CN-121998050-A - Large model optimization method and device for power industry

CN121998050ACN 121998050 ACN121998050 ACN 121998050ACN-121998050-A

Abstract

The embodiment of the application provides a power industry large model optimization method and device, which comprise the steps of obtaining set role information, input data, task information and output information, generating an input template according to the role information, the input data, the task information and the output information and preset template rules, analyzing the input template to obtain an analysis result, converting the analysis result into a feature vector, inputting the feature vector into a pre-built power industry large model, and outputting an output result corresponding to the output information by the power industry large model based on the feature vector. The method and the device can improve the accuracy and reliability of the model for processing the power service and support the multi-field power service scene.

Inventors

WANG HAO
GE HUALI
XU HENGYAN
LI YANG
XU CHUANBO
Hao Mengning
LI BO

Assignees

北京中电普华信息技术有限公司

Dates

Publication Date: 20260508
Application Date: 20251210

Claims (10)

1. The large model optimization method for the power industry is characterized by comprising the following steps of: acquiring set role information, input data, task information and output information; generating an input template according to the character information, the input data, the task information and the output information and a preset template rule; Analyzing the input template to obtain an analysis result, and converting the analysis result into a feature vector; and inputting the characteristic vector into a pre-constructed power industry large model, and outputting an output result corresponding to the output information based on the characteristic vector by the power industry large model.
2. The method of claim 1, wherein before the acquiring the set character information, the input data, the task information, and the output information, further comprising: according to the role information, a corresponding professional knowledge base is called; and training and adjusting the pre-trained general large model based on the professional knowledge base to obtain the large model of the power industry.
3. The method of claim 1, wherein the template rules include a role name tag, a role field tag, a power parameter constraint tag, an original input data tag, an input data type tag, an output keyword tag, and a verification control tag, wherein the role information includes a role name and a role field, wherein the task information includes a power parameter and a verification switch state, and wherein the output information includes an output type and an output keyword; According to the character information, the input data, the task information and the output information, an input template is generated according to a preset template rule, and the method comprises the following steps: setting the character name label and the character domain label according to the character name and the character domain; Setting the original input data tag and the input data type tag according to the input data, wherein the original input data tag is the input data, and the input data type tag is the type of the input data; setting the power parameter constraint tag according to the power parameter; setting an output type label and an output keyword label according to the output type and the output keyword; and setting the check control tag according to the check switch state.
4. A method according to claim 3, wherein parsing the input template to obtain a parsed result and converting the parsed result into feature vectors comprises: when the input data type label is a long text, segmenting the input data by utilizing a sliding window to obtain short texts of a plurality of paragraphs; Extracting key information from each short text by using a preset electric power pre-training language model; fusing key information of each short text to construct a long text semantic map; and converting the long text semantic graph into corresponding feature vectors.
5. The method of claim 3, wherein the verification switch comprises a parameter verification switch, the method further comprising: When the verification control tag is turned on for parameter verification, the output result is verified according to a preset parameter rule, and the verified output result is obtained.
6. The method of claim 5, wherein the parameter rules include parameter threshold rules, business logic rules; And verifying the output result according to a preset parameter rule to obtain a verified output result, wherein the verification comprises the following steps: And checking the output result according to the parameter threshold rule and the business logic rule, and correcting the output result according to the parameter threshold rule and/or the business logic rule if the output result does not accord with the parameter threshold rule and/or the business logic rule.
7. The method of claim 6, wherein modifying the output result according to the parameter threshold rule and/or business logic rule comprises: If the output result has parameter threshold errors, business logic contradiction errors and flow specification errors; and correcting parameter threshold class errors, business logic contradiction class errors and flow specification class errors in sequence.
8. The method of claim 3, wherein the check switch comprises a safety check switch, the method further comprising: When the verification control tag is opened for safety verification, detecting whether sensitive information exists in the input data and the output result according to a preset safety rule; if the sensitive information exists, desensitizing treatment is carried out on the sensitive information, and input data and output results after security verification are obtained.
9. The method according to any one of claims 3 to 8, wherein outputting, by the electric power industry large model, an output result corresponding to the output information based on the feature vector, includes: And outputting a standardized output result comprising the output keywords corresponding to the output keyword labels according to the output types corresponding to the output type labels.
10. The utility model provides a power industry big model optimizing apparatus which characterized in that includes: The acquisition module is used for acquiring the set role information, input data, task information and output information; The generating module is used for generating an input template according to the character information, the input data, the task information and the output information and preset template rules; the analysis module is used for analyzing the input template to obtain an analysis result and converting the analysis result into a feature vector; And the output module is used for inputting the characteristic vector into a pre-constructed power industry large model, and outputting an output result corresponding to the output information based on the characteristic vector by the power industry large model.

Description

Large model optimization method and device for power industry Technical Field The embodiment of the application relates to the technical field of artificial intelligence, in particular to a large model optimization method and device in the power industry. Background With the development of artificial intelligence technology, the large model is widely applied to various industries, for example, in the power industry, the functions of fault diagnosis, resource scheduling and the like can be realized based on the large model, the stable operation of a power system is ensured, and the energy utilization efficiency is improved. However, because the large model does not deeply integrate the knowledge of the power industry, lacks a view angle of the power industry, and is easy to cause problems of misjudgment, deviation and the like when processing the data of the power industry, the accuracy of business decision is affected, and the model cannot accurately understand the data of the power industry due to the fact that a prompt word template is not specifically designed. Disclosure of Invention In view of the above, an objective of the embodiments of the present application is to provide a method and an apparatus for optimizing a large model in the power industry. Based on the above object, the embodiment of the application provides a power industry large model optimization method, which comprises the following steps: acquiring set role information, input data, task information and output information; generating an input template according to the character information, the input data, the task information and the output information and a preset template rule; Analyzing the input template to obtain an analysis result, and converting the analysis result into a feature vector; and inputting the characteristic vector into a pre-constructed power industry large model, and outputting an output result corresponding to the output information based on the characteristic vector by the power industry large model. Optionally, before the set role information, the input data, the task information and the output information are obtained, the method further includes: according to the role information, a corresponding professional knowledge base is called; and training and adjusting the pre-trained general large model based on the professional knowledge base to obtain the large model of the power industry. Optionally, the template rule comprises a role name tag, a role field tag, a power parameter constraint tag, an original input data tag, an input data type tag, an output keyword tag and a verification control tag, wherein the role information comprises a role name and a role field, the task information comprises a power parameter and a verification switch state, and the output information comprises an output type and an output keyword; According to the character information, the input data, the task information and the output information, an input template is generated according to a preset template rule, and the method comprises the following steps: setting the character name label and the character domain label according to the character name and the character domain; Setting the original input data tag and the input data type tag according to the input data, wherein the original input data tag is the input data, and the input data type tag is the type of the input data; setting the power parameter constraint tag according to the power parameter; setting an output type label and an output keyword label according to the output type and the output keyword; and setting the check control tag according to the check switch state. Optionally, analyzing the input template to obtain an analysis result, and converting the analysis result into a feature vector, including: when the input data type label is a long text, segmenting the input data by utilizing a sliding window to obtain short texts of a plurality of paragraphs; Extracting key information from each short text by using a preset electric power pre-training language model; fusing key information of each short text to construct a long text semantic map; and converting the long text semantic graph into corresponding feature vectors. Optionally, the check switch comprises a parameter check switch, and the method further comprises: When the verification control tag is turned on for parameter verification, the output result is verified according to a preset parameter rule, and the verified output result is obtained. Optionally, the parameter rule includes a parameter threshold rule and a business logic rule; And verifying the output result according to a preset parameter rule to obtain a verified output result, wherein the verification comprises the following steps: And checking the output result according to the parameter threshold rule and the business logic rule, and correcting the output result according to the parameter threshold rule and/or the business logic rule if the outpu