Electric power system self-asking and self-answering online monitoring method and system based on large language model

Abstract

The invention discloses an electric power system self-asking and self-answering online monitoring method and system based on a large language model, and the technical key points are as follows: constructing a knowledge base of an electric power communication large model, forming a physical topology model, constructing a general large model, constructing a fine-tuning large model, and constructing a self-asking and self-answering online monitoring system based on the fine-tuning large model. The retrieval enhancement generation technology is adopted to construct the power system self-asking and self-answering model of the heterogeneous graph neural network based on word segmentation, and monitoring of operation data of the power system is more timely and accurate. Meanwhile, a prediction model and a fine-tuning prediction model are adopted to realize safer online monitoring of the power system.

Description

Technical Field

[0001] The present invention relates to a new self-questioning and self-answering power monitoring method and system, and in particular to a self-questioning and self-answering online monitoring method and system based on fine-tuning + retrieval enhancement to generate a hybrid large model. Background Art

[0002] Existing large-scale models only address customer-initiated questions and answers, resolving specialized knowledge issues. They fail to consider the integration of power systems and language models. Power system troubleshooting and maintenance are essentially a process where humans ask questions based on existing power network conditions and then seek solutions. Given the nature of this process, it's tempting to consider large-scale language models, which follow the same question-and-answer model, to investigate whether they could enable the power system to self-question and answer questions based on existing detection data technology, thereby achieving self-inspection. Summary of the Invention

[0003] 1. Technical core of the present invention First, it is reflected in the improvement of the large model, adopting a hybrid mode of fine-tuning + retrieval enhancement generation, which is reflected in the complementary advantages of the two solutions; Secondly, it is reflected in the architecture of the power communication large model knowledge base data base required for large model improvement, realizing the dual topology network structure of equipment and data; Third, we use a dual-topology network structure to construct a heterogeneous graph neural network for word segmentation. We also use node aggregation to perform similarity analysis on question sentences, realizing a question-answering model based on the retrieval enhancement generation of the heterogeneous graph neural network for word segmentation and improving the large language model. Fourth, based on the real-time monitoring data of the power system, the system continuously asks itself questions and obtains the real-time status and prediction of the dual topology network structure of the power system based on the above question-answering model.

[0004] 2. One of the main technical solutions of the present invention: method To implement the above core technology, the present invention provides a self-questioning and self-answering online monitoring method for a power system based on a large language model, comprising the following steps: S1 builds a data base for the power communication large model knowledge base and forms a physical topology model, which is used to realize a dual topology network structure of device entities and data on the physical topology model through dynamic operation data drive. The physical topology model includes a data model of devices and optical cables in each area (i.e., a model that includes dual existence of entities of devices and optical cables and data assigned to the entities), a physical topology data model of device data and optical cable data in multiple areas (i.e., the physical topology of the dual existence model), and forms a graph model based on the physical topology model. Then, the data is stored as graph data based on the graph model, thereby realizing graph data with the dual topology network structure. S2 builds a fine-tuned large model based on the power communication large model knowledge base data base of S1 and the general large model; S3 uses the fine-tuned large model and adopts retrieval-enhanced generation technology to build a power system self-questioning and self-answering model based on word segmentation and heterogeneous graph neural network; S4 forms questions based on real-time data in the graph data having the dual topology network structure at a specified frequency, and inputs the questions into the power system self-questioning and self-answering model constructed in S3 to obtain answers; S5 evaluates and predicts the current operating status of the power system based on the answers.

[0005] 3. Detailed description of one of the main technical solutions Specifically, the method for constructing the data base of the S1 power communication large model knowledge base includes: S1-1 Build a content system for the power backbone communication knowledge base, including static data represented by management data, standards and specifications, resource data, and dynamic data represented by equipment operation data. At the same time, clean, integrate, convert, and evaluate data quality for static and dynamic data to achieve standardization of multi-source heterogeneous data. S1-2 performs entity mapping on the data standardized in S1-1 according to the equipment list and optical cable list of each region within the geographical scope of power service, and obtains the equipment data model and optical cable data model; S1-3 maps the device data model and the optical cable data model to regional connection relationships based on the connection relationships between multiple regions within the power service geographical scope to form a physical topology data model, and then models the device data model, the optical cable data model, and the physical topology data model into a graph model based on the geographical coordinates within the power service geographical scope; S1-4 extracts real-time dynamic operation data from the standardized data of S1-1, builds relational data storage, and implements data-driven storage in the graph model in S1-3 to form graph data, so as to finally complete the construction of the data base of the power communication large model knowledge base.

[0006] The specific steps of S2 include: obtaining a general large model, using the power communication large model knowledge base data base built in S1 to fine-tune the general large model, and forming a power communication large model.

[0007] Optionally, fine-tuning methods include: S2-1 extracts static data from the data base of the power communication large model knowledge base; S2-2 converts static data into text and performs text segmentation to form text blocks; S2-3 divides the text blocks into a training set and a validation set, both of which contain question input and question output text blocks. The text blocks in the training and validation sets are labeled with their relevance to the power system field. The general large language model is supervised for training. The matching degree between the question input and question output is determined based on the output of the general large language model, and the attention mechanism parameters are fine-tuned and optimized. S2-4 continues to train the general large language model according to S2-3 until the verification set is used to verify that the matching degree reaches the highest level. Training is stopped and fine-tuning is completed.

[0008] Step S3 specifically includes: S3-1 loads documents from the power communication large model knowledge base data base, converts them into text, and performs word segmentation processing; S3-2 builds a heterogeneous graph neural network for word segmentation based on a dual topology network structure, performs node-level aggregation, and obtains the importance distribution between each word and its neighbors at each level; S3-3 performs text segmentation according to word meaning, generates text blocks with length, uses application embedding model to vectorize the text blocks, and stores them in the knowledge vector database. S3-4 retrieves historical questions and performs word segmentation processing. Similarly, the embedding model is used to vectorize the historical questions to form a vector matrix. An importance analysis is performed between the knowledge vector and each element of the question vector based on the importance distribution. The text block corresponding to the knowledge vector with the greatest importance (i.e., the highest similarity) to the question vector is extracted as the key text block. S3-5 fuses the key text blocks and the text blocks corresponding to the vectors in the vector matrix according to the prompt word template to form prompt words, and inputs them into the fine-tuning large model to obtain question and answer results containing knowledge of the power communication industry, thereby completing the construction of the self-question and self-answer model.

[0009] Among them, the heterogeneous graph neural network for word segmentation is constructed based on the dual topology network structure, and the specific method for node-level aggregation is: S3-2-1 uses the segmented words in the area represented by the nodes in the dual topology network structure as nodes to form a heterogeneous graph neural network for segmentation. The heterogeneous graph neural network for segmentation is a subgraph of the dual topology network structure. The same segmented word is regarded as a different node in different heterogeneous graph neural network nodes. For example, segmented words such as current and voltage represent different nodes of heterogeneous graph neural networks in different areas. S3-2-2 Given different word pairs connected by paths formed by the edges of heterogeneous graph neural networks ,importance (1), where They are word segmentation vectors, It is an attention mechanism in the fine-tuned language model, wherein there is a device and / or optical cable connection, and / or communication between two nodes of an edge.

[0010] It should be understood that, conversely, there is no device and / or optical cable connection, and / or no edge between the two communicating nodes, so the heterogeneous graph neural network for word segmentation is a subgraph of the dual topology network structure.

[0011] S3-2-3 traverses all neighboring word j of target word i through softmax function, and calculates the importance Do normalization processing to get word segmentation pairs The normalized importance between (2), where neighbor is the relationship between the text block where the target segmentation i is located and the segmentation j of other text blocks. The greater the normalized importance, the higher the neighbor level, indicating closer proximity.

[0012] It should be understood that the normalized importance is essentially the total probability of the target word and the neighbor word appearing in pairs in the same text vector and in different text vectors.

[0013] S4 specifically comprises: forming a question sentence based on the real-time data in the graph data having the dual topology network structure; S4-1 collects historical operation data from the graph data, and constructs multiple question templates based on abnormal data in the historical operation data and corresponding processing methods, as well as normal data and corresponding adjustment processing methods; S4-2 collects the current running data in the graph data in real time, and compares the similarity with the abnormal data and the normal data that has undergone the adjustment processing (i.e., divides the two data). When the similarity exceeds a threshold (optional range 90-95%), multiple question templates are called to form questions.

[0014] Optionally, the multiple question templates include "Is X currently abnormal? Does it need to be adjusted? What are the reasons?", "Does Y processing method need to be used for X?", "Based on the current X, is there a risk of Z failure in the future?", "Based on your judgment that there is a risk of Z failure, based on the current X, what is the recommended processing method?", "For your selected Y, what are the possible positive and negative consequences?", and "For your selected Z, what historical processing methods have been used?", where X is real-time operating data automatically filled in by the system, Y is a historical processing method type for the system to automatically select one by one, and Z is a historical fault type for the system to automatically select one by one. The method of calling the multiple question templates to form the questions is to call all the aforementioned question templates, automatically fill in X, and fill in Y one by one to form the multiple questions.

[0015] Optionally, the prescribed frequency of executing step S4 is every 5 seconds to once per hour.

[0016] S5 evaluates and predicts the current operating status of the power system based on the answers, specifically including: S5-1 When the answer is determined to be necessary, a human confirms whether to perform Y process again; S5-2 When the answer result indicates fault Z, the current real-time operating data is input into the pre-trained long-short memory model to obtain the predicted fault risk. If it is consistent with the answer result, the recommended processing method described in the answer result is retrieved, and a manual decision is made whether to execute the recommended processing method; if it is inconsistent, manual questioning is intervened, and the self-questioning and self-answering model enters the manual questioning mode to answer how to deal with the predicted fault risk.

[0017] It should be emphasized that the self-questioning and S5 steps are performed on all nodes of the dual-topology network structure, not just on a specific node. The pre-trained long-short-term memory model can be constructed for the entire dual-topology network structure or for a specific node.

[0018] 4. The second main technical solution of the present invention: system The present invention also provides an online monitoring system for self-questioning and self-answering of an electric power system based on a large language model for implementing the above-mentioned method, specifically comprising: an electric power system within a geographical scope, an operation data monitoring system for monitoring the operation parameters of the electric power system, multiple servers communicating with the operation data monitoring system for providing the static data and dynamic data required for building a data base for a knowledge base of an electric power communication large model, and at least one general server for building a data base for a knowledge base of an electric power communication large model, building a fine-tuning large model and a self-questioning and self-answering model, and realizing real-time self-questioning and self-answering of the self-questioning and self-answering model at a prescribed frequency, and evaluating and predicting the current operation status of the electric power system based on the answers, wherein multiple servers are distributed in different nodes of a dual topology network structure, and at least one general server is distributed in nodes of a maximum of three dual topology network structures. Beneficial effects

[0019] This invention utilizes graph data from a dual-topology network structure built on a large-scale power communication model knowledge base. It employs a hybrid model of fine-tuning and retrieval-enhanced generation, improving the general language model. It also utilizes a heterogeneous graph neural network with word segmentation to implement a system self-questioning and self-answering monitoring model based on retrieval-enhanced generation, enabling more timely and accurate monitoring of operational data. Furthermore, it employs a predictive model and a fine-tuned large-scale prediction model to achieve safer online monitoring of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of graph data of the dual topology network structure of embodiment 1 of the present invention, Figure 2 Flowchart of the method for constructing the data base of the power communication large model knowledge base, Figure 3A for Figure 1 The zoomed-in details of the circled node I in the Central and Western Areas include the data models related to the equipment and optical cables in the cells in the corresponding area. Figure 3B for Figure 1 Zoomed-in details of the circled node II in the Central and Western Areas, including data models related to the power plant's cluster equipment and optical cables. Figure 3C The node network structure of the intra-regional connection and inter-regional power line represented by node I is the node-specific structure of the dual topology network structure. Figure 4 Flowchart of the general large model construction and fine-tuning method, Figure 5 The flowchart of the self-questioning and self-answering model of the power system based on word segmentation heterogeneous graph neural network is constructed by using retrieval enhancement generation technology. Figure 6 Text vector element importance analysis as a graphical illustration of similarity comparison, Figure 7 Real-time data self-questioning and self-answering flow chart, Figure 8 Based on the answers, the flow chart of evaluating and predicting the current operating status of the power system is presented. Figure 9 Schematic diagram of the distribution of servers and main servers in the large language model-based self-questioning and self-answering online monitoring system for the power system according to embodiment 2 of the present invention, which implements the method of embodiment 1. DETAILED DESCRIPTION

[0021] Example 1

[0022] This example describes the method in detail.

[0023] like Figure 1 As shown, the method is implemented within the power service geographical range shown in the figure. Specifically, two areas are formed on the east and west sides of the main river, and the method is constructed separately. Specifically, it includes: Detailed explanation of S1 Build a data base for the knowledge base of the power communication model to form a physical topology model, which is used to realize the dual topology network structure of device entities and data on the physical topology model through dynamic operation data drive. For specific reference Figure 2The physical topology model includes a data model of the devices and optical cables in the two areas (i.e., a model including the dual existence of the entities of the devices and optical cables and the data assigned to the entities), and a physical topology data model of the device data and optical cable data in the two areas (i.e., the physical topology of the dual existence model), and a graph model is formed based on the physical topology model, and then the data is formed into a graph data storage based on the graph model, thereby realizing the graph data having the dual topology network structure; like Figure 2 As shown in FIG, the method for constructing the data base of the S1 electric power communication large model knowledge base includes: S1-1 Build a content system for the power backbone communication knowledge base, including static data represented by management data, standards and specifications, resource data, and dynamic data represented by equipment operation data. At the same time, clean, integrate, convert, and evaluate data quality for static and dynamic data to achieve standardization of multi-source heterogeneous data. S1-2 performs entity mapping on the data standardized in S1-1 according to the equipment list and optical cable list of each region within the geographical scope of power service, and obtains the equipment data model and optical cable data model; S1-3 maps the device data model and the optical cable data model to regional connection relationships based on the connection relationships between multiple regions within the power service geographical scope to form a physical topology data model, and then models the device data model, the optical cable data model, and the physical topology data model into a graph model based on the geographical coordinates within the power service geographical scope; S1-4 extracts real-time dynamic operation data from the standardized data of S1-1, builds relational data storage, and realizes data-driven, and stores the graph model in S1-3 to form graph data (i.e. Figure 1 ) to finally complete the construction of the data base of the power communication large model knowledge base.

[0024] Figure 1 The local topological forms of the physical topology model are shown in the east and west areas, respectively, as a dual topology network structure. Circles (○) and triangles (△) represent nodes in the dual topology network structure, respectively. The nodes contain models of device entities (i.e., equipment in the power system), models of optical cable entities (for monitoring system communications), and dynamic and static data stored on a server (see Example 2 for details) assigned to the entity models, forming a data model. The "dual" designation refers to the dual model identity of the entity model and the assigned data model.

[0025] Figure 3A Given Figure 1 The circled node I in the Middle East area has specific equipment entity models a and b (existing in cell B, cell A is represented similarly but not shown), and an optical cable entity model (represented by the word "optical cable"). The circled node II is the power plant ( Figure 3B), which includes the equipment model c of the cluster and the physical model of the optical cable (still represented by the word "optical cable"). Figure 3C It provides the intra-area connections of devices within the area and the power lines between areas (actually an abstraction of the node network for the power and optical cable connections of the internal devices of cells B and A). Figure 1 The locally connected nodes in the physical topology model represent the presence of power line connections between devices and / or optical cables for communication.

[0026] Detailed explanation of S2 S2 is based on the power communication big model knowledge base data base of S1, as well as the general big model to build a fine-tuned big model.

[0027] like Figure 4 As shown, the fine-tuning methods include: S2-1 extraction Figure 2 Static data in the data base of the power communication large model knowledge base; S2-2 converts static data into text and performs text segmentation to form text blocks; S2-3 divides the text blocks into training sets and validation sets, both of which contain question input and question output text blocks. The text blocks in the training set and validation set are labeled with their relevance to the power system field, and the general large language model is supervised for training.

[0028] It should be noted that even in texts belonging to static data, there are segmented words that can be used in other fields, such as cut off, close, gate, operation and maintenance, and fault, which are not highly related to the power system field; but some segmented words, such as current, voltage, switch, transformer, insulation terminal, etc., are highly related to the power system field.

[0029] Specifically, Figure 4 In the paper, the general large language model includes two sub-networks, input embedding and output embedding, which correspond to the text blocks of questions and answers respectively. The outputs of the sub-networks are respectively input into the question input network and the question output network (both of which include attention mechanisms). The output result of the question input network is input into the question output network, and together with the output embedding result, it is calculated by the question output network and output to the classification machine (including classification functions such as softmax function). Finally, the matching degree classification of the question input and question output text blocks is obtained.

[0030] pass Figure 4 The continuous input and output of questions are trained to continuously improve the matching degree. If the highest matching degree is not reached (actually represented by the stabilization of the matching degree), the matching degree of question input and question output is classified according to the output of the general large language model, and the attention mechanism parameters in the question input network and the question output network are fine-tuned and optimized, and supervised training is continued; S2-4 continues to train the general large language model according to S2-3 until the verification set is used to verify that the matching degree reaches the highest level. Training is stopped and fine-tuning is completed.

[0031] Detailed explanation of S3 S3 uses the fine-tuned large model and adopts retrieval enhancement generation technology to build a power system self-questioning and self-answering model based on word segmentation heterogeneous graph neural network.

[0032] Step S3 specifically includes: like Figure 5 As shown in the figure, the circled Arabic numerals represent the process sequence numbers.

[0033] S3-1 from Figure 2 The document is loaded into the data base of the electric power communication large model knowledge base, converted into text, and word segmentation is performed; S3-2 builds a heterogeneous graph neural network for word segmentation based on a dual topology network structure, performs node-level aggregation, and obtains the importance distribution between each word and its neighbors at each level; Figure 5 In accordance with Figure 1 The physical topology model of the western area is locally isomorphic, and a heterogeneous graph neural network for word segmentation is constructed, where the edges represent the edges between nodes with power connections and / or optical cable communications.

[0034] It should be understood that Figure 5 It is just a simulation of the heterogeneous graph neural network of word segmentation for the sake of convenience. The actual heterogeneous graph neural network of word segmentation belongs to the logic diagram of abstract mathematics and only has a topological structure, not a Figure 5 The local structure of the concrete physical topological model.

[0035] The figure illustrates the existence of segmentation nodes, using two bold circles as an example: segmentation I, segmentation II, and so on, up to segmentation n; and segmentation x, segmentation y, and so on, up to segmentation z. (Note: I, II, n, x, y, and z represent definition numbers given to facilitate the distinction between segmentations and do not represent the specific content of each segmentation.) The neighboring levels of the target segmentation machine are visualized in the figure through an analysis of importance distribution. The farther the dotted circle is from the target segmentation, the lower the level, i.e., the less important it is. Therefore, the three dotted circles represent three different levels of importance, e1, e2, and e3. e1, e2, and e3 represent the set of importance values ​​at their respective levels, each with a certain range of values ​​(i.e., within the threshold range), but with differences between them that far exceed the threshold range. The figure only shows three neighboring levels; a three-dot ellipsis indicates that more levels of neighbors may exist, further from the target segmentation circle.

[0036] Among them, the heterogeneous graph neural network for word segmentation is constructed based on the dual topology network structure, and the specific method for node-level aggregation is: S3-2-1 uses the word segmentations in the areas represented by the nodes in the dual topology network structure as nodes to form a heterogeneous graph neural network for word segmentation. The heterogeneous graph neural network for word segmentation is a subgraph of the dual topology network structure. The same word segmentation is regarded as a different node in different heterogeneous graph neural network nodes.

[0037] For example, in the two example nodes shown in bold circles, if the word "i" and the word "x" both represent electric current, since they are assigned static data and reside in different nodes, they belong to different word nodes, even though the word content is electric current. The edge between the two word nodes (that is, an edge in the heterogeneous graph neural network for word segmentation) is like the line connecting the two example nodes shown in the figure. An edge actually represents the connection relationship of the network path.

[0038] S3-2-2 Given different word pairs connected by paths formed by the edges of heterogeneous graph neural networks ,importance (1), where They are word segmentation vectors, It is the attention mechanism in the fine-tuned amplified language model, where the two nodes with an edge are connected by devices and optical cables.

[0039] S3-2-3 traverses all neighboring word j of target word i through softmax function, and calculates the importance Do normalization processing to get word segmentation pairs The normalized importance between (2), where neighbor is the relationship between the text block where the target segmentation i is located and the segmentation j of other text blocks. The greater the normalized importance, the higher the neighbor level, indicating closer proximity.

[0040] S3-3, segment the text according to word meaning, generate text blocks with length, use the application embedding model to vectorize the text blocks, and store them in the knowledge vector database. Figure 6 , describes a graphical representation of a knowledge vector library, showing a knowledge vector where one element is Zi, Chou, Yin, or Mao (just symbols representing the elements). Knowledge vector elements are composed of the text blocks formed from static or dynamic data. Multiple knowledge vectors constitute a knowledge vector database.

[0041] S3-4 retrieves historical questions and performs word segmentation processing. It also uses the application embedding model to vectorize the historical questions and form the following Figure 6A vector matrix, where each element is a historical question vector. An importance analysis is performed between the knowledge vector and each element of the question vector based on the importance distribution. The text block corresponding to the knowledge vector with the greatest question vector importance is extracted as the key text block. As shown in the vector matrix, the importance of the elements between a historical question vector (containing elements numbered A, B, C, D, etc.) and each knowledge vector in the knowledge vector library is analyzed. For example, the importance between element A and Zi, Chou, Yin, and Mao is given in the importance distribution as eI, eII, eIII, eIV, etc. For element B, multiple important properties are obtained. The maximum importance corresponding to each element A, B, C, D, etc. is selected, and the average of these maximum values ​​is calculated. The text block corresponding to the knowledge vector with the greatest average importance between the knowledge vector library and each historical question vector is extracted as the key text block.

[0042] S3-5, based on the prompt word template, fuses the key text block and the text block corresponding to the vector in the vector matrix to form a prompt word. Figure 5 , input the prompt words into the fine-tuned large model completed above to obtain the question and answer results containing knowledge of the power communication industry, and complete the construction of the self-question and self-answer model.

[0043] Detailed explanation of S4 like Figure 7 As shown, S4 forms questions based on the real-time data in the graph data with the dual topology network structure at a frequency of once every 10 seconds, and inputs them into the power system self-questioning and self-answering model constructed by S3 to obtain answers.

[0044] The forming of question sentences based on the real-time data in the graph data having the dual topology network structure specifically includes: S4-1 collects historical operation data from the graph data, and constructs multiple question templates based on abnormal data in the historical operation data and corresponding processing methods, as well as normal data and corresponding adjustment processing methods; S4-2 collects the current running data in the graph data in real time, and compares the similarity with the abnormal data and the normal data that has undergone the adjustment processing, that is, divides the two data. When the similarity exceeds a threshold of 93%, multiple question templates are called to form a question. Otherwise, no processing is performed (Null).

[0045] Multiple question templates include "Is X currently abnormal? Does it require adjustment? What are the reasons?", "Does Y treatment method need to be used for X?", "Based on the current X, is there a risk of Z failure in the future?", "Based on your judgment of the risk of Z failure, what is the recommended treatment method based on the current X?", "For your selected Y, what are the possible positive and negative outcomes?", and "For your selected Z, what historical treatment methods have been used?", where X is real-time operating data automatically entered by the system, Y is the historical treatment method type automatically selected by the system, and Z is the historical fault type automatically selected by the system. To form questions using multiple question templates, call all of the aforementioned question templates, automatically enter X, and then enter Y one by one to form multiple questions.

[0046] It should be understood that Y and Z can be any processing methods and fault types in the prior art in the power system, and are all applicable to the method of the present invention.

[0047] Detailed explanation of S5 like Figure 8 As shown, S5 evaluates and predicts the current operating status of the power system based on the answers, specifically including: If the result of step S5-1 is determined to require Y-processing, a human will confirm whether to execute Y-processing. Otherwise, it is determined whether there is a fault and the process goes to step S5-2. S5-2 When the answer result indicates fault Z, the current real-time operating data is input into the pre-trained long-short memory model to obtain the predicted fault risk. If it is consistent with the answer result, the recommended processing method described in the answer result is retrieved, and a human decides whether to execute the recommended processing method. If they are inconsistent, manual questioning is intervened, and the self-questioning and self-answering model enters the manual questioning mode to answer how to deal with the predicted fault risk. If there is no fault, no processing is performed (Null).

[0048] This demonstrates that the self-questioning and self-answering model is not simply designed to allow the power system to ask and answer questions on its own. When the fault risk predictions from the pre-trained long-short memory model are inconsistent with the answers, the self-questioning and self-answering model can be used to intervene manually. This dual guarantee of a predictive model and a fine-tuned large oracle model allows for safer online monitoring of the power system.

[0049] Example 2 This embodiment illustrates a power system self-questioning and self-answering online monitoring system based on a large language model that can implement the method of embodiment 1. Figure 9As shown, it specifically includes: a power system within a geographical scope, an operation data monitoring system for monitoring the operating parameters of the power system, multiple servers communicating with the operation data monitoring system, 3 servers each for the local physical topology model of the east and west areas, and marked with a bold √ sign, used to provide the static data and dynamic data required for building the power communication large model knowledge base data base, and one general server each for the local physical topology model of the east and west areas, used to build the power communication large model knowledge base data base, build a fine-tuning large model and a self-questioning and self-answering model, and realize real-time self-questioning and self-answering of the self-questioning and self-answering model at a specified frequency, and evaluate and predict the current power system operation status according to the answers, wherein multiple servers are distributed in different dual topology network structure nodes, and at least one general server is distributed in a maximum of three dual topology network structure nodes.

Claims

1. A self-questioning and self-answering online monitoring method for power systems based on a large language model, characterized by: The steps include: S1 builds a data base for the power communication large model knowledge base and forms a physical topology model, which is used to realize a dual topology network structure of device entities and data on the physical topology model through dynamic operation data drive. The physical topology model includes a data model of devices and optical cables in each area, and a physical topology data model of device data and optical cable data in multiple areas. A graph model is formed based on the physical topology model, and then the data is stored as graph data based on the graph model, thereby realizing graph data with the dual topology network structure. S2 builds a fine-tuned large model based on the power communication large model knowledge base data base of S1 and the general large model; S3 is based on the fine-tuned large model and adopts retrieval-enhanced generation technology to build a power system self-questioning and self-answering model based on word segmentation and heterogeneous graph neural network; S4 forms questions based on real-time data in the graph data having the dual topology network structure at a specified frequency, and inputs the questions into the power system self-questioning and self-answering model constructed in S3 to obtain answers; S5 evaluates and predicts the current operating status of the power system based on the answers.

2. The method according to claim 1, characterized in that The method for constructing the data base of the S1 power communication large model knowledge base includes: S1-1 Build a content system for the power backbone communication knowledge base, including static data represented by management data, standards and specifications, resource data, and dynamic data represented by equipment operation data. At the same time, clean, integrate, convert, and evaluate data quality for static and dynamic data to achieve standardization of multi-source heterogeneous data. S1-2 performs entity mapping on the data standardized in S1-1 according to the equipment list and optical cable list of each region within the geographical scope of power service, and obtains the equipment data model and optical cable data model; S1-3 maps the device data model and the optical cable data model to regional connection relationships based on the connection relationships between multiple regions within the power service geographical scope to form a physical topology data model, and then models the device data model, the optical cable data model, and the physical topology data model into a graph model based on the geographical coordinates within the power service geographical scope; S1-4 extracts real-time dynamic operation data from the standardized data of S1-1, builds relational data storage, and implements data-driven storage in the graph model in S1-3 to form graph data, so as to finally complete the construction of the data base of the power communication large model knowledge base.

3. The method according to claim 2, characterized in that The specific steps of S2 include: obtaining a general large model, using the power communication large model knowledge base data base built in S1 to fine-tune the general large model, and forming a power communication large model.

4. The method according to claim 3, characterized in that Fine-tuning methods include: S2-1 extracts static data from the data base of the power communication large model knowledge base; S2-2 converts static data into text and performs text segmentation to form text blocks; S2-3 divides the text blocks into a training set and a validation set, both of which contain question input and question output text blocks. The text blocks in the training and validation sets are labeled with their relevance to the power system field. The general large language model is supervised for training. The matching degree between the question input and question output is determined based on the output of the general large language model, and the attention mechanism parameters are fine-tuned and optimized. S2-4 continues to train the general large language model according to S2-3 until the verification set is used to verify that the matching degree reaches the highest level. Training is stopped and fine-tuning is completed.

5. The method according to claim 3 or 4, characterized in that Step S3 specifically includes: S3-1 loads documents from the power communication large model knowledge base data base, converts them into text, and performs word segmentation processing; S3-2 builds a heterogeneous graph neural network for word segmentation based on a dual topology network structure, performs node-level aggregation, and obtains the importance distribution between each word and its neighbors at each level; S3-3 performs text segmentation according to word meaning, generates text blocks with length, uses application embedding model to vectorize the text blocks, and stores them in the knowledge vector database. S3-4 retrieves historical questions and performs word segmentation processing. Similarly, the application embedding model is used to vectorize the historical questions to form a vector matrix. An importance analysis based on the importance distribution is performed between the knowledge vector and each element of the question vector. The text block corresponding to the knowledge vector with the greatest importance to the question vector is extracted as the key text block. S3-5 fuses the key text blocks and the text blocks corresponding to the vectors in the vector matrix according to the prompt word template to form prompt words, and inputs them into the fine-tuning large model to obtain question and answer results containing knowledge of the power communication industry, thereby completing the construction of the self-question and self-answer model.

6. The method according to claim 5, characterized in that Based on the dual topology network structure, a heterogeneous graph neural network for word segmentation is constructed. The specific method for node-level aggregation is as follows: S3-2-1 takes the word segments in the area represented by the nodes in the dual topology network structure as nodes to form a heterogeneous graph neural network for word segmentation. The heterogeneous graph neural network for word segmentation is a subgraph of the dual topology network structure. The same word segmentation is regarded as a different node in different heterogeneous graph neural network nodes, and represents different nodes of different heterogeneous graph neural networks in different areas. S3-2-2 Given different word pairs connected by paths formed by the edges of heterogeneous graph neural networks ,importance (1), where They are word segmentation vectors, is the attention mechanism in the fine-tuned amplified language model, wherein there is a device and / or optical cable connection, and / or communication between two nodes of the edge, S3-2-3 traverses all neighboring word j of target word i through softmax function, and calculates the importance Do normalization processing to get word segmentation pairs The normalized importance between (2), where neighbor is the relationship between the text block where the target segmentation i is located and the segmentation j of other text blocks. The greater the normalized importance, the higher the neighbor level, indicating closer proximity.

7. The method according to claim 6, characterized in that S4 specifically comprises: forming a question sentence based on the real-time data in the graph data having the dual topology network structure; S4-1 collects historical operation data from the graph data, and constructs multiple question templates based on abnormal data in the historical operation data and corresponding processing methods, as well as normal data and corresponding adjustment processing methods; S4-2 collects the current running data in the graph data in real time, compares the similarity with the abnormal data and the normal data that has undergone the adjustment process, and when the similarity exceeds a threshold, calls multiple question templates to form a question.

8. The method according to claim 7, characterized in that Multiple question templates include "Is X currently abnormal? Does it require adjustment? What are the reasons?", "Is Y treatment method required for X?", "Based on the current X, is there a risk of Z failure in the future?", "Based on your judgment of the risk of Z failure, what is the recommended treatment method based on the current X?", "For your selected Y, what are the possible positive and negative consequences?", and "For your selected Z, what historical treatment methods have been used?", where X is real-time operating data automatically entered by the system, Y is a historical treatment method type automatically selected by the system, and Z is a historical fault type automatically selected by the system. To form questions by calling multiple question templates, call all the aforementioned question templates, automatically enter X, and then enter Y one by one to form multiple questions.

9. The method according to claim 8, characterized in that S5 evaluates and predicts the current operating status of the power system based on the answers, specifically including: S5-1 When the answer is determined to be necessary, a human confirms whether to perform Y process again; S5-2 When the answer result indicates fault Z, the current real-time operating data is input into the pre-trained long-short memory model to obtain the predicted fault risk. If it is consistent with the answer result, the recommended processing method described in the answer result is retrieved, and a manual decision is made whether to execute the recommended processing method; if it is inconsistent, manual questioning is intervened, and the self-questioning and self-answering model enters the manual questioning mode to answer how to deal with the predicted fault risk.

10. A large language model-based power system self-questioning and self-answering online monitoring system implementing the method according to any one of claims 1 to 9, characterized in that: Specifically include: An electric power system within a geographical scope, an operation data monitoring system for monitoring the operating parameters of the electric power system, multiple servers communicating with the operation data monitoring system for providing static data and dynamic data required for building a data base for a knowledge base of a large electric power communication model, and at least one main server for building a data base for a knowledge base of a large electric power communication model, building a fine-tuning large model and a self-questioning and self-answering model, and realizing real-time self-questioning and self-answering of the self-questioning and self-answering model at a prescribed frequency, and evaluating and predicting the operating status of the current electric power system based on the answers, wherein multiple servers are distributed in different nodes of a dual topology network structure, and at least one main server is distributed in a maximum of three nodes of a dual topology network structure.

Images