A Power Grid Section Quota Reasoning and Early Warning Method Based on Knowledge Graph Distributed Representation

Abstract

This invention relates to a method for power grid cross-section quota reasoning and early warning based on knowledge graph distributed representation, comprising the following steps: Step 1, inputting a dataset from an intelligent dispatch control system; Step 2, performing data preprocessing; Step 3, extracting knowledge from the cross-section quota information in Step 2; Step 4, linking three types of entities in the extracted knowledge according to prescribed event relationships to form sub-knowledge graph units, and establishing a cross-section quota knowledge graph by connecting the various sub-units; Step 5, based on the knowledge extraction model established in Step 3 and the knowledge graph generated in Step 4, reasoning out the stable quota value of a specific power grid cross-section through intelligent cross-section quota reasoning; Step 6, providing early warning of exceeding the limit based on the stable quota value of the power grid cross-section derived in Step 5. This invention can improve the efficiency of power grid steady-state cross-section monitoring and the level of intelligent early warning, ensuring the safe and stable operation of the power grid.

Description

Technical Field

[0001] This invention belongs to the field of power system technology and relates to a method for inferring and warning of grid section quotas, and more particularly to a method for inferring and warning of grid section quotas based on knowledge graph distributed representation. Background Technology

[0002] The structure of AC / DC hybrid power grids is becoming increasingly complex, with flexible and varied operating modes. The amount of information in power grid stability regulations is surging, making it difficult to cope with the increased volume of information by manually establishing cross-section safety margin assessment models. With the construction of new power systems dominated by new energy sources, the grid connection of a large number of new energy power plants has led to new time-varying characteristics in power sources, loads, and the grid structure. Limit monitoring models established based on offline, periodic stability regulations cannot meet the needs of multiple power grid operating modes. Current online analysis and calculation of transmission section risks largely rely on manual methods, resulting in complex calculation processes, difficulty in accurately controlling control errors, and the potential for errors and omissions in multi-section safety constraints. Therefore, an online, rapid, and electronic stability regulation cross-section limit model is of great significance for intelligent calculation and online monitoring of cross-section limits in power grids operating under multiple modes. Knowledge graphs, a semantic network describing the relationships between objective physical knowledge, have a knowledge organization closer to the form of human cognition of the world. They can represent complex knowledge relationships at the semantic level and are very suitable for characterizing the knowledge model structure and knowledge reasoning process of power grid stability limits.

[0003] However, the current cross-sectional safety margin assessment models established by manually analyzing stability procedures cannot meet the needs of frequent changes in procedures. Furthermore, online analysis and calculation of power grid cross-sectional risks rely heavily on manual labor, resulting in complex calculation processes, difficulty in accurately controlling control errors, and the potential for errors and omissions due to multi-section safety constraints.

[0004] A search revealed no publicly available literature of the same or similar prior art as this invention. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a power grid section limit reasoning and early warning method based on knowledge graph distributed representation. Based on knowledge graph technology, a stable limit model of the power grid section is constructed. On the basis of the established stable limit model, the section limit can be intelligently calculated and reasoned in combination with the real-time operation mode of the power grid, thereby improving the monitoring efficiency of the steady-state section of the power grid and the level of intelligent early warning, and ensuring the safe and stable operation of the power grid.

[0006] The present invention solves its practical problem by adopting the following technical solution:

[0007] A method for power grid section quota reasoning and early warning based on knowledge graph distributed representation includes the following steps:

[0008] Step 1: Input the intelligent dispatching and control system dataset;

[0009] Step 2: Perform data preprocessing on the intelligent dispatching and control system dataset input in Step 1;

[0010] Step 3: Using the information obtained in Step 2, extract knowledge from the cross-sectional quota information in Step 2;

[0011] Step 4: Using the cross-sectional quota knowledge extracted in Step 3, link the three types of entities in the extracted knowledge according to the prescribed event relationships to form sub-knowledge graph units. Connect the sub-units to establish the cross-sectional quota knowledge graph.

[0012] Step 5: Based on the knowledge extraction model established in Step 3 and the knowledge graph generated in Step 4, the stability limit value of a specific power grid section is inferred through intelligent reasoning of the section limit.

[0013] Step 6: Issue an over-limit warning based on the grid section stability limit value derived in Step 5.

[0014] Furthermore, the intelligent dispatch and control system dataset in step 1 includes: stability procedure text, power grid model equipment name, unique identifier, measurement and status;

[0015] Furthermore, the specific method for step 2 is as follows:

[0016] Based on regular expressions, a language recognition template for cross-section quota description under multiple operating modes in the stable procedure text is developed, and irrelevant content is removed.

[0017] Furthermore, the specific steps for knowledge extraction of the cross-sectional quota information in step 2 in step 3 include:

[0018] (1) Labeling the stability protocol language. Define entity categories in the stability protocol language, including three entities: grid section, grid operation mode, and stability quota. Define entity type identification labels for each entity category, including three entity types and one non-entity type, for a total of four label data representation formats. Specifically, the first character of the grid section is labeled "B-section", and the remaining characters of the grid section are labeled "I-section"; the first character of the grid operation mode is labeled "B-oper_m", and the remaining characters of the grid operation mode are labeled "I-oper_m"; the first character of the stability quota is labeled "B-quota", and the remaining characters of the stability quota are labeled "I-quota"; non-entity components are labeled "O".

[0019] (2) Identification of cross-section quota knowledge labels based on Transformer Bidirectional Encoder Representation Technology - Conditional Random Field (BERT-CRF). Based on the BERT pre-trained model, the stability procedure language is transformed into word vectors that can represent power industry terminology. The initial parameters of the BERT model are fine-tuned by incorporating the stability procedure language, improving its representational capability. CRF is used to represent the dependency relationship between character labels in stability procedure language entities and global character labels, thereby predicting the character labels of entities such as optimal grid cross-sections, operating modes, and stability quotas. The main parameters of the BERT model include: an embedding layer size of 128, a Transformer layer size of 12, a hidden layer size of 768, an attention head size of 12, and the AdamW optimizer. The parameters adjusted during the training of the fault handling plan are as follows: learning rate 10... -5 The batch size is 16, the number of iterations is 10, and the function loss is calculated using cross-entropy.

[0020] (3) The BERT model uses 12 multi-head attention mechanisms to calculate the degree of association between each character in the input text and all other characters, and uses the Softmax function to calculate the global representation vector of each character, as shown in the following expression:

[0021]

[0022] In the formula: q is a character query vector, obtained by inputting the initial encoding matrix X of the text. e With query vector matrix W q A linear transformation is performed to obtain k; each character key vector is obtained by inputting the initial encoding matrix X of the text. e With the key vector matrix W k A linear transformation is performed to obtain v; each character value vector is obtained by inputting the initial encoding matrix X of the text. e And value vector matrix W v The linear transformation yields d. k The dimension of the key vector;

[0023] (4) Calculate the inner product of the q-vector and k-vector of each character in the input text. The larger the calculated value, the greater the correlation. By scaling the correlation and applying Softmax normalization, the correlation between each character and other characters in the input text is obtained. Finally, this correlation is weighted and fused with the v-vectors of each character to calculate the global semantic vector of each character. Through multiple attention calculations, the results of each character vector in the input text are concatenated, and then processed by matrix W. 0 The vector is transformed to the same dimension as the initial input character vector, thus completing the dynamic semantic encoding of the initial character vector, as shown in the following expression:

[0024] MutiHead(Q,P,V)=Concat(head1,...,head h )·W 0 (2)

[0025] In the formula: W 0 For multi-head attention splicing matrix; head h For each semantic word vector computed through multi-attention;

[0026] (5) The CRF is used to calculate the relationship between stable procedure language entity labels and global character labels, predicting the optimal label for each entity character. The CRF calculates the entity character label score for the stable procedure language by calculating the relationship between the entity label score output by the BERT model and the transition score between adjacent character labels. The calculation formula is shown below:

[0027]

[0028] In the formula: It is the transition score between adjacent character tags of stable procedure language entities. It is the stable procedure language entity label score;

[0029] (6) The probability of character prediction in a stable procedure language entity is calculated by dividing the total score of correctly labeled characters in the stable procedure language by the total score of all possible labels. The expression is as follows:

[0030]

[0031] In the formula: Y X This represents the character annotations for all possible stable procedural language entities. When the value of p(y|X) is close to 1, it indicates that the predicted entity labels and annotation results are consistent, and the stable procedural language entity recognition performance is optimal at this point.

[0032] (7) CRF uses logarithmic probability to optimize the correct label sequence. For the training samples, the following formula applies:

[0033]

[0034] (8) The label sequence with the highest score is obtained through the maximum likelihood estimation function as the final predicted output sequence, and the formula is as follows:

[0035]

[0036] (9) Text similarity technology is used to align quota knowledge with grid section information and operating modes of the dispatch control system. Vector cosine values ​​are used to calculate the similarity distance between stability specification quota knowledge and standard grid sections, operating modes, and operating states. Entity alignment is performed when the similarity distance is greater than a threshold of 0.95 to achieve semantic alignment of quota knowledge. The similarity calculation of quota knowledge includes: grid section information in the stability specification language and grid section information in the dispatch control system; equipment names and operating states in the stability specification language and equipment names and operating states in the dispatch control system. The vector cosine value calculation expression is shown below:

[0037]

[0038] In the formula: x i This represents the entity vector in the contingency plan entity library; y i This represents the standard name vector for the power grid model. Furthermore, the specific steps for establishing the cross-sectional quota knowledge graph in step 4 include:

[0039] (1) Define entities linked by the event relationship <operation mode, grid section, section limit> to form a section limit sub-knowledge graph unit. The operation mode attribute includes: equipment composition, equipment name, and equipment operation status; the grid section attribute includes: equipment composition, equipment name, and section power flow; the section limit attribute includes: upper limit value and lower limit value.

[0040] (2) Link the M cross-section quota sub-knowledge graph units according to the multiple operation modes of the power grid <operation mode 1, operation mode 2, ..., operation mode M> to finally form a knowledge graph containing cross-section quota knowledge.

[0041] Furthermore, the specific steps in step 5 of reasoning out the stability limit value of a specific power grid section through intelligent reasoning of the section limit include:

[0042] (1) Identify the entity information of the operation mode by identifying the composition of the power grid equipment, the equipment name, and the equipment operation status information; identify the entity information of the power grid section by identifying the composition of the power grid equipment, the equipment name, and the cross-sectional power flow.

[0043] (2) Reconstruct the standard information of the identified power grid operation mode entities and power grid section entities, and convert it into Cypher query language;

[0044] (3) Perform high-speed query and matching of the operating mode entity and the power grid section entity in the distributed representation event triple of the cross-section quota knowledge graph <operating mode, power grid section, cross-section quota>;

[0045] (4) When the operating mode and the grid section are matched, the stability limit value of the specific grid section under this operating mode can be automatically derived.

[0046] Furthermore, the specific method for step 6 is as follows:

[0047] Based on the grid section stability limit value derived in step 5, the real-time power flow value of each section under the grid operation mode is compared with the inferred stability limit value. If the real-time power flow value of the section reaches more than 80% of the limit but less than 100%, an overload warning is issued. If the real-time power flow value of the section reaches 100% or more of the limit, an over-limit warning is issued. At the same time, the monitoring and warning information is sent to the intelligent dispatch and control system.

[0048] Advantages and beneficial effects of the present invention:

[0049] 1. This invention proposes a method for power grid section quota reasoning and early warning based on knowledge graph distributed representation. The section quota knowledge extraction module proposes a stability procedure language tagging method, defines four types of entity tags, and proposes to identify section quota knowledge tags based on BERT-CRF neural network. Text similarity technology is used to align quota knowledge with equipment names and unique identifiers, which improves the accuracy of quota knowledge recognition and helps to improve the accuracy of knowledge graph.

[0050] 2. The intelligent reasoning module for transmission section limits in this invention utilizes the identified standard information of power grid operation mode entities and power grid section entities, converting it into Cypher query language. Using a pre-constructed knowledge graph, it can quickly match power grid sections under a given operation mode and then obtain the stable limit value for the power grid section. This invention transforms relational languages ​​such as power grid operation modes and section information into graph language, improving the rapid reasoning of transmission section limits and solving the problem of untimely early warning for transmission sections under multiple operation modes in the current system.

[0051] 3. This invention first interacts with the intelligent dispatch and control system through a data preprocessing module. Then, it extracts knowledge and constructs a knowledge graph of the data section limit. Under a given power grid operation mode, it quickly matches the power grid section based on the constructed knowledge graph and infers the section limit. Finally, according to different limits, it sends relevant monitoring and early warning information to the intelligent dispatch and control system to ensure the stable and safe operation of the power grid.

[0052] 4. The difference between this invention and previous methods for power grid cross-section quota reasoning and early warning lies in its use of a power grid cross-section quota reasoning and early warning system based on a knowledge graph-based distributed representation. The key aspects of this invention are the cross-section quota knowledge extraction technology and the method for constructing the cross-section quota knowledge graph. This invention establishes a cross-section quota knowledge graph through a deep neural network, which can automatically convert the current stability regulations text into a computer-executable language, solving the problems of low efficiency and error-proneness of manual data entry.

[0053] 5. The model used for knowledge extraction in this invention is different from the previous one. This invention uses a BERT-CRF neural network, which improves the accuracy of entity recognition and relation extraction, and improves the quality of the cross-sectional quota knowledge graph, so as to better provide basic services for cross-sectional quota reasoning.

[0054] 6. The method for constructing the cross-sectional quota knowledge graph in this invention differs from previous methods. It proposes using <operation mode, grid cross-section, cross-sectional quota> as a linking relationship, serving as sub-knowledge graph units. By connecting multiple sub-units, a grid cross-sectional knowledge graph is constructed. This invention integrates stable quota sub-knowledge graph units to form a knowledge graph, establishing a rapid construction, updating, and decoupling mechanism for cross-sectional quota knowledge. This solves the problem of chaotic and difficult-to-update local stable quotas across multiple operation modes. Attached Figure Description

[0055] Figure 1 This is a diagram illustrating the structure of the power grid section quota reasoning and early warning system based on knowledge graph distributed representation of the present invention. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to the accompanying drawings:

[0057] A method for power grid section quota reasoning and early warning based on knowledge graph distributed representation includes the following steps:

[0058] Step 1: Input the intelligent dispatching and control system dataset;

[0059] The intelligent dispatch and control system dataset in step 1 includes: stability procedure text, power grid model equipment name, unique identifier, measurement and status;

[0060] Step 2: Perform data preprocessing on the intelligent dispatching and control system dataset input in Step 1;

[0061] The specific method for step 2 is as follows:

[0062] Based on regular expressions, a language recognition template for cross-section quota description under multiple operating modes in the stable procedure text is developed, and irrelevant content is removed.

[0063] In this embodiment, the description language includes: power grid section information, operating mode information, and stability limit information. Information such as the name, unique identifier, measurement data, operating status, and section composition of the power grid model equipment are read from the intelligent dispatch and control system dataset.

[0064] Step 3: Using the information obtained in Step 2, extract knowledge from the cross-section quota information in Step 2. Identify the cross-section quota knowledge in the stability procedure text based on deep neural network, and use text similarity technology to align the quota knowledge with the grid cross-section, operation mode and other information of the dispatch control system.

[0065] The specific steps for knowledge extraction of the cross-sectional quota information in step 2 in step 3 include:

[0066] (1) Labeling the stability protocol language. Define entity categories in the stability protocol language, including three entities: grid section, grid operation mode, and stability quota. Define entity type identification labels for each entity category, including three entity types and one non-entity type, for a total of four label data representation formats. Specifically, the first character of the grid section is labeled "B-section", and the remaining characters of the grid section are labeled "I-section"; the first character of the grid operation mode is labeled "B-oper_m", and the remaining characters of the grid operation mode are labeled "I-oper_m"; the first character of the stability quota is labeled "B-quota", and the remaining characters of the stability quota are labeled "I-quota"; non-entity components are labeled "O".

[0067] (2) Identification of cross-section quota knowledge labels based on Transformer Bidirectional Encoder Representation Technology - Conditional Random Field (BERT-CRF). Based on the BERT pre-trained model, the stability procedure language is transformed into word vectors that can represent power industry terminology. The initial parameters of the BERT model are fine-tuned by incorporating the stability procedure language, improving its representational capability. CRF is used to represent the dependency relationship between character labels in stability procedure language entities and global character labels, thereby predicting the character labels of entities such as optimal grid cross-sections, operating modes, and stability quotas. The main parameters of the BERT model include: an embedding layer size of 128, a Transformer layer size of 12, a hidden layer size of 768, an attention head size of 12, and the AdamW optimizer. The parameters adjusted during the training of the fault handling plan are as follows: learning rate 10... -5 The batch size is 16, the number of iterations is 10, and the function loss is calculated using cross-entropy.

[0068] (3) The BERT model uses 12 multi-head attention mechanisms to calculate the degree of association between each character in the input text and all other characters, and uses the Softmax function to calculate the global representation vector of each character, as shown in the following expression:

[0069]

[0070] In the formula: q is a character query vector, obtained by inputting the initial encoding matrix X of the text. e With query vector matrix Wq A linear transformation is performed to obtain k; each character key vector is obtained by inputting the initial encoding matrix X of the text. e With the key vector matrix W k A linear transformation is performed to obtain v; each character value vector is obtained by inputting the initial encoding matrix X of the text. e And value vector matrix W v The linear transformation yields d. k The dimension of the key vector;

[0071] (4) Calculate the inner product of the q-vector and k-vector of each character in the input text. The larger the calculated value, the greater the correlation. By scaling the correlation and applying Softmax normalization, the correlation between each character and other characters in the input text is obtained. Finally, this correlation is weighted and fused with the v-vectors of each character to calculate the global semantic vector of each character. Through multiple attention calculations, the results of each character vector in the input text are concatenated, and then processed by matrix W. 0 The vector is transformed to the same dimension as the initial input character vector, thus completing the dynamic semantic encoding of the initial character vector, as shown in the following expression:

[0072] MutiHead(Q,P,V)=Concat(head1,...,head h )·W 0 (2)

[0073] In the formula: W 0 For multi-head attention splicing matrix; head h For each semantic word vector computed through multi-attention;

[0074] (5) The CRF is used to calculate the relationship between stable procedure language entity labels and global character labels, predicting the optimal label for each entity character. The CRF calculates the entity character label score for the stable procedure language by calculating the relationship between the entity label score output by the BERT model and the transition score between adjacent character labels. The calculation formula is shown below:

[0075]

[0076] In the formula: It is the transition score between adjacent character tags of stable procedure language entities. It is the stable procedure language entity label score;

[0077] (6) The probability of character prediction in a stable procedure language entity is calculated by dividing the total score of correctly labeled characters in the stable procedure language by the total score of all possible labels. The expression is as follows:

[0078]

[0079] In the formula: Y X This represents the character annotations for all possible stable procedural language entities. When the value of p(y|X) is close to 1, it indicates that the predicted entity labels and annotation results are consistent, and the stable procedural language entity recognition performance is optimal at this point.

[0080] (7) CRF uses logarithmic probability to optimize the correct label sequence. For the training samples, the following formula applies:

[0081]

[0082] (8) The label sequence with the highest score is obtained through the maximum likelihood estimation function as the final predicted output sequence, and the formula is as follows:

[0083]

[0084] (9) Text similarity technology is used to align quota knowledge with grid section information and operating modes of the dispatch control system. Vector cosine values ​​are used to calculate the similarity distance between stability specification quota knowledge and standard grid sections, operating modes, and operating states. Entity alignment is performed when the similarity distance is greater than a threshold of 0.95 to achieve semantic alignment of quota knowledge. The similarity calculation of quota knowledge includes: grid section information in the stability specification language and grid section information in the dispatch control system; equipment names and operating states in the stability specification language and equipment names and operating states in the dispatch control system. The vector cosine value calculation expression is shown below:

[0085]

[0086] In the formula: x i This represents the entity vector in the contingency plan entity library; y i This represents the standard name vector for the power grid model.

[0087] Step 4: Using the cross-sectional quota knowledge extracted in Step 3, link the three types of entities in the extracted knowledge according to the prescribed event relationships to form sub-knowledge graph units. After connecting the sub-units, establish the cross-sectional quota knowledge graph and store it in the form of an attribute graph data model.

[0088] The specific steps for establishing the cross-sectional quota knowledge graph in step 4 include:

[0089] (1) Define entities linked by the event relationship <operation mode, grid section, section limit> to form a section limit sub-knowledge graph unit. The operation mode attribute includes: equipment composition, equipment name, and equipment operation status; the grid section attribute includes: equipment composition, equipment name, and section power flow; the section limit attribute includes: upper limit value and lower limit value.

[0090] (2) Link the M cross-section quota sub-knowledge graph units according to the multiple operation modes of the power grid <operation mode 1, operation mode 2, ..., operation mode M> to finally form a knowledge graph containing cross-section quota knowledge.

[0091] Step 5: Based on the knowledge extraction model established in Step 3 and the knowledge graph generated in Step 4, identify the entity information of the power grid operation mode and the entity information of the power grid section generated by the power grid regulation and operation. Then, convert the identified standard information of the power grid operation mode entity and the power grid section entity into Cypher query language, perform high-speed query and matching on the operation mode entity and the power grid section entity, and infer the stability limit value of the specific power grid section through intelligent reasoning of the section limit.

[0092] The specific steps in step 5, which involve using intelligent reasoning to deduce the stability limit value for a specific power grid section, include:

[0093] (1) Identify the entity information of the operation mode by identifying the composition of the power grid equipment, the equipment name, and the equipment operation status information; identify the entity information of the power grid section by identifying the composition of the power grid equipment, the equipment name, and the cross-sectional power flow.

[0094] (2) Reconstruct the standard information of the identified power grid operation mode entities and power grid section entities, and convert it into Cypher query language;

[0095] (3) Perform high-speed query and matching of the operating mode entity and the power grid section entity in the distributed representation event triple of the cross-section quota knowledge graph <operating mode, power grid section, cross-section quota>;

[0096] (4) When the operating mode and the grid section are matched, the stability limit value of the specific grid section under this operating mode can be automatically derived.

[0097] Step 6: Based on the grid section stability limit value derived in Step 5, compare the real-time power flow value of each section under the grid operation mode with the inferred stability limit value. If the real-time power flow value of the section reaches more than 80% of the limit but less than 100%, an overload warning is issued. If the real-time power flow value of the section reaches 100% or more of the limit, an over-limit warning is issued. At the same time, the monitoring and warning information is sent to the intelligent dispatch and control system.

[0098] In the task of grid section limit reasoning and early warning, most approaches rely on physical formulas and expert strategies, with little research at the knowledge graph level. This invention proposes a grid section limit reasoning and early warning method based on distributed representation of knowledge graphs. It employs deep neural networks to identify section limit knowledge within the stability protocol language, links this knowledge according to event relationships to form a knowledge graph, and, by sensing grid operation modes and power flow changes at the sections, infers section stability limits under multiple operation modes based on the distributed representation of the knowledge graph. By comparing section stability limits and power flow, the degree of early warning is calculated, improving the efficiency of grid steady-state section monitoring and the intelligence level of early warning, thus ensuring the safe and stable operation of the grid.

[0099] The working principle of this invention is:

[0100] The purpose of this invention is to address the problems that rely on manual establishment of cross-section safety margin assessment models, which are difficult to cope with the surge in the amount of information in regulations; that rely on offline periodic stability regulations to establish quota monitoring models, which cannot meet the needs of multiple power grid operation modes; and that the current online analysis and calculation of transmission section risks rely heavily on manual labor, resulting in complex calculation processes, difficulty in accurately grasping control errors, and the tendency for errors and omissions to occur due to multi-section safety constraints.

[0101] This invention proposes a method for grid section quota reasoning and early warning based on knowledge graph distributed representation. It employs a deep neural network to identify section quota knowledge within the stability protocol language, links this knowledge according to event relationships to form a knowledge graph, and infers section stability quotas under multiple operating modes based on the distributed representation of the knowledge graph by sensing grid operation modes and section power flow changes. The degree of early warning for each section is calculated by comparing the section stability quotas and the section power flow, thereby improving the efficiency of grid steady-state section monitoring and the level of intelligent early warning, and ensuring the safe and stable operation of the grid.

[0102] This invention provides a power grid section quota limit reasoning and early warning system based on knowledge graph distributed representation, such as... Figure 1As shown, it is divided into five modules, including: data preprocessing, cross-section quota knowledge extraction, cross-section quota knowledge graph establishment, cross-section quota intelligent reasoning, and cross-section quota monitoring and early warning. The data preprocessing module interacts with the intelligent dispatch control system, obtaining data from it including: stability procedure text, power grid model equipment names, unique identifiers, measurements, and status. Data sent to the intelligent dispatch control system includes: cross-section monitoring and early warning information. The cross-section quota knowledge extraction module uses a deep neural network to identify cross-section quota knowledge within the stability procedure text, aligning the quota knowledge with equipment names and unique identifiers using text similarity technology. The cross-section quota knowledge graph establishment module links the extracted cross-section quota knowledge according to the <operating mode, cross-section, cross-section quota> event relationship to form a cross-section quota knowledge graph. The cross-section quota intelligent reasoning module senses the power grid operating mode and the text language of the operating equipment in real time, matching the operating mode entity and the power grid cross-section entity in the knowledge graph based on distributed representation, and infers the cross-section stability quota through the defined <operating mode, cross-section, cross-section quota> event triples. The cross-section limit monitoring and early warning module compares the real-time operating value of each cross section with the inferred stable limit value, and provides early warning information based on different degrees of exceeding the limit. The monitoring and early warning information is then sent to the intelligent dispatch and control system.

[0103] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A method for power grid section quota reasoning and early warning based on knowledge graph distributed representation, characterized in that: Includes the following steps: Step 1: Input the intelligent dispatching and control system dataset; Step 2: Perform data preprocessing on the intelligent dispatching and control system dataset input in Step 1; Step 3: Using the information obtained in Step 2, extract knowledge from the cross-sectional quota information in Step 2; Step 4: Using the cross-sectional quota knowledge extracted in Step 3, link the three types of entities in the extracted knowledge according to the prescribed event relationships to form sub-knowledge graph units. Connect the sub-units to establish the cross-sectional quota knowledge graph. Step 5: Based on the knowledge extraction model established in Step 3 and the knowledge graph generated in Step 4, the stability limit value of a specific power grid section is inferred through intelligent reasoning of the section limit. Step 6: Issue an over-limit warning based on the grid section stability limit value derived in Step 5; The specific steps for establishing the cross-sectional quota knowledge graph in step 4 include: (1) Define entities linked by the event relationship <operation mode, grid section, section limit> to form a section limit sub-knowledge graph unit; where the operation mode attribute includes: equipment composition, equipment name, equipment operation status; the grid section attribute includes: equipment composition, equipment name, section power flow; the section limit attribute includes: upper limit value, lower limit value; (2) Link the M cross-section quota sub-knowledge graph units according to the multiple operation modes of the power grid <operation mode 1, operation mode 2, ..., operation mode M> to finally form a knowledge graph containing cross-section quota knowledge; The specific steps in step 5, which involve using intelligent reasoning to deduce the stability limit value for a specific power grid section, include: (1) Identify the entity information of the operation mode by identifying the composition of the power grid equipment, the equipment name, and the equipment operation status information; identify the entity information of the power grid section by identifying the composition of the power grid equipment, the equipment name, and the cross-sectional power flow. (2) Reconstruct the standard information of the identified power grid operation mode entities and power grid section entities, and convert it into Cypher query language; (3) Perform high-speed query and matching of the operating mode entity and the power grid section entity in the distributed representation event triple of the cross-section quota knowledge graph <operating mode, power grid section, cross-section quota>; (4) When the operating mode and the grid section are matched, the stability limit value of the specific grid section under this operating mode can be automatically derived.

2. The method for power grid section quota reasoning and early warning based on knowledge graph distributed representation according to claim 1, characterized in that: The intelligent dispatch and control system dataset in step 1 includes: stability procedure text, power grid model equipment name, unique identifier, measurement and status.

3. The method for power grid section quota reasoning and early warning based on knowledge graph distributed representation according to claim 1, characterized in that: The specific method of step 2 is as follows: based on regular expressions, write a cross-section quota description language recognition template for multiple operating modes in the stable procedure text, and remove irrelevant content.

4. The method for power grid section quota reasoning and early warning based on knowledge graph distributed representation according to claim 1, characterized in that: The specific steps for knowledge extraction of the cross-sectional quota information in step 2 in step 3 include: (1) Mark the stability procedure language; (2) Based on the bidirectional encoder characterization technology of converter - conditional random field identification of cross-sectional quota knowledge tags; (3) The BERT model calculates the degree of association between each character in the input text and all characters through 12 multi-head attention mechanisms, and uses the Softmax function to calculate the global representation vector of each character, as shown in the following expression: In the formula: q is a character query vector, obtained by inputting the initial encoding matrix X of the text. e With query vector matrix W q A linear transformation is performed to obtain k; each character key vector is obtained by inputting the initial encoding matrix X of the text. e With the key vector matrix W k A linear transformation is performed to obtain v; each character value vector is obtained by inputting the initial encoding matrix X of the text. e And value vector matrix W v The linear transformation yields d. k The dimension of the key vector; (4) Calculate the inner product of the q vector and k vector of each character in the input text. The larger the calculated value, the greater the correlation. By scaling the correlation and normalizing with Softmax, the correlation between each character in the input text and other characters is obtained. Finally, the correlation is weighted and fused with the v vector of each character to calculate the semantic word vector of each character globally. Through multiple attention calculations, the results of each word vector in the input text are concatenated, and then the matrix W is used to calculate the semantic word vector of each character globally. 0 The vector is transformed to the same dimension as the initial input character vector, thus completing the dynamic semantic encoding of the initial character vector, as shown in the following expression: MutiHead(Q,P,V)=Concat(head1,...,head h )·W 0 (2) In the formula: W 0 For multi-head attention splicing matrix; head h For each semantic word vector computed through multi-attention; (5) The relationship between stable procedure language entity labels and global character labels is calculated using CRF to predict the optimal label for the entity character. CRF calculates the entity character label score of stable procedure language by calculating the relationship between the entity label score output by the BERT model and the transition score between adjacent character labels. The calculation formula is as follows: In the formula: It is the transition score between adjacent character tags of stable procedure language entities. It is the stable procedure language entity label score; (6) The probability of character prediction in a stable procedure language entity is calculated by dividing the total score of correctly labeled characters in the stable procedure language by the total score of all possible labels. The expression is as follows: In the formula: Y X It represents all possible stable procedural language entity character annotations; when the value of p(y|X) is close to 1, it indicates that the predicted entity label and the annotation result are consistent, and the stable procedural language entity recognition effect is the best at this time; (7) CRF uses logarithmic probability to optimize the correct label sequence. For the training samples, the following formula applies: (8) The label sequence with the highest score is obtained through the maximum likelihood estimation function as the final predicted output sequence, and the formula is as follows: (9) Text similarity technology is used to align the quota knowledge with the grid section information and operation mode of the dispatch control system; the similarity distance between the stability procedure quota knowledge and the standard grid section, operation mode and operation status is calculated using vector cosine value. If the similarity distance is greater than the threshold, entity alignment is performed to achieve semantic alignment of quota knowledge. The vector cosine value calculation expression is as follows: In the formula: x i y represents the entity vector in the contingency plan entity library; i This represents the standard name vector for the power grid model.

5. The method for power grid section quota reasoning and early warning based on knowledge graph distributed representation according to claim 1, characterized in that: The specific method for step 6 is as follows: Based on the grid section stability limit value derived in step 5, the real-time power flow value of each section under the grid operation mode is compared with the inferred stability limit value. If the real-time power flow value of the section reaches more than 80% of the limit but less than 100%, an overload warning is issued. If the real-time power flow value of the section reaches 100% or more of the limit, an over-limit warning is issued. At the same time, the monitoring and warning information is sent to the intelligent dispatch and control system.

Images