Large model knowledge retrieval enhancement method and related system oriented to power dispatching field

Abstract

The invention belongs to the field of large-model retrieval, and discloses a large-model knowledge retrieval enhancement method and a related system for the field of power dispatching. A dynamic weighting mechanism of time sensitivity and credibility factors is introduced on the basis of a traditional knowledge retrieval enhancement RAG retrieval framework, so that a retrieval result not only depends on semantic similarity, but also depends on credibility factors; and the timeliness and the source reliability of the information can be comprehensively considered. In traditional knowledge retrieval enhancement (RAG), document sorting often mainly depends on semantic similarity, and differentiation processing of content release time and credibility is lacked, so that the problem that information is outdated although the information is similar or the content source is unreliable easily occurs in a scheduling scene. According to the method, after keyword retrieval and vector retrieval are combined to obtain initial semantic related data, the timeliness weight is calculated through the interval between the timestamp of the query content and the current time, and it is ensured that high-timeliness knowledge such as real-time trend data and an operation mode is preferentially presented in sorting.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of large model retrieval, and particularly relates to a large model knowledge retrieval enhancement method for the power dispatch field and a related system. BACKGROUND

[0002] Large language model technology has begun to be applied in the field of power system dispatch operation. The purpose of this technology is to utilize the powerful nonlinear fitting and massive data processing capabilities of the model to improve the intelligent level of power grid operation. Typical practices include the "Guangming Power Large Model" proposed by State Grid Corporation and the "Large Watt" model system developed by China Southern Power Grid. These industry-specific large models learn deeply from massive historical operation data, dispatching regulations, fault cases, and weather information, aiming to realize the automation and optimization of specific dispatching businesses.

[0003] Knowledge Retrieval Augmentation (RAG) is a key technology paradigm for solving the problems of static built-in knowledge, fact illusion, and lack of source basis in the output of large language models (LLM). When applied in the field of power grid dispatching, the state of the power system, operation regulations, and fault information have strong timestamp attributes. An operation ticket from last year and the state of the power grid an hour ago have very different reference values. Compared with other general knowledge bases (such as Wikipedia and general web pages), the knowledge base in the field of power dispatching has special properties (high timeliness, high security, strong structure, and strong process) when generating vector libraries, so it is necessary to optimize the general RAG link.

[0004] The core of traditional knowledge vector retrieval is to calculate the cosine similarity between the query (Query) and the knowledge document (Document) in the semantic space. This method only cares about whether the semantics are related, but ignores the properties of the knowledge itself. SUMMARY

[0005] The purpose of the present application is to overcome the problem of insensitivity to time and credibility in traditional RAG document retrieval, and to provide a large model knowledge retrieval enhancement method for the field of power dispatching and a related system.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows: In a first aspect, the present application provides a large model knowledge retrieval enhancement method for the field of power dispatching, comprising the following steps: Obtain a query instruction, and according to the query instruction, use a combination of keyword retrieval and vector retrieval to query the required content, process the query content, and obtain the initial semantic related data of the query content; obtaining the interval between the timestamp of the query content and the current query time to obtain the timeliness data of the query content; According to the preset rule, the credibility score is given to the query content, the credibility scores of all query contents are normalized to obtain the credibility data of the query content; According to the query requirement, the initial semantic correlation data, timeliness data and credibility data of the query content are given corresponding weights to obtain the dynamic score of each query content; According to the dynamic score of each query content, the query content is reordered, and the query content with the highest score is taken as the retrieval output.

[0007] The further improvement of the application is that the query instruction is obtained, the required content is queried by combining keyword retrieval and vector retrieval according to the query instruction, the query content is processed, and the specific method for obtaining the initial semantic correlation data of the query content is as follows: The query instruction is obtained, the query instruction is parsed, and the keywords and query vectors corresponding to the query instruction are obtained; According to the keywords and query vectors, the query instruction is retrieved to obtain a plurality of candidate documents; The cosine similarity of all candidate documents is calculated, and all candidate documents are reordered according to the cosine similarity to obtain the initial semantic correlation data of the query content.

[0008] The further improvement of the application is that the interval between the timestamp of the query content and the current query time is obtained to obtain the timeliness data of the query content, and the specific method is as follows: The interval between the timestamp of the query content and the current query time is obtained, and the timeliness data of the query content is obtained according to the following formula :

[0009] Among them, is the timeliness of the query content , and is the decay coefficient.

[0010] The further improvement of the application is that according to the preset rule, the credibility score is given to the query content, the credibility scores of all query contents are normalized to obtain the credibility data of the query content, and the specific method is as follows: Obtaining a plurality of source knowledge assets, giving a credibility level to the metadata in the plurality of source knowledge assets, and taking the plurality of source knowledge assets with the credibility level as the preset rule; The credibility score corresponding to the query content is mapped in the preset rule, the credibility score is normalized to obtain the credibility data of the query content.

[0011] ​The further improvement of the present application is that according to the query requirement, corresponding weights are given to the initial semantic correlation data, timeliness data and credibility data of the query content, and the specific method for obtaining the dynamic score of each query content is as follows: According to the query requirement, corresponding weights are given to the initial semantic correlation data, timeliness data and credibility data of the query content; According to the initial semantic correlation data, timeliness data and credibility data, corresponding weights are given to calculate the dynamic score of each query content : =α· β· γ·

[0012] Wherein, alpha is the importance weight of the initial semantic correlation data, beta is the importance weight of the timeliness data, gamma is the importance weight of the credibility data, and alpha+beta+gamma=1.

[0013] In the second aspect, the present application provides a large model knowledge retrieval enhancement system for the power dispatching field, comprising: An initial semantic correlation data acquisition module is used to acquire a query instruction, query the required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, process the query content, and obtain the initial semantic correlation data of the query content; A timeliness data acquisition module is used to acquire the interval between the timestamp of the query content and the current query time, and obtain the timeliness data of the query content; A credibility data acquisition module is used to give a credibility score to the query content according to a preset rule, normalize the credibility scores of all query contents, and obtain the credibility data of the query content; A weighting module is used to give corresponding weights to the initial semantic correlation data, timeliness data and credibility data of the query content according to the query requirement, and obtain the dynamic score of each query content; A reordering module is used to reorder the query content according to the dynamic score of each query content, and the query content with the highest score is taken as the retrieval output.

[0014] The further improvement of the present application is that the function of the initial semantic correlation data acquisition module is realized by the following method: The query instruction is acquired, the query instruction is parsed, and the keywords and query vectors corresponding to the query instruction are obtained; The query instruction is retrieved according to the keyword pair and the query vector, and a plurality of candidate documents are obtained; The cosine similarity of all candidate documents is calculated, all candidate documents are reordered according to the cosine similarity, and the initial semantic correlation data of the query content is obtained.

[0015] The further improvement of the application is that the function of the timeliness data acquisition module is realized by the following method: The interval between the timestamp of the query content and the current query time is obtained, and the timeliness data of the query content is obtained according to the following formula :

[0016] Among them, is the timeliness representation of the query content , and is the decay coefficient.

[0017] The further improvement of the application is that the function of the credibility data acquisition module is realized by the following method: Obtain multi-source knowledge assets, assign a credibility level to the metadata in the multi-source knowledge assets, and use the multi-source knowledge assets with the assigned credibility level as a preset rule; Map the credibility score corresponding to the query content in the preset rule, normalize the credibility score, and obtain the credibility data of the query content.

[0018] The further improvement of the application is that the function of the weighting module is realized by the following method: According to the query requirement, the initial semantic correlation data, timeliness data and credibility data corresponding to the query content are obtained, and corresponding weights are assigned; According to the initial semantic correlation data, timeliness data and credibility data assigned with corresponding weights, the dynamic score of each query content is calculated :

[0019] Among them, α is the importance weight of the initial semantic correlation data, β is the importance weight of the timeliness data, γ is the importance weight of the credibility data, and α+β+γ=1.

[0020] In a third aspect, the application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the large model knowledge retrieval enhancement method for the power dispatch field when executing the computer program.

[0021] In a fourth aspect, the application provides a storage medium having a computer program stored thereon, wherein the computer program implements the steps of the large model knowledge retrieval enhancement method for the power dispatch field when executed by a processor.

[0022] Compared with the prior art, the application has the following beneficial effects: The application introduces a dynamic weighting mechanism of time sensitivity and credibility factor on the basis of traditional knowledge retrieval enhanced RAG retrieval framework, so that the retrieval result not only depends on semantic similarity, but also considers the timeliness and source reliability of information. In traditional knowledge retrieval enhanced RAG, document sorting often mainly relies on semantic similarity, lacks differentiated processing of content publishing time and credibility, and leads to the problem that in the dispatching scene, similar information may be outdated or the content source may be unreliable. The application calculates the timeliness weight by using the interval between the timestamp of the query content and the current time after obtaining the initial semantic related data by combining keyword retrieval and vector retrieval, so as to ensure that real-time trend data, operation mode and other high timeliness knowledge are presented in priority in sorting. At the same time, by assigning a credibility score to each piece of search content according to the preset rule and normalizing, the risk that unknown sources, insufficient verification or low-quality data occupy a higher position in the search results can be significantly suppressed. The initial semantic correlation, timeliness weight and credibility weight are weighted and fused to form a dynamic score, and then reordering is performed, which can effectively improve the freshness and reliability of information required for dispatching decision while maintaining semantic correlation. For power dispatching which has very high requirements for real-time and accuracy, the application can reduce the operation risk caused by using outdated or untrusted data, improve the support ability of search results for actual dispatching decision, and thus significantly improve the practicability and safety of traditional knowledge retrieval enhanced RAG in professional field application. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a flowchart of the application; Figure 2 is a system diagram of the application; Figure 3 is a flowchart of examples 1-4; Figure 4 is a system diagram of example 5. DETAILED DESCRIPTION

[0024] In order to further understand the content of the application, the application will be described in detail below in combination with the drawings and specific examples. It should be understood that the examples are only used to explain the application and are not limited.

[0025] Referring to Figure 1 , the large model knowledge retrieval enhancement method for the field of power dispatching includes the following steps: S1, obtaining a query instruction, according to the query instruction, using a combination of keyword retrieval and vector retrieval to query the required content, processing the query content to obtain initial semantic related data of the query content.

[0026] S2, obtaining the interval between the timestamp of the query content and the current query time to obtain the timeliness data of the query content.

[0027] S3, according to the preset rule, the query content is given a credibility score, and the credibility scores of all query contents are normalized to obtain the credibility data of the query content.

[0028] S4, according to the query requirement, the initial semantic correlation data, timeliness data and credibility data of the query content are given corresponding weights to obtain the dynamic score of each query content.

[0029] S5, according to the dynamic score of each query content, the query content is reordered, and the query content with the highest score is taken as the retrieval output.

[0030] Referring to Figure 2 , the large model knowledge retrieval enhancement system for the power dispatching field comprises: An initial semantic correlation data acquisition module is configured to acquire a query instruction, query required content by combining keyword retrieval and vector retrieval according to the query instruction, process the query content, and obtain initial semantic correlation data of the query content.

[0031] A timeliness data acquisition module is configured to acquire the interval between the timestamp of the query content and the current query time, and obtain the timeliness data of the query content.

[0032] A credibility data acquisition module is configured to give a credibility score to the query content according to a preset rule, normalize the credibility scores of all query contents, and obtain the credibility data of the query content.

[0033] A weighting module is configured to give corresponding weights to the initial semantic correlation data, timeliness data and credibility data of the query content according to the query requirement, and obtain the dynamic score of each query content.

[0034] A reordering module is configured to reorder the query content according to the dynamic score of each query content, and take the query content with the highest score as the retrieval output.

[0035] Embodiment 1: Referring to Figure 3 , the embodiment further limits the method of step S1 and the initial semantic correlation data acquisition module based on the above content, and specifically as follows: S11, acquiring a query instruction, analyzing the query instruction to obtain the keywords and query vectors corresponding to the query instruction.

[0036] S12, retrieving the query instruction according to the keyword pair and the query vector to obtain a plurality of candidate documents.

[0037] S13, calculate the cosine similarity of all candidate documents, reorder all candidate documents according to the cosine similarity, and obtain the initial semantic related data of the query content.

[0038] Further, the embodiment belongs to the power dispatch knowledge base, the query instruction Q is: the reason of B103 line overload last night and the basis of dispatching treatment, the current time is 2024-06-11, and the specific method is as follows: S11: acquire and analyze the query instruction, the keyword extraction is: B103, line overload, treatment, last night (among them, B103, line overload, treatment are strong business terms). Query vector: encode the query instruction Q into a vector .

[0039] S12: retrieve candidate documents according to keywords and query vectors, keyword retrieval (inverted and BM25) recalls documents containing B103, line overload, and treatment; vector retrieval (ANN, HNSW, FAISS) recalls semantic similar documents; after merging and deduplication, 3 candidates are obtained: Candidate document D1: “Last night tide analysis: relationship between B103 line overload and high temperature load growth” Candidate document D2: “B123 line operation mode adjustment record (June)” Candidate document D3: “Analysis points of transmission line overload and reactive power distribution” Set the document vector: ; ; .

[0040] S13: calculate the cosine similarity and reorder to obtain the initial semantic related data, the calculation formula is as follows: .

[0041] Calculation result: Candidate document D1: dot product 0.7000, ,Similarity 0.700. Candidate document D2: dot product 0.6200,

[0042] , Similarity 0.650. Candidate document D3: dot product 0.8000.

[0043] , Similarity 0.800. According to the cosine similarity in descending order, the initial semantic related data is obtained:

[0044] Candidate document D3 (0.800), candidate document D1 (0.700), candidate document D2 (0.650).

[0045] Embodiment 2: Referring to Figure 3 , the embodiment further limits the method of step S2 and the timeliness data acquisition module based on the above content, as follows: The interval between the timestamp of the query content and the current query time is obtained, and the timeliness data of the query content is obtained according to the following formula :

[0046] Among them, is the timeliness representation of the query content , and is the decay coefficient.

[0047] For each regulatory knowledge text, an explicit timestamp must be extracted. This timestamp is divided into: document creation / modification date (such as the release date of the operation procedure), event occurrence time (fault occurrence time recorded in the fault report), data recording time (such as the recording time in the scheduling log), and accurate time of the identification report.

[0048] Further, when querying the timeliness data of the query content according to the interval between the timestamp of the query content and the current query time, the document type is set as follows: Real-time power flow: (high timeliness) Operation record: (medium timeliness) Basic theory: (low timeliness) Calculate the timeliness data of the query content : Candidate document D1: days, , ; Candidate document D2: days, , ; Candidate document D3: days, , .

[0049] For power grid operation mode, real-time power flow data analysis and other timeliness strong knowledge, set a larger value, for basic electrical theory, long-term unchanged accident plan, set a small value, because their value does not change much over time.

[0050] Embodiment 3 Referring to Figure 3 On the basis of the above, the embodiment further limits the method of step S3 and the credibility data acquisition module, as follows: S31, acquire multi-source knowledge assets, and assign a credibility level to metadata in the multi-source knowledge assets. The multi-source knowledge assets after the credibility level is assigned are used as preset rules.

[0051] S32, map out a credibility score corresponding to the query content in the preset rules, normalize the credibility score, and obtain the credibility data of the query content.

[0052] The multi-source knowledge assets include various documents in the power dispatching field, such as operation procedures, accident plans, historical event reports, expert experience summaries, and real-time dispatching logs. For each knowledge asset, key metadata is automatically or semi-automatically extracted, such as source_type (source type: national regulation, company regulation, expert experience, etc.), source_level (source level), and timestamp (creation / publishing timestamp).

[0053] The control knowledge text is annotated and classified, and is labeled by using automatic recognition + manual correction, and the data source type and source level are labeled, which provides support for subsequent knowledge retrieval.

[0054] Table 1 Credibility level comparison table

[0055] Further, according to the multi-source knowledge assets, the credibility level of the candidate document is normalized:

[0056] The calculation result is obtained: Candidate document D1: 5 / 5 = 1.0; Candidate document D2: 4 / 5 = 0.8; Candidate document D3: 3 / 5 = 0.6.

[0057] Embodiment 4 Referring to Figure 3 On the basis of the above, the embodiment further limits the method of step S4 and the weighting module, as follows: S41, according to the query requirement, the initial semantic relevance data, timeliness data and credibility data corresponding to the query content are acquired, and corresponding weights are assigned.

[0058] S42, calculating a dynamic score of each query content according to the corresponding weight of the initial semantic correlation data, timeliness data and credibility data : =α· β· γ·

[0059] Wherein, alpha is the importance weight of the initial semantic correlation data, beta is the importance weight of the timeliness data, gamma is the importance weight of the credibility data, and alpha+beta+gamma=1.

[0060] Further, according to the query requirement, the corresponding weight is set as:

[0061] Substituting the data in the above embodiment, we get: Candidate document D1=0.5x0.700+0.3x0.905+0.2x1.0=0.8215; Candidate document D2=0.5x0.650+0.3x0.232+0.2x0.8=0.5546; Candidate document D3=0.5x0.800+0.3x0.775+0.2x0.6=0.7525.

[0062] According to the descending order of FinalScore, the final ranking is obtained: Candidate document D1, candidate document D3, candidate document D2.

[0063] The settings of the three hyperparameters in this embodiment are dynamically adjusted according to the question and answer. When the query contains words such as "latest" and "today", the value of beta can be dynamically increased, if the query involves "safety red line" and "operation ticket", the value of beta can be dynamically increased, and for general knowledge queries, the default value can be used.

[0064] The application is aimed at the demand of fast iteration of scheduling text knowledge update and high accuracy of answer in question and answer scene. On the basis of traditional document semantic vector comparison, time sequence and credibility evaluation are added, and the recall chunk is reordered according to the score, which effectively improves the accuracy of model answer. According to the business demand of scheduling field, the weighting parameters are dynamically adjusted for different business question and answer scenes, and reasonable parameter values are verified for various question and answer scenes. On the basis of fine-tuning of large model itself, controllable question and answer debugging mechanism is added.

[0065] Embodiment 5: Referring to Figure 4The application further provides an electronic device 100 for a large model knowledge retrieval enhancement method in the power dispatch field; the electronic device 100 comprises a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0066] The memory 101 can be used to store the computer program 103, and the processor 102 can realize the steps of the large model knowledge retrieval enhancement method in the power dispatch field by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 can mainly comprise a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data (such as audio data) created according to the use of the electronic device 100. In addition, the memory 101 can comprise a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0067] The at least one processor 102 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The processor 102 can be a microprocessor or any conventional processor, etc. The processor 102 is the control center of the electronic device 100, and is connected with various parts of the electronic device 100 through various interfaces and lines.

[0068] The memory 101 in the electronic device 100 stores a plurality of instructions to realize the large model knowledge retrieval enhancement method in the power dispatch field, and the processor 102 can execute the plurality of instructions to realize the following functions: Obtaining a query instruction, and querying required content according to the query instruction by using a keyword search and a vector search in combination, processing the query content, and obtaining initial semantic related data of the query content; Obtaining a time interval between a time stamp of the query content and a current query time, and obtaining timeliness data of the query content; According to a preset rule, a credibility score is given to the query content, and the credibility scores of all the query contents are normalized to obtain credibility data of the query content; According to a query requirement, corresponding weights are given to the initial semantic related data, the timeliness data and the credibility data of the query content to obtain a dynamic score of each query content; According to the dynamic score of each query content, the query content is reordered, and the query content with the highest score is taken as a search output.

[0069] Embodiment 6: The modules / units of the electronic device 100, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory and read-only memory (ROM, Read-Only Memory) capable of carrying the computer program code.

[0070] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0071] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0072] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0073] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 Figure 1 one or more flow or blocks.

[0074] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, but are not intended to limit the present application. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced, and any modification or replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A large model knowledge retrieval enhancement method for the power dispatch field, characterized in that, The method comprises the following steps: Obtaining a query instruction, and querying required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, processing the query content, and obtaining initial semantic correlation data of the query content; Obtaining the interval between the timestamp of the query content and the current query time, and obtaining the timeliness data of the query content; According to the preset rule, the credibility score of the query content is obtained, and the credibility score of all query contents is normalized to obtain the credibility data of the query content; According to the query requirement, the initial semantic correlation data, timeliness data and credibility data of the query content are given corresponding weights to obtain the dynamic score of each query content; According to the dynamic score of each query content, the query content is reordered, and the query content with the highest score is taken as the retrieval output.

2. The method of claim 1, wherein the method is enhanced by a large model knowledge retrieval for the power dispatch domain. The specific method for obtaining the initial semantic correlation data of the query content is as follows: Obtaining a query instruction, and querying required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, processing the query content, and obtaining initial semantic correlation data of the query content; Obtaining a query instruction, and querying required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, processing the query content, and obtaining initial semantic correlation data of the query content; According to the preset rule, the credibility score of the query content is obtained, and the credibility score of all query contents is normalized to obtain the credibility data of the query content; 3. The method of claim 1, wherein the method is enhanced by a large model knowledge retrieval for power dispatch domain. The specific method for obtaining the initial semantic correlation data of the query content is as follows: The interval between the time stamp of the query content and the current query time is obtained, and the timeliness data of the query content is obtained according to the following formula : wherein, is a representation of the timeliness of the query content , is a decay coefficient.

4. The method of claim 1, wherein the method is enhanced by a large model knowledge retrieval for power dispatch domain. Obtaining a query instruction, and querying required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, processing the query content, and obtaining initial semantic correlation data of the query content; Obtaining a query instruction, and querying required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, processing the query content, and obtaining initial semantic correlation data of the query content; According to the preset rule, the credibility score of the query content is obtained, and the credibility score of all query contents is normalized to obtain the credibility data of the query content; 5. The method of claim 1, wherein the method is enhanced by a large model knowledge retrieval for the power dispatch domain. The specific method for obtaining the initial semantic correlation data of the query content is as follows: According to the query requirement, the initial semantic correlation data, timeliness data and credibility data corresponding to the query content are given corresponding weights; According to initial semantic correlation data, timeliness data and credibility data, corresponding weights are given to calculate the dynamic score of each query content : = a · β· γ· Wherein, α is the importance weight of the initial semantic correlation data, β is the importance weight of the timeliness data, and γ is the importance weight of the credibility data, and α+β+γ=1.

6. The large model knowledge retrieval enhancement system for the power dispatch field, characterized in that, It comprises: An initial semantic correlation data acquisition module is configured to obtain a query instruction, and query required content by adopting a combination of keyword retrieval and vector retrieval according to the query instruction, process the query content, and obtain initial semantic correlation data of the query content; A timeliness data acquisition module is configured to obtain the interval between the timestamp of the query content and the current query time, and obtain the timeliness data of the query content; A credibility data acquisition module is configured to obtain the credibility score of the query content according to the preset rule, and normalize the credibility score of all query contents to obtain the credibility data of the query content; The weighting module is configured to assign corresponding weights to the initial semantic relevance data, timeliness data and credibility data of the query content according to the query requirement, and obtain a dynamic score of each query content. The reordering module is configured to reorder the query content according to the dynamic score of each query content, and output the query content with the highest score as the search result.

7. The large model knowledge retrieval enhancement system for the power dispatch field according to claim 6, wherein, The function of the initial semantic relevance data acquisition module is implemented by the following method: Obtain the query instruction, analyze the query instruction to obtain the keywords and query vector corresponding to the query instruction; According to the keywords and query vector, the query instruction is searched to obtain a plurality of candidate documents; Calculate the cosine similarity of all candidate documents, reorder all candidate documents according to the cosine similarity, and obtain the initial semantic relevance data of the query content.

8. The large model knowledge retrieval enhancement system for the power dispatch field according to claim 6, wherein, The function of the timeliness data acquisition module is implemented by the following method: An interval between a time stamp of the query content and a current query time is obtained, and the timeliness data of the query content is obtained according to the following formula : wherein is a representation of the timeliness of the query content , is a decay coefficient.

9. The large model knowledge retrieval enhancement system for the power dispatch field according to claim 6, wherein, The function of the credibility data acquisition module is implemented by the following method: Obtain the multi-source knowledge asset, assign a credibility level to the metadata in the multi-source knowledge asset, and use the multi-source knowledge asset with the assigned credibility level as the preset rule; Map the credibility score corresponding to the query content in the preset rule, normalize the credibility score, and obtain the credibility data of the query content.

10. The large model knowledge retrieval enhancement system for the power dispatch domain according to claim 6, wherein, The function of the weighting module is implemented by the following method: According to the query requirement, the initial semantic relevance data, timeliness data and credibility data corresponding to the query content are assigned corresponding weights. According to initial semantic correlation data, timeliness data and credibility data, corresponding weights are given to calculate the dynamic score of each query content : Wherein, α is the importance weight of the initial semantic relevance data, β is the importance weight of the timeliness data, γ is the importance weight of the credibility data, and α+β+γ=1.

11. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to realize the steps of the large model knowledge retrieval enhancement method for the power dispatching field in any one of claims 1 to 5.

12. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the large model knowledge retrieval enhancement method for the power dispatching field in any one of claims 1 to 5.

Images