Large electric power model performance optimization method and system based on declarative self-prompt

Abstract

The invention discloses an electric power large model performance optimization method and system based on declarative self-prompt. The method comprises the following steps of performing mathematical modeling on a large model performance optimization task; constructing a task optimizer based on the large model performance optimization task; performing parameter optimization by using a bootstrap random optimization method, a module-level prompt optimization method and a multi-prompt instruction optimization method in sequence on the basis of the training set through a step task optimizer to obtain optimal parameters; and completing the performance optimization of the large electric power model by adopting the optimal parameters. According to the method, an optimized LLM calling scheme and corresponding prompts are automatically generated by utilizing a programming thought, and the system automatically learns and adjusts the prompts to adapt to different tasks and context environments without manual intervention, so that the application range of the LLM is widened, and the robustness and generalization ability of the LLM are enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of artificial intelligence and power system, and particularly relates to a power large model performance optimization method and system based on declarative self-prompting. BACKGROUND

[0002] The rise of large language models (LLMs) marks a major advance in the field of natural language processing (NLP). These models, due to their massive parameter size and powerful pre-training capabilities, allow researchers to work at a higher level of abstraction when developing NLP applications, reducing the reliance on large amounts of labeled data. In practice, this means that developers can build complex multi-stage pipelines, where each stage can be completed by calling an LLM for a specific task, simplifying the overall design of the system.

[0003] To enable LLMs to efficiently perform these tasks, researchers need to carefully design "prompt words". An effective prompt is like providing LLM with a clear set of instructions on how to generate the desired output based on the input. For example, in a translation task, a good prompt can help the model understand the relationship between the source language and the target language; while in a code writing task, it can guide the model to generate program code that meets the syntax and logic requirements.

[0004] However, current power large model prompts often rely on static, manually set prompt templates, which have certain limitations. First, fixed prompts are difficult to adapt to changes in task requirements, as they are created based on best practices in specific scenarios. Second, if the prompt is not clear or misleading, it may lead to inaccurate or even completely wrong results for the power LLM. In addition, for new or different power LLMs, existing prompts may not achieve the expected effect, which limits the versatility and flexibility of the model. SUMMARY

[0005] To address the shortcomings of the prior art, one of the purposes of the present application is to provide a power large model performance optimization method based on declarative self-prompting, which automatically generates optimized LLM calling schemes and corresponding prompts using programming ideas, allowing the system to automatically learn and adjust prompts without human intervention, to adapt to different tasks and context environments, widening the application range of LLMs, and enhancing their robustness and generalization ability.

[0006] The second purpose of the present application is to provide a system for implementing the power large model performance optimization method based on declarative self-prompting.

[0007] The present invention provides a method for optimizing the performance of a large power model based on declarative self-prompting, comprising the following steps:

[0008] S1. Mathematical modeling of large model performance optimization tasks;

[0009] S2. Build a task optimizer based on large-scale model performance optimization tasks.

[0010] S3. The task optimizer obtained in step S2 is optimized based on the training set using the bootstrap random optimization method, the module-level prompt optimization method, and the multi-prompt instruction optimization method to obtain the optimal parameters;

[0011] S4. Use the optimal parameters obtained in step S3 to complete the performance optimization of the large power model.

[0012] Step S1 includes the following steps:

[0013] Define the composition and operation process of the large power model;

[0014] Based on the data definition of the large power model, mathematical modeling is performed on the large model performance optimization task.

[0015] Step S1 is specifically as follows:

[0016] Define the power model program as , consists of m submodules; each module i is protected by a prompt word template with a set of variables v Define; define V as a large model program A collection of all variables used for prompt word templates in Indicates that the program runs under the assignment condition of assigning a string S from the variable set V; each prompt word template p includes 1 instruction variable and K example variables, and the other variables are set to constants;

[0017] Based on the definition of the power big model, the goal of the big model performance optimization task is: In the training set D composed of In the indicator performance on;

[0018] The goal of the large model performance optimization task uses the objective function express:

[0019] in, is to maximize the function value under the assignment condition of the string S from the variable set V; is the modulus of the training set D; x is an input data in the input data set X; Metadata set A metadata in .

[0020] In step S2, the task optimizer includes an example bootstrapping module and a credit allocation module;

[0021] The sample bootstrapping module uses the rejection sampling algorithm to produce the sample set;

[0022] The credit allocation module determines the contribution of specific hyperparameters to the power model program based on the example sample set, implements credit allocation for different hyperparameter combinations, and completes the performance optimization of the power model.

[0023] The operation of the example bootstrap module is as follows:

[0024] From the power-related training set The sampling input x is input into the power model program to run, and the input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples;

[0025] Traverse the training set When all inputs or sampling times in reach the preset threshold, a valid sample set is obtained.

[0026] The operation of the credit allocation module specifically includes the following steps:

[0027] Define the credit allocation target EI;

[0028] Use kernel density estimation to model the probability distribution of the objective function value under given hyperparameters;

[0029] The hyperparameter combination that currently maximizes the credit allocation target is selected as the next evaluation point, and iterative calculations are performed to obtain the optimal hyperparameter combination.

[0030] The operation of the credit allocation module is specifically as follows:

[0031] Define the credit allocation target EI, using the following formula:

[0032] in, is the observation threshold; is the objective function value, i.e., the indicator score; Given the hyperparameters Lower objective function value The conditional probability distribution of ; For a given hyperparameter Lower credit allocation target value;

[0033] The training set Divided into two groups according to the preset ratio and ; Among them, the training set Sort by the objective function value from large to small. For the front The hyperparameter combination of For the remaining The hyperparameter combination of and The probability is modeled using the following formula:

[0034] in, for Kernel density estimation of ; for Kernel density estimation of ; for and The proportion of training set D; is the probability distribution function of the objective function value;

[0035] The EI value is calculated explicitly using the following formula:

[0036] in, yes The mean of

[0037] Select The largest hyperparameter combination x is used as the next evaluation point, and iterative calculation is performed to finally obtain the optimal hyperparameter combination.

[0038] Step S3 is specifically as follows:

[0039] The bootstrap random optimization method is used to use the example bootstrap module in the task optimizer to obtain the power-related training set. The sampling instruction input x is input into the power model program for execution, and an input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples; the all values ​​include the parameters of the operation process of each submodule and the final score result;

[0040] Traverse the training set All inputs or sampling times in reach a preset threshold, and an example sample set is obtained;

[0041] A module-level prompt optimization method is adopted, and the Qwen-Plus large model is introduced as a judge. For the input and corresponding score of each submodule stored in the sample set, the judge generates new instruction inputs, which are input into the corresponding submodule and run to obtain the result score. The new instructions and corresponding scores are also stored in the sample set and classified and stored to obtain the instruction set of each submodule.

[0042] A multi-hint instruction optimization method is adopted. According to the instruction set of each sub-module, the parameter instructions and examples of each sub-module are calculated using the credit allocation component. Finally, the parameter with the highest probability is the optimal parameter.

[0043] The module-level prompt optimization method comprises the following steps:

[0044] Obtain a sample set obtained by a bootstrap random optimization method; the sample set includes several instruction inputs and the scores of each submodule after the instruction input large model is run;

[0045] Set the Qwen-Plus large model as the judge, input the instruction input and score of each submodule in the sample set to the judge, and the judge generates a new instruction input for each submodule based on the input;

[0046] Input the instruction input generated by the judge into the corresponding submodule of the large model program, and store the instruction input and the corresponding score according to different submodules to obtain the instruction set of each submodule;

[0047] The above steps are repeated based on the instruction set of each submodule, and new instruction inputs and corresponding scores are continuously generated and stored and updated in the instruction set of each submodule until the number of iterations reaches the preset maximum number of iterations.

[0048] The multi-hint instruction optimization method comprises the following steps:

[0049] The instruction set of each submodule is obtained by the module-level hint optimization method. Each module contains T instructions, and K task examples are generated for each submodule using the example bootstrapping module.

[0050] For the instructions and examples of each submodule, the credit allocation module is used to calculate the iterations. After the final convergence or reaching the preset maximum number of iterations, the parameter instruction with the highest probability is selected as the optimal parameter.

[0051] The present invention also provides a system for implementing the power large model performance optimization method based on declarative self-prompt, comprising a mathematical modeling module, a task optimizer construction module, an optimal parameter acquisition module and a power large model optimization module;

[0052] The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads the data to the task optimizer construction module;

[0053] The task optimizer construction module optimizes the task based on the received data and the large model performance, builds the task optimizer, and uploads the data to the optimal parameter acquisition module;

[0054] The optimal parameter acquisition module uses the task optimizer to optimize the parameters based on the received data using the bootstrap random optimization method, module-level prompt optimization method, and multi-prompt instruction optimization method based on the training set to obtain the optimal parameters and upload the data to the power large model optimization module;

[0055] The power model optimization module uses the optimal parameters based on the received data to complete the performance optimization of the power model.

[0056] The present invention discloses a method and system for optimizing the performance of a large power model based on declarative self-prompting. The method utilizes programming ideas to automatically generate optimized LLM calling schemes and corresponding prompts. Without the need for human intervention, the system automatically learns and adjusts prompts to adapt to different tasks and contextual environments, thereby expanding the application scope of LLM and enhancing its robustness and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the process of the present invention;

[0058] Figure 2 Schematic diagram of the structure of the system of the present invention;

[0059] Figure 3 This is a flowchart of an example bootstrap module in the task optimizer of the present invention. DETAILED DESCRIPTION

[0060] The present invention provides a method for optimizing the performance of a large power model based on declarative self-prompt, the flow chart of which is as follows: Figure 1 As shown, the following steps are included:

[0061] S1. Mathematical modeling of large model performance optimization tasks;

[0062] Step S1 includes the following steps:

[0063] Define the composition and operation process of the large power model;

[0064] Based on the data definition of the large power model, mathematical modeling is performed on the large model performance optimization task.

[0065] Step S1 is specifically as follows:

[0066] Define the power model program as , consists of m submodules; each module i is protected by a prompt word template with a set of variables v Define; define V as a large model program A collection of all variables used for prompt word templates in Indicates that the program runs under the assignment condition of assigning a string S from the variable set V; each prompt word template p includes 1 instruction variable and K example variables, and the other variables are set to constants;

[0067] Based on the definition of the power big model, the goal of the big model performance optimization task is: In the training set D composed of In the indicator performance on;

[0068] The goal of the large model performance optimization task uses the objective function express:

[0069] in, is to maximize the function value under the assignment condition of the string S from the variable set V; is the modulus of the training set D; x is an input data in the input data set X; Metadata set A metadata in .

[0070] S2. Build a task optimizer based on large-scale model performance optimization tasks.

[0071] In step S2, the task optimizer includes an example bootstrapping module and a credit allocation module;

[0072] The example bootstrap module uses the rejection sampling algorithm to generate the example sample set. The operation flow diagram is as follows: Figure 3 As shown;

[0073] The credit allocation module determines the contribution of specific hyperparameters to the power model program based on the example sample set, implements credit allocation for different hyperparameter combinations, and completes the performance optimization of the power model.

[0074] The example bootstrap module operates as follows:

[0075] From the power-related training set The sampling input x is input into the power model program to run, and the input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples;

[0076] Traverse the training set When all inputs or sampling times in reach the preset threshold, a valid sample set is obtained.

[0077] The operation of the credit allocation module specifically includes the following steps:

[0078] Define the credit allocation target EI;

[0079] Use kernel density estimation to model the probability distribution of the objective function value under given hyperparameters;

[0080] The hyperparameter combination that currently maximizes the credit allocation target is selected as the next evaluation point, and iterative calculations are performed to obtain the optimal hyperparameter combination.

[0081] The operation of the credit allocation module is specifically as follows:

[0082] Define the credit allocation target EI, using the following formula:

[0083] in, is the observation threshold; is the objective function value, i.e., the indicator score; Given the hyperparameters Lower objective function value The conditional probability distribution of ; For a given hyperparameter Lower credit allocation target value;

[0084] The training set Divided into two groups according to the preset ratio and ; Among them, the training set Sort by the objective function value from large to small. For the front The hyperparameter combination of For the remaining The hyperparameter combination of and The probability is modeled using the following formula:

[0085] in, for Kernel density estimation of ; for Kernel density estimation of ; for and The proportion of training set D; is the probability distribution function of the objective function value;

[0086] The EI value is calculated explicitly using the following formula:

[0087] in, yes The mean of

[0088] Select The largest hyperparameter combination x is used as the next evaluation point, and iterative calculation is performed to finally obtain the optimal hyperparameter combination.

[0089] S3. The task optimizer obtained in step S2 is optimized based on the training set using the bootstrap random optimization method, the module-level prompt optimization method, and the multi-prompt instruction optimization method to obtain the optimal parameters;

[0090] Step S3 is specifically as follows:

[0091] The bootstrap random optimization method is used to use the example bootstrap module in the task optimizer to obtain the power-related training set. The sampling instruction input x is input into the power model program for execution, and an input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples; the all values ​​include the parameters of the operation process of each submodule and the final score result;

[0092] Traverse the training set All inputs or sampling times in reach a preset threshold, and an example sample set is obtained;

[0093] A module-level prompt optimization method is adopted, and the Qwen-Plus large model is introduced as a judge. For the input and corresponding score of each submodule stored in the sample set, the judge generates new instruction inputs, which are input into the corresponding submodule and run to obtain the result score. The new instructions and corresponding scores are also stored in the sample set and classified and stored to obtain the instruction set of each submodule.

[0094] A multi-hint instruction optimization method is adopted. According to the instruction set of each sub-module, the parameter instructions and examples of each sub-module are calculated using the credit allocation component. Finally, the parameter with the highest probability is the optimal parameter.

[0095] The module-level prompt optimization method comprises the following steps:

[0096] Obtain a sample set obtained by a bootstrap random optimization method; the sample set includes several instruction inputs and the scores of each submodule after the instruction input large model is run;

[0097] Set the Qwen-Plus large model as the judge, input the instruction input and score of each submodule in the sample set to the judge, and the judge generates a new instruction input for each submodule based on the input;

[0098] Input the instruction input generated by the judge into the corresponding submodule of the large model program, and store the instruction input and the corresponding score according to different submodules to obtain the instruction set of each submodule;

[0099] The above steps are repeated based on the instruction set of each submodule, and new instruction inputs and corresponding scores are continuously generated and stored and updated in the instruction set of each submodule until the number of iterations reaches the preset maximum number of iterations.

[0100] The multi-hint instruction optimization method comprises the following steps:

[0101] The instruction set of each submodule is obtained by the module-level hint optimization method. Each module contains T instructions, and K task examples are generated for each submodule using the example bootstrapping module.

[0102] For the instructions and examples of each submodule, the credit allocation module is used to calculate the iterations. After the final convergence or reaching the preset maximum number of iterations, the parameter instruction with the highest probability is selected as the optimal parameter.

[0103] S4. Use the optimal parameters obtained in step S3 to complete the performance optimization of the large power model.

[0104] The present invention also provides a system for realizing the power large model performance optimization method based on declarative self-prompt, the structural diagram of which is shown as follows: Figure 2 As shown, it includes a mathematical modeling module, a task optimizer construction module, an optimal parameter acquisition module and a power large model optimization module;

[0105] The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads the data to the task optimizer construction module;

[0106] The task optimizer construction module optimizes the task based on the received data and the large model performance, builds the task optimizer, and uploads the data to the optimal parameter acquisition module;

[0107] The optimal parameter acquisition module uses the task optimizer to optimize the parameters based on the received data using the bootstrap random optimization method, module-level prompt optimization method, and multi-prompt instruction optimization method based on the training set to obtain the optimal parameters and upload the data to the power large model optimization module;

[0108] The power model optimization module uses the optimal parameters based on the received data to complete the performance optimization of the power model.

[0109] The method of the present invention is further described below with reference to an embodiment:

[0110] Taking the fault diagnosis task for power relay protection as an example, a large-scale diagnostic program was constructed to solve the target task. This program includes a retrieval module, a thought chain module, and a generation module. The prompt word for each module will contain three variables: instructions, examples, and target questions, where the target question is optional. The training set D uses the relay protection diagnostic knowledge base constructed by this invention.

[0111] This task uses three evaluation metrics. The retrieval module uses knowledge coverage, which measures the extent to which the text returned by the search covers the query. The thought chain module uses logic, which measures the rationality of the thought process. The generation module uses accuracy, which measures whether the generated content accurately answers the question. A task optimizer is constructed based on these evaluation metrics.

[0112] First, using a bootstrap random optimization method, we sample an input command from the training set of the relay protection diagnosis knowledge base: "CT disconnection alarm signal appears." This command is executed by the large model diagnostic program, and the parameters and final score of the retrieval module, thought chain module, and generation module are recorded. This is stored as a sample. Sampling from the training set is repeated N times to obtain N groups of samples.

[0113] The Qwen-Plus large model is then introduced as the judge. For the input instruction of the retrieval module "CT disconnection alarm signal appears" and the corresponding retrieval result score, the judge generates new instructions: "The alarm signal indicates that there is a CT disconnection", "The system detects the CT disconnection and sends an alarm signal", "What should be done when the alarm signal indicating CT disconnection appears". Similarly, new instructions are generated and run for the thinking chain module and the generation module to obtain the result score.

[0114] Finally, we used a multi-hint instruction optimization method. Based on the instruction set of each submodule, the optimizer's example bootstrapping component generated multiple examples. Taking the Thinking Chain module as an example, two examples are shown in Table 1 below. The credit assignment component calculated the parameter instructions and examples for each submodule, and the parameter with the highest probability was considered the optimal parameter.

[0115] Table 1 Two examples of sample bootstrap component generation

[0116]

Claims

1. The present invention provides a method for optimizing the performance of a large power model based on declarative self-prompting, characterized in that: The following steps are involved: S1. Mathematical modeling of large model performance optimization tasks; S2. Build a task optimizer based on large-scale model performance optimization tasks. S3. The task optimizer obtained in step S2 is optimized based on the training set using the bootstrap random optimization method, the module-level prompt optimization method, and the multi-prompt instruction optimization method to obtain the optimal parameters; S4. Using the optimal parameters obtained in step S3, complete the power model performance optimization; In step S2, the task optimizer includes an example bootstrapping module and a credit allocation module; The sample bootstrapping module uses the rejection sampling algorithm to produce the sample set; The credit allocation module determines the contribution of specific hyperparameters to the power model program based on the example sample set, implements credit allocation for different hyperparameter combinations, and completes the performance optimization of the power model.

2. The power large model performance optimization method based on declarative self-prompt according to claim 1 is characterized in that: Step S1 includes the following steps: Define the composition and operation process of the power model; Based on the data definition of the large power model, mathematical modeling is performed on the large model performance optimization task.

3. The method for optimizing the performance of a large power model based on declarative self-prompting according to claim 2 is characterized in that: Step S1 is specifically as follows: Define the power model program as , consists of m submodules; each module i is protected by a prompt word template with a set of variables v Define; define V as a large model program A collection of all variables used for prompt word templates in Indicates that the program runs under the assignment condition of assigning a string S from the variable set V; each prompt word template p includes 1 instruction variable and K example variables, and the other variables are set to constants; Based on the definition of the power big model, the goal of the big model performance optimization task is: In the training set D composed of In the indicator performance on; The goal of the large model performance optimization task uses the objective function express: in, is to maximize the function value under the assignment condition of the string S from the variable set V; is the modulus of the training set D; x is an input data in the input data set X; Metadata set A metadata in .

4. The method for optimizing the performance of a large power model based on declarative self-prompting according to claim 1 is characterized in that: The operation of the example bootstrap module is as follows: From the power-related training set The sampling input x is input into the power model program to run, and the input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples; Traverse the training set When all inputs or sampling times in reach the preset threshold, a valid sample set is obtained.

5. The method for optimizing the performance of a large power model based on declarative self-prompt according to claim 1 is characterized in that: The operation of the credit allocation module is specifically as follows: Define the credit allocation target EI, using the following formula: in, is the observation threshold; is the objective function value; Given the hyperparameters Lower objective function value The conditional probability distribution of ; For a given hyperparameter Lower credit allocation target value; The training set Divided into two groups according to the preset ratio and ; Among them, the training set Sort by the objective function value from large to small. For the front The hyperparameter combination of For the remaining The hyperparameter combination of and The probability is modeled using the following formula: in, for Kernel density estimation of ; for Kernel density estimation of ; for and The proportion of training set D; is the probability distribution function of the objective function value; The EI value is calculated explicitly using the following formula: in, yes The mean of Select The largest hyperparameter combination x is used as the next evaluation point, and iterative calculation is performed to finally obtain the optimal hyperparameter combination.

6. The method for optimizing the performance of a large power model based on declarative self-prompt according to claim 1, characterized in that: Step S3 is specifically as follows: The bootstrap random optimization method is used to use the example bootstrap module in the task optimizer to obtain the power-related training set. The sampling instruction input x is input into the power model program for execution, and an input-output trajectory of each submodule in the power model program is generated, and all values ​​on the trajectory are stored as example samples; the all values ​​include the parameters of the operation process of each submodule and the final score result; Traverse the training set All inputs or sampling times in reach a preset threshold, and an example sample set is obtained; A module-level prompt optimization method is adopted, and the Qwen-Plus large model is introduced as a judge. For the input and corresponding score of each submodule stored in the sample set, the judge generates new instruction inputs, which are input into the corresponding submodule and run to obtain the result score. The new instructions and corresponding scores are also stored in the sample set and classified and stored to obtain the instruction set of each submodule. A multi-hint instruction optimization method is adopted. According to the instruction set of each sub-module, the parameter instructions and examples of each sub-module are calculated using the credit allocation component. Finally, the parameter with the highest probability is the optimal parameter.

7. The method for optimizing the performance of a large power model based on declarative self-prompt according to claim 6, characterized in that: The module-level prompt optimization method comprises the following steps: Obtain a sample set obtained by a bootstrap random optimization method; the sample set includes several instruction inputs and the scores of each submodule after the instruction input large model is run; Set the Qwen-Plus large model as the judge, input the instruction input and score of each submodule in the sample set to the judge, and the judge generates a new instruction input for each submodule based on the input; Input the instruction input generated by the judge into the corresponding submodule of the large model program, and store the instruction input and the corresponding score according to different submodules to obtain the instruction set of each submodule; The above steps are repeated based on the instruction set of each submodule, and new instruction inputs and corresponding scores are continuously generated and stored and updated in the instruction set of each submodule until the number of iterations reaches the preset maximum number of iterations.

8. The method for optimizing the performance of a large power model based on declarative self-prompt according to claim 6, characterized in that: The multi-hint instruction optimization method comprises the following steps: The instruction set of each submodule is obtained by the module-level hint optimization method. Each module contains T instructions, and K task examples are generated for each submodule using the example bootstrapping module. For the instructions and examples of each submodule, the credit allocation module is used to calculate the iterations. After the final convergence or reaching the preset maximum number of iterations, the parameter instruction with the highest probability is selected as the optimal parameter.

9. A system for implementing the power large model performance optimization method based on declarative self-prompt according to any one of claims 1 to 8, characterized in that: It includes mathematical modeling module, task optimizer construction module, optimal parameter acquisition module and power large model optimization module; The mathematical modeling module performs mathematical modeling on the large model performance optimization task and uploads the data to the task optimizer construction module; The task optimizer construction module optimizes the task based on the received data and the large model performance, builds the task optimizer, and uploads the data to the optimal parameter acquisition module; The optimal parameter acquisition module uses the task optimizer to optimize the parameters based on the received data using the bootstrap random optimization method, module-level prompt optimization method, and multi-prompt instruction optimization method based on the training set to obtain the optimal parameters and upload the data to the power large model optimization module; The power model optimization module uses the optimal parameters based on the received data to complete the performance optimization of the power model.

Images