A power knowledge answering method and device based on large model step-by-step collaborative solving, a terminal device, and a storage medium

Abstract

This invention discloses a method, apparatus, terminal device, and storage medium for answering power knowledge questions based on a large-model step-by-step collaborative solution, belonging to the field of natural language processing technology. The method involves: acquiring the original question text input by the user for querying power-related issues, extracting information and identifying sub-questions; constructing a first prompt word, and inputting the sub-questions and the first prompt word into a preset dependency graph generation model to generate a dependency graph; then, solving and sorting all sub-questions according to the dependency graph to obtain a sorted result, and answering each sub-question according to the sorted result to obtain several sub-answers; finally, synthesizing all sub-answers to obtain the final answer to the original question text. By implementing this invention, the problem of inaccurate question-and-answer results caused by directly processing with single-step solution logic in existing technologies, which leads to deviations from the core requirements of the question, can be solved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method, apparatus, terminal device, and storage medium for solving power knowledge based on a step-by-step collaborative solution of a large model. Background Technology

[0002] With the accelerated construction of new power systems, the continuous expansion of the power grid, and the increasing penetration rate of new energy sources, the complexity and real-time requirements of power dispatching scenarios have significantly increased. Dispatchers need to process massive amounts of operational data and respond to multi-dimensional business needs on a daily basis. Therefore, quickly obtaining professional answers and assisting decision-making through intelligent systems has become a key support for improving dispatching efficiency and ensuring the safe and stable operation of the power grid.

[0003] However, in actual production processes, problems in power dispatch scenarios often have complex characteristics such as "multiple conditions related, logical nesting, and cross-domain integration." Existing technologies usually process these problems directly with single-step solution logic, which leads to the output results deviating from the core requirements of the problem and resulting in inaccurate answers. Summary of the Invention

[0004] This invention provides a method, device, terminal equipment, and storage medium for answering power knowledge based on step-by-step collaborative solution of a large model. It can solve the problem in the prior art that the output results deviate from the core requirements of the problem and the answer results are inaccurate when the processing is carried out directly by single-step solution logic.

[0005] One embodiment of the present invention provides a method for solving power knowledge based on a step-by-step collaborative solution of a large model, comprising: Obtain the original question text input by the user for querying electricity-related issues; Entities and key information are extracted from the original question text, and semantic tags are added to the original question text based on the entities and key information to obtain the preprocessed question text. The preprocessed question text is input into the preset sub-question recognition model to extract several sub-questions corresponding to the original question text. Based on the original problem text and several sub-problems, a first prompt word is constructed to generate the dependency relationship between the sub-problems. The several sub-problems and the first prompt word are then input into a preset dependency graph generation model to generate a dependency graph that represents the solution dependency relationship between the sub-problems. In the dependency graph, the nodes represent sub-problems, and the directed edges represent the solution dependency relationship between the corresponding sub-problems. Based on the above dependency graph, all sub-problems are solved and sorted to obtain the sorting results. Then, each sub-problem is solved according to the sorting results to obtain several sub-answers. By combining all the sub-answers, the final solution to the original question text is obtained.

[0006] Furthermore, entities and key information are extracted from the original question text, including: The original question text and the preset entity prompt words are input into the preset entity extraction model so that the preset entity extraction model can extract entities from the original question text based on the preset entity prompt words. The original question text and preset key information prompts are input into the preset key information extraction model, so that the preset key information extraction model can extract key information representing the implicit inquiry points of the original question text from the original question text based on the preset key information prompts.

[0007] Furthermore, based on the aforementioned dependency graph, all subproblems are solved and sorted to obtain the sorted results, including: For each node in the above dependency graph, the number of initial directed edges pointing to the current node is counted. Based on the initial number of directed edges and the dependency graph described above, the sorting operation is repeated to obtain the sorting result. The sorting operations mentioned above include: Get all currently unsorted nodes and the current sorting queue; initially, the currently unsorted nodes are all nodes in the dependency graph above, and the initial sorting queue is empty; For each node to be sorted, obtain the current number of directed edges pointing to the current node; where the initial number of directed edges is the same as the initial number of directed edges mentioned above. The node with 0 directed edges is selected as the current selected node and added to the current sorting queue to obtain the updated sorting queue. At the same time, the current selected node is removed from the current unsorted nodes to obtain the updated unsorted nodes. In the updated sorting queue, all currently selected nodes belong to the same order. Determine whether the currently updated sorted queue contains all the nodes in the above dependency graph. If so, use the currently updated sorted queue as the sorting result. Otherwise, based on the dependency graph above, determine the current successor node corresponding to the currently selected node, and decrement the number of directed edges of the current successor node in the currently updated unsorted nodes by 1 to obtain the updated number of directed edges.

[0008] Furthermore, the above-mentioned answer synthesis of all sub-answers yields the final solution to the original question text, including: Construct an answer template based on the original question text above; Fill all the sub-answers into the answer template above to obtain the final solution result; or, All sub-answers are input into a preset answer synthesis model, so that the preset answer synthesis model generates a natural language response based on all sub-answers, and the natural language response is used as the final answer.

[0009] Furthermore, after obtaining the final solution result mentioned above, it also includes: Retrieve all historical sub-problems and count the number of times each historical sub-problem was queried; The historical sub-questions and their answers that have exceeded a preset threshold number of times are stored so that, after obtaining a new original question text, the similarity between the new original question text and all the stored historical sub-questions can be calculated based on the new original question text and all the stored historical sub-questions. Obtain the maximum similarity among all similarities. If the maximum similarity is greater than a preset threshold, retrieve the historical sub-answers of the historical sub-question corresponding to the maximum similarity, and use them as the final solution to the new original question text.

[0010] Furthermore, the training of the aforementioned pre-defined sub-problem identification model includes: Obtain several power problem samples with a first real label; wherein the first real label is used to represent the real sub-problem corresponding to the power problem sample; The above-mentioned power problem samples are input into the sub-problem identification model to be trained for iterative training until the first loss function converges, thereby generating the above-mentioned preset sub-problem identification model; In each iteration of training, a current predicted sub-problem is generated based on the current power problem sample; based on the current predicted sub-problem and the corresponding first true label, the current first loss function is calculated, and it is determined whether the current first loss function has converged; if the current first loss function has converged, the current sub-problem identification model is used as the above-mentioned preset sub-problem identification model; otherwise, the model parameters in the current sub-problem identification model are adjusted, and training continues.

[0011] Furthermore, the training of the aforementioned pre-defined dependency graph generation model includes: Based on the real sub-problems corresponding to the above power problem samples, sub-problem samples with second real labels are generated; wherein, the second real labels are used to represent the solution dependencies between the sub-problem samples; Based on the above sub-problem samples, construct a second cue word to generate the dependency relationship between the sub-problem samples; The above sub-problem samples and the second prompt word are input into the dependency graph generation model to be trained for iterative training until the second loss function converges, thereby generating the above-mentioned preset dependency graph generation model. In each iteration of training, a current predicted dependency graph is generated based on the current sub-problem samples and the second prompt word; the current second loss function is calculated based on the current predicted dependency graph and the corresponding second true label, and it is determined whether the current second loss function has converged; if it has converged, the current dependency graph generation model is used as the above-mentioned preset dependency graph generation model; otherwise, the model parameters of the current dependency graph generation model are adjusted, and training continues.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments; This invention provides a power knowledge solving device based on a large model with step-by-step collaborative solution, comprising: The module includes a question text acquisition module, a text preprocessing module, a sub-question extraction module, a dependency graph generation module, a sub-answer generation module, and a final solution result generation module. The aforementioned question text acquisition module is used to acquire the original question text input by the user for querying electricity-related issues; The text preprocessing module described above is used to extract entities and key information from the original question text, and to annotate the original question text with semantic tags based on the entities and key information, so as to obtain the preprocessed question text. The aforementioned sub-problem extraction module is used to input the preprocessed problem text into a preset sub-problem recognition model and extract several sub-problems corresponding to the original problem text. The aforementioned dependency graph generation module is used to construct a first prompt word for generating the dependency relationship between the sub-problems based on the original problem text and several sub-problems, and input the several sub-problems and the first prompt word into a preset dependency graph generation model to generate a dependency graph representing the solution dependency relationship between the sub-problems; wherein, the nodes in the dependency graph represent sub-problems, and the directed edges in the dependency graph represent the solution dependency relationship between the corresponding sub-problems; The aforementioned sub-answer generation module is used to solve and sort all sub-problems according to the aforementioned dependency graph, obtain the sorting result, and solve each sub-problem according to the sorting result to obtain several sub-answers; The aforementioned final answer generation module is used to synthesize answers from all sub-answers to obtain the final answer to the original question text.

[0013] Based on the above method embodiments, the present invention provides a corresponding terminal device embodiment; The present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power knowledge solution method based on step-by-step collaborative solution of a large model as described in any embodiment of the present invention.

[0014] Based on the above method embodiments, the present invention provides a corresponding storage medium embodiment; The present invention provides a storage medium including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power knowledge solution method based on step-by-step collaborative solution of a large model as described in any embodiment of the present invention.

[0015] The embodiments of the present invention have the following beneficial effects: This invention provides a method, apparatus, terminal device, and storage medium for solving power knowledge problems through step-by-step collaborative solution based on a large model. The method includes: acquiring the original question text input by a user for querying power-related questions; extracting entities and key information from the original question text, and labeling the original question text with semantic tags based on the entities and key information to obtain a preprocessed question text; inputting the preprocessed question text into a preset sub-question recognition model to extract several sub-questions corresponding to the original question text; constructing a first prompt word for generating dependencies between sub-questions based on the original question text and the several sub-questions, and inputting the several sub-questions and the first prompt word into a preset dependency graph generation model to generate a dependency graph representing the solution dependencies between sub-questions; wherein, nodes in the dependency graph represent sub-questions, and directed edges in the dependency graph represent the solution dependencies between corresponding sub-questions; sorting all sub-questions according to the dependency graph to obtain a sorting result, and solving each sub-question according to the sorting result to obtain several sub-answers; finally, synthesizing all sub-answers to obtain the final solution result of the original question text. Therefore, in this invention, the user's original problem is first broken down to obtain multiple sub-problems implied in the original problem, and then solved one by one in sequence, and finally the solution is synthesized. This makes the whole solution process more fully understand the core requirements of the original problem, and the solution is more complete, thus improving the accuracy of the solution. Attached Figure Description

[0016] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a power knowledge solution method based on a step-by-step collaborative solution of a large model, provided by an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of a power knowledge solving device based on a large model step-by-step collaborative solution, provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0021] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0024] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0025] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0026] See Figure 1 To address the problem that existing technologies, which directly use single-step solution logic, lead to output results that deviate from the core requirements of the problem and result in inaccurate answers, an embodiment of this invention provides a method for answering power knowledge based on a large-model step-by-step collaborative solution, comprising: Step S101: Obtain the original question text input by the user for querying electricity issues; Specifically, the original question text mentioned above is a complex question about power dispatch that the user inputs through a natural language interface. For example, the original question text could be: "Which regions have seen record-high loads on their power grid systems this year?"

[0027] Step S102: Extract entities and key information from the original question text, and label the original question text with semantic tags based on the entities and key information to obtain the preprocessed question text; Specifically, the original question text mentioned above involves some entities and key information. For example, in the question "Which regions have seen record-high loads in their power grid systems this year?", the entities involved include: time entity ("this year"), location entity ("regions"), and professional domain entity ("system load", "record-high"), etc. The key information mentioned above refers to the questions that need to be answered implicitly in the original question text, such as the core inquiry points. For example, in the original question texts listed above, the key information involved includes: core inquiry points ("Which ones"), constraints ("this year", "regions", "record-high"), and implicit comparison objects ("last year").

[0028] Specifically, based on the identified entities and key information, the original question text is annotated to obtain the preprocessed question text.

[0029] Preferably, by adding semantically rich representations to the original question text, a foundation is laid for subsequent deep semantic decomposition.

[0030] In a preferred embodiment, entities and key information are extracted from the original question text, including: The original question text and the preset entity prompt words are input into the preset entity extraction model so that the preset entity extraction model can extract entities from the original question text based on the preset entity prompt words. Specifically, the aforementioned pre-defined entity extraction model is a large language model. This model extracts entities from the original question text using pre-defined entity prompts. For example, the pre-defined entity prompts could be: "Please perform named entity recognition on the following question and label it by category. Named entity categories: 1. Time entities: words representing time, such as year, time period, etc.; 2. Location entities: words representing geographical regions, such as province, city, etc.; 3. Professional domain entities: professional terms representing specific industries or technologies. Input question: Which regions have seen record-breaking loads in their power grid systems this year?"

[0031] The original question text and preset key information prompts are input into the preset key information extraction model so that the preset key information extraction model can extract key information representing the implicit question points of the original question text based on the preset key information prompts. Specifically, the pre-set key information extraction model is also a large language model. It extracts key information from the original question text based on pre-set key information prompts, revealing the implicit inquiry points. For example, the pre-set key information prompts could be: "Please extract key information from the following questions, categorized into three types: core inquiry points, constraints, and implicit comparison objects. Classification criteria: 1. Core inquiry points: The core demand of the question, i.e., the specific content the user wants to know; 2. Constraints: Words that limit the scope of the answer, such as time, location, attributes, etc.; 3. Implicit comparison objects: Comparison objects not explicitly mentioned in the question but which can be inferred. Input question: Which regions' power grid systems have reached record high loads this year?"

[0032] Specifically, the identified entities and key information are output in a structured form.

[0033] In this preferred embodiment, information is extracted from the original question text to obtain the preprocessed question text.

[0034] Step S103: Input the preprocessed question text into the preset sub-question recognition model to extract several sub-questions corresponding to the original question text; Specifically, because the original question text implicitly contains sub-questions requiring multiple steps to solve, it is necessary to analyze its grammatical structure and semantic connotation beforehand, breaking down the question into multiple sub-questions and analyzing and solving them sequentially. For example, the original question text "Which regions have achieved record-high power grid system loads this year?" can be broken down into the following sub-questions: Q1: "Which cities / prefectures are included in each region?" This is a sub-question of basic geographical knowledge; Q2: "What was the maximum system load for each city / prefecture last year?" This is a sub-question requiring historical data lookup; Q3: "What is the maximum system load for each city / prefecture this year?" This is a sub-question requiring current data lookup; Q4: "For each city / prefecture, is its maximum system load this year greater than last year's maximum?" This is a sub-question requiring comparative calculation; Q5: "List all cities / prefectures that meet the condition 'this year's load > last year's load'." This is a sub-question requiring result summarization.

[0035] Preferably, after these sub-problems are broken down, atomic problems are generated. These atomic problems are complete in themselves and can be solved directly by querying a knowledge base, database, or calling a simple computational model.

[0036] Preferably, semantic understanding technology is used to automatically identify the hidden sub-problems and their logical levels in complex problems, thereby achieving accurate decomposition of multi-condition, nested or cross-domain problems and overcoming the limitation of existing technologies that can only handle simple problems.

[0037] In a preferred embodiment, the training of the aforementioned preset sub-problem identification model includes: Obtain several power problem samples with a first real label; wherein the first real label is used to represent the real sub-problem corresponding to the power problem sample; Specifically, the above-mentioned power problem samples were all obtained from power dispatch-related texts (such as power technical documents, dispatch logs, etc.).

[0038] The above-mentioned power problem samples are input into the sub-problem identification model to be trained for iterative training until the first loss function converges, thereby generating the above-mentioned preset sub-problem identification model; In each iteration of training, a current predicted sub-problem is generated based on the current power problem sample; based on the current predicted sub-problem and the corresponding first true label, the current first loss function is calculated, and it is determined whether the current first loss function has converged; if the current first loss function has converged, the current sub-problem identification model is used as the above-mentioned preset sub-problem identification model; otherwise, the model parameters in the current sub-problem identification model are adjusted, and training continues.

[0039] Specifically, the cross-entropy loss function is used as the first loss function mentioned above, and its expression is: In the formula, Let T represent the value of the first loss function, T represent the total time length covered by the problem, and t represent the t-th time period. This represents the probability, as predicted by the model, that the actual subproblem occurs under the predicted subproblem x. Represents the true subproblem at the current moment. Indicates model parameters, This represents all subproblems up to the current time t.

[0040] Preferably, by decomposing complex problems into sub-problems, accurate decomposition of complex problems is achieved, avoiding irrelevant answers. Secondly, by integrating professional knowledge from the power dispatching field for model training, the model can deeply analyze the semantic connotation and implicit logic of the problem, accurately decompose it into atomic sub-problems covering the core requirements, avoid the deviation in answers caused by incomplete decomposition in existing technologies, and also provide an efficient foundation for subsequent collaborative solving, improving the overall solution efficiency and accuracy.

[0041] In this preferred embodiment, the preset sub-problem identification model is trained using several power problem samples with first real labels.

[0042] Step S104: Based on the original problem text and several sub-problems, construct a first prompt word for generating the dependency relationship between the sub-problems, and input the several sub-problems and the first prompt word into a preset dependency graph generation model to generate a dependency graph representing the solution dependency relationship between the sub-problems; wherein, the nodes in the dependency graph represent sub-problems, and the directed edges in the dependency graph represent the solution dependency relationship between the corresponding sub-problems; Specifically, the first prompt word mentioned above guides the preset dependency graph generation model to generate a dependency graph. A dependency graph is a directed acyclic graph where nodes are subproblems, and directed edges represent the "dependency" relationship between connected nodes. Illustratively, if "Q1→Q2" exists in the dependency graph, it means that subproblem Q2 depends on Q1; that is, to solve Q2, Q1 must first be solved.

[0043] Indicatively, the first prompt word mentioned above could be: #Given: Original problem: {Q} After problem analysis, the following subproblems are obtained: {Q1, Q2, ..., Qn}.

[0044] #Refer to rule Q3, which requires Q1 and Q2 to derive; Q4 and Q5, which require Q3 to derive; ...

[0045] #Task: Based on the above known information and reference rules, please represent the dependencies of the known subproblems using a directed acyclic graph.

[0046] #Requirements: Return results strictly according to the template format below, without any analysis process: "Q1→Q2, Q2→Q3, Q3→Q4, ...". The dependency graph information in the returned results must cover all sub-problems without omission.

[0047] As an illustration, among the five sub-questions broken down from "Which regions have seen record-high loads in their power grid systems this year?", the solution to Q4 depends on the results of Q2 (last year's load) and Q3 (this year's load), while the solutions to Q2 and Q3 depend on the results of Q1 (the list of cities). Finally, the solution to Q5 depends entirely on the results of Q4. Therefore, the model can construct a dependency graph by analyzing these contents.

[0048] Preferably, by constructing a dependency graph between sub-problems, the data flow and logical connections between steps are clarified, avoiding reasoning errors or contradictory answers caused by missing dependencies, thus enhancing the logical coherence and reliability of the solution process.

[0049] In a preferred embodiment, training the aforementioned preset dependency graph generation model includes: Based on the real sub-problems corresponding to the above power problem samples, sub-problem samples with second real labels are generated; wherein, the second real labels are used to represent the solution dependencies between the sub-problem samples; Based on the above sub-problem samples, construct a second cue word to generate the dependency relationship between the sub-problem samples; Specifically, each power problem sample has a set of real subproblems. Therefore, this set of real subproblems can generate a second real label to represent the solution dependencies between real subproblems within the set, as well as a second cue word.

[0050] The above sub-problem samples and the second prompt word are input into the dependency graph generation model to be trained for iterative training until the second loss function converges, thereby generating the above-mentioned preset dependency graph generation model. In each iteration of training, a current predicted dependency graph is generated based on the current sub-problem samples and the second prompt word; the current second loss function is calculated based on the current predicted dependency graph and the corresponding second true label, and it is determined whether the current second loss function has converged; if it has converged, the current dependency graph generation model is used as the above-mentioned preset dependency graph generation model; otherwise, the model parameters of the current dependency graph generation model are adjusted, and training continues.

[0051] Specifically, the second loss function mentioned above is also a type of cross-entropy loss function, and its expression is as follows: In the formula, This represents the value of the second loss function. This represents the probability predicted by the model of the corresponding second true label appearing under the predicted dependency graph x'. The second real label represents the current moment. 'Represents the model parameters for generating the dependency graph model. This represents all sub-problem samples up to the current time t.

[0052] In this preferred embodiment, the training of the preset dependency graph generation model is achieved based on the sub-question samples with second real labels and the corresponding second prompt words.

[0053] Step S105: Solve and sort all sub-problems according to the above dependency graph, obtain the sorting result, and solve each sub-problem according to the sorting result to obtain several sub-answers; Specifically, the dependency graph represents the solution dependencies between subproblems. In order to solve the problem more efficiently, these subproblems need to be sorted for solution. For example, in the sorting results, subproblems that belong to the same solution order can be solved in parallel to reduce solution time and speed up the overall solution process.

[0054] Specifically, during the solution process, different solvers are invoked for different sub-problems. This solver is a general-purpose solver applicable to the field of power dispatching. Taking the aforementioned sub-problems as an example, the solution steps can generally be divided into: a database query interface (for querying historical load data), a knowledge graph query engine (for querying static knowledge such as regional information), a computational model (for performing comparisons and arithmetic operations), and another dedicated simple question-and-answer language model (for answering factual questions).

[0055] Preferably, during the solution process, the solution result of each subproblem can be stored in a shared cache database after it is generated, and its corresponding subproblem's hash value or semantic code can be used as the key for querying. This avoids redundant calculations; that is, when subsequent subproblems or future user questions require the same intermediate result, it can be directly read from the cache, greatly improving efficiency. Secondly, it also supports breakpoint continuation; if the solution process of the original complex problem is interrupted, the intermediate results in the cache can be retained, without having to start from the beginning. Finally, it can serve as known items to support subsequent solutions. For example, after the answer A1 (the list of cities) of Q1 is cached, it becomes a "known condition" or input parameter for solving Q2 and Q3. In this way, redundant calculations can be reduced, and the cached content can serve as a historical knowledge base for quick response to similar questions, improving the overall efficiency of the system.

[0056] Preferably, by understanding the semantics of the problem and based on the dependencies, a problem-solving mode that allows for both serial and parallel processing can be implemented simultaneously, significantly improving the problem-solving speed.

[0057] In a preferred embodiment, the above-mentioned sorting of all subproblems based on the dependency graph to obtain the sorting result includes: For each node in the above dependency graph, the number of initial directed edges pointing to the current node is counted. Specifically, the "initial number of directed edges pointing to the current node" mentioned above is the in-degree of the corresponding node.

[0058] Based on the initial number of directed edges and the dependency graph described above, the sorting operation is repeated to obtain the sorting result. The sorting operations mentioned above include: Get all currently unsorted nodes and the current sorting queue; initially, the currently unsorted nodes are all nodes in the dependency graph above, and the initial sorting queue is empty; For each node to be sorted, obtain the current number of directed edges pointing to the current node; where the initial number of directed edges is the same as the initial number of directed edges mentioned above. The node with 0 directed edges is selected as the current selected node and added to the current sorting queue to obtain the updated sorting queue. At the same time, the current selected node is removed from the current unsorted nodes to obtain the updated unsorted nodes. In the updated sorting queue, all currently selected nodes belong to the same order. Determine whether the currently updated sorted queue contains all the nodes in the above dependency graph. If so, use the currently updated sorted queue as the sorting result. Otherwise, based on the dependency graph above, determine the current successor node corresponding to the currently selected node, and decrement the number of directed edges of the current successor node in the currently updated unsorted nodes by 1 to obtain the updated number of directed edges.

[0059] Specifically, when adding the currently selected node to the sorting queue, it is done according to the iterative batch. That is to say, the generated sorting queue is not necessarily arranged in different orders for each subproblem. If the number of currently selected nodes is not less than 2, then in the queuing sequence, this batch of currently selected nodes is arranged in the same order in a synchronous form. Therefore, in the sorting result, each subproblem has a corresponding order, and if the orders are the same, they can be solved at the same time.

[0060] To illustrate, if there are five subproblems: Q1, Q2, Q3, Q4, and Q5, and according to the dependency graph, initially, Q1 has an in-degree of 0 (no dependency), Q2 has an in-degree of 1 and depends on Q1, Q3 has an in-degree of 1 and depends on Q1, Q4 has an in-degree of 2 and depends on both Q2 and Q3, and Q5 has an in-degree of 1 and depends on Q4, then during the first sorting operation, since Q1 has an in-degree of 0, Q1 is added to the sorting queue as the currently selected node, resulting in the updated sorting queue. Then, the in-degrees of the subproblems dependent on Q1 (i.e., Q2 and Q3) are decremented by 1, and the subproblems to be sorted are updated to Q2, Q3, Q4, and Q5 before entering the next loop operation. During the second sorting operation, Q2 and Q3 have an in-degree of 0. Therefore, they are added to the sorting queue as the currently selected nodes, and since they belong to the same order, the sorting queue is updated. Then, the in-degree of Q4 is decremented by 2 (because the in-degree needs to be decremented by 1 when adding Q2, and also when adding Q3). The subproblems to be sorted are updated to Q4 and Q5, and the next loop operation begins. This process continues until all subproblems have been added to the sorting queue, resulting in the sorted result.

[0061] In this preferred embodiment, all subproblems are sorted using a dependency graph to obtain the sorted order of solutions.

[0062] Step S106: Combine all sub-answers to obtain the final solution to the original question text.

[0063] Specifically, through the aforementioned steps, the answers to each sub-question are obtained. Therefore, to obtain the answer to the original question required by the user, the answers to these sub-questions need to be synthesized.

[0064] Preferably, after replying to the user with the final solution, the solution path can be optimized based on the user's feedback (such as acceptance rate and satisfaction level) to improve the accuracy and reliability of the answer.

[0065] Preferably, by comprehensively aggregating the solutions to each sub-problem, and combining domain knowledge (such as historical power grid load data) for consistency verification and logical integration, a complete, accurate, and interpretable answer to complex problems can be generated, avoiding one-sided or fragmented responses.

[0066] In a preferred embodiment, the above-described process of synthesizing answers from all sub-answers to obtain the final solution to the original question text includes: Construct an answer template based on the original question text above; Fill all the sub-answers into the answer template above to obtain the final solution result; or, All sub-answers are input into a preset answer synthesis model, so that the preset answer synthesis model generates a natural language response based on all sub-answers, and the natural language response is used as the final answer.

[0067] Specifically, the final answer can be obtained through two methods: template filling or natural language generation. For template filling, a predefined template (i.e., the answer template mentioned above) is used to fill in the answers to the sub-questions to obtain the final answer. For example, the answer template for the original question text Q could be "According to the query, the regions with record-high system load this year are: [X]." Replacing [X] with the sub-answer [A5] corresponding to sub-question Q5 will yield the final answer: "According to the query, the regions with record-high system load this year are: [A5]..."

[0068] Specifically, for natural language generation, the model utilizes the generation capabilities of a large language model, such as a pre-defined answer synthesis model, to smoothly organize structured sub-answers into a concise and professional natural language response presented to the user. For example, given the answers to the sub-questions {Q1, Q2, Q3, Q4, Q5} as {A1, A2, A3, A4, A5}, the large model would compile the final solution to the original question text Q as: "According to the query, the cities that have achieved record-high system loads this year are: [A5]..."

[0069] In this preferred embodiment, the final solution to the original question text is obtained by synthesizing the answers from all the sub-answers.

[0070] In another preferred embodiment, after obtaining the above final solution result, the method further includes: Retrieve all historical sub-problems and count the number of times each historical sub-problem was queried; The historical sub-questions and their answers that have exceeded a preset threshold number of times are stored so that, after obtaining a new original question text, the similarity between the new original question text and all the stored historical sub-questions can be calculated based on the new original question text and all the stored historical sub-questions. Obtain the maximum similarity among all similarities. If the maximum similarity is greater than a preset threshold, retrieve the historical sub-answers of the historical sub-question corresponding to the maximum similarity, and use them as the final solution to the new original question text.

[0071] Specifically, if a certain historical sub-problem is found to be queried multiple times, this historical sub-problem and its corresponding historical sub-answer can be stored. When a similar problem is encountered in the future, this historical sub-answer can be directly called to provide a response, thereby achieving dynamic optimization of the solution path.

[0072] In this preferred embodiment, by storing frequently asked historical sub-questions, the corresponding historical sub-answers can be directly retrieved as answers to the same questions encountered in the future.

[0073] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0074] like Figure 2 As shown, one embodiment of the present invention provides a power knowledge solving device based on a large model step-by-step collaborative solution, comprising: The module includes a question text acquisition module, a text preprocessing module, a sub-question extraction module, a dependency graph generation module, a sub-answer generation module, and a final solution result generation module. The aforementioned question text acquisition module is used to acquire the original question text input by the user for querying electricity-related issues; Specifically, the original question text mentioned above is a complex question about power dispatch that the user inputs through a natural language interface. For example, the original question text could be: "Which regions have seen record-high loads on their power grid systems this year?"

[0075] The text preprocessing module described above is used to extract entities and key information from the original question text, and to annotate the original question text with semantic tags based on the entities and key information, so as to obtain the preprocessed question text. Specifically, the original question text mentioned above involves some entities and key information. For example, in the question "Which regions have seen record-high loads in their power grid systems this year?", the entities involved include: time entity ("this year"), location entity ("regions"), and professional domain entity ("system load", "record-high"), etc. The key information mentioned above refers to the questions that need to be answered implicitly in the original question text, such as the core inquiry points. For example, in the original question texts listed above, the key information involved includes: core inquiry points ("Which ones"), constraints ("this year", "regions", "record-high"), and implicit comparison objects ("last year").

[0076] Specifically, based on the identified entities and key information, the original question text is annotated to obtain the preprocessed question text.

[0077] Preferably, by adding semantically rich representations to the original question text, a foundation is laid for subsequent deep semantic decomposition.

[0078] The aforementioned sub-problem extraction module is used to input the preprocessed problem text into a preset sub-problem recognition model and extract several sub-problems corresponding to the original problem text. Specifically, because the original question text implicitly contains sub-questions requiring multiple steps to solve, it is necessary to analyze its grammatical structure and semantic connotation beforehand, breaking down the question into multiple sub-questions and analyzing and solving them sequentially. For example, the original question text "Which regions have achieved record-high power grid system loads this year?" can be broken down into the following sub-questions: Q1: "Which cities / prefectures are included in each region?" This is a sub-question of basic geographical knowledge; Q2: "What was the maximum system load for each city / prefecture last year?" This is a sub-question requiring historical data lookup; Q3: "What is the maximum system load for each city / prefecture this year?" This is a sub-question requiring current data lookup; Q4: "For each city / prefecture, is its maximum system load this year greater than last year's maximum?" This is a sub-question requiring comparative calculation; Q5: "List all cities / prefectures that meet the condition 'this year's load > last year's load'." This is a sub-question requiring result summarization.

[0079] Preferably, after these sub-problems are broken down, atomic problems are generated. These atomic problems are complete in themselves and can be solved directly by querying a knowledge base, database, or calling a simple computational model.

[0080] Preferably, semantic understanding technology is used to automatically identify the hidden sub-problems and their logical levels in complex problems, thereby achieving accurate decomposition of multi-condition, nested or cross-domain problems and overcoming the limitation of existing technologies that can only handle simple problems.

[0081] The aforementioned dependency graph generation module is used to construct a first prompt word for generating the dependency relationship between the sub-problems based on the original problem text and several sub-problems, and input the several sub-problems and the first prompt word into a preset dependency graph generation model to generate a dependency graph representing the solution dependency relationship between the sub-problems; wherein, the nodes in the dependency graph represent sub-problems, and the directed edges in the dependency graph represent the solution dependency relationship between the corresponding sub-problems; Specifically, the first prompt word mentioned above guides the preset dependency graph generation model to generate a dependency graph. A dependency graph is a directed acyclic graph where nodes are subproblems, and directed edges represent the "dependency" relationship between connected nodes. Illustratively, if "Q1→Q2" exists in the dependency graph, it means that subproblem Q2 depends on Q1; that is, to solve Q2, Q1 must first be solved.

[0082] As an illustration, among the five sub-questions broken down from "Which regions have seen record-high loads in their power grid systems this year?", the solution to Q4 depends on the results of Q2 (last year's load) and Q3 (this year's load), while the solutions to Q2 and Q3 depend on the results of Q1 (the list of cities). Finally, the solution to Q5 depends entirely on the results of Q4. Therefore, the model can construct a dependency graph by analyzing these contents.

[0083] The aforementioned sub-answer generation module is used to solve and sort all sub-problems according to the aforementioned dependency graph, obtain the sorting result, and solve each sub-problem according to the sorting result to obtain several sub-answers; The aforementioned final answer generation module is used to synthesize answers from all sub-answers to obtain the final answer to the original question text.

[0084] In a preferred embodiment, the text preprocessing module includes: Entity extraction unit and key information extraction unit; The aforementioned entity extraction unit is used to input the aforementioned original question text and preset entity prompt words into the preset entity extraction model, so that the preset entity extraction model extracts entities from the aforementioned original question text based on the aforementioned preset entity prompt words; Specifically, the aforementioned pre-defined entity extraction model is a large language model that extracts entities from the original question text using pre-defined entity prompts.

[0085] The key information extraction unit is used to input the original question text and preset key information prompts into the preset key information extraction model, so that the preset key information extraction model can extract key information representing the implicit inquiry point of the original question text from the original question text based on the preset key information prompts.

[0086] Specifically, the pre-defined key information extraction model is also a large language model. Based on pre-defined key information prompts, it extracts key information about the implicit question points from the original question text. The identified entities and key information can be output in a structured form.

[0087] In another preferred embodiment, the above-mentioned sub-answer generation module includes: Initial directed edge count unit and sorting unit; The aforementioned initial directed edge count unit is used to count the number of initial directed edges pointing to the current node for each node in the aforementioned dependency graph. Specifically, the "initial number of directed edges pointing to the current node" mentioned above is the in-degree of the corresponding node.

[0088] The sorting unit described above is used to repeatedly perform the sorting operation based on the initial number of directed edges and the dependency graph described above, to obtain the sorting result described above. The sorting operations mentioned above include: Get all currently unsorted nodes and the current sorting queue; initially, the currently unsorted nodes are all nodes in the dependency graph above, and the initial sorting queue is empty; For each node to be sorted, obtain the current number of directed edges pointing to the current node; where the initial number of directed edges is the same as the initial number of directed edges mentioned above. The node with 0 directed edges is selected as the current selected node and added to the current sorting queue to obtain the updated sorting queue. At the same time, the current selected node is removed from the current unsorted nodes to obtain the updated unsorted nodes. In the updated sorting queue, all currently selected nodes belong to the same order. Determine whether the currently updated sorted queue contains all the nodes in the above dependency graph. If so, use the currently updated sorted queue as the sorting result. Otherwise, based on the dependency graph above, determine the current successor node corresponding to the currently selected node, and decrement the number of directed edges of the current successor node in the currently updated unsorted nodes by 1 to obtain the updated number of directed edges.

[0089] Specifically, when the sorting unit adds the currently selected node to the sorting queue, it does so according to the iterative batch. That is to say, the generated sorting queue is not necessarily arranged in different orders for each subproblem. If the number of currently selected nodes is not less than 2, then in the queuing sequence, this batch of currently selected nodes is arranged in the same order in a synchronous manner. Therefore, in the sorting result, each subproblem has a corresponding order, and if the orders are the same, they can be solved simultaneously.

[0090] To illustrate, if there are five subproblems: Q1, Q2, Q3, Q4, and Q5, and according to the dependency graph, initially, Q1 has an in-degree of 0 (no dependency), Q2 has an in-degree of 1 and depends on Q1, Q3 has an in-degree of 1 and depends on Q1, Q4 has an in-degree of 2 and depends on both Q2 and Q3, and Q5 has an in-degree of 1 and depends on Q4, then during the first sorting operation, since Q1 has an in-degree of 0, Q1 is added to the sorting queue as the currently selected node, resulting in the updated sorting queue. Then, the in-degrees of the subproblems dependent on Q1 (i.e., Q2 and Q3) are decremented by 1, and the subproblems to be sorted are updated to Q2, Q3, Q4, and Q5 before entering the next loop operation. During the second sorting operation, Q2 and Q3 have an in-degree of 0. Therefore, they are added to the sorting queue as the currently selected nodes, and since they belong to the same order, the sorting queue is updated. Then, the in-degree of Q4 is decremented by 2 (because the in-degree needs to be decremented by 1 when adding Q2, and also when adding Q3). The subproblems to be sorted are updated to Q4 and Q5, and the next loop operation begins. This process continues until all subproblems have been added to the sorting queue, resulting in the sorted result.

[0091] In another preferred embodiment, the final solution generation module includes: Answer template construction unit and response generation unit; The above answer template construction unit is used to construct an answer template based on the above original question text; The aforementioned response generation unit is used to fill all sub-responses into the aforementioned answer template to obtain the aforementioned final solution result; or, All sub-answers are input into a preset answer synthesis model, so that the preset answer synthesis model generates a natural language response based on all sub-answers, and the natural language response is used as the final answer.

[0092] Specifically, the final answer can be obtained through two methods: template filling or natural language generation. For template filling, a predefined template (i.e., the answer template mentioned above) is used to fill in the answers to the sub-questions to obtain the final answer. For example, the answer template for the original question text Q could be "According to the query, the regions with record-high system load this year are: [X]." Replacing [X] with the sub-answer [A5] corresponding to sub-question Q5 will yield the final answer: "According to the query, the regions with record-high system load this year are: [A5]..."

[0093] Specifically, for natural language generation, the model utilizes the generation capabilities of a large language model, such as a pre-defined answer synthesis model, to smoothly organize structured sub-answers into a concise and professional natural language response presented to the user. For example, given the answers to the sub-questions {Q1, Q2, Q3, Q4, Q5} as {A1, A2, A3, A4, A5}, the large model would compile the final solution to the original question text Q as: "According to the query, the cities that have achieved record-high system loads this year are: [A5]..."

[0094] In another preferred embodiment, it further includes: Answer matching module; The above-mentioned answer matching module is used to obtain all historical sub-questions after obtaining the final answer result, and to count the number of times each historical sub-question was asked. The historical sub-questions and their answers that have exceeded a preset threshold number of times are stored so that, after obtaining a new original question text, the similarity between the new original question text and all the stored historical sub-questions can be calculated based on the new original question text and all the stored historical sub-questions. Obtain the maximum similarity among all similarities. If the maximum similarity is greater than a preset threshold, retrieve the historical sub-answers of the historical sub-question corresponding to the maximum similarity, and use them as the final solution to the new original question text.

[0095] Specifically, if a certain historical sub-problem is found to be queried multiple times, this historical sub-problem and its corresponding historical sub-answer can be stored. When a similar problem is encountered in the future, this historical sub-answer can be directly called to provide a response, thereby achieving dynamic optimization of the solution path.

[0096] Based on the above method embodiments, the present invention provides corresponding terminal device embodiments.

[0097] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power knowledge solution method based on step-by-step collaborative solution of a large model as described in any embodiment of the present invention.

[0098] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the device. The aforementioned terminal devices may be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These devices may include, but are not limited to, processors and memory. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the device, connecting various parts of the device via various interfaces and lines. The aforementioned memory can be used to store the aforementioned computer programs and / or modules. The aforementioned processor implements various functions of the aforementioned device by running or executing the computer programs and / or modules stored in the aforementioned memory, and by calling data stored in the memory. The aforementioned memory may mainly include a program storage area and a data storage area, wherein the program storage area may store the operating system, at least one application program required for a function, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0099] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0100] Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the power knowledge solution method based on step-by-step collaborative solution of a large model as described in any embodiment of the present invention.

[0101] In this embodiment, the storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0102] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for solving power knowledge based on a large-scale model through step-by-step collaborative solution, characterized in that, include: Obtain the original question text input by the user for querying electricity-related issues; Entities and key information are extracted from the original question text, and semantic tags are added to the original question text based on the entities and key information to obtain the preprocessed question text; The preprocessed question text is input into a preset sub-question recognition model to extract several sub-questions corresponding to the original question text; Based on the original problem text and several sub-problems, a first prompt word is constructed to generate the dependency relationship between the sub-problems. The several sub-problems and the first prompt word are then input into a preset dependency graph generation model to generate a dependency graph representing the solution dependency relationship between the sub-problems. Nodes in the dependency graph represent sub-problems, and directed edges in the dependency graph represent the solution dependency relationship between the corresponding sub-problems. Based on the dependency graph, all sub-problems are solved and sorted to obtain a sorting result. Then, each sub-problem is solved according to the sorting result to obtain several sub-answers. The answers to all sub-answers are combined to obtain the final solution to the original question text.

2. The method for solving power knowledge based on a large model with step-by-step collaborative solution as described in claim 1, characterized in that, The extraction of entities and key information from the original question text includes: The original question text and preset entity hints are input into a preset entity extraction model, so that the preset entity extraction model extracts entities from the original question text based on the preset entity hints. The original question text and preset key information prompts are input into a preset key information extraction model, so that the preset key information extraction model can extract key information representing the implicit inquiry point of the original question text from the original question text based on the preset key information prompts.

3. The method for solving power knowledge based on a large model with step-by-step collaborative solution as described in claim 2, characterized in that, The step of solving and sorting all subproblems according to the dependency graph to obtain the sorting result includes: For each node in the dependency graph, the number of initial directed edges pointing to the current node is counted. Based on the initial number of directed edges and the dependency graph, the sorting operation is repeated to obtain the sorting result; The sorting operation includes: Get all currently unsorted nodes and the current sorting queue; where, initially, the currently unsorted nodes are all nodes in the dependency graph, and the initial sorting queue is empty; For each node to be sorted, obtain the current number of directed edges pointing to the current node; wherein, the initial number of directed edges is the initial number of directed edges. The node with 0 directed edges is selected as the current selected node and added to the current sorting queue to obtain the updated sorting queue. At the same time, the current selected node is removed from the current unsorted nodes to obtain the updated unsorted nodes. In the updated sorting queue, all currently selected nodes belong to the same order. Determine whether the currently updated sorted queue contains all nodes in the dependency graph. If so, use the currently updated sorted queue as the sorting result. Otherwise, based on the dependency graph, determine the current successor node corresponding to the currently selected node, and decrement the number of directed edges of the current successor node in the currently updated unsorted nodes by 1 to obtain the updated number of directed edges.

4. The method for solving power knowledge based on a large model with step-by-step collaborative solution as described in claim 3, characterized in that, The process of synthesizing all sub-answers to obtain the final solution to the original question text includes: Construct an answer template based on the original question text; Fill all the sub-answers into the answer template to obtain the final solution result; or, All sub-answers are input into a preset answer synthesis model, so that the preset answer synthesis model generates a natural language response result based on all sub-answers, and the natural language response result is used as the final answer result.

5. The method for solving power knowledge based on a large model through step-by-step collaborative solution as described in claim 4, characterized in that, After obtaining the final solution result, the following is also included: Retrieve all historical sub-problems and count the number of times each historical sub-problem was queried; The historical sub-questions and their answers that have appeared more than a preset threshold are stored so that, after obtaining a new original question text, the similarity between the new original question text and all the stored historical sub-questions can be calculated based on the new original question text and all the stored historical sub-questions. Obtain the maximum similarity among all similarities. If the maximum similarity is greater than a preset threshold, retrieve the historical sub-answers of the historical sub-question corresponding to the maximum similarity as the final solution result of the new original question text.

6. The method for solving power knowledge based on a large model with step-by-step collaborative solution as described in claim 5, characterized in that, The training of the preset sub-problem identification model includes: Obtain several power problem samples with a first real label; wherein, the first real label is used to represent the real sub-problem corresponding to the power problem sample; The power problem sample is input into the sub-problem identification model to be trained for iterative training until the first loss function converges, thereby generating the preset sub-problem identification model; In each iteration of training, a current predicted sub-problem is generated based on the current power problem sample; a current first loss function is calculated based on the current predicted sub-problem and the corresponding first true label, and it is determined whether the current first loss function has converged; if the current first loss function has converged, the current sub-problem identification model is used as the preset sub-problem identification model; otherwise, the model parameters in the current sub-problem identification model are adjusted, and training continues.

7. The method for solving power knowledge based on a large model through step-by-step collaborative solution as described in claim 6, characterized in that, The training of the preset dependency graph generation model includes: Based on the actual sub-problems corresponding to the power problem samples, sub-problem samples with second real labels are generated; wherein, the second real label is used to represent the solution dependency relationship between each sub-problem sample; Based on the sub-problem samples, construct a second cue word for generating the dependency relationships between the sub-problem samples; The sub-problem samples and the second prompt words are input into the dependency graph generation model to be trained for iterative training until the second loss function converges, thereby generating the preset dependency graph generation model. In each iteration of training, a current predicted dependency graph is generated based on the current sub-problem samples and the second prompt word; the current second loss function is calculated based on the current predicted dependency graph and the corresponding second true label, and it is determined whether the current second loss function has converged; if it has converged, the current dependency graph generation model is used as the preset dependency graph generation model; otherwise, the model parameters of the current dependency graph generation model are adjusted, and training continues.

8. A power knowledge solving device based on step-by-step collaborative solution of a large model, characterized in that, include: The module includes a question text acquisition module, a text preprocessing module, a sub-question extraction module, a dependency graph generation module, a sub-answer generation module, and a final solution result generation module. The question text acquisition module is used to acquire the original question text input by the user for querying power-related issues; The text preprocessing module is used to extract entities and key information from the original question text, and to annotate the original question text with semantic tags based on the entities and key information to obtain the preprocessed question text. The sub-question extraction module is used to input the preprocessed question text into a preset sub-question recognition model to extract several sub-questions corresponding to the original question text; The dependency graph generation module is used to construct a first prompt word for generating the dependency relationship between the sub-problems based on the original problem text and several sub-problems, and input the several sub-problems and the first prompt word into a preset dependency graph generation model to generate a dependency graph for representing the solution dependency relationship between the sub-problems; wherein, the nodes in the dependency graph are used to represent sub-problems, and the directed edges in the dependency graph are used to represent the solution dependency relationship between the corresponding sub-problems; The sub-answer generation module is used to solve and sort all sub-problems according to the dependency graph, obtain the sorting result, and solve each sub-problem according to the sorting result to obtain several sub-answers; The final answer generation module is used to synthesize answers from all sub-answers to obtain the final answer to the original question text.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement a power knowledge solution method based on a large model step-by-step collaborative solution as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute a power knowledge solution method based on a large model step-by-step collaborative solution as described in any one of claims 1 to 7.

Images