Carbon emission intelligent auditing system based on large model

Abstract

The invention relates to a carbon emission intelligent auditing system based on a large model, which comprises a data layer, a model layer, an application layer and an interaction layer, and during carbon emission data auditing, the application layer adopts three-step auditing. According to the method, the mode of'application + large model + industry knowledge base 'is used, the general capability and model fine tuning of the large model, the Prompt instruction word technology and the analysis of unstructured data to push the vector embedding technology are utilized, a high-performance vectorization database is built, similarity search and semantic understanding are efficiently processed, and the carbon emission intelligent auditing system is more intelligent and higher in accuracy.

Description

Technical Field

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

[0002] Carbon emissions trading is a key policy tool for incentivizing carbon emitters to take measures to reduce emissions, promoting synergistic improvements in pollution reduction and carbon reduction, and actively addressing climate change. Therefore, effective management of carbon emitters and carbon emissions trading is crucial. Currently, emitters submit reports and various information through the carbon emissions trading management platform. Relevant departments manually review these submitted information and uploaded attachments, resulting in a cumbersome and demanding review process, low accuracy, and low efficiency.

[0003] Currently, AI-powered carbon emission auditing systems are in the exploratory stage and are largely undeveloped. Based on the aforementioned pain points, building an intelligent carbon emission auditing system with carbon emission auditing as the application scenario is crucial. This system, built on large-scale model technology and targeted at various industries, third-party verification agencies, and environmental regulatory authorities, meets the demand for precise carbon emission control. This system is crucial for ensuring the technical support of carbon emission trading. Summary of the Invention

[0004] The purpose of this invention is to propose a carbon emission intelligent audit system based on a large model. By utilizing artificial intelligence, big data analysis and professional carbon accounting knowledge, it aims to make the complex, tedious and error-prone audit process intelligent and automated, and to ensure the authenticity, accuracy and compliance of the data submitted by enterprises from the source collection to the final accounting results.

[0005] To achieve the above objectives, the present invention provides a carbon emission intelligent audit system based on a large model, comprising:

[0006] Data layer, including carbon audit big data, vector knowledge base, case library, and rule library;

[0007] The model layer is used to identify various large models for unstructured data and audit, including planning models, multimodal large models for task execution, large models for report generation, and prompt projects.

[0008] The application layer includes user applications and operational applications. The user applications are used to identify and interactively process user input data, while the operational applications are used to analyze and optimize the carbon emission reporting data obtained in real time by the data layer based on the one-way data information and interactive information received from the user application services.

[0009] The interaction layer is used to complete the interaction between the user and the system, including the device or mobile device used by the user;

[0010] When conducting carbon emission data audit, the application layer adopts a three-step audit, which includes:

[0011] First step review: By calling the large model workflow in the model layer, the planning model plans the review task, matches industry review rules and standards based on the basic enterprise information, and inputs them into the task execution multimodal large model. The task execution multimodal large model combines the review rules in the rule library to identify the attached image information and the enterprise's reported information, conducts the first step of the review of the carbon emission data, and obtains the preliminary review results.

[0012] Second step review: Based on the results of the first step review, the review results are re-evaluated and optimized through the knowledge base of rewards and penalties with a large amount of human feedback in the vector knowledge base, filtering out low-accuracy results to obtain the results of the second step review;

[0013] The third step is review: Based on the results of the second step, the report generation model generates an audit report, and the prompt is used to optimize the report format and style. After the results are output, the user application provides interactive feedback to the user interaction layer. The user manually corrects or ignores the audit results through the interaction layer, and provides feedback on key points that have not been reviewed, forming a data reflux result. The feedback result is saved in the knowledge base for model effect iteration.

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

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

[0016] Preferably, the large model includes a planning model, a task execution multimodal large model, a report generation large model, and a prompt project.

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

[0018] Step S1: The operator plays the role of system administrator, fills in the audit rules through the audit scale block, and saves the audit rules to the rule library in the data layer;

[0019] Step S2: The carbon audit data reported by the user is obtained in real time and stored in the carbon emission database. The user selects the enterprise carbon emission reporting data to be audited in the interactive layer;

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

[0021] Step S4: Based on the audit results from the first audit 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 returns the optimized audit results to the user application layer and displays them in the interaction layer;

[0022] Step S5: Based on the results of the three-step review, the report results are saved in the case library and pushed to the intelligent review module in the user application layer for display in the interactive layer;

[0023] Step S6: The user fills in the unidentified audit questions through manual processing based on the audit results displayed by the interactive layer, and saves the results to the knowledge base.

[0024] Preferably, in step S4, the step of calling the reward knowledge base and the penalty knowledge base to optimize the audit results includes:

[0025] Step S4.1: The first step of the review result will be accompanied by an accuracy rate indicator value. If the accuracy rate is lower than the user-set accuracy rate, the penalty knowledge base will be called;

[0026] Step S4.2: Perform similarity matching between the low accuracy results and the penalty knowledge base;

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

[0028] Step S4.4: calling 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, and combine them with the audit content prompt words to call the large model to adjust and re-audit the first step audit results;

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

[0031] Preferably, in step S5, the steps of synthesizing the results of the three-step review and saving the report results to the case library and simultaneously pushing them to the intelligent review module in the user application layer include:

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

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

[0034] Step S5.3: The user provides feedback on unreviewed issues at the interactive layer;

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

[0036] Based on the above technical solution, the advantages of the present invention are:

[0037] This invention uses industry data to fine-tune and reinforce the learning of a large model, and uses vector embedding technology to parse unstructured data, allowing the large model to efficiently complete report audit tasks. Furthermore, combined with prompt tuning technology, it outputs audit reports that meet requirements. During use, the carbon emissions intelligent audit system can accumulate user professional experience, generating reliable and rich training sample data over a long period of time. This in turn forms the input data for the large model, making the carbon emissions intelligent audit system more intelligent and more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0039] Figure 1 This is the architecture diagram of the carbon emission intelligent audit system;

[0040] Figure 2 A flowchart of the steps for the carbon emission intelligent audit system to audit carbon emission data;

[0041] Figure 3 is a detailed flow chart of step S4;

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

[0043] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0044] The present invention provides a carbon emission intelligent audit system based on a large model, which uses the "application + large model + industry knowledge base" model, including data layer, model layer, application layer, and interaction layer. Specifically, Figure 1 Shown, including:

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

[0046] The model layer is used to identify various large models for unstructured data and auditing, including planning models, multimodal large models for task execution, large models for report generation, and prompt projects.

[0047] The application layer includes user applications and operation applications. The user application is used to identify and interactively process data input by users, and the operation application is used to analyze and optimize the carbon emission reporting data obtained in real time by the data layer based on the one-way data information and interactive information received in the user application service.

[0048] The interaction layer is used to complete the interaction between the user and the system, including the device or mobile device used by the user.

[0049] Furthermore, when conducting carbon emission data audit, the application layer adopts a three-step audit, which includes:

[0050] First step review: By calling the big model workflow, the planning model plans the review task, matches the industry review rules and standards based on the basic information of the enterprise, and inputs them into the task execution multimodal big model. The task execution multimodal big model combines the review rules in the rule library to identify the attached image information and the enterprise's reported information, conducts the first step of the review of the carbon emission data, and obtains the preliminary review results.

[0051] Among them, the basic information of the enterprise mainly includes the enterprise name, region code, industry name and other information. During the review process, the review 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 auditing involves multiple industries and types of data. The application of vectorized database construction in carbon emission auditing is mainly to solve the audit of text, image, and stamped report data. The following are the business audit rules and standards:

[0053] 1. Review data integrity

[0054] For example, all emission units are required to complete an electricity usage report and upload the corresponding attachments. Attachment requirements include: Ⅰ. Purchased thermal power (purchased electricity) - Monthly electricity purchase invoice, electricity consumption statistics ledger, electricity usage data from the platform, etc. Any supporting documentation is acceptable; upload any one of these is sufficient; Ⅱ. Purchased green electricity - Green electricity consumption voucher + settlement statement is mandatory; Ⅲ. Self-generated green electricity - Statistics ledger, data from the platform, etc. Any supporting documentation is acceptable; upload any one of these is sufficient.

[0055] Emission units submitting the final version of the document must upload the stamped document.

[0056] 2. Review data consistency

[0057] For example, check whether 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] If you upload a stamped version of the carbon emission verification report, the cover must be stamped by the company and a third-party organization (interleaved stamp + complete official seal, a total of 4). If you upload a stamped version of the carbon emission report, the cover must be signed and stamped (interleaved stamp + complete official seal). The authenticity statement on the last page must be signed / stamped by the legal person or representative + the company's official seal.

[0060] The planning model utilizes a SFT+DPO approach for model training, with task decomposition and execution planning as core capabilities. It aims to break down complex tasks into multiple subtasks and rationally arrange execution order and resource allocation. In carbon emissions audits, the large model understands the different entities mentioned in the text and the relationships between them. Based on this understanding, the model rationally plans tasks, determines the dependencies and timelines of each subtask, and leverages relevant knowledge about carbon emissions from audit cases to enhance the rationality and accuracy of planning.

[0061] Multimodal Big Model for Task Execution: Leveraging the multimodal big model's multi-source data processing, cross-modal data understanding, and reasoning capabilities, along with knowledge augmentation technology, audits are performed. Audit tasks within the multimodal big model include not only textual information but also image recognition and text comparison. In carbon emission audit scenarios, this model not only understands the literal meaning of carbon emission regulations but also accurately parses the logical relationships and industry background within the text.

[0062] Second step review: Based on the results of the first step review, with the goal of fully covering all audit issues, semantic analysis is used to match the audit results with the audit issues. Results with low similarity are filtered out, while those with high similarity are retained. The problem detection rate of the filtered results is calculated. Problem detection rates below 80% are re-evaluated and optimized using the vector knowledge base, which contains a large amount of human feedback and rewards and penalties. This results in the second step review.

[0063] Similarity matching involves keyword-based matching of each audit result in the first step against the issues in the audit rules. Keywords and keyword counts are pre-defined in the audit rules. When matching, each audit result must meet the criteria for keyword matching and keyword coverage exceeding 90% (inclusive) to qualify and be included in the next step of the problem detection rate calculation. Conversely, results with keyword matching and keyword coverage below 90% are considered unqualified and will be filtered out.

[0064] Calculate the problem detection rate (problem detection rate = detected problems / audited problems) and set a problem detection rate limit. A problem detection rate below 80% means that 20% of the audit results were not audited. The audit results need to be re-evaluated and optimized using the reward and penalty knowledge base of a large amount of human feedback in the vector knowledge base to obtain the results of the second step of the audit.

[0065] The third step is review: Based on the results of the second step, the report generation model generates an audit report, and the prompt is used to optimize the report format and style. After the results are output, the user application provides interactive feedback to the user interaction layer. The user manually corrects or ignores the audit results through the interaction layer, and provides feedback on key points that have not been reviewed, forming a data reflux result. The feedback result is saved in the knowledge base for model effect iteration.

[0066] Adopting the above-mentioned technical solution, the present invention adopts a "three-step audit" process, namely: preliminary audit based on a large model, optimized audit based on a knowledge base, and manual review and confirmation. Among them, the large model recall based on the case base + rule base is to audit structured and unstructured data according to the audit rules; the optimized audit based on the knowledge base is a process of secondary adjustment and optimization of the preliminary audit results to improve accuracy; the third step is to output the audit report, which finally needs to be manually reviewed and confirmed, which is a correction and supplement to the audit results. Manual review and confirmation are performed by carbon emission audit professionals to ensure the accuracy of the sedimentation results and the correctness of the knowledge base data.

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

[0068] System login: Log in to the system page with your account and password. Users who pass the system authentication can access the system.

[0069] Overview Module: Dynamically displays company information and status on the map, allowing direct query by company name or filtering by reporting type or reporting time. Supports statistics on the industry to which companies belong, displaying the number of companies in each industry. Supports statistics on the number of reports by reporting type and time range: 1. Counts the number of reports pending submission, pending review, and approved; 2. Counts the number of companies audited by AI, and also displays statistical information on various types of issues identified after AI audit.

[0070] Intelligent Audit Module: The system automatically audits carbon emission reports. The audit content mainly includes: 1. Auditing data integrity: Checking whether required items in the form are missing; checking whether relevant certificates for electricity use are missed or transmitted incorrectly, and whether the carbon emission report / verification report uploaded by the enterprise is signed and stamped. 2. Auditing data consistency: Checking whether the fossil fuel consumption and electricity consumption data between different forms are consistent, and identifying data filling and statistical errors. 3. Identifying outliers: Comparing the company's historical data, identifying consumption data with a large degree of fluctuation, and marking them as outliers. After the audit is completed, the audit results are displayed in the dialogue, and the identified problems are prompted on the relevant forms to assist auditors in quickly locating the problems.

[0071] Report query module: realizes monthly reporting, initial reporting, and verification reporting information query, can filter according to company name, industry, reporting time, AI review status, and report status, display corresponding records according to the filtering conditions, and support exporting list information.

[0072] In the embodiment, specifically, the login includes an input module for the user's login account and login password; the one-picture overview module includes statistics on the industry to which the enterprise belongs, statistics on the enterprise's audit status, a map overview, and a display of the audit status; the intelligent audit module includes AI one-click audit, audit status and problem overview, and report query.

[0073] Preferably, the operation application includes a knowledge base module, a case base module, a data analysis module, and an audit rule module.

[0074] Knowledge base module: In the embodiment, specifically, the knowledge base module includes a reward knowledge base and a penalty knowledge base. The reward knowledge base stores records that users believe are correct in the review, and performs deep learning and analysis on similar review data in subsequent reviews. The penalty knowledge base stores issues that have not been reviewed. Issues that have not been reviewed are manually labeled and injected into the large model again. In subsequent reviews, the model's ability to identify and judge such issues is improved, thereby improving the accuracy of model reviews.

[0075] Case library module: Provides various types of cases for large models through prompt projects to improve the problem identification and judgment capabilities of large models.

[0076] Data Parsing Module: This module parses user-uploaded attachments, calling upon the large model embedded in the model layer to analyze the attachments and extract key features. The data parsing module can extract key audit points from attachment data based on audit rules and case studies, resulting in analysis results that better meet user needs.

[0077] Audit rule module: Users can preset audit rules in advance and combine them with prompt word engineering to obtain the first step of the audit results.

[0078] Furthermore, the large model includes a planning model (SFT+DPO), a task execution model, and a report generation model. The planning model utilizes SFT+DPO. Focused on task decomposition and execution planning, it aims to break down complex tasks into multiple subtasks and rationally arrange execution order and resource allocation. In carbon emission auditing, the model uses this to understand the different entities mentioned in the text and the relationships between them. Based on this understanding, the model rationally plans tasks, determines the dependencies and timelines for each subtask, and leverages relevant carbon emission knowledge from audit cases to improve the rationality and accuracy of the planning. The multimodal large model for task execution leverages multi-source data processing and cross-modal data understanding and reasoning capabilities. Leveraging knowledge augmentation technology, it incorporates extensive case information during pre-training. In carbon emission auditing scenarios, the model not only understands the literal meaning of carbon emission-related regulations but also leverages relevant knowledge from audit cases to more accurately parse logical relationships within the text. Prompt engineering optimizes input instructions to guide the large model to generate output that meets requirements, improving processing efficiency and quality. This model is applied to carbon emission audit report generation and data question-and-answer scenarios.

[0079] Specifically, the DPO large model is trained based on the Direct Preference Optimization (DPO) method, primarily used during the fine-tuning phase of large models. Its core approach is to directly leverage user preference data or specific preference strategies to optimize the model's output, making it more aligned with the needs of the target user. This approach does not rely on traditional supervisory signals or reward functions, but instead directly adjusts the model's output based on preference data to achieve higher user satisfaction.

[0080] Supervised Fine-Tuning (SFT) is a key step in the large-scale model training process. SFT+DPO, based on the pre-trained large-scale model, uses a high-quality, labeled dataset and supervised learning algorithms to further optimize model parameters, making it more suitable for the specific task of carbon emissions data auditing.

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

[0082] Prompt engineering refers to the technology of designing and optimizing text prompts (prompts) to guide large language models to generate more targeted responses. It is a key skill for unlocking the potential of large models and is widely used in scenarios such as natural language processing, content generation, reasoning, and code development. Prompt engineering essentially means "talking to AI in a language AI understands." Through techniques such as detailed requirements, pre-set rules, and scenario simulations, the model's capabilities are precisely directed to practical problems, much like equipping AI with a "navigation system." By optimizing input instructions (prompts), AI models are guided to generate outputs that better meet user needs.

[0083] The interaction layer includes the user's device or mobile device, which is 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 used by the user, which provides the user with a page equipped with the application and provides a portal for user data entry. The user enters data through the page to realize information transmission between the interaction layer and the application layer.

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

[0085] Step S1: The operator plays the role of system administrator, fills in the audit rules through the audit scale block, and saves the audit rules to the rule library in the data layer;

[0086] Step S2: The carbon audit data reported by the user is obtained in real time and stored in the carbon emission database. The user selects the enterprise carbon emission reporting data to be audited in the interactive layer;

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

[0088] Step S4: Based on the audit results from the first audit 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 returns the optimized audit results to the user application layer and displays them in the interaction layer;

[0089] Step S5: Based on the results of the three-step review, the report results are saved in the case library and pushed to the intelligent review module in the user application layer for display in the interactive layer;

[0090] Step S6: The user fills in the unidentified audit questions through manual processing based on the audit results displayed by the interactive layer, and saves the results to the knowledge base.

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

[0092] Step S4.1: The first step of the review result will be accompanied by an accuracy rate indicator value. If the accuracy rate is lower than the user-set accuracy rate, the penalty knowledge base will be called;

[0093] Step S4.2: Perform similarity matching between the low accuracy results and the penalty knowledge base;

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

[0095] Step S4.4: calling the reward knowledge base according to the carbon emission reporting type;

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

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

[0098] For high or moderate similarity, this invention employs vector retrieval to improve search efficiency and accuracy, ensuring users quickly access the most relevant information. The core process is as follows: User query → Multi-strategy retrieval → Data integration → Reranking → Output of TopK results.

[0099] The key to vector search technology lies in the aforementioned process. The specific technical solutions involved are well known to those skilled in the art. Here, only some of the technical implementations involved are listed, including:

[0100] BM25 algorithm: keyword relevance ranking.

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

[0102] Metadata filtering: Quickly filter by structured information.

[0103] Rerank: Comprehensive score optimization result sorting.

[0104] Further, if Figure 4 As shown, in step S5, the steps of summarizing the results of the three-step review and saving the report results to the case library and pushing them to the intelligent review module in the user application layer include:

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

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

[0107] Step S5.3: The user provides feedback on unreviewed issues at the interactive layer;

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

[0109] The present invention is an important application of large models in the field of ecological environmental carbon emission data auditing, which can significantly improve audit efficiency and audit accuracy. Through the innovative three-step audit, the method improves accuracy, superimposes the audit results on manual experience and precipitates them into cases, enriches the large model training data type, and improves the application effect of the large model.

[0110] Through the technical solution of the present invention, the audit effect of the carbon emission intelligent audit system based on the large model is calculated, and the accuracy rate reaches about 75%. The accuracy rate of the large model can be further improved in the future through the accumulated feedback data.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention. They should all be included in the scope of the technical solution for protection of the present invention.

Claims

1. A carbon emission intelligent audit system based on a large model, characterized by: include: Data layer, including carbon audit big data, vector knowledge base, case library, and rule library; The model layer is used to identify various large models for unstructured data and audit, including planning models, multimodal large models for task execution, and large models for report generation; The application layer includes user applications and operational applications. The user applications are used to identify and interactively process user input data, while the operational applications are used to analyze and optimize the carbon emission reporting data obtained in real time by the data layer based on the one-way data information and interactive information received from the user application services. The interaction layer is used to complete the interaction between the user and the system, including the device or mobile device used by the user; When conducting carbon emission data audit, the application layer adopts a three-step audit, which includes: First step of review: By calling the large model workflow in the model layer, the planning model plans the review task, matches the industry review rules and standards based on the basic information of the enterprise, and inputs them into the multimodal large model for task execution. The multimodal large model for task execution combines the review rules in the rule library to identify the attached image information and the enterprise's reported information, conducts the first step of review on the carbon emission data, and obtains the preliminary review results. Second step review: Based on the results of the first step review, the review results are re-evaluated and optimized through the knowledge base of rewards and penalties with a large amount of human feedback in the vector knowledge base, and low-accuracy results are filtered out to obtain the results of the second step review; The third step is review: Based on the results of the second step, the report generation model generates an audit report, and the prompt project is combined to optimize the report format and style. After the results are output, the user application provides interactive feedback to the user interaction layer. The user manually corrects or ignores the audit results through the interaction layer, and provides feedback on key points that have not been reviewed, forming a data reflux result. The feedback result is saved in the knowledge base for model effect iteration.

2. The carbon emission intelligent auditing system according to claim 1, characterized in that: The user application includes login, one-picture overview module, intelligent audit module, and report query module.

3. The carbon emission intelligent auditing system according to claim 1, characterized in that: The operation application includes a knowledge base module, a case library module, a data analysis module, and a rule review module.

4. The carbon emission intelligent auditing system according to claim 1, characterized in that: The planning model is trained using the SFT+DPO method; the task execution multimodal large model has the ability to process multi-source data and understand and reason across modal data; the report generation large model has the ability to generate text for the final audit report generation; the Prompt project guides the large model to generate output that meets the requirements by optimizing input instructions.

5. The carbon emission intelligent auditing system according to claim 1 is characterized by: The steps of the carbon emission intelligent audit system for auditing carbon emission data include: Step S1: The operator plays the role of system administrator, fills in the audit rules through the audit scale block, and saves the audit rules to the rule library in the data layer; Step S2: The carbon audit data reported by the user is obtained in real time and stored in the carbon emission database. The user selects the enterprise carbon emission reporting data to be audited in the interactive layer; Step S3: The application layer accepts the input data from the user interaction layer, performs the first step of auditing the target audit data, and returns the audit result; Step S4: Based on the audit results from the first audit 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 returns the optimized audit results to the user application layer and displays them in the interaction layer; Step S5: Based on the results of the three-step review, the report results are saved in the case library and pushed to the intelligent review module in the user application layer for display in the interactive layer; Step S6: The user fills in the unidentified audit questions through manual processing based on the audit results displayed by the interactive layer, and saves the results to the knowledge base.

6. The carbon emission intelligent auditing system according to claim 1, characterized in that: In step S4, the steps of calling the reward knowledge base and the penalty knowledge base to optimize the audit results include: Step S4.1: The first step of the review result will be accompanied by an accuracy rate indicator value. If the accuracy rate is lower than the user-set accuracy rate, the penalty knowledge base will be called; Step S4.2: Perform similarity matching between the low accuracy results and the penalty knowledge base; Step S4.3: Filter and penalize results with high similarity in the knowledge base; Step S4.4: calling the reward knowledge base according to the carbon emission reporting type; Step S4.5: Match the results with high similarity in the reward knowledge base, and combine them with the audit content prompt words to call the large model to adjust and re-audit the first step audit results; Step S4.6: Return the second step review result.

7. The carbon emission intelligent auditing system according to claim 1, characterized in that: In step S5, the steps of summarizing the results of the three-step review and saving the report results to the case library and pushing them to the intelligent review module in the user application layer include: Step S5.1: The application layer returns the audit results to the interaction layer, and the user manually reviews the audit results; Step S5.2: The user manually determines whether to ignore the audit result; Step S5.3: The user provides feedback on unreviewed issues at the interactive layer; Step S5.4: Save the user's feedback results to the knowledge base.

Images