A large model-based carbon emission intelligent auditing system

Abstract

The application relates to a large model-based carbon emission intelligent auditing system, which comprises a data layer, a model layer, an application layer and an interaction layer; when carbon emission data is audited, the application layer adopts three-step auditing. The application uses an application+large model+industry knowledge base mode, utilizes general capacity and model fine tuning of a large model, Prompt instruction word technology, analyzes unstructured data to promote vector embedding technology, builds a high-performance vector database, efficiently processes similarity search and semantic understanding, and makes the carbon emission intelligent auditing system more intelligent and higher in accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the application of artificial intelligence technology in the field of environmental protection, in particular to a carbon emission intelligent auditing system based on a large model. BACKGROUND

[0002] Carbon emission trading is an important policy tool to encourage carbon emission units to take measures to reduce carbon emissions, promote pollution reduction and carbon reduction synergy, and actively respond to climate change. Therefore, it is important to do a good job in the management of carbon emission units and the related work of carbon emission trading. The current business is that the emission units report relevant reports and fill in various information through the carbon emission trading management platform. The information filled in and the attachments uploaded are manually checked by the competent departments. The auditing process is complicated, the auditing task is heavy, the accuracy is low, and the efficiency is low.

[0003] At present, the artificial intelligence carbon emission auditing system is in the exploratory stage and is blank. Based on the above pain points, the carbon emission intelligent auditing system is constructed based on the application scene of carbon emission auditing. According to the demand for precise control of enterprise carbon emission, the large model intelligent auditing system based on large model technology for various types of enterprises, third-party verification institutions and environmental protection supervision departments is very important for the technical support of carbon emission trading. SUMMARY

[0004] The purpose of the present application is to provide a carbon emission intelligent auditing system based on a large model. By using artificial intelligence, big data analysis and professional carbon accounting knowledge, the complex, tedious and error-prone auditing process is intelligentized and automated, and the authenticity, accuracy and compliance of the enterprise's reported data from the source to the final accounting result are ensured.

[0005] To achieve the above purpose, the present application provides a carbon emission intelligent auditing system based on a large model, comprising:

[0006] A data layer comprising carbon auditing big data, vector knowledge base, case base and rule base;

[0007] A model layer for identifying unstructured data and various large models for auditing, including planning model, task execution multi-modal model, report generation model and Prompt engineering;

[0008] An application layer comprising user application and operation application, wherein the user application is used for identifying and interacting with the data input by the user, and the operation application is used for analyzing and optimizing the carbon emission reporting data obtained by the data layer in real time according to the one-way data information and interactive information received in the user application service;

[0009] An interaction layer for completing the interaction between the user and the system, comprising devices or mobile devices used by the user;

[0010] In the process of carbon emission data auditing, the application layer adopts a three-step auditing, which includes:

[0011] The first step of auditing: by calling the large model workflow in the model layer, the planning model plans the auditing task, matches the industry auditing rules and standards according to the basic information of the enterprise, and inputs the task execution multi-modal large model; the task execution multi-modal large model combines the auditing rules in the rule library, identifies the attachment image information and the enterprise filling information, and performs the first step of auditing on the carbon emission data, to obtain the preliminary audit result;

[0012] The second step of auditing: based on the result of the first step of auditing, the result of the auditing is further judged and optimized through the reward and punishment knowledge base of a large number of human feedback in the vector knowledge base, and the results with low accuracy are filtered out, to obtain the result of the second step of auditing;

[0013] The third step of auditing: based on the result of the second step of auditing, the report generation large model generates an auditing report, and the prompt optimizes the report format and style; after the result is output, the user application performs interactive feedback to the user interaction layer, and the user manually corrects or ignores the auditing result through the interaction layer, and feeds back the un-audited points, to form a data backflow result, and the feedback result is saved to the knowledge base for model effect iteration.

[0014] Preferably, the user application includes login, one-picture overview module, intelligent auditing module, and report query module.

[0015] Preferably, the operation application includes knowledge base module, case base module, data analysis module, and rule auditing module.

[0016] Preferably, the large model includes a planning model, a task execution multi-modal large model, a report generation large model, and a prompt engineering.

[0017] Preferably, the steps of the carbon emission intelligent auditing system for auditing carbon emission data include:

[0018] Step S1: the role of the operation personnel is the system administrator role, fills in the auditing rules through the auditing scale block, and saves the auditing rules to the rule library in the data layer;

[0019] Step S2: by real-time acquisition of the carbon auditing data filled by the user, the carbon auditing data is stored in the carbon emission database, and the user selects the enterprise carbon emission filling data to be audited in the interaction layer;

[0020] Step S3: the application layer accepts the input data from the user interaction layer, performs the first step of auditing on the target auditing data, and returns the auditing result;

[0021] Step S4: Based on the feedback of the first step audit result, the knowledge base in the application layer starts to work, calls the reward knowledge base and the punishment knowledge base to optimize the audit result, returns the optimized audit result to the user application layer, and displays in the interaction layer;

[0022] Step S5: The results of the three-step audit are integrated, the report results are saved to the case library and pushed to the intelligent audit module in the user application layer, and displayed in the interaction layer;

[0023] Step S6: The user fills in the un-identified audit problems according to the audit results displayed in the interaction layer through manual processing, and saves the results to the knowledge base.

[0024] Preferably, in step S4, the step of calling the reward knowledge base and the punishment knowledge base to optimize the audit result comprises:

[0025] Step S4.1: The first step audit result will have an accuracy index value, and when the accuracy is lower than the accuracy set by the user, the punishment knowledge base is called;

[0026] Step S4.2: The low accuracy result is matched with the punishment knowledge base according to similarity;

[0027] Step S4.3: Filter the results with high similarity in the punishment knowledge base;

[0028] Step S4.4: Call the reward knowledge base according to the carbon emission reporting type;

[0029] Step S4.5: Match the results with high similarity in the reward knowledge base, combine the audit content prompt words, call the large model to adjust and re-audit the first step audit result;

[0030] Step S4.6: Return the second step audit result.

[0031] Preferably, in step S5, the step of integrating the results of the three-step audit and saving the report results to the case library while pushing them to the intelligent audit module in the user application layer comprises:

[0032] Step S5.1: The application layer returns the obtained audit result to the interaction layer, and the user manually reviews the audit result;

[0033] Step S5.2: The user manually judges whether to ignore the audit result;

[0034] Step S5.3: The user feeds back the un-audited problems in the interaction layer;

[0035] Step S5.4: Save the user's feedback results to the knowledge base.

[0036] Based on the above technical scheme, the application has the advantages of:

[0037] The application fine-tunes and reinforces learning of the large model based on industry data, analyzes unstructured data based on vector embedding technology, and efficiently completes the report auditing task by the large model; meanwhile, the prompt optimization technology is combined to output the required auditing report. In the use process, the carbon emission intelligent auditing system can deposit the professional experience of the user, form reliable and rich training sample data in the long term, and in turn form the input data of the large model, so that the carbon emission intelligent auditing system is more intelligent and has higher accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings, which are included to provide a further understanding of the application and constitute a part of this application, illustrate certain illustrative embodiments of the application and together with the description serve to explain the application. In the drawings:

[0039] Figure 1 The figure is a framework diagram of the carbon emission intelligent auditing system;

[0040] Figure 2 The figure is a step flowchart of the carbon emission intelligent auditing system for auditing carbon emission data;

[0041] Figure 3 The figure is a detailed flowchart of step S4;

[0042] Figure 4 The figure is a detailed flowchart of step S5. DETAILED DESCRIPTION

[0043] The technical scheme of the application will be further described in detail below with the aid of the drawings and embodiments.

[0044] The application provides a carbon emission intelligent auditing system based on a large model, which uses an "application+large model+industry knowledge base" mode and includes a data layer, a model layer, an application layer, and an interaction layer. Specifically, as shown in the figure, Figure 1 The figure includes:

[0045] The data layer includes carbon auditing big data, a vector knowledge base, a case base, and a rule base.

[0046] The model layer is used to identify unstructured data and various large models for auditing, including a planning model, a task execution multi-modal large model, a report generation large model, and a Prompt engineering.

[0047] The application layer includes user applications and operation applications, wherein the user applications are used to identify and interact with the data input by the user, and the operation applications are used to analyze and optimize the carbon emission reporting data obtained by the data layer in real time according to the one-way data information and interaction information received in the user application service.

[0048] An interaction layer for completing the interaction between the user and the system, including the device or mobile device used by the user.

[0049] Further, in the carbon emission data audit, the application layer adopts a three-step audit, which includes:

[0050] The first step of the audit: by calling a large model workflow, the planning model plans the audit task, matches the industry audit rules and standards according to the basic information of the enterprise, and inputs the task execution multi-modal large model; the task execution multi-modal large model combines the audit rules in the rule base, identifies the attachment image information and enterprise reporting information, and performs the first step of the audit on the carbon emission data to obtain the preliminary audit result.

[0051] Among them, the basic information of the enterprise mainly includes the enterprise name, the regional code, the industry name and other information. During the audit process, the audit rules will be matched according to the basic information of the enterprise and the industry to which it belongs.

[0052] Audit rules and standards: Carbon emission audit business involves multiple industries and multiple types of data. The application of vector database construction in carbon emission audit is mainly to solve the audit of text, picture and stamped report data. The following are the business audit rules and standards:

[0053] 1. Audit data integrity

[0054] For example, all emission units need to report the power usage table and upload the corresponding attachments. The attachment requirements are as follows: I. Purchased thermal power (purchased electricity) - monthly purchase invoice, electricity consumption statistical account, electricity consumption on the platform, etc. Any proof can be uploaded. II. Green power purchased - green power consumption certificate + settlement sheet, which must be uploaded. III. Self-generated green power - statistical account, platform data, etc. Any proof can be uploaded.

[0055] The emission unit submits the final version of the file, which needs to upload the stamped file.

[0056] 2. Audit data consistency

[0057] For example, check if the fossil fuel consumption and electricity consumption data between different forms are consistent, and identify data reporting and statistical errors.

[0058] 3. Identify data outliers

[0059] For example, upload the stamped version of the carbon emission verification report; the cover needs to be stamped by the enterprise and the third-party institution (saddle stitch + complete seal, a total of 4); upload the stamped version of the carbon emission report; the cover needs to be signed and stamped (saddle stitch + complete seal); the last page of the true statement needs to be signed and stamped by the legal person or representative + enterprise seal.

[0060] Among them, the planning model: adopts SFT+DPO mode for model training, takes task decomposition and execution planning as the core ability, and aims to decompose complex tasks into multiple sub-tasks and reasonably arrange the execution order and resource allocation. In the carbon emission audit business, the large model understands different entities mentioned in the text and their relationships. Based on these understandings, the model reasonably plans the task, determines the dependency relationship and time node of each sub-task, and fully utilizes the knowledge about carbon emissions in the audit case to improve the rationality and accuracy of the planning.

[0061] Task execution multi-modal large model: based on multi-modal large model multi-source data processing and cross-modal data understanding and reasoning ability, while taking advantage of knowledge enhancement technology to perform audit. The audit task of the task execution multi-modal large model not only contains text information, but also involves image recognition and text comparison. In the carbon emission audit business scenario, not only can the literal meaning of the carbon emission related regulations be understood, but also the logical relationship and industry background information in the text can be accurately analyzed.

[0062] Second step audit: based on the results of the first step audit, the goal is to audit all audit problems, and through semantic analysis, the similarity of the audit results and the audit problems is matched, low similarity results are filtered out, and high similarity results are retained. Calculate the problem detection rate of the filtered results, and if the problem detection rate is less than 80%, the vector knowledge base with a large number of human feedback rewards and penalties in the knowledge base is used to re-judge and optimize the audit results, and the second step audit results are obtained.

[0063] Among them, similarity matching refers to traversing and matching the keywords of each audit result in the first step with the problems in the audit rules. The keywords and the number of keywords will be built in the audit rules in advance. When matching, the keyword matching degree and the number of covered keywords of each audit result are more than 90% (including 90%), which is a qualified result and can be included in the next problem detection rate calculation. Otherwise, the keyword matching degree and the number of covered keywords are less than 90%, which is an unqualified result and needs to be filtered out.

[0064] Calculate the problem detection rate (problem detection rate = detected problem / audit problem), set the problem detection rate limit, and if the problem detection rate is less than 80%, it means that 20% of the problems in the audit results have not been detected. It needs to be re-judged and optimized through the reward and punishment knowledge base with a large number of human feedback in the vector knowledge base, and the results of the second step audit are obtained;

[0065] The third step of the review: Based on the results of the second step of the review, the report generation model generates a review report, and at the same time, the report format and style are optimized in conjunction with the prompt. After the results are output, the user application provides interactive feedback to the user interaction layer. Through the interaction layer, users can manually correct or ignore the review results, and at the same time provide feedback on the key points that were not reviewed, forming a data feedback result. The feedback result is saved to the knowledge base for model effect iteration.

[0066] Using the above technical solution, the "three-step review" adopted in this invention are: preliminary review based on a large model, optimized review based on a knowledge base, and manual review and confirmation. Specifically, the large model recall based on a case library and rule base reviews structured and unstructured data according to review rules; the optimized review based on the knowledge base is a process of secondary adjustment and optimization of the preliminary review results to improve accuracy; the third step is to output a review report, which finally requires manual review and confirmation to correct and supplement the review results. Manual review and confirmation are performed by carbon emission audit professionals to ensure the accuracy of the results and the correctness of the knowledge base data.

[0067] Preferably, the user application includes a system login module, a one-map overview module, an intelligent audit module, and a report query module.

[0068] System Login: Log in to the system page using your account and password. Users who have passed system authentication can access the system.

[0069] The overview module dynamically displays company information and status on a map, allowing users to directly search for companies by name or filter them by reporting type and reporting time. It supports industry statistics for each company, displaying the number of companies in each industry. It also supports reporting quantity statistics by reporting type and time range: 1. Statistics on the number of reports pending submission, pending review, and approved reports; 2. Statistics on the number of companies using AI review, along with statistics on various types of issues identified after AI review.

[0070] Intelligent Audit Module: The system automatically audits carbon emission reports. The audit content mainly includes: 1. Auditing data completeness: Checking whether required fields in the forms are missing; checking whether electricity usage-related vouchers are omitted or incorrectly submitted; and whether the carbon emission reports / verification reports uploaded by the company are signed and stamped. 2. Auditing data consistency: Checking whether fossil fuel consumption and electricity consumption data are consistent across different forms, and identifying data entry and statistical errors. 3. Identifying outliers: Comparing with the company's historical data, identifying consumption data with large fluctuations and marking them as outliers. After the audit is completed, the audit results are displayed in a dialog box, and identified problems are highlighted on the relevant forms to help auditors quickly locate issues.

[0071] Report query module: Enables querying of monthly, initial, and verification report information. It allows filtering by company name, industry, report time, AI review status, and report status, and displays the corresponding records based on the filtering criteria. It also supports exporting list information.

[0072] In this embodiment, the login module includes a user login account and password input module; the overview module includes statistics on the company's industry, statistics on the company's review status, a map overview, and a display of the review status; the intelligent review module includes AI one-click review, an overview of the review status and issues, and a report query.

[0073] Preferably, the operational application includes a knowledge base module, a case library module, a data parsing module, and an audit rules module.

[0074] Knowledge Base Module: In this embodiment, the knowledge base module specifically includes a reward knowledge base and a penalty knowledge base. The reward knowledge base stores records that users believe are correctly reviewed. In subsequent reviews, similar review data is processed through deep learning and analysis. The penalty knowledge base stores issues that were not reviewed. Issues that were not reviewed are manually labeled and then reinjected into the large model. In subsequent reviews, this improves the model's ability to identify and judge such issues and increases the accuracy of the model's review.

[0075] Case Library Module: Provides various types of cases for large models through the prompt project, improving the ability to identify and judge problems in large models.

[0076] The data parsing module can parse user-uploaded attachments, call upon the large models built into the model layer to analyze the attachments, and extract key features. Based on review rules and case studies, the data parsing module can extract key review points from the attachment data, resulting in analysis results that better meet user needs.

[0077] Review rules module: Users can pre-set review rules and combine them with prompt words to obtain the first step of the review results.

[0078] Furthermore, the large-scale model includes a planning model (SFT+DPO), a task execution model, and a report generation model. The planning model, employing SFT+DPO, focuses on task decomposition and execution planning, aiming to break down complex tasks into multiple sub-tasks and rationally arrange the execution order and resource allocation. In carbon emission auditing, the model leverages this understanding to comprehend different entities mentioned in the text and the relationships between them. Based on this understanding, the model rationally plans tasks, determines the dependencies and time nodes of each sub-task, and fully utilizes carbon emission-related knowledge from audit cases to improve the rationality and accuracy of the planning. The multimodal task execution model is based on multimodal data processing, cross-modal data understanding, and reasoning capabilities. It utilizes knowledge augmentation technology to integrate a large amount of case information during pre-training. In carbon emission auditing scenarios, it can not only understand the literal meaning of carbon emission-related regulations but also, with the help of relevant knowledge from audit cases, more accurately analyze the logical relationships in the text. The Prompt project optimizes input instructions to guide the large-scale model to generate outputs that meet requirements, improving processing efficiency and quality, and is applied to carbon emission audit report generation and data question-and-answer scenarios.

[0079] Specifically, the DPO large model refers to training a large model based on the Direct Preference Optimization (DPO) method, primarily used in the fine-tuning stage of large models. Its core is to directly utilize user preference data or specific preference strategies to optimize the model's output, making it more aligned with the needs of the target users. This method does not rely on traditional supervision signals or reward functions, but rather adjusts the model's generated results directly based on preference data to achieve higher user satisfaction.

[0080] Supervised Fine-Tuning (SFT) is a crucial step in the training process of large models. SFT+DPO further optimizes the model parameters based on the pre-trained large model using labeled high-quality datasets and supervised learning algorithms, making it more suitable for the specific task of carbon emission data verification.

[0081] Embedding is a technique in machine learning and deep learning that transforms discrete data (such as text, images, user IDs, etc.) into dense vectors in a continuous vector space. Its core idea is to use mathematical mapping to transform high-dimensional, sparse original data into low-dimensional, dense feature vectors, so that the distances between vectors (such as cosine similarity or Euclidean distance) can reflect the semantic or structural relationships of the original data.

[0082] Prompt engineering refers to the technique of designing and optimizing text prompts to guide large language models to generate more expected responses. It is a key skill for unlocking the potential of large models and is widely used in natural language processing, content generation, reasoning, and code development. Essentially, prompt engineering is about "communicating with AI using language that AI understands." By refining requirements, pre-setting rules, and simulating scenarios, it precisely directs the model's capabilities towards real-world problems, similar to equipping AI with a "navigation system." By optimizing input prompts, it guides AI models to generate outputs that better meet user needs.

[0083] The interaction layer includes the user's device or mobile device, used to complete the interaction between the user and the system. In some embodiments, the user interaction layer refers to the PC or mobile APP program used by the user, providing the user with a page that carries the application, and providing an entry point for the user to input data. The user enters data through the page to realize the information exchange between the interaction layer and the application layer.

[0084] like Figure 2 As shown, the steps of the carbon emission intelligent auditing system in auditing carbon emission data include:

[0085] Step S1: The operator's role is that of a system administrator. They fill in the audit rules through the audit scale block and save the audit rules to the rule base in the data layer.

[0086] Step S2: By acquiring the carbon audit data submitted by users in real time, the data is stored in the carbon emission database. Users select the carbon emission data submitted by the enterprise to be audited in the interaction layer.

[0087] Step S3: The application layer receives input data from the user interaction layer, performs the first step of review on the target review data, and returns the review result;

[0088] Step S4: Based on the audit results from the first step, the knowledge base in the application layer starts working, calling the reward knowledge base and the penalty knowledge base to optimize the audit results, and then returning the optimized audit results to the user application layer and displaying them in the interaction layer;

[0089] Step S5: Based on the results of the three-step review, save the report results to the case library and push them to the intelligent review module in the user application layer for display in the interaction layer;

[0090] Step S6: Based on the review results displayed in the interaction layer, the user manually fills in the unrecognized review questions and saves the results to the knowledge base.

[0091] like Figure 3As shown, in step S4, the steps of optimizing the audit results by calling the reward knowledge base and the penalty knowledge base include:

[0092] Step S4.1: The first step of the review will include an accuracy rate indicator. If the accuracy rate is lower than the user-set accuracy rate, the penalty knowledge base will be invoked.

[0093] Step S4.2: Match the results with low accuracy to the penalty knowledge base based on similarity.

[0094] Step S4.3: Filter and penalize results with high similarity in the knowledge base;

[0095] Step S4.4: Retrieve the reward knowledge base based on the carbon emission reporting type;

[0096] Step S4.5: Match results with high similarity in the reward knowledge base, and combine them with the prompts for the review content, call the large model to adjust and re-review the review results from the first step;

[0097] Step S4.6: Return to the second step audit result.

[0098] For determining high or relatively high similarity, this invention employs vector retrieval to improve retrieval efficiency and accuracy, ensuring users quickly obtain the most relevant information. The core process is as follows: User query → Multi-strategy retrieval → Data integration → Re-sorting → Outputting TopK results.

[0099] The key to vector retrieval technology lies in the aforementioned process, the specific technical solutions of which are well known to those skilled in the art. Only some of the technical implementation methods are listed here, including:

[0100] BM25 Algorithm: Keyword Relevance Ranking.

[0101] Vector similarity retrieval: vector calculation based on semantic matching.

[0102] Metadata filtering: Quickly filter information using structured data.

[0103] Rerank: Optimizes the ranking of results based on comprehensive scores.

[0104] Furthermore, such as Figure 4 As shown, in step S5, the steps of combining the results of the three-step review and saving the report results to the case library while simultaneously pushing them to the intelligent review module in the user application layer include:

[0105] Step S5.1: The application layer returns the obtained review results to the interaction layer, where the user manually reviews the review results;

[0106] Step S5.2: The user manually determines whether to ignore the review result;

[0107] Step S5.3: Users provide feedback on unresolved issues at the interaction layer;

[0108] Step S5.4: Save the user's feedback results to the knowledge base.

[0109] This invention represents an important application of large-scale models in the field of ecological and environmental carbon emission data verification. It can significantly improve verification efficiency and accuracy. Through an innovative three-step verification method, the accuracy is improved. The verification results are superimposed with human experience to form case studies, enriching the data types of large-scale model training and enhancing the application effect of large-scale models.

[0110] The technical solution of this invention shows that the auditing effect of the carbon emission intelligent auditing system based on a large model is measured with an accuracy of about 75%. The accuracy of the large model can be further improved by the accumulated feedback data.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A large model-based carbon emission intelligent auditing system, characterized in that: Comprise: A data layer including carbon audit big data, vector knowledge base, case base, and rule base; A model layer for identifying unstructured data and various types of audit big models, including a planning model, a task execution multi-modal big model, and a report generation big model; The planning model is trained in an SFT+DPO manner, the planning model decomposes complex tasks into multiple subtasks, reasonably arranges the execution order and resource allocation, determines the dependency relationship and time node of each subtask, and uses the carbon emission related knowledge in the audit case to improve the rationality and accuracy of the planning; The task execution multi-modal big model has multi-source data processing, cross-modal data understanding and reasoning ability, the audit task of the task execution multi-modal big model includes identification of text information, image and text comparison, understanding of the literal meaning in the carbon emission related regulations, and analysis of the logical relationship and industry background information in the text; The report generation big model has text generation capability and is used for generating the final audit report; Through Prompt engineering optimization input instruction, guide the big model to generate the output meeting the demand; An application layer including user applications and operation applications, wherein the user applications are used for identifying and interacting with user input data, and the operation applications are used for analyzing and optimizing the carbon emission reporting data obtained by the data layer in real time according to the one-way data information and interaction information received by the user application service; An interaction layer for completing the interaction between the user and the system, including the devices or mobile devices used by the user; When performing carbon emission data audit, the application layer adopts three-step audit, the three-step audit comprises: First step audit: calling the big model workflow in the model layer, planning the audit task by the planning model, matching the industry audit rules and standards according to the enterprise basic information, and inputting to the task execution multi-modal big model; The task execution multi-modal big model identifies the attachment image information and enterprise reporting information according to the audit rules in the rule base, and performs the first step audit on the carbon emission data to obtain the preliminary audit result; Wherein, the steps of calling the reward knowledge base and the punishment knowledge base to optimize the audit result comprise: Step S4.1: The first step audit result has a correct rate index value, and when the correct rate is lower than the correct rate set by the user, the punishment knowledge base is called; Step S4.2: The result with low correct rate is matched with the punishment knowledge base in similarity; Step S4.3: Filter the results with similarity greater than or equal to 90% in the punishment knowledge base; Step S4.4: Call the reward knowledge base according to the carbon emission reporting type; Step S4.5: Match the results with similarity greater than or equal to 90% in the reward knowledge base, and call the big model to adjust and re-audit the first step audit result combined with the audit content prompt word; Step S4.6: Return the second step audit result; Second step audit: based on the result of the first step audit, the reward and punishment knowledge base in the vector knowledge base is used to judge and optimize the audit result again, and the result with low accuracy is filtered to obtain the second step audit result; The third step of auditing: based on the results of the second step of auditing, the report generation large model generates an audit report, and the report format and style are optimized by the Prompt engineering; after the results are output, the user application performs interactive feedback to the user interaction layer, and the user manually corrects or ignores the audit results through the interaction layer, and feedbacks the unreviewed points to form data backflow results, and the feedback results are saved to the knowledge base for model effect iteration.

2. The carbon emission intelligent auditing system according to claim 1, wherein: The user application includes login, a general overview module, an intelligent auditing module, and a report query module.

3. The carbon emission intelligent auditing system as claimed in claim 1, wherein: The operation application includes a knowledge base module, a case base module, a data analysis module, and a rule auditing module.

4. The carbon emission intelligent auditing system as claimed in claim 1, wherein: The steps of the carbon emission intelligent auditing system for auditing carbon emission data include: Step S1: The role of the operation personnel is the system administrator role, and the auditing rules are filled in the auditing scale block and saved to the rule base in the data layer; Step S2: Real-time carbon auditing data filled by the user is obtained and stored in the carbon emission database, and the user selects the enterprise carbon emission reporting data to be audited in the interaction layer; Step S3: The application layer accepts input data from the user interaction layer, performs the first step of auditing on the target auditing data, and returns the auditing results; Step S4: Based on the auditing results fed back by the first step of auditing, the knowledge base in the application layer starts to work, calls the reward knowledge base and the punishment knowledge base to optimize the auditing results, and returns the optimized auditing results to the user application layer and displays them in the interaction layer; Step S5: The results of the three steps of auditing are integrated, the report results are saved to the case base and pushed to the intelligent auditing module in the user application layer, and displayed in the interaction layer; Step S6: The user manually processes the unreviewed auditing problems according to the audit results displayed in the interaction layer, and saves the results to the knowledge base.

5. The carbon emission intelligent auditing system as claimed in claim 1, wherein: In step S5, the step of integrating the results of the three steps of auditing, saving the report results to the case base, and pushing them to the intelligent auditing module in the user application layer includes: step S5.1: the application layer returns the obtained auditing results to the interaction layer, and the user manually reviews the auditing results; Step S5.2: The user manually judges whether to ignore the auditing results; Step S5.3: The user feeds back the unreviewed problems in the interaction layer; Step S5.4: Save the user's feedback results to the knowledge base.

Images