CN-122019135-A - Electric power complex task execution method and device based on large model task planning

CN122019135ACN 122019135 ACN122019135 ACN 122019135ACN-122019135-A

Abstract

The invention provides a method and a device for executing an electric power complex task based on large model task planning, wherein the method comprises the steps of obtaining an electric power task to be executed, and outputting a subtask sequence of the electric power task to be executed based on a task planning model; the method comprises the steps of inputting background knowledge matched with any subtask and any subtask into a task decision model to obtain a subtask execution decision of any subtask output by the task decision model, and integrating the subtask execution decisions of all subtasks in a subtask sequence to obtain a complete task execution scheme. The method provided by the invention not only utilizes the semantic understanding and generating capacity of the large language model, but also introduces domain knowledge retrieval on subtask granularity, thereby automatically generating the electric complex task solution which has reasonable structure, complete steps, accords with regulations and has extremely high executable performance.

Inventors

KONG QINGCHAO
ZENG DAJUN
SUN YIWEI
CAO YILIN
YANG DAI
CHEN KUN
GU CHAO
ZHANG YINGQIANG
LIU ZIHAN
XU JIANNAN

Assignees

中国科学院自动化研究所
国网山东省电力公司电力科学研究院
中国电力科学研究院有限公司
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. The electric power complex task execution method based on the large model task planning is characterized by comprising the following steps of: acquiring an electric power task to be executed; inputting the electric power task to be executed into a task planning model to obtain a subtask sequence output by the task planning model; Inputting background knowledge matched with any subtask in the subtask sequence and the any subtask into a task decision model to obtain a subtask execution decision of the any subtask output by the task decision model; and integrating subtask execution decisions of all subtasks in the subtask sequence to obtain a complete task execution scheme, wherein the task planning model and the task decision model are obtained based on general large language model training.
2. The method for performing complex tasks on power based on large model task planning according to claim 1, wherein the step of determining background knowledge matched with any one of the sub-tasks comprises: Constructing a search query based on the electric power task to be executed and any sub-task; And carrying out semantic retrieval in a background knowledge base based on the retrieval query to obtain background knowledge matched with any subtask.
3. The method for performing complex tasks on electric power based on large model task planning according to claim 1, wherein the subtask sequence includes a plurality of subtasks and task dependencies among the subtasks; the task dependency relationship is used for determining the execution sequence of each subtask; the integrating the subtask execution decisions of all the subtasks in the subtask sequence to obtain a complete task execution scheme comprises the following steps: and integrating the subtask execution decisions of all the subtasks according to the execution sequence to obtain the complete task execution scheme.
4. A method of performing a complex task of power based on large model mission planning as claimed in any one of claims 1 to 3, wherein the training steps of the mission planning model and the mission decision model include: Acquiring an initial comprehensive model, wherein the initial comprehensive model comprises an initial task planning model and an initial task decision model, and the initial task planning model and the initial task decision model are obtained based on the general large language model through supervised fine tuning training; obtaining a sample track group corresponding to a sample task based on the initial comprehensive model, wherein the sample track group comprises a plurality of sample subtask sequences and a plurality of sample complete execution decisions; calculating to obtain an optimization target value based on the average track rewards of the sample track groups and the track rewards of the sample tracks under the sample track groups, wherein the track rewards of any sample track comprise task planning rewards and strategy rewards; And carrying out parameter adjustment on the initial comprehensive model based on the optimization target value to obtain the task planning model and the task decision model.
5. The method for performing complex tasks on power based on large model task planning according to claim 4, wherein the step of determining the policy rewards for any of the sample trajectories comprises: calculating a decision matching degree based on the complete execution decision label corresponding to the sample task and the complete execution decision of the sample in any sample track; calculating the decision security of the complete execution decision of the sample in any sample track; And obtaining the strategy rewards based on the decision matching degree and the decision security weighting calculation.
6. The method for performing complex tasks on power based on large model task planning according to claim 4, wherein the step of determining task planning rewards for any of the sample trajectories comprises: performing semantic matching based on a subtask label sequence corresponding to the sample task and a sample subtask sequence in any sample track, and calculating to obtain an F1 index; And taking the F1 index as the task planning reward.
7. An electric power complex task execution device based on large model task planning is characterized by comprising: The acquisition unit acquires an electric power task to be executed; the task planning unit inputs the electric power task to be executed into a task planning model to obtain a subtask sequence output by the task planning model; The decision unit inputs the background knowledge matched with any subtask in the subtask sequence and the any subtask into a task decision model to obtain a subtask execution decision of the any subtask output by the task decision model; The scheme integrating unit integrates subtask execution decisions of all subtasks in the subtask sequence to obtain a complete task execution scheme, and the task planning model and the task decision model are obtained based on general large language model training.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the large model mission planning based power complex mission execution method of any of claims 1 to 6 when the computer program is executed.
9. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the power complex task execution method based on large model task planning according to any of claims 1 to 6.
10. A computer program product comprising a computer program which, when executed by a processor, implements a method of performing electrically complex tasks based on large model task planning as claimed in any one of claims 1 to 6.

Description

Electric power complex task execution method and device based on large model task planning Technical Field The invention relates to the technical field of electric power task execution, in particular to an electric power complex task execution method and device based on large model task planning. Background In the power industry, as the power grid scale expands and the operation and inspection traffic continues to grow, the traditional operation mode which depends on manual experience and fixed flow has difficulty in coping with increasingly complex tasks such as equipment overhaul, fault handling and the like. The appearance of large language models provides a new technical approach for natural language understanding and scheme generation of complex tasks. The large model is used for scenes such as power question-answering, rule retrieval, bill auxiliary writing and the like at present, and the specificity of generating the lifting reply is enhanced through retrieval. However, most of these applications consider the service request as a one-time question and answer, and in the long-chain and multi-constraint complex tasks, problems such as unreasonable task disassembly, missing key steps, and wrong execution sequence are easy to occur. Disclosure of Invention The invention provides a method and a device for executing an electric power complex task based on large model task planning, which are used for solving the defects that in the prior art, when a large model directly applies a long-link and multi-constraint electric power complex task, task disassembly is unreasonable, key steps are lost, execution sequence is wrong and the like. The invention provides a power complex task execution method based on large model task planning, which comprises the following steps: acquiring an electric power task to be executed; inputting the electric power task to be executed into a task planning model to obtain a subtask sequence output by the task planning model; Inputting background knowledge matched with any subtask in the subtask sequence and the any subtask into a task decision model to obtain a subtask execution decision of the any subtask output by the task decision model; and integrating subtask execution decisions of all subtasks in the subtask sequence to obtain a complete task execution scheme, wherein the task planning model and the task decision model are obtained based on general large language model training. According to the electric power complex task execution method based on large model task planning provided by the invention, the determining step of the background knowledge matched with any subtask comprises the following steps: Constructing a search query based on the electric power task to be executed and any sub-task; And carrying out semantic retrieval in a background knowledge base based on the retrieval query to obtain background knowledge matched with any subtask. According to the electric power complex task execution method based on the large model task planning, the subtask sequence comprises a plurality of subtasks and task dependency relations among the subtasks; the task dependency relationship is used for determining the execution sequence of each subtask; the integrating the subtask execution decisions of all the subtasks in the subtask sequence to obtain a complete task execution scheme comprises the following steps: and integrating the subtask execution decisions of all the subtasks according to the execution sequence to obtain the complete task execution scheme. According to the electric power complex task execution method based on large model task planning, the training steps of the task planning model and the task decision model comprise: Acquiring an initial comprehensive model, wherein the initial comprehensive model comprises an initial task planning model and an initial task decision model, and the initial task planning model and the initial task decision model are obtained based on the general large language model through supervised fine tuning training; obtaining a sample track group corresponding to a sample task based on the initial comprehensive model, wherein the sample track group comprises a plurality of sample subtask sequences and a plurality of sample complete execution decisions; calculating to obtain an optimization target value based on the average track rewards of the sample track groups and the track rewards of the sample tracks under the sample track groups, wherein the track rewards of any sample track comprise task planning rewards and strategy rewards; And carrying out parameter adjustment on the initial comprehensive model based on the optimization target value to obtain the task planning model and the task decision model. According to the method for executing the electric complex task based on the large model task planning, which is provided by the invention, the determining step of the strategy rewards of any sample track comprises the following steps: ca