Panoramic intelligent navigation system and method for power grid fault disposal

Abstract

The invention discloses a power grid fault disposal panoramic intelligent navigation system, which comprises a fault disposal navigation affair mechanism model building module, which is used for carrying out text preprocessing on acquired power grid fault disposal data, then carrying out coding, identifying fault disposal entities, relationships and events, and carrying out data processing. The real-time alignment information is fused to establish a fault handling navigation affair mechanism model; the fault handling panoramic information reasoning module extracts fault handling entity information according to real-time information, transmits the fault handling entity information to a fault handling navigation affair mechanism model to generate a fault handling low-dimensional vector, and reasones an optimal fault handling measure by analyzing a power grid fault event and combining a power grid operation mode and power grid flow information; and the fault handling panoramic intelligent navigation module performs reasoning to obtain fault handling information, risk information, a risk control strategy and power restoration decision key factors, and generates a power restoration result in combination with on-site in-station and line patrol conditions. According to the invention, the accuracy and efficiency of power grid fault handling are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system fault handling, and in particular to a panoramic intelligent navigation system and method for power grid fault handling. Background Art

[0002] At present, my country's AC / DC hybrid large power grid is gradually taking shape, the grid structure is becoming increasingly complex, and the operation mode is flexible and changeable, which leads to increasingly complex control business. If a single power grid fault is not handled in time, it will trigger a chain reaction of faults between the AC and DC systems, posing a serious threat to the safe and stable operation of the power grid. The existing power grid fault handling scheme relies on dispatchers to judge and analyze the faults, use manual records for reporting, and communicate information through manual telephones. According to statistics, the current manual handling of faults takes an average of tens of minutes, which increases the risk of further chain failures and cannot meet the requirements of timely handling of power grid faults in complex AC / DC hybrid power grids.

[0003] With the rapid development of computer and artificial intelligence technology, comprehensive technologies with machine learning, deep learning, and big data methods as the core have gradually taken shape, and have achieved remarkable application results in many fields such as computer vision, natural language processing, and voice interaction. Artificial intelligence technology will be the core technology for realizing intelligent power grid control. Research on intelligent power grid fault handling based on real-time power grid fault information and prior knowledge of fault handling is still in the theoretical research and preliminary exploration stage. How to combine multi-source heterogeneous power grid fault data to simulate the thinking of dispatchers, automatically guide the power grid fault handling process, and realize intelligent power grid fault handling is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a panoramic intelligent navigation system and method for power grid fault handling, which improves the accuracy and efficiency of power grid fault handling.

[0005] To achieve this purpose, the present invention designs a panoramic intelligent navigation system for power grid fault handling, which includes:

[0006] The fault handling navigation mechanism model construction module is used to perform text preprocessing on the acquired power grid fault handling data, identify fault handling entities, relationships and events from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and integrate and align the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data, thereby unifying the same type of entities in the power grid fault handling data and establishing a fault handling navigation mechanism model;

[0007] The fault handling panoramic information reasoning module is used in the power grid dispatching and control system to extract fault handling entity information based on power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications, and transmit the fault handling entity information to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained;

[0008] The fault handling panoramic intelligent navigation module is used to obtain the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process through the graph attention reasoning model, and generate the power restoration result in combination with the on-site station and line patrol conditions.

[0009] Preferably, the specific method for constructing the fault handling navigation mechanism model is:

[0010] Obtain power grid fault handling data from the power grid dispatching control center and operation management department. The power grid fault handling data includes power grid operation business data, dispatching fault handling principles and fault handling experience of domain experts. Perform text preprocessing and fault handling entity tagging on the power grid fault handling data. The fault handling entity tagging is used for knowledge extraction of power grid fault handling data.

[0011] The power grid operation business data is directly extracted to obtain the entity information of the power grid equipment. The dispatch fault handling principle is extracted by rule-based keyword extraction to obtain the fault entities, relations and events in the accident plan. The domain expert fault handling experience is extracted by a sequence annotation model based on knowledge enhanced semantic representation (ERNIE2.0-BiLSTM-CRF) to obtain the fault entities and relations. Based on the above knowledge extraction training process, a trained key knowledge extraction model for power grid fault handling is obtained.

[0012] Centered on power grid equipment, the fault entities in power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications are integrated and aligned to unify entities of the same type in power grid fault handling data;

[0013] Combined with the entities, relationships and events obtained by knowledge extraction of power grid fault handling data, the power grid fault handling process is generated based on the fault handling logic and fault handling rules. According to the above fault handling rules and power grid fault handling process, the extracted fault knowledge is constructed into a basic structure of "concept-entity-rule-process" and stored in a graph database to construct a power grid fault handling navigation mechanism model.

[0014] Preferably, the text preprocessing includes text cleaning, text standardization, text segmentation and removal of stop words.

[0015] Preferably, the fault handling entity information is extracted based on the power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and is used as the input of the fault handling navigation mechanism model, and the fault handling low-dimensional vector is generated by the fault handling navigation mechanism model;

[0016] The low-dimensional vector of fault handling is used as the input of the Graph Transformer framework combined with the self-attention mechanism. The Graph Transformer framework is used to analyze the power grid fault events, power grid operation mode and power grid flow information, and the panoramic information reasoning process of fault handling is carried out to obtain the optimal power grid fault handling measures.

[0017] Based on the trained key knowledge extraction model for power grid fault handling, the fault entities and relationships obtained by knowledge extraction of power grid fault events, power grid operation modes, power grid flow information, power grid equipment information, power grid model information, operation information, stability regulations, accident plans and operation control applications are tripled according to the basic knowledge structure in the form of "concept, entity, rule, process", and multiple types of end-to-end triples are obtained, including "fault relationship, plan information", "plan information, relationship, risk information", "risk information, relationship, risk control strategy", and "risk control strategy, relationship, key factors for power restoration decision-making". The multiple types of end-to-end triples are used as the input of the graph attention reasoning model. The power grid dispatching and control system obtains the fault handling information, risk information, risk control strategy and key factors for power restoration decision-making required for the power grid fault handling process through the graph attention reasoning model, and generates the power restoration result in combination with the on-site station and line patrol conditions.

[0018] Preferably, the specific process of aligning the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications using the VR-GCN model with a high-speed gate mechanism added is:

[0019] The fault entities contained in the power grid equipment information, power grid model information, operation information, stability procedures, accident plans and operation control applications are represented as nodes, and the relationships are represented as edges to construct the corresponding graph structure data; vectorized relational convolution operations are performed on each node to aggregate the features of neighboring nodes and update the node representation; a high-speed gating mechanism is introduced to set a gating value for each node or edge according to the structural characteristics, and dynamically determine which nodes or edges are filtered during the message propagation process according to the gating value; VR-GCN model training and inference are performed, and the similarity between similar nodes in different graphs is calculated through the node representation output by the VR-GCN model, and the node pairs are determined to be aligned according to the set threshold. The greedy matching method is used to determine the alignment results of various fault entities;

[0020] Based on the alignment relationship between the knowledge points covered by power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, as well as the fault entities, redundant entities are automatically identified and eliminated, and multiple referential items of the same entity object are set to a unified entity name and integrated into information related to power grid fault handling.

[0021] Preferably, the implementation process of the fault handling panoramic information reasoning includes:

[0022] Use power grid fault information to perform fault semantic search and obtain related fault knowledge; build a fault joint reasoning space; use the Graph Transformer framework combined with the self-attention mechanism to perform fault reasoning;

[0023] The specific process of fault reasoning using the Graph Transformer framework combined with the self-attention mechanism is as follows:

[0024] Obtain entity embedding n for each node v and its neighbor node u based on node knowledge v , n u ;

[0025] Incorporate the conditional probability feature p between nodes v and u uv , the calculation formula is as follows:

[0026]

[0027] Where n is the number of occurrences of the corresponding node event counted in the historical fault corpus, and p uv is the conditional probability of v and u, which is used to calculate the probability that the fault corresponding to node v also occurs when the fault corresponding to neighboring node u occurs, thereby measuring the size of the correlation between different fault handling measures and fault events;

[0028] By p uv Calculate the relationship embedding r between the two uv, the calculation formula is as follows:

[0029] r uv =f r (n u ,n v ,e uv ,p uv ) (2)

[0030] In the formula, f r is a multi-layer perceptron network; uv is a unique hot vector reflecting the relationship between node u and node v;

[0031] Calculate the attention weight of neighbor nodes to their own nodes, and use the attention mechanism to aggregate information:

[0032] For each node, its entity embedding is mapped to the query vector Q, key vector K and value vector V. Taking node v as an example, the calculation formula is as follows:

[0033] Q v =W Q n v ,K v =W K n v ,V v =W V n v (3)

[0034] Where W Q , W K , W V is the learnable parameter matrix, W Q is the parameter matrix corresponding to the query vector; W K is the parameter matrix corresponding to the key vector; W V is the parameter matrix corresponding to the value vector; Q v K represents the query vector of node v; v represents the key vector of node v; V v Represents the value vector of node v; n v is the embedding vector representation of any node v;

[0035] Calculate the attention weight α of each neighbor node to its own node uv , the calculation formula is as follows:

[0036]

[0037] Where, d k The dimension of the key vector obtained by mapping the entity embedding of node v, Q u is the query vector of neighbor node u, is the transposed vector of the key vector of node v. The softmax function converts the input vector into a probability distribution so that the value of each element is between 0 and 1, and the sum of all elements is 1.

[0038] The attention weight value is used to perform weighted averaging on the value vectors of all neighbor nodes to achieve information aggregation of neighbor nodes. The calculation formula is as follows:

[0039]

[0040] Where N(v) is the number of neighbor nodes of node v, l represents the lth layer of the Graph Transformer framework, is the embedding vector representation of any node v in the lth layer, V u is the value vector of neighbor node u;

[0041] The output nodes of each layer are input into the two-layer feedforward neural network, and the ReLU function is used for nonlinear activation. The calculation formula is as follows:

[0042]

[0043] Where W1 is the weight parameter matrix of the first layer of feedforward neural network; W2 is the weight parameter matrix of the second layer of feedforward neural network; It represents the node feature representation after being processed by two layers of feedforward neural network;

[0044] At the end of the model, the representation of each node is updated through layer normalization operations and residual connections. The calculation formula is as shown in formula (7). The message is passed between nodes through the attention mechanism until convergence is reached:

[0045]

[0046] In the formula, LayerNorm is the layer normalization operation, is the embedding vector representation of node v after the forward propagation update of the Graph Transformer framework;

[0047] The breadth-first traversal algorithm is used to track the high attention weight transmission path from the fault information node W to the fault phenomenon node, then to the fault type node, and all possible next-step fault handling measures nodes. After the message transmission is completed, the final representation of each node is returned, and the weighted summation is performed through the multi-layer perceptron layer to obtain each possible next-step fault handling measure and their respective probabilities, and the optimal fault handling measure is inferred.

[0048] Preferably, the graph attention reasoning model is a graph attention reasoning model based on the ConvKB embedding model and combined with node neighborhood information. The specific reasoning process of obtaining the fault handling information, risk information, risk control strategy, and key factors of power restoration decision required for the power grid fault handling process through the graph attention reasoning model is:

[0049] The random initialization method is used to vectorize the fault entities and relations based on the triple structure, and the initialization embedding vector, i.e., the low-dimensional vector, is obtained.

[0050] The known triples are recorded as: T = {(h, r, t) | h, t∈E, r∈R}, where h and t are the head entity vector and the tail entity vector respectively; r is the relation vector;

[0051] Update the representation vector of the central node and the head entity vector and tail entity vector of the known fault triplet through the neighborhood-level attention method;

[0052] Taking the fault information node Z as the central node z, the direct neighborhood triples around it can be expressed as:

[0053]

[0054] Where z is the representation vector corresponding to the head entity; r k is the representation vector corresponding to the relationship; e j is the representation vector corresponding to the tail entity; [] represents the vector concatenation operation; W1 is the weight parameter matrix for the linear transformation of the triplet;

[0055] The LeakyReLU function is used for nonlinear transformation to obtain the attention value of the triplet:

[0056]

[0057] In the formula, s zjk is the attention value of the triple; W2 is the linear transformation matrix parameterized by weight; is the direct neighborhood triplet around the central node z;

[0058] Use the Softmax function to normalize all attention values ​​to obtain the relative attention value β zjk :

[0059]

[0060] In the formula, E z is the set of tail entities in all direct neighborhood triples around the central entity Z; R z is the set of relations in all direct neighborhood triples around the central entity Z; exp(s zjk) Get the value of the triplet’s attention value after taking the natural exponential function; s inr It refers to the attention value of the ith direct neighborhood triplet around the fault information node Z, i represents the head entity of the triplet, n represents the tail entity of the triplet, and r represents the relationship between the triples;

[0061] Combine equations (8) and (10) to aggregate the fault information of the immediate neighborhood around the central entity and update the representation vector of z:

[0062]

[0063] In the formula, is the set of all direct neighborhood triplets around the central node z; β zjk is the relative attention value of the triple; σ is the sigmoid activation function of the graph attention network;

[0064] Att(z) is the updated central node representation vector, which is still recorded as z in subsequent calculation steps;

[0065] Similarly, the embedding vectors of the head entity and the tail entity in the known fault triple are also updated according to the above method;

[0066] Read the training data set and generate negative samples through the negative sampling strategy. Use uniform sampling or Bernoulli sampling to randomly replace the head entity or tail entity or relationship in the positive sample triple with other entities or relationships. The set of negative example triples is defined as follows:

[0067]

[0068] In the formula, h′ is the head entity after random replacement; r′ is the relation after random replacement; t′ is the tail entity after random replacement; T′ is the set of negative triples;

[0069] The convolution feature vector is converted into a scoring function through the fully connected layer of the ConvKB embedding model to predict the effectiveness of known triples, which is defined as follows:

[0070]

[0071] Where s is the score; M is the number of convolution layers; * represents the convolution operation; ε n represents the nth layer of convolution; W is the linear transformation matrix used to convert and obtain the final score;

[0072] Based on the ConvKB embedding model framework, in order to compare positive and negative samples to update model parameters and obtain more comprehensive fault semantic information, the loss function is constructed by minimizing the logarithmic loss to optimize the model, which is defined as follows:

[0073]

[0074] Where L is the model function value; T ∪T′ is the set of positive triples and negative triples, that is, the set of all triple samples; ω (h,r,t) is the sample label. If the triple is a positive example, that is, (h, r, t)∈T, then ω (h,r,t) =1; if the triple is a negative example, that is, (h, r, t)∈T′, then ω (h,r,t) =-1; α is the regularization coefficient, which is used to control the contribution of the regularization term to the loss value;

[0075] The stochastic gradient descent optimization method is used to update the fault entity vector and relationship vector to minimize the loss function; the above steps are repeated until the model convergence condition is reached and the algorithm stops; the information corresponding to the final node representation output by the graph attention reasoning model is the fault handling information; the triples of "plan information, relationship, risk information", "risk information, relationship, risk control strategy", and "risk control strategy, relationship, key factors for power restoration decision-making" are used as the input of the graph attention reasoning model in turn, and the above steps are repeated to continue to infer the risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process.

[0076] Preferably, the power grid flow information refers to the data of electrical characteristics of each node and branch and their mutual relationship obtained through flow calculation during the operation of the power system, including current, voltage, power, frequency, flow direction and network topology.

[0077] A panoramic intelligent navigation method for power grid fault handling, comprising the following steps:

[0078] The acquired power grid fault handling data is subjected to text preprocessing, and fault handling entities, relationships and events are identified from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data are integrated and aligned, thereby unifying the same type of entities in the power grid fault handling data, and then establishing a fault handling navigation mechanism model;

[0079] In the power grid dispatching and control system, the fault handling entity information is extracted based on the power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and the fault handling entity information is transmitted to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained;

[0080] Based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process, the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process are obtained through the graph attention reasoning model. Combined with the optimal power grid fault handling measures, on-site station and line patrol conditions, the power restoration results are generated.

[0081] A computer program product comprises a computer program, wherein the computer program implements the steps of the above method when executed by a processor.

[0082] Beneficial effects of the present invention:

[0083] The present invention improves the accuracy and efficiency of power grid fault handling. Through the implementation of the present invention, the intelligence level of power grid fault handling is significantly improved, and power grid faults can be quickly and accurately identified and handled, reducing the impact of faults on power grid operation, helping to improve the safety and reliability of power grid operation, and ensuring the stability and continuity of power supply; at the same time, the intelligent fault handling method can reduce manual intervention, reduce the labor intensity and workload of operation and maintenance personnel, help reduce power grid operation and maintenance costs, improve the economic benefits of power grid operation, and provide strong support for the safe, reliable and efficient operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a structural schematic diagram of the present invention;

[0085] Figure 2 It is a partial entity labeled diagram of the present invention;

[0086] Figure 3 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0087] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0088] Example 1

[0089] A panoramic intelligent navigation system for power grid fault handling, such as Figure 1 , 2 As shown, it includes:

[0090] The fault handling navigation mechanism model construction module is used to perform text preprocessing on the acquired power grid fault handling data, identify fault handling entities, relationships and events from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and integrate and align the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data, thereby unifying the same type of entities in the power grid fault handling data, and then establishing a fault handling navigation mechanism model. This design can improve the efficiency of fault handling and enhance the accuracy of fault handling;

[0091] The fault handling panoramic information reasoning module is used in the power grid dispatching and control system to extract fault handling entity information based on power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications, and transmit the fault handling entity information to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained. This design is used to optimize power grid fault handling measures, enhance the safety and stability of power grid operation, and promote the application of intelligent technology in power grid dispatching;

[0092] The fault handling panoramic intelligent navigation module is used to obtain the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process through the graph attention reasoning model, and generate power restoration results in combination with the on-site station and line patrol conditions. This design can improve the intelligence level of fault handling, optimize the fault handling process, improve risk control capabilities, enhance the scientific nature of power restoration decisions, and promote the intelligent development of power grids.

[0093] In the above technical solution, the specific method for constructing the fault handling navigation mechanism model is:

[0094] Obtain power grid fault handling data from the power grid dispatching control center and operation management department. The power grid fault handling data includes power grid operation business data, dispatching fault handling principles and fault handling experience of domain experts. Perform text preprocessing and fault handling entity tagging on the power grid fault handling data. The fault handling entity tagging is used for knowledge extraction of power grid fault handling data.

[0095] The power grid operation business data (structured data reflecting the real-time operation status of power grid equipment) is directly extracted (i.e., directly using the power grid operation section data to identify and match adjacent topological nodes, combining the equipment name and the area to which it belongs, and automatically extracting the entity information of the power grid equipment) to obtain the entity information of the power grid equipment (the operation status, topological structure and power grid flow information of the power grid equipment). The dispatching fault handling principles (the guidelines and operating specifications that dispatchers follow when handling power grid faults, including power grid fault phenomena, stability procedures and accident plans, which have good compilation specifications, and the text expression and editing format are consistent, and belong to semi-structured data) are extracted by rule-based keywords (according to the fault case name and corresponding handling principles in the power grid fault handling terminology library, the normalized information is locked, the field content before and after the normalized information is identified, the key information is matched using the regular expression method, the redundant characters in the text are removed, the key information is extracted and merged, and the fault entities, relations and events in the accident plan are obtained) to obtain the fault entities, relations and events in the accident plan. The fault handling experience of domain experts ( Experts with rich knowledge and practical experience in the power field have acquired relevant knowledge, insights and skills in the work of handling power grid faults; this type of data lacks a fixed format and is presented in the form of oral communication, text records and emails, and is unstructured data) adopts a sequence labeling model based on knowledge-enhanced semantic representation (i.e., ERNIE2.0-BiLSTM-CRF model). The model principle and process are as follows: input the domain expert experience text, perform text segmentation through the model input layer, and obtain the tokens corresponding to the entity words; the ERNIE2.0 pre-training model is pre-trained through a masked language model and a variety of knowledge enhancement tasks, and the tokens are encoded in combination with contextual semantic information to obtain an embedded vector representation; the embedded vector is input into the BiLSTM model to capture the forward and backward semantic information of the entity word, model long-distance dependencies, and obtain the final representation of each word; the output features of the BiLSTM model are used as the input of the CRF layer to model the dependency relationship of the label sequence, and the Viterbi algorithm is used to obtain the global optimal label sequence path, and the predicted entity label sequence and relationship label sequence are output. According to the obtained label sequence, the continuous words are identified as an entity, and the relationship type and direction are identified by combining the entity position and context information to perform knowledge extraction to obtain fault entities and relationships; based on the above knowledge extraction training process, a trained key knowledge extraction model for power grid fault handling is obtained;

[0096] Centered on power grid equipment, the fault entities in power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications are integrated and aligned to unify entities of the same type in power grid fault handling data;

[0097] Combined with the entities, relations and events obtained by knowledge extraction of power grid fault handling data, a power grid fault handling process is generated based on fault handling logic and fault handling rules. According to the above fault handling rules and power grid fault handling process, the extracted fault knowledge is constructed into a basic structure of "concept-entity-rule-process" and stored in a graph database to build a power grid fault handling navigation mechanism model.

[0098] The above design, by constructing a grid fault handling navigation mechanism model, can provide dispatchers and on-site personnel with intelligent navigation and auxiliary decision-making support, carry out early warning and prevention of potential fault risks, and enhance the safety and stability of grid operation.

[0099] In the above technical scheme, the text preprocessing includes text cleaning (removing irrelevant special characters, extra spaces, etc. in the dispatching fault handling principles and the experience of domain experts to ensure that the fault text content is compact and consistent), text standardization (unifying the text case, standardizing inconsistent abbreviations and variable units in the text, and standardizing digital expressions, for example, unifying "kilovolts" as "kV"), text segmentation (building a professional dictionary based on the professional terminology of power grid fault handling and the entity labels of plants, units, lines, and substations defined in advance by the dispatcher, and using the Jieba word segmenter for Chinese word segmentation to facilitate subsequent knowledge extraction tasks) and removing stop words (building a stop word list in combination with the State Grid corporate standards and the power industry-specific dictionary to remove modifying words and meaningless punctuation marks in the corpus); on the corpus after text preprocessing, the BIO annotation mode is adopted to mark the power grid fault handling entities according to the "concept-entity-rule-process" framework. The entity labels used are determined according to the fault logical relationship in the corpus, specifically including: plants, units, lines, substations, voltage levels, fault events, and handling operations. Some annotation information such as Figure 2 As shown in the figure, each row represents an entity word and its corresponding entity label; in the entity label, SB represents the plant, UB represents the unit, RB represents the line, LB represents the voltage level, TB represents the substation, EB represents the fault event, and DB represents the disposal operation; the above design can improve the accuracy and consistency of information, improve the efficiency of text processing, and enhance the model recognition ability by performing text preprocessing on the power grid fault disposal data.

[0100] In the above technical solution, fault handling entity information is extracted based on power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and used as input to the fault handling navigation mechanism model, which generates a fault handling low-dimensional vector;

[0101] The low-dimensional vector of fault handling is used as the input of the Graph Transformer framework combined with the self-attention mechanism. The Graph Transformer framework is used to analyze the power grid fault events, power grid operation mode and power grid flow information, and the panoramic information reasoning process of fault handling is carried out to obtain the optimal power grid fault handling measures.

[0102] Based on the trained key knowledge extraction model for power grid fault handling, the fault entities and relationships obtained by knowledge extraction of power grid fault events, power grid operation modes, power grid flow information, power grid equipment information, power grid model information, operation information, stability regulations, accident plans and operation control applications are tripled according to the basic knowledge structure in the form of "concept, entity, rule, process", and multiple types of end-to-end triples are obtained, 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". The multiple types of end-to-end triples are used as the input of the graph attention reasoning model. The power grid dispatching and control system obtains the fault handling information, risk information, risk control strategy and key factors for power restoration decision-making required for the power grid fault handling process through the graph attention reasoning model, and generates the power restoration result in combination with the on-site station and line patrol conditions;

[0103] The above design achieves comprehensive and in-depth processing of power grid fault handling by integrating a variety of advanced technologies and methods, improves the intelligence level of fault handling, optimizes the fault handling process, enhances risk control and power restoration decision-making capabilities, improves knowledge extraction and structured processing capabilities, and promotes the intelligent development of power grids.

[0104] In the above technical solution, the specific process of aligning the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications using the VR-GCN (Vectorized Relational Graph Convolutional Network) model with a high-speed gate mechanism is as follows:

[0105] The fault entities contained in the power grid equipment information, power grid model information, operation information, stability procedures, accident plans and operation control applications are represented as nodes, and the relationships are represented as edges to construct the corresponding graph structure data; vectorized relational convolution operations are performed on each node (through vectorized relational convolution operations, the neighbor node information of each node is passed to itself, and the node representation is updated, which is equivalent to combining rich contextual semantic information when performing vector representation on each fault entity) to aggregate the features of neighbor nodes and update the node representation; a high-speed gating mechanism is introduced to set a gating value for each node or edge according to the structural characteristics ("gating value" refers to a dynamic weight that controls the transmission or aggregation of information, which is set according to the structural characteristics of the node or edge and is used to selectively control the propagation of information; the gating mechanism can better cope with the problem of information transmission in complex structures, ensuring that the model captures global information while not losing local key features, thereby effectively improving the accuracy and efficiency of the model ), dynamically determine which nodes or edges are filtered during the message propagation process according to the gating value; perform VR-GCN model training and inference, calculate the similarity between similar nodes in different graphs through the node representation output by the VR-GCN model, and determine whether the node pair (a "node pair" refers to two different nodes of the same type) is aligned (compare the calculated similarity value between nodes with the set threshold value, if the similarity value is greater than or equal to the threshold value, the two nodes are aligned; if the similarity value is less than the threshold value, the two nodes are not aligned), and use the greedy matching method to determine the alignment results of various fault entities;

[0106] Automatically identify and eliminate redundant entities based on the alignment relationship between the knowledge points and fault entities covered by power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and set multiple referents of the same entity object as a unified entity name to merge into power grid fault handling related information;

[0107] The above design improves the accuracy and efficiency of information alignment by introducing a VR-GCN model with a high-speed gate mechanism for information alignment.

[0108] In the above technical solution, the implementation process of the fault handling panoramic information reasoning includes:

[0109] Use power grid fault information to perform fault semantic search and obtain related fault knowledge (input the power grid operation mode and power grid flow information (complete fault information) when the fault occurs into the semantic search pre-training model (BERT), and update the network parameters through pre-training tasks (masked language model and next sentence prediction); use the Jieba word segmentation method based on dictionary and statistics to segment the input text, and vectorize it through the word embedding layer of the BERT model and convert it into 768 dimensions (768 dimensions are determined by the structure of the BERT model. The number of hidden units of each layer of the Transformer encoder of the BERT-base model is 768, which means that each word is embedded in a 768-dimensional vector space , which is a compromise between complexity and computing resources in the design of the BERT model) and sums up the Token, Segment and Position embedding vectors, which are input into the encoding layer of the BERT model. After the information transmission and interaction of the n-layer Transformer structure, the language modeling is completed, and the feature encoding of a knowledge unit in the fault information and fault handling navigation mechanism model is output; the two feature encodings are input into two consecutive fully connected layers, and the similarity between the two is calculated to obtain a two-dimensional vector representing the similarity and dissimilarity (the specific processing flow of obtaining the two-dimensional vector representing the similarity and dissimilarity is as follows: the two feature encodings are transformed through consecutive fully connected layers respectively, and the corresponding feature F is output. A and F B ; Calculate the cosine similarity of the two features as the similarity score S, and calculate the Euclidean distance between the two as the dissimilarity score D; S and D are combined into a two-dimensional vector V = (S, D)), and similar fault phenomenon nodes and fault handling current step nodes are returned accordingly, and the breadth-first search algorithm is used to find the fault type nodes that contain multiple similar fault phenomena and all possible next-step fault handling measure nodes, thereby obtaining associated fault knowledge); Construct a fault joint reasoning space (embed the complete fault information into a fault knowledge node W through the BERT model encoding layer, and use two new relationships to connect the node W with the fault type node found in the previous step and all possible next-step fault handling measure nodes, respectively, to construct a fault joint reasoning space, combining the complete fault information and associated fault knowledge); Use the Graph Transformer framework combined with the self-attention mechanism for fault reasoning;

[0110] The specific process of fault reasoning using the Graph Transformer framework combined with the self-attention mechanism is as follows:

[0111] Obtain entity embedding n for each node v and its neighbor node u based on node knowledge v , n u ;

[0112] Incorporate the conditional probability feature p between nodes v and uuv , the calculation formula is as follows:

[0113]

[0114] Where n is the number of occurrences of the corresponding node event counted in the historical fault corpus, and p uv is the conditional probability of v and u, which is used to calculate the probability that the fault corresponding to node v also occurs when the fault corresponding to neighboring node u occurs, thereby measuring the size of the correlation between different fault handling measures and fault events;

[0115] By p uv Calculate the relationship embedding r between the two uv , the calculation formula is as follows:

[0116] r uv =f r (n u ,n v ,e uv ,p uv ) (2)

[0117] In the formula, f r is a multi-layer perceptron network; uv is a unique hot vector reflecting the relationship between node u and node v;

[0118] Calculate the attention weight of neighbor nodes to their own nodes, and use the attention mechanism to aggregate information:

[0119] For each node, its entity embedding is mapped to the query vector Q, key vector K and value vector V. Taking node v as an example, the calculation formula is as follows:

[0120] Q v =W Q n v ,K v =W K n v ,V v =W V n v (3)

[0121] Where W Q , W K , W V is the learnable parameter matrix, W Q is the parameter matrix corresponding to the query vector, W K is the parameter matrix corresponding to the key vector, W V is the parameter matrix corresponding to the value vector, Q v represents the query vector of node v, K v represents the key vector of node v, V vRepresents the value vector of node v, n v is the embedding vector representation of any node v;

[0122] Calculate the attention weight α of each neighbor node to its own node uv , the calculation formula is as follows:

[0123]

[0124] Where, d k The dimension of the key vector obtained by mapping the entity embedding of node v, Q u is the query vector of neighbor node u, is the transposed vector of the key vector of node v. The softmax function converts the input vector into a probability distribution so that the value of each element is between 0 and 1, and the sum of all elements is 1.

[0125] The attention weight value is used to perform weighted averaging on the value vectors of all neighbor nodes to achieve information aggregation of neighbor nodes. The calculation formula is as follows:

[0126]

[0127] Where N(v) is the number of neighbor nodes of node v, l represents the lth layer of the Graph Transformer framework, is the embedding vector representation of any node v in the lth layer, V u is the value vector of neighbor node u;

[0128] The output nodes of each layer are input into the two-layer feedforward neural network, and the ReLU function is used for nonlinear activation. The calculation formula is as follows:

[0129]

[0130] Where W1 is the weight parameter matrix of the first layer of feedforward neural network; W2 is the weight parameter matrix of the second layer of feedforward neural network; It represents the node feature representation after being processed by two layers of feedforward neural network;

[0131] At the end of the model, the representation of each node is updated through layer normalization operations and residual connections. The calculation formula is as shown in formula (7). The message is passed between nodes through the attention mechanism until convergence is reached:

[0132]

[0133] In the formula, LayerNorm is the layer normalization operation, is the embedding vector representation of node v after the forward propagation update of the Graph Transformer framework;

[0134] The breadth-first traversal algorithm is used to track the high attention weight transmission path from the fault information node W to the fault phenomenon node, then to the fault type node, and all possible next-step fault handling measures nodes. After the message transmission is completed, the final representation of each node is returned, and the weighted sum is performed through the multi-layer perceptron layer to obtain each possible next-step fault handling measure and their respective probabilities, and the best fault handling measure is obtained by reasoning;

[0135] The above design realizes the reasoning of panoramic information of fault handling and improves the accuracy and efficiency of fault handling.

[0136] In the above technical solution, the graph attention reasoning model is a graph attention reasoning model based on the ConvKB embedding model and combined with node neighborhood information. The specific reasoning process of obtaining the fault handling information, risk information, risk control strategy, and key factors of power restoration decision required by the power grid fault handling process through the graph attention reasoning model is as follows:

[0137] The random initialization method is used to vectorize the fault entities and relations based on the triple structure, and the initialization embedding vector, i.e., the low-dimensional vector, is obtained.

[0138] The known triples are recorded as: T = {(h, r, t) | h, t∈E, r∈R}, where h and t are the head entity vector and the tail entity vector respectively; r is the relation vector;

[0139] Update the representation vector of the central node and the head entity vector and tail entity vector of the known fault triplet through the neighborhood-level attention method;

[0140] Taking the fault information node Z as the central node z, the direct neighborhood triples around it can be expressed as:

[0141]

[0142] Where z is the representation vector corresponding to the head entity, r k is the representation vector corresponding to the relationship, e j is the representation vector corresponding to the tail entity; [] represents the vector concatenation operation; W1 is the weight parameter matrix for the linear transformation of the triplet;

[0143] The LeakyReLU function is used for nonlinear transformation to obtain the attention value of the triplet:

[0144]

[0145] In the formula, s zjkis the attention value of the triple; W2 is the linear transformation matrix parameterized by weight; is the direct neighborhood triplet around the central node z;

[0146] Use the Softmax function to normalize all attention values ​​to obtain the relative attention value β zjk :

[0147]

[0148] In the formula, E z is the set of tail entities in all direct neighborhood triples around the central entity Z; R z is the set of relations in all direct neighborhood triples around the central entity Z; exp(s zjk ) Get the value of the triplet’s attention value after taking the natural exponential function; s inr It refers to the attention value of the ith direct neighborhood triplet around the fault information node Z, i represents the head entity of the triplet, n represents the tail entity of the triplet, and r represents the relationship between the triples;

[0149] Combine equations (8) and (10) to aggregate the fault information of the immediate neighborhood around the central entity and update the representation vector of z:

[0150]

[0151] In the formula, is the set of all direct neighborhood triplets around the central node z; β zjk is the relative attention value of the triple; σ is the sigmoid activation function of the graph attention network;

[0152] Att(z) is the updated central node representation vector, which is still recorded as z in subsequent calculation steps;

[0153] Similarly, the embedding vectors of the head entity and the tail entity in the known fault triple are also updated according to the above method;

[0154] Read the training data set and generate negative samples through the negative sampling strategy. Use uniform sampling or Bernoulli sampling to randomly replace the head entity or tail entity or relationship in the positive sample triple with other entities or relationships. The set of negative example triples is defined as follows:

[0155]

[0156] In the formula, h′ is the head entity after random replacement, r′ is the relation after random replacement, t′ is the tail entity after random replacement; T′ is the set of negative triples;

[0157] The convolution feature vector is converted into a scoring function through the fully connected layer of the ConvKB embedding model to predict the effectiveness of known triples, which is defined as follows:

[0158]

[0159] Where s is the score; M is the number of convolution layers; * represents the convolution operation; ε n represents the nth layer of convolution; W is the linear transformation matrix used to convert and obtain the final score;

[0160] Based on the ConvKB embedding model framework, in order to compare positive and negative samples to update model parameters and obtain more comprehensive fault semantic information, the loss function is constructed by minimizing the logarithmic loss to optimize the model, which is defined as follows:

[0161]

[0162] Where L is the model function value; T ∪T′ is the set of positive triples and negative triples, that is, the set of all triple samples; ω (h,r,t) is the sample label. If the triple is a positive example, that is, (h, r, t)∈T, then ω (h,r,t) =1; if the triple is a negative example, that is, (h, r, t)∈T′, then ω (h,r,t) =-1; α is the regularization coefficient, which is used to control the contribution of the regularization term to the loss value;

[0163] Adopt the stochastic gradient descent optimization method to update the fault entity vector and relationship vector to minimize the loss function; repeat the above steps until the model convergence condition is reached and the algorithm stops; the information corresponding to the final node representation output by the graph attention reasoning model is the fault handling information; take the triples of "plan information-relationship-risk information", "risk information-relationship-risk control strategy", and "risk control strategy-relationship-key factors for power restoration decision-making" as the input of the graph attention reasoning model in turn, repeat the above steps, and continue to infer the risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process;

[0164] The above design realizes the precise reasoning of the information required in the power grid fault handling process by constructing a graph attention reasoning model based on the ConvKB embedding model and combining the node neighborhood information, improves the accuracy of information reasoning, enhances the interpretability of the model, improves the learning efficiency of the model, supports the processing of complex fault scenarios, and promotes the intelligent operation and maintenance of the power grid.

[0165] In the above technical solution, the power grid flow information refers to the data of electrical characteristics of each node and branch and their mutual relationship obtained through flow calculation during the operation of the power system, including current, voltage, power, frequency, flow direction and network topology.

[0166] Example 2

[0167] A panoramic intelligent navigation method for power grid fault handling, such as Figure 3 As shown, the fault handling process performs text preprocessing on the acquired power grid fault handling data, and then encodes it to identify fault handling entities, relationships and events, and integrates and aligns real-time information to establish a fault handling navigation mechanism model; fault handling entity information is extracted based on real-time information, and transmitted to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector, and the best fault handling measures are inferred by analyzing power grid fault events, combining power grid operation mode and power grid flow information; fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making are inferred, and the power restoration results are generated in combination with the on-site station and line patrol conditions.

[0168] The specific method of panoramic intelligent navigation for power grid fault handling includes the following steps:

[0169] The acquired power grid fault handling data is subjected to text preprocessing, and fault handling entities, relationships and events are identified from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data are integrated and aligned, thereby unifying the same type of entities in the power grid fault handling data, and then establishing a fault handling navigation mechanism model;

[0170] In the power grid dispatching and control system, the fault handling entity information is extracted based on the power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and the fault handling entity information is transmitted to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained;

[0171] Based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process, the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process are obtained through the graph attention reasoning model, and the power restoration results are generated in combination with the on-site station and line patrol conditions.

[0172] Example 3

[0173] A computer program product includes a computer program, characterized in that when the computer program is executed by a processor, the steps of the method described in Example 2 are implemented.

[0174] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A panoramic intelligent navigation system for power grid fault handling, characterized in that: It includes: The fault handling navigation mechanism model construction module is used to perform text preprocessing on the acquired power grid fault handling data, identify fault handling entities, relationships and events from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and integrate and align the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data, thereby unifying the same type of entities in the power grid fault handling data and establishing a fault handling navigation mechanism model; The fault handling panoramic information reasoning module is used in the power grid dispatching and control system to extract fault handling entity information based on power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications, and transmit the fault handling entity information to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained; The fault handling panoramic intelligent navigation module is used to obtain the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process through the graph attention reasoning model, and generate the power restoration result in combination with the on-site station and line patrol conditions.

2. The panoramic intelligent navigation system for power grid fault handling according to claim 1 is characterized by: The specific method for constructing the fault handling navigation mechanism model includes: Obtain power grid fault handling data from the power grid dispatching control center and operation management department. The power grid fault handling data includes power grid operation business data, dispatching fault handling principles and fault handling experience of domain experts. Perform text preprocessing and fault handling entity tagging on the power grid fault handling data. The fault handling entity tagging is used for knowledge extraction of power grid fault handling data. The power grid operation business data is directly extracted to obtain the power grid equipment entity information, the dispatch fault handling principle is extracted by rule-based keyword extraction to obtain the fault entities, relations and events in the accident plan, and the domain expert fault handling experience is extracted by a sequence annotation model based on knowledge enhanced semantic representation to obtain the fault entities and relations; based on the above knowledge extraction training process, a trained key knowledge extraction model for power grid fault handling is obtained; Centered on power grid equipment, the fault entities in power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications are integrated and aligned to unify entities of the same type in power grid fault handling data; Combined with the entities, relations and events obtained by knowledge extraction of power grid fault handling data, the power grid fault handling process is generated based on the fault handling logic and fault handling rules. According to the above fault handling rules and power grid fault handling process, the extracted fault knowledge is constructed into a basic structure of "concept, entity, rule and process" and stored in the graph database to build a power grid fault handling navigation mechanism model.

3. The panoramic intelligent navigation system for power grid fault handling according to claim 2 is characterized by: The text preprocessing includes text cleaning, text standardization, text segmentation and removal of stop words.

4. The panoramic intelligent navigation system for power grid fault handling according to claim 1 is characterized by: According to the grid equipment information, grid model information, operation information, stability regulations, accident plans, and operation control applications, the fault handling entity information is extracted and used as the input of the fault handling navigation mechanism model, and the fault handling low-dimensional vector is generated by the fault handling navigation mechanism model; The low-dimensional vector of fault handling is used as the input of the Graph Transformer framework combined with the self-attention mechanism. The Graph Transformer framework is used to analyze the power grid fault events, power grid operation mode and power grid flow information, and the panoramic information reasoning process of fault handling is carried out to obtain the optimal power grid fault handling measures. Based on the trained key knowledge extraction model for power grid fault handling, the fault entities and relationships obtained by knowledge extraction of power grid fault events, power grid operation modes, power grid flow information, power grid equipment information, power grid model information, operation information, stability regulations, accident plans and operation control applications are tripled according to the basic knowledge structure in the form of "concept, entity, rule, process", and multiple types of end-to-end triples are obtained, 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". The multiple types of end-to-end triples are used as the input of the graph attention reasoning model. The power grid dispatching and control system obtains the fault handling information, risk information, risk control strategy and key factors for power restoration decision-making required for the power grid fault handling process through the graph attention reasoning model, and generates the power restoration result in combination with the on-site station and line patrol conditions.

5. The panoramic intelligent navigation system for power grid fault handling according to claim 1 is characterized by: The specific process of aligning the grid equipment information, grid model information, operation information, stability procedures, accident plans, and operation control applications using the VR-GCN model with a high-speed gate mechanism includes: The fault entities contained in the power grid equipment information, power grid model information, operation information, stability procedures, accident plans and operation control applications are represented as nodes, and the relationships are represented as edges to construct the corresponding graph structure data; vectorized relational convolution operations are performed on each node to aggregate the features of neighboring nodes and update the node representation; a high-speed gating mechanism is introduced to set a gating value for each node or edge according to the structural characteristics, and dynamically determine which nodes or edges are filtered during the message propagation process according to the gating value; VR-GCN model training and inference are performed, and the similarity between similar nodes in different graphs is calculated through the node representation output by the VR-GCN model, and the node pairs are determined to be aligned according to the set threshold. The greedy matching method is used to determine the alignment results of various fault entities; Based on the alignment relationship between the knowledge points covered by power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, as well as the fault entities, redundant entities are automatically identified and eliminated, and multiple referential items of the same entity object are set to a unified entity name and integrated into information related to power grid fault handling.

6. The panoramic intelligent navigation system for power grid fault handling according to claim 1 is characterized by: The implementation process of the fault handling panoramic information reasoning includes: Use power grid fault information to perform fault semantic search and obtain related fault knowledge; build a fault joint reasoning space; use the Graph Transformer framework combined with the self-attention mechanism to perform fault reasoning; The specific process of fault reasoning using the Graph Transformer framework combined with the self-attention mechanism includes: Obtain entity embedding n for each node v and its neighbor node u based on node knowledge v , n u ； Incorporate the conditional probability feature p between nodes v and u uv , the calculation formula is as follows: Where n is the number of occurrences of the corresponding node event counted in the historical fault corpus, and p uv is the conditional probability of v and u, which is used to calculate the probability that the fault corresponding to node v also occurs when the fault corresponding to neighboring node u occurs, thereby measuring the size of the correlation between different fault handling measures and fault events; By p uv Calculate the relationship embedding r between the two uv , the calculation formula is as follows: r uv =f r (n u ,n v ,e uv ,p uv ) (2) In the formula, f r is a multi-layer perceptron network; uv is a unique hot vector reflecting the relationship between node u and node v; Calculate the attention weight of neighbor nodes to their own nodes, and use the attention mechanism to aggregate information: For each node, its entity embedding is mapped to the query vector Q, key vector K and value vector V. Taking node v as an example, the calculation formula is as follows: Q v =W Q n v ,K v =W K n v ,V v =W V n v (3) Where W Q , W K , W V is the learnable parameter matrix, W Q is the parameter matrix corresponding to the query vector, W K is the parameter matrix corresponding to the key vector, W V is the parameter matrix corresponding to the value vector, Q v represents the query vector of node v, K v represents the key vector of node v, V v Represents the value vector of node v, n v is the embedding vector representation of any node v; Calculate the attention weight α of each neighbor node to its own node uv , the calculation formula is as follows: Where, d k The dimension of the key vector obtained by mapping the entity embedding of node v, Q u is the query vector of neighbor node u, is the transposed vector of the key vector of node v. The softmax function converts the input vector into a probability distribution so that the value of each element is between 0 and 1, and the sum of all elements is 1. The attention weight value is used to perform weighted averaging on the value vectors of all neighbor nodes to achieve information aggregation of neighbor nodes. The calculation formula is as follows: Where N(v) is the number of neighbor nodes of node v, l represents the lth layer of the Graph Transformer framework, is the embedding vector representation of any node v in the lth layer, V u is the value vector of neighbor node u; The output nodes of each layer are input into the two-layer feedforward neural network, and the ReLU function is used for nonlinear activation. The calculation formula is as follows: Where W1 is the weight parameter matrix of the first layer of feedforward neural network, W2 is the weight parameter matrix of the second layer of feedforward neural network, It represents the node feature representation after being processed by two layers of feedforward neural network; At the end of the model, the representation of each node is updated through layer normalization operations and residual connections. The calculation formula is as shown in formula (7). The message is passed between nodes through the attention mechanism until convergence is reached: In the formula, LayerNorm is the layer normalization operation, is the embedding vector representation of node v after the forward propagation update of the Graph Transformer framework; The breadth-first traversal algorithm is used to track the high attention weight transmission path from the fault information node W to the fault phenomenon node, then to the fault type node, and all possible next-step fault handling measures nodes. After the message transmission is completed, the final representation of each node is returned, and the weighted summation is performed through the multi-layer perceptron layer to obtain each possible next-step fault handling measure and their respective probabilities, and the optimal fault handling measure is inferred.

7. The panoramic intelligent navigation system for power grid fault handling according to claim 4 is characterized by: The graph attention reasoning model is a graph attention reasoning model based on the ConvKB embedding model and combined with node neighborhood information. The specific reasoning process of obtaining the fault handling information, risk information, risk control strategy, and key factors of power restoration decision required by the power grid fault handling process through the graph attention reasoning model includes: The random initialization method is used to vectorize the fault entities and relations based on the triple structure, and the initialization embedding vector, i.e., the low-dimensional vector, is obtained. The known triples are recorded as: T = {(h, r, t) | h, t∈E, r∈R}, where h and t are the head entity vector and the tail entity vector respectively; r is the relation vector; Update the representation vector of the central node and the head entity vector and tail entity vector of the known fault triplet through the neighborhood-level attention method; Taking the fault information node Z as the central node z, the direct neighborhood triples around it can be expressed as: Where z is the representation vector corresponding to the head entity, r k is the representation vector corresponding to the relationship, e j is the representation vector corresponding to the tail entity; [] represents the vector concatenation operation; W1 is the weight parameter matrix for the linear transformation of the triplet; The LeakyReLU function is used for nonlinear transformation to obtain the attention value of the triplet: In the formula, s zjk is the attention value of the triple; W2 is the linear transformation matrix parameterized by weight; is the direct neighborhood triplet around the central node z; Use the Softmax function to normalize all attention values ​​to obtain the relative attention value β zjk : In the formula, E z is the set of tail entities in all direct neighborhood triples around the central entity Z; R z is the set of relations in all direct neighborhood triples around the central entity Z; exp(s zjk ) Get the value of the triplet’s attention value after taking the natural exponential function; s inr It refers to the attention value of the ith direct neighborhood triplet around the fault information node Z, i represents the head entity of the triplet, n represents the tail entity of the triplet, and r represents the relationship between the triples; Combine equations (8) and (10) to aggregate the fault information of the immediate neighborhood around the central entity and update the representation vector of z: In the formula, is the set of all direct neighborhood triplets around the central node z; β zjk is the relative attention value of the triple; σ is the sigmoid activation function of the graph attention network; Att(z) is the updated central node representation vector, which is still recorded as z in subsequent calculation steps; Similarly, the embedding vectors of the head entity and the tail entity in the known fault triple are also updated according to the above method; Read the training data set and generate negative samples through the negative sampling strategy. Use uniform sampling or Bernoulli sampling to randomly replace the head entity or tail entity or relationship in the positive sample triple with other entities or relationships. The set of negative example triples is defined as follows: In the formula, h′ is the head entity after random replacement, r′ is the relation after random replacement, t′ is the tail entity after random replacement; T′ is the set of negative triples; The convolution feature vector is converted into a scoring function through the fully connected layer of the ConvKB embedding model to predict the effectiveness of known triples, which is defined as follows: Where s is the score; M is the number of convolution layers; * represents the convolution operation; ε n represents the nth layer of convolution; W is the linear transformation matrix used to convert and obtain the final score; Based on the ConvKB embedding model framework, in order to compare positive and negative samples to update model parameters and obtain more comprehensive fault semantic information, the loss function is constructed by minimizing the logarithmic loss to optimize the model, which is defined as follows: Where L is the model function value; T∪T′ is the set of positive triples and negative triples; ω (h,r,t) is the sample label. If the triple is a positive example, that is, (h, r, t)∈T, then ω (h,r,t) =1; if the triple is a negative example, that is, (h, r, t)∈T′, then ω (h,r,t) =-1; α is the regularization coefficient, which is used to control the contribution of the regularization term to the loss value; The stochastic gradient descent optimization method is used to update the fault entity vector and relationship vector to minimize the loss function; the above steps are repeated until the model convergence condition is reached and the algorithm stops; the information corresponding to the final node representation output by the graph attention reasoning model is the fault handling information; the "plan information-relationship-risk information", "risk information-relationship-risk control strategy", and "risk control strategy-relationship-key factors for power restoration decision-making" triples are used as the input of the graph attention reasoning model in turn, and the above steps are repeated to continue to infer the risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process.

8. The panoramic intelligent navigation system for power grid fault handling according to claim 1 is characterized by: The power grid flow information refers to the data of electrical characteristics of each node and branch and their mutual relationship obtained through flow calculation during the operation of the power system, including current, voltage, power, frequency, flow direction and network topology.

9. A panoramic intelligent navigation method for power grid fault handling, characterized in that: It includes the following steps: The acquired power grid fault handling data is subjected to text preprocessing, and fault handling entities, relationships and events are identified from the text preprocessed power grid fault handling data based on a deep hybrid neural network, and the power grid equipment information, power grid model information, operation information, stability procedures, accident plans, and operation control applications in the power grid fault handling data are integrated and aligned, thereby unifying the same type of entities in the power grid fault handling data, and then establishing a fault handling navigation mechanism model; In the power grid dispatching and control system, the fault handling entity information is extracted based on the power grid equipment information, power grid model information, operation information, stability regulations, accident plans, and operation control applications, and the fault handling entity information is transmitted to the fault handling navigation mechanism model to generate a fault handling low-dimensional vector. By analyzing the power grid fault events, power grid operation mode and power grid flow information, and combining the fault handling low-dimensional vector to perform the fault handling panoramic information reasoning process, the optimal power grid fault handling measures are obtained; Based on the fault handling navigation mechanism model and the fault handling panoramic information reasoning process, the fault handling information, risk information, risk control strategy, and key factors for power restoration decision-making required for the power grid fault handling process are obtained through the graph attention reasoning model, and the power restoration results are generated in combination with the on-site station and line patrol conditions.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in claim 9 are implemented.

Images