the Neo4jBrowser interface, and finally the generated full process information of power grid fault disposal operation is converted into a power grid fault disposal navigation map.

12. The intelligent auxiliary analysis system for power grid faults driven by complementary graphics and models according to claim 7, characterized in that: Integrate large-scale language models with the knowledge graph of power grid faults to conduct power grid fault handling decision reasoning and generate power grid fault handling solutions, including: Based on the KG-GPT framework, the large-scale language model in the field of power grid control and the knowledge graph of power grid faults are integrated to form a KG-GPT model. Through the knowledge graph of power grid fault events, the entities, relationships and attribute information related to the fault events are extracted and input into the knowledge layer of the KG-GPT model; Through the knowledge layer, the entities, relations and attribute information related to the fault event are connected to the fault alarm sentences in the alarm event information to form a sentence tree structure; The sentence tree structure is converted into a vector sequence through the embedding layer of the KG-GPT model and input into the Transformer decoder layer of the KG-GPT model; The Transformer decoder layer associates the vector sequence according to the process sequence of power grid fault handling, completes fault information processing, and generates fault analysis briefings and power grid fault handling plans.

13. A computer storage medium, wherein the computer readable storage medium stores a computer program, characterized in that: When the computer program is executed by a processor, the method for intelligent auxiliary analysis of power grid faults driven by complementary graphics and models as described in any one of claims 1 to 6 is implemented.

14. An electronic device, characterized in that: It includes a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the intelligent auxiliary analysis method for power grid faults driven by complementary graphics and models as described in any one of claims 1 to 6 is implemented.

Images