Greenhouse gas emission analysis method and system based on multi-modal agent

Abstract

The application discloses a greenhouse gas emission analysis method and system based on a multi-modal intelligent agent, and belongs to the fields of greenhouse gas emission data analysis, large models and multi-agents. The application firstly processes report documents, time series data and geospatial data, constructs a special knowledge base and a standardized data set; then uses a hybrid directed acyclic graph to establish a collaborative architecture containing six intelligent agents, supports parallel processing of chapter tasks; finally generates a greenhouse gas emission analysis report according to user natural language input, and can be dynamically adjusted according to modification opinions, iteratively optimized until the requirements are met. The application automatically locates the problem source based on artificial review results, optimizes three types of structured parameters, improves the quality of generated reports, realizes intelligent and professional generation of greenhouse gas emission analysis reports, and solves the problems of long report writing period, insufficient professionalism and difficult data integration in the prior art.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of greenhouse gas emission data analysis, large models and multi-agent, and particularly relates to a greenhouse gas emission analysis method and system based on multi-modal agents. BACKGROUND

[0002] With the rapid development of deep learning and artificial intelligence technology, the powerful reasoning and understanding capabilities of large language models and the powerful visual analysis capabilities of multi-modal large models provide a new technical path for intelligent analysis and automatic report generation of greenhouse gas emissions. However, the existing single large language model has technical bottlenecks such as limited output length, insufficient global planning ability, and limited multi-modal information processing ability when directly generating professional analysis reports. Multi-agent systems can effectively overcome the above limitations through task decomposition and parallel processing, significantly improving the processing efficiency and accuracy of complex environmental data analysis tasks, and laying a technical foundation for building an end-to-end intelligent analysis report generation system.

[0003] In summary, the existing greenhouse gas emission analysis methods have the following problems:

[0004] First, structured report generation relies on human work, and the standardization degree is low. Traditional methods mainly rely on manual writing of analysis reports, which not only consumes time and effort, but also easily produces subjective bias, inconsistent formats, and uneven content quality, making it difficult to meet the needs of large-scale, standardized decision support. And the existing single large language model directly generates professional reports also has multiple technical bottlenecks, which are: (1) limited output length, constrained by the upper limit of the number of tokens in a single output, unable to generate a complete long professional analysis report at once; (2) insufficient global planning ability, lack of overall control of the report structure, difficult to maintain a unified analysis framework and logical argument; (3) limited multi-modal information processing capability, unable to simultaneously process and deeply integrate text, images, data and other information types.

[0005] Second, the task decomposition and demand analysis ability is insufficient. The existing methods lack intelligent understanding and automatic decomposition ability for user complex query requirements, and cannot split comprehensive analysis tasks into executable sub-tasks, resulting in insufficient pertinence and completeness of the analysis results.

[0006] Third, the integration efficiency of multi-source information related to greenhouse gas emissions is low. Existing research mainly relies on a single data source for analysis, lacks parallel search and correlation analysis capabilities for multi-dimensional professional information such as climate policy documents, emission monitoring reports, international agreement progress, and emission reduction technology information, and cannot build a comprehensive greenhouse gas emission knowledge system.

[0007] Fourth, the greenhouse gas emission data visualization analysis capability is missing. The existing method lacks intelligent graphical generation and deep visual analysis capability for greenhouse gas emission time trend, spatial distribution pattern, and it is difficult to provide intuitive emission change rule and regional difference analysis for climate policy makers. SUMMARY

[0008] The purpose of the present application is to solve the problems existing in the prior art and provide a greenhouse gas emission analysis method and system based on a multi-modal agent.

[0009] In order to achieve the above-mentioned purpose of the application, the following technical solutions are specifically adopted:

[0010] In the first aspect, the present application provides a greenhouse gas emission analysis method based on a multi-modal agent, which comprises the following steps:

[0011] S1: Obtain report document data related to greenhouse gas emissions, greenhouse gas emission time series data, and geographic spatial data in the target area, and process each of the three data. After processing, the first data, the second data and the third data are formed, and a special knowledge base in the field of greenhouse gas emission is constructed based on the first data, and a standardized data set supporting multi-modal data visualization is constructed based on the third data;

[0012] S2: Construct a hybrid directed acyclic graph based on the logical dependency relationship and parallel processing requirements of the greenhouse gas emission analysis report generation, and construct a multi-agent collaborative architecture using the hybrid directed acyclic graph, and establish a state transfer mechanism between agents; the multi-agent collaborative architecture includes six agents, namely semantic analysis, knowledge retrieval, data visualization, image understanding, content fusion and report generation. The data visualization agent generates a visual image based on the standardized data set and / or the second data, the knowledge retrieval agent performs knowledge retrieval based on the special knowledge base, and the greenhouse gas emission analysis report sub-chapter tasks are processed in parallel by calling the knowledge retrieval agent and the content fusion agent multiple times;

[0013] S3: Obtain the natural language question related to greenhouse gas emission input by the user, call the multi-agent collaborative architecture to generate a greenhouse gas emission analysis report based on the input natural language question, and feed back the report to the user in a predetermined manner. When receiving the user input modification opinion text, a new multi-agent collaborative architecture is created to regenerate the greenhouse gas emission analysis report until the user's demand is met.

[0014] On the basis of the above-mentioned scheme, each step can be realized in the following preferred specific manner.

[0015] As a preferred embodiment of the first aspect mentioned above, in S1, the data processing procedure specifically includes: extracting text from the report document data using optical character recognition technology and cleaning it with regular expressions to generate standardized corpus, which is then used as the first data; dividing the standardized corpus into text blocks and using a large language model to perform entity recognition and relation extraction from the text blocks to construct an entity relation graph, thereby forming a dedicated knowledge base; filling missing values ​​in the greenhouse gas emission time series data by averaging before and after and calculating the annual change gradient, with the processed greenhouse gas emission time series data serving as the second data; and applying spatial indexing technology to the geospatial data to establish a mapping relationship between target area codes and geographic coordinates, with the processed geospatial data serving as the third data.

[0016] As a preferred embodiment of the first aspect mentioned above, the specific functions of each agent in the S2 multi-agent cooperative architecture are implemented as follows:

[0017] The semantic parsing agent serves as the starting agent. It employs hierarchical intent recognition and, based on a pre-designed semantic parsing function prompt word template, first extracts key entities from the input user question. Then, it combines the extracted key entities to determine the user's analysis needs through intent classification. The task of generating the greenhouse gas emission analysis report is decomposed into three types of structured parameters: visualization requirement parameters, sub-chapter task parameters, and report structure parameters.

[0018] The data visualization agent receives visualization requirement parameters based on pre-designed visualization function prompt templates, determines specific visualization tasks, and calls the corresponding image generation tools to automatically extract secondary data and / or standardized datasets to generate statistical images that meet professional standards.

[0019] The knowledge retrieval agent receives task parameters from each sub-chapter based on pre-designed retrieval function prompt word templates, executes the retrieval tasks of each sub-chapter in parallel, and uses a multi-source parallel retrieval strategy to obtain relevant background knowledge from the constructed dedicated knowledge base and real-time web page information that has been evaluated for credibility, and outputs the knowledge retrieval results of each sub-chapter.

[0020] The image understanding agent utilizes the visual-text cross-modal understanding capability of a multimodal large language model to perform in-depth analysis on generated statistical images based on pre-designed image understanding function prompt word templates. It identifies the statistical image type, extracts data trends from the statistical images, analyzes data change characteristics, forms analysis results, and converts them into structured text format. This text, along with the statistical image file path and statistical image type, serves as the output information of the image understanding agent.

[0021] The content fusion agent employs entity alignment and semantic relationship reasoning. Based on a pre-designed content fusion function prompt word template, it fuses the knowledge retrieval results of each sub-chapter with the output information of the image understanding agent. Through entity recognition and standardization processing, it unifies the entity representations from different agents. By calling multiple times to process the fusion tasks of different sub-chaps in parallel, it generates the structured content of each sub-chapter.

[0022] As a terminal agent, the report generation agent receives report structure parameters based on a pre-designed report generation function prompt template to determine the overall structure of the report. It then organizes the structured content of each sub-chapter output by the content fusion agent according to the specified chapter logical order to generate a logically coherent and complete greenhouse gas emission analysis report.

[0023] As a preferred option for the first aspect mentioned above, the visualization requirements parameters include data dimensions, image type, data range, and visualization objectives; the task parameters for each sub-chapter include chapter theme, key search terms, analysis dimensions, and output requirements; and the report structure parameters include a list of chapter titles, logical order of chapters, overall report framework, and format requirements.

[0024] As a preferred option for the first aspect mentioned above, the functional prompt word template for each agent includes role definition, task description, input format, output requirements, and example display.

[0025] As a preferred embodiment of the first aspect mentioned above, when evaluating the credibility of the i-th real-time webpage information, a preset authority weight is used to weight the authority score of the i-th real-time webpage information, a preset timeliness weight is used to weight the timeliness score of the i-th real-time webpage information, and a preset content relevance weight is used to weight the content relevance score of the i-th real-time webpage information. The weighted authority score, timeliness score, and content relevance score are added together to obtain the credibility score of the i-th real-time webpage information. When the credibility score of the i-th real-time webpage information is greater than the preset credibility score threshold, it indicates that it has passed the credibility evaluation. Among them, the timeliness score is obtained by multiplying the initial timeliness score by an exponential decay term. The exponential decay term is a power operation with the natural constant as the base, and the exponent is the negative of the time decay coefficient multiplied by the time interval.

[0026] Furthermore, the method for evaluating the credibility of real-time webpage information is shown in the following formula:

[0027]

[0028] in, Indicates authority weight. Indicates the weight of timeliness. Indicates the content relevance weight; , , and The first The scores for the authority, timeliness, content relevance, and credibility of real-time web page information are as follows:

[0029] Furthermore, the timeliness score is calculated as follows:

[0030]

[0031] in, This serves as the initial timeliness score; Indicates the exponentially decaying term; It is a natural constant; For time intervals; This is the time decay coefficient.

[0032] As a preferred embodiment of the first aspect mentioned above, in the multi-agent cooperative architecture of S2, one agent... The level value is equal to the maximum level value among all predecessor agents of this agent plus 1, as shown in the following formula:

[0033]

[0034] in, Represents intelligent agents The level value; This represents the set of all directed edges in a mixed directed acyclic graph. Indicates from the intelligent agent Pointing to intelligent agents A directed edge; Represents a set Any one of the points to the intelligent agent The directed edges; Represents intelligent agents Precursor intelligent agent The level value; Indicates taking the intelligent agent The highest level value among all predecessor agents.

[0035] As a preferred option in the first aspect mentioned above, the specific hierarchical structure of the multi-agent collaborative architecture is as follows: the first layer is a semantic parsing agent, the second layer is a data visualization agent and a knowledge retrieval agent, the third layer is an image understanding agent, the fourth layer is a content fusion agent, and the fifth layer is a report generation agent.

[0036] As a preferred embodiment of the first aspect mentioned above, in the multi-agent collaborative architecture of S2, each agent maintains four execution states: waiting state, running state, completed state, and failed state. The waiting state indicates that an agent is waiting for the predecessor agent to complete; the running state indicates that the agent is executing a task; the completed state indicates that the agent has successfully completed the task and can pass the result to the successor agent; and the failed state indicates that an error occurred during the execution of the agent.

[0037] As a preferred embodiment of the first aspect, in the multi-agent collaborative architecture of S2, the state transition mechanism between agents is as follows: when all predecessor agents of an agent have reached the completion state, the agent automatically transitions from the waiting state to the running state; when the agent successfully completes the task, it automatically notifies all its successor agents to check the pre-built startup conditions; when the agent fails to complete the task, it initiates the error propagation mechanism, sets all its successor agents to the waiting state, and triggers a retry or degradation processing strategy.

[0038] Secondly, the present invention provides a greenhouse gas emission analysis system based on a multimodal intelligent agent, comprising:

[0039] The data acquisition module is used to acquire natural language questions related to greenhouse gas emissions input by the user;

[0040] The report generation module is used to process the natural language questions acquired by the data acquisition module according to the greenhouse gas emission analysis method based on multimodal intelligent agents described in any of the above-mentioned first aspects, and generate and visualize a greenhouse gas emission analysis report that meets the user's needs.

[0041] Thirdly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0042] The memory is used to store computer programs;

[0043] The processor is configured to, when executing the computer program, implement the greenhouse gas emission analysis method based on multimodal intelligent agents as described in any of the first aspects above.

[0044] When using the method of this invention to generate a greenhouse gas emission analysis report, the following effects are achieved:

[0045] 1) This invention achieves structured and efficient generation of greenhouse gas emission analysis reports. By constructing a multi-agent collaborative architecture based on a DAG structure, this invention overcomes the bottlenecks of the output length limitation of a single large language model and insufficient global planning capabilities, solving the efficiency and quality problems of traditional manual writing methods, and realizing the generation of logically coherent and structurally complete professional analysis reports. Through the "inter-layer serial, intra-layer parallel" execution mode, the processing efficiency of multi-sub-chapter tasks is significantly improved compared with the traditional sequential execution method.

[0046] 2) This invention achieves intelligent decomposition and accurate demand identification for greenhouse gas emission analysis tasks. Through hierarchical intent recognition by a semantic parsing intelligent agent, this invention can automatically decompose complex user queries into three types of structured parameters: visualization requirements, sub-chapter tasks, and report structure. This solves the problem of existing methods lacking intelligent understanding and automatic decomposition capabilities, improving the relevance and completeness of the analysis results.

[0047] 3) This invention achieves intelligent integration and quality assurance of multi-source information in the field of greenhouse gas emissions. By constructing a dedicated knowledge base based on knowledge graphs, this invention enables the correlated querying of multi-dimensional professional information such as climate policy documents, progress of international agreements, and emission reduction technology information, solving the problems of scattered and unrelated information in traditional knowledge bases; by establishing a dynamic weighted credibility assessment mechanism, it ensures the reliability of web page information sources, significantly improving the comprehensiveness and accuracy of information acquisition compared to single data source analysis methods.

[0048] 4) This invention enables specialized multimodal analysis of greenhouse gas emission data. It generates professional statistical images through a data visualization agent, performs visual-text cross-modal analysis through an image understanding agent, and integrates multi-source information through a content fusion agent, effectively solving the problem of existing methods lacking visualization analysis and multimodal information processing capabilities. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0050] Figure 2 This is a schematic diagram of the multi-agent collaborative architecture of the method of the present invention;

[0051] Figure 3 This is a schematic diagram illustrating the intelligent generation of greenhouse gas emission analysis reports using the method of the present invention;

[0052] Figure 4 This is a trend diagram of greenhouse gas emissions generated in a certain region by the method of the present invention;

[0053] Figure 5 This is a system block diagram of the present invention;

[0054] Figure 6This is a schematic diagram of the computer electronic device components of the present invention. Detailed Implementation

[0055] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0056] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0057] like Figure 1 As shown, in a preferred embodiment of the present invention, the greenhouse gas emission analysis method based on multimodal intelligent agents includes the following steps S1 to S3. The specific implementation process of each step will be described in detail below.

[0058] S1: Acquire report documents related to greenhouse gas emissions, time-series data of greenhouse gas emissions, and geospatial data within the target area. Process each of the three types of data to form the first data, the second data, and the third data. Based on the first data, construct a dedicated knowledge base in the field of greenhouse gas emissions, and based on the third data, construct a standardized dataset that supports multimodal data visualization.

[0059] It should be noted that, in S1 of this invention, the data processing process is as follows: Text is extracted from the report document data using optical character recognition technology and cleaned using regular expressions to generate standardized corpus. This standardized corpus is then used as the first data. The standardized corpus is further divided into text blocks, and entity recognition and relation extraction are performed from the text blocks using a large language model to construct an entity relation graph, thereby forming a dedicated knowledge base. For greenhouse gas emission time-series data, missing values ​​are filled using pre- and post-average methods, and the annual change gradient is calculated. The processed greenhouse gas emission time-series data is used as the second data. For geospatial data, spatial indexing technology is applied to establish a mapping relationship between target area codes and geographic coordinates. The processed geospatial data is used as the third data.

[0060] It should be noted that there are many sources of relevant data collected in step S1 of this invention. The corpus required for the construction of the dedicated knowledge base can be relevant documents published by authoritative organizations such as the United Nations Environment Programme, the World Meteorological Organization, and the International Energy Agency. Time series data of greenhouse gas emissions of countries / regions can be collected by the World Bank and Our World in Data. World geographic boundary vector data can be collected by public map datasets such as Natural Earth. No restrictions are imposed in this invention.

[0061] In this embodiment, S1, reports related to greenhouse gas emissions issued by authoritative organizations such as the United Nations Environment Programme, the World Meteorological Organization, and the International Energy Agency are collected, along with time-series data on greenhouse gas emissions from countries / regions and world geographic boundary vector data. The data is then processed as follows: For the reports, text is extracted using Optical Character Recognition (OCR) technology and cleaned using regular expressions to generate standardized corpora, thus constructing a dedicated knowledge base for greenhouse gas emissions; for the time-series data, missing values ​​are filled using averaging and annual variation gradients are calculated to identify peak characteristics and inflection points in emission reduction trends; and for the geospatial data, a mapping relationship between country / region codes and geographic coordinates is established using spatial indexing technology. The first data refers to the processed report document data, the second data refers to the processed greenhouse gas emission time-series data, and the third data refers to the processed geospatial data.

[0062] Specifically, the data processing procedures for the above three types of data are as follows:

[0063] S11: Report document data processing. The report document data is in PDF format. It is filtered and the core content chapters are retained. Editable text format is extracted based on OCR optical character recognition technology. Redundant characters and invalid identifiers are deleted through regular expressions. Standardization processing is performed using UTF-8 unified encoding format to generate standardized corpus in the field of greenhouse gas emissions.

[0064] S12: Construct a dedicated knowledge base for the field of greenhouse gas emissions, design dedicated entity recognition rules, segment standardized corpora into text blocks, perform entity recognition and relation extraction from text blocks using a large language model, construct an entity relation graph, and thus form a dedicated knowledge base.

[0065] In this embodiment S12, considering the characteristics of greenhouse gas emissions, such as complex entity relationships (e.g., multiple associations between driving factors, countries / regions, and policy measures) and strong knowledge hierarchy (covering multiple levels such as international conventions, national policies, and industry standards), a dedicated knowledge base is constructed in the form of a knowledge graph to fully preserve the relationships between entities. Compared with traditional knowledge bases, this construction method can support associated queries and path reasoning, avoiding the problems of scattered information and lack of association in traditional knowledge bases.

[0066] S13: Processing of time-series greenhouse gas emissions data for countries / regions. Missing time points are identified based on a data integrity verification algorithm, and missing values ​​are imputed using forward and backward averaging, as shown in the following formula:

[0067]

[0068] in, Interpolation results representing missing data, This represents the value at the moment immediately preceding the missing point. This represents the value at the moment following the missing point.

[0069] The annual gradient of greenhouse gas emissions for each country / region is calculated using a gradient calculation formula, as shown in the following formula. The year in which greenhouse gas emissions peak and the inflection point of the emission reduction trend are identified by the changes in the above gradient.

[0070]

[0071] in, Represents the annual variation gradient, Represents emissions for the current year. This represents the emissions from the previous year.

[0072] S14: Geospatial data preprocessing, establishing a mapping relationship between ISO A3 country / region codes and geographic coordinates based on spatial indexing technology, and constructing a standardized dataset that supports multimodal data visualization.

[0073] S2: Based on the logical dependencies and parallel processing requirements of the greenhouse gas emission analysis report generation, a hybrid directed acyclic graph is constructed. A multi-agent collaborative architecture is built using the hybrid directed acyclic graph, and a state transition mechanism between agents is established. The multi-agent collaborative architecture includes six agents: semantic parsing, knowledge retrieval, data visualization, image understanding, content fusion, and report generation. Among them, the data visualization agent generates visualization images based on standardized datasets and / or secondary data, and the knowledge retrieval agent performs knowledge retrieval based on a dedicated knowledge base. By calling the knowledge retrieval agent and the content fusion agent multiple times, the tasks of each sub-section of the greenhouse gas emission analysis report are processed in parallel.

[0074] It should be noted that in step S2 of this invention, to address the need for structured generation of greenhouse gas emission analysis reports, a hybrid directed acyclic graph (DAG) is used to construct a multi-agent collaborative architecture, such as... Figure 2 As shown, this architecture supports a multi-level parallel execution mechanism. The data visualization agent and the knowledge retrieval agent for each sub-chapter execute in parallel; the content fusion agent for each sub-chapter processes content in parallel, achieving data flow correctness and consistency through a distributed state synchronization mechanism. Compared to traditional sequential execution or simple parallelism, the DAG structure maximizes processing efficiency while ensuring logical correctness. Furthermore, each agent in this architecture adopts a unified structural design: a reasoning engine based on a large language model, responsible for performing natural language understanding, logical reasoning, and text generation tasks; a tool call interface module, encapsulating the calling methods of external tools such as database queries, web crawling, and image rendering; and a context state manager, maintaining the intermediate results and state information of the agent during execution.

[0075] It should be noted that in step S2 of this invention, the specific functions of each agent in the multi-agent cooperative architecture are implemented as follows:

[0076] The semantic parsing agent serves as the initial agent, employing hierarchical intent recognition. Based on pre-designed semantic parsing function prompt word templates, it first extracts key entities (such as country names and time ranges) from the input user query. Then, combining the extracted key entities, it determines the user's analytical needs through intent classification, decomposing the greenhouse gas emission analysis report generation task into three types of structured parameters, namely visualization requirement parameters. Parameters of each sub-chapter task and report structure parameters ;

[0077] The data visualization agent receives visualization requirement parameters based on pre-designed visualization function prompt templates, determines specific visualization tasks, and calls the corresponding image generation tools to automatically extract secondary data and / or standardized datasets to generate statistical images that meet professional standards.

[0078] The knowledge retrieval agent receives task parameters from each sub-chapter based on pre-designed retrieval function prompt word templates, executes the retrieval tasks of each sub-chapter in parallel, and uses a multi-source parallel retrieval strategy to obtain relevant background knowledge from the constructed dedicated knowledge base and real-time web page information that has been evaluated for credibility, and outputs the knowledge retrieval results of each sub-chapter.

[0079] The image understanding agent utilizes the visual-text cross-modal understanding capability of a multimodal large language model to perform in-depth analysis on generated statistical images based on pre-designed image understanding function prompt word templates. It identifies the statistical image type, extracts data trends from the statistical images, analyzes data change characteristics, forms analysis results, and converts them into structured text format. This text, along with the statistical image file path and statistical image type, serves as the output information of the image understanding agent.

[0080] The content fusion agent employs entity alignment and semantic relationship reasoning. Based on a pre-designed content fusion function prompt word template, it fuses the knowledge retrieval results of each sub-chapter with the output information of the image understanding agent. Through entity recognition and standardization processing, it unifies the entity representations from different agents. By calling multiple times to process the fusion tasks of different sub-chaps in parallel, it generates the structured content of each sub-chapter.

[0081] As a terminal agent, the report generation agent receives report structure parameters based on a pre-designed report generation function prompt template to determine the overall structure of the report. It then organizes the structured content of each sub-chapter output by the content fusion agent according to the specified chapter logical order to generate a logically coherent and complete greenhouse gas emission analysis report.

[0082] It should be noted that in this embodiment, the starting agent is the first agent to execute the task, and the terminal agent is the last agent to execute the task.

[0083] It should be noted that, in this embodiment, a parameter passing mechanism is established between agents. Specifically, the parameter passing mechanism involves a semantic parsing agent breaking down the complex report generation task into three types of structured parameters: visualization requirement parameters, sub-chapter task parameters, and report structure parameters. The visualization requirement parameters, including data dimensions, image types, data range, and visualization objectives, are passed to the data visualization agent. The sub-chapter task parameters, including chapter topics, key search terms, analysis dimensions, and output requirements, are passed to the corresponding knowledge retrieval agent. The report structure parameters, including a list of chapter titles, the logical order of chapters, the overall report framework, and format requirements, are passed to the report generation agent.

[0084] It should be noted that this embodiment, tailored to the characteristics of the greenhouse gas emissions field and its professional knowledge system and report writing standards, designs hierarchical functional prompt word templates for each agent. These templates include multiple aspects such as role definition, task description, input format, output requirements, and example demonstrations, enabling the agents to accurately understand task requirements and generate results that meet standards. Each agent's functional prompt word template has pre-set corresponding placeholders for automatic filling, achieving function-oriented personalized task execution. In subsequent steps, the knowledge retrieval agent, data visualization agent, and report generation agent, based on their respective functional prompt word templates, receive structured parameters generated by the semantic parsing agent to execute their respective tasks; the image understanding agent and content fusion agent, based on their functional prompt word templates, receive the output results from the preceding agents to execute their tasks.

[0085] It should be noted that in the knowledge retrieval agent of this embodiment, the specific process of obtaining relevant background knowledge from the constructed dedicated knowledge base using a multi-source parallel retrieval strategy to form knowledge retrieval results is as follows: When a query instruction input by the user is obtained, the semantic similarity between the text description of each entity in the above entity relationship graph and the query instruction is calculated to obtain the entity associated with the query instruction, thereby obtaining a complete knowledge fragment containing contextual information. Finally, the structured knowledge in the entity relationship graph is converted into natural language form to form knowledge retrieval results and returned to the user.

[0086] It should be noted that, in this invention, the first... When evaluating the credibility of real-time webpage information, a preset authority weight is used to assess the credibility of the first webpage. The authority score of each real-time webpage is weighted, and a preset timeliness weight is used to evaluate the information. The timeliness score of each real-time webpage is weighted and calculated using a preset content relevance weight. The content relevance scores of each real-time webpage are weighted, and the weighted authority score, timeliness score, and content relevance score are summed to obtain the result. The credibility score of the real-time webpage information, when the first When the credibility score of a real-time webpage is greater than the preset credibility score threshold, it means that it has passed the credibility assessment.

[0087] In the knowledge retrieval agent of this embodiment, real-time web page information can be obtained through various news API tools. Since most news retrieval tools currently only provide credibility assessment of content similarity, this invention proposes a method for credibility assessment of real-time web page information, as shown in the following formula:

[0088]

[0089] in, , , As weight, Specifically, it represents the authority weight. Specifically, it represents the weight of timeliness. Specifically, this represents the weighting of content relevance. In this embodiment, the three weights are fixed within the same sub-chapter task, while the three weights in different sub-chapter tasks are dynamically adjusted based on the analysis dimension of each sub-chapter task parameter. For example, when the analysis dimension of each sub-chapter task parameter is "policy effect," the authority weight is increased; when the analysis dimension of each sub-chapter task parameter is "latest trend," the timeliness weight is increased; and when the analysis dimension of each sub-chapter task parameter is "emission reduction technology," the content relevance weight is increased. , , and The first The scores for the authority, timeliness, content relevance, and credibility of real-time web page information are as follows:

[0090] Furthermore, the timeliness score is obtained by multiplying the initial timeliness score by an exponential decay term. The exponential decay term is a power operation with the natural constant as the base, and the exponent is the negative of the time decay coefficient multiplied by the time interval.

[0091] In this embodiment, the timeliness score is calculated as follows:

[0092]

[0093] in, This serves as the initial timeliness score; Indicates the exponentially decaying term; It is a natural constant; For time intervals, when At that time, the timeliness score is the initial timeliness score, that is ; This is a time decay coefficient, dynamically determined based on the type of webpage information: policy analysis information decays slowly, technology development information decays moderately, and data statistics information decays quickly. In the above calculation of timeliness scoring, the timeliness of webpage information decays exponentially with time, ensuring that the most up-to-date webpage information receives a higher weight.

[0094] Furthermore, the authority score is calculated based on a source type classification method. Specifically, the large language model automatically determines the source type of a website by identifying website domain characteristics and content characteristics, thereby outputting a corresponding score, i.e., the authority score.

[0095] Furthermore, the content relevance score is shown in the following formula:

[0096]

[0097] in, This is a numerical feature vector of the key search terms in the task parameters of each sub-chapter, after semantic encoding. It is a numerical feature vector of real-time web page information content after semantic encoding.

[0098] The semantic encoding process here converts the text content into a multi-dimensional numerical array, where each dimension represents a different semantic feature. This embodiment measures the semantic matching degree between the query question and the real-time webpage information content by calculating the cosine similarity between two numerical feature vectors. The cosine similarity value ranges from [value missing]. The closer the value is The higher the semantic relevance, the better.

[0099] It should be noted that in step S2 of this invention, in the multi-agent cooperative architecture, the agents... The level value of an agent is equal to the maximum level value among all its predecessor agents plus 1, thereby ensuring that the level of each agent is strictly greater than the level of all its predecessor agents.

[0100] In this embodiment, a hierarchical parallel execution strategy is used to determine the execution level of the agent. Topological sorting is used to allocate agents to different execution layers, achieving an efficient processing mode of serial processing between layers and parallel processing within layers. Figure 2 As shown. The hierarchy of each agent is determined by its dependencies. An agent can only execute after all its predecessor agents have completed. Agents within the same layer can execute in parallel, as shown in the following equation:

[0101]

[0102] in, Represents intelligent agents The level value; This represents the set of all directed edges in a mixed directed acyclic graph. Indicates from the intelligent agent Pointing to intelligent agents A directed edge; Represents a set Any one of the points to the intelligent agent The directed edges; Represents intelligent agents Precursor intelligent agent The level value; Indicates taking the intelligent agent The highest level value among all predecessor agents.

[0103] In the dependency graph of a multi-agent system, the predecessor agent and the successor agent define the execution order constraints between agents. Specifically, the predecessor agent refers to the agent that... If there exists a directed edge in the set of directed edges Then the other agent connected by the directed edge Then it is The precursor intelligent agent, representing It needs to rely on when performing tasks The output result; the successor agent refers to the agent's .... If there exists a directed edge in the set of directed edges ,but for The successor intelligent agent, representing Will use The output is used as input. Based on this dependency, the execution of the agent must satisfy the topological constraint: the predecessor agent must complete its execution before the successor agent.

[0104] It should be noted that in step S2 of this invention, the specific hierarchical structure of the multi-agent collaborative architecture is as follows: the first layer is the semantic parsing agent, the second layer is the data visualization agent and the knowledge retrieval agent, the third layer is the image understanding agent, the fourth layer is the content fusion agent, and the fifth layer is the report generation agent.

[0105] In this embodiment, the first layer is a semantic parsing agent, responsible for decomposing the input user query into structured parameters; the second layer consists of a data visualization agent and a knowledge retrieval agent, the former generating statistical images and the latter retrieving the knowledge required for each sub-chapter in parallel; the third layer is an image understanding agent, which performs in-depth analysis on the generated statistical images; the fourth layer is a content fusion agent, which fuses the retrieved knowledge with the statistical image analysis results to achieve parallel fusion of sub-chapter content; and the fifth layer is a report generation agent, which organizes all fusion results into a complete report according to a specified structure.

[0106] The core advantage of this multi-agent collaborative architecture lies in the high parallelism of its second and fourth layers, enabling it to process multiple sub-tasks simultaneously. Compared to sequential execution, this reduces processing time significantly. Reduce to ,in Number of sub-chaps; Indicates constant time complexity; This indicates linear time complexity.

[0107] It should be noted that in step S2 of this invention, each agent maintains four execution states: waiting state, running state, completed state, and failed state. The waiting state indicates that an agent is waiting for the predecessor agent to complete; the running state indicates that the agent is executing a task; the completed state indicates that the agent has successfully completed the task and can pass the result to the successor agent; and the failed state indicates that an error occurred during the execution of the agent.

[0108] It should be noted that in step S2 of this invention, the state transition mechanism between agents is used to ensure the controllability and reliability of the report generation process. When all predecessor agents of an agent reach the completion state, the agent automatically transitions from the waiting state to the running state; when the agent successfully completes the task, it automatically notifies all its successor agents to check the pre-built startup conditions; when the agent fails to complete the task, the error propagation mechanism is activated, setting all its successor agents to the waiting state and triggering a retry or degradation processing strategy to ensure the robustness and fault tolerance of the system.

[0109] S3: Obtain the user's input of a natural language question related to greenhouse gas emissions, call the multi-agent collaborative architecture to generate a greenhouse gas emissions analysis report based on the input natural language question, and feed the report back to the user in a preset manner. When the user's input of modification suggestions is received, a new multi-agent collaborative architecture is created to regenerate the greenhouse gas emissions analysis report until the user's needs are met.

[0110] In this embodiment, the specific process for generating a greenhouse gas emission analysis report when used by users is as follows: Figure 3As shown, in brief, the process is as follows: Users input a natural language question related to greenhouse gas emissions, such as "Analysis of Global Greenhouse Gas Emission Change Rates in 2022-2023". The semantic parsing agent breaks down this natural language question into three types of structured parameters: visualization requirement parameters, sub-chapter task parameters, and report structure parameters. After the visualization requirement parameters are input into the data visualization agent, a global 2022-2023 greenhouse gas emission change rate distribution map is generated. The sub-chapter task parameters are input into each knowledge retrieval agent, generating corresponding knowledge background information, such as the global 2022-2023 greenhouse gas emission background, the spatial distribution of global 2022-2023 greenhouse gas emissions, influencing factors and driving mechanisms, emission reduction measures and policy analysis, and future prospects and policy recommendations. The images generated by the data visualization agent are input into the image understanding agent to extract key feature information. Multiple content fusion agents rely on the outputs of the corresponding knowledge retrieval agents and image understanding agents. The report structure parameters and all content generated by the fusion agent are output to the report generation agent to generate a structured greenhouse gas emission analysis report. This report will include an overview of the background and current status, emission data and spatial distribution visualization, influencing factors and driving mechanisms, emission reduction measures and policy analysis, future outlook and policy recommendations, etc.

[0111] It should be noted that after generating the greenhouse gas emission analysis report, it can be visualized and returned to the user through preset methods, such as graphical display and push in a pre-set functional interface module, or push in the form of system message, etc., which are not limited in this invention.

[0112] Furthermore, in this embodiment, to address the need for report quality optimization, a versioned multi-agent collaborative architecture execution mechanism is adopted to establish a complete quality assessment and optimization process. Each report generation process forms an independent DAG structure execution instance. After the user reviews the report and provides specific modification suggestions (including image adjustment suggestions, specific problem annotations for each sub-section, and overall report improvement suggestions, etc.), the optimization process triggered by manual review is realized by creating a new DAG structure execution instance: The method of this invention first locates the source of the problem and automatically optimizes three types of structured parameters, adjusts the visualization requirement parameters according to the image adjustment suggestions, adjusts the report structure parameters according to the overall report improvement suggestions, and precisely adjusts the task parameters of each sub-section according to the specific problem annotations of each sub-section, while creating a new DAG structure execution instance; then, the new DAG structure execution instance re-executes the process based on the optimized structured parameters, avoiding the formation of circular dependencies in a single execution instance, ensuring the acyclic nature of the system architecture, the determinism of the execution process, and improving the quality of the generated report.

[0113] To better demonstrate the specific implementation and technical effects of the present invention, the greenhouse gas emission analysis method based on multimodal intelligent agents shown in steps S1 to S3 of the above preferred implementation is applied to a specific example.

[0114] Example

[0115] The implementation process of the greenhouse gas emission analysis method based on multimodal intelligent agents used in this embodiment is as described above and will not be repeated here.

[0116] To verify the feasibility of the method of the present invention, this embodiment takes "analysis of greenhouse gas emission trends and emission reduction policy effects in a certain region" as an example for verification. The specific implementation method is as shown above and will not be repeated here.

[0117] In this embodiment, the three types of structured parameters decomposed by the first-layer semantic parsing agent are shown in Tables 1, 2, and 3. In Table 3, the task parameters for each sub-chapter are illustrated using Chapter 3, "The Impact of Emission Reduction Policies on Industries," as an example. The second-layer data visualization agent receives the visualization requirement parameters, calls the trend chart generation tool, and generates greenhouse gas emission trend charts for the aforementioned regions based on the processed national / regional greenhouse gas emission time-series data, as shown below. Figure 4 As shown, the data used here comes from the Our World In Data database.

[0118] Table 1. Visual Display of Requirement Parameters

[0119]

[0120] Table 2. Report Structure Parameters Display

[0121]

[0122] Table 3. Display of Task Parameters for Each Sub-Chapter

[0123]

[0124] In addition, the aforementioned dedicated knowledge base is constructed using a knowledge graph. Given the complex relationships between entities in the greenhouse gas emissions field (multiple connections exist between driving factors, countries / regions, and policy measures) and the strong hierarchical nature of knowledge (covering multiple levels such as international conventions, national policies, and industry standards), the knowledge graph can effectively preserve the relationships between entities and support relational queries and path reasoning.

[0125] For webpage credibility assessment, this embodiment uses the Tavily search engine to obtain real-time webpage information and combines it with its built-in content relevance scoring to construct a multi-dimensional evaluation system. The large language model scores webpage information based on three dimensions: website authority (information source, publishing institution, domain authority), information timeliness (an exponential decay model based on publication time), and content relevance, with scores ranging from 0 to 1. After scoring each dimension, this embodiment sets corresponding weights according to the analysis needs of different sub-sections and calculates the final comprehensive score of the webpage information. Here, a credibility score threshold of 0.6 is set; only when a webpage information score exceeds this threshold is it included in the knowledge background reference of the corresponding sub-section.

[0126] The third-layer image understanding agent utilizes the visual-text cross-modal understanding capabilities of a multimodal large language model to perform deep analysis on the generated images. The specific process is as follows: First, the image files are converted to Base64 encoding format to ensure that the image data can be effectively processed by the multimodal model; then, visual analysis is used to identify image types, extract data trends, and analyze changing features; finally, the visual information is converted into a structured text format, containing metadata such as image file path, image type, and feature information.

[0127] The fourth-layer content fusion agent uses entity alignment and semantic relationship reasoning to fuse the knowledge retrieval results of each sub-chapter with the output information of the image understanding agent. Through entity recognition and standardization processing, the entity representations from different agents are unified to generate structured content output for each sub-chapter.

[0128] The fifth-layer report-generating agent receives the report structure parameters, organizes all fusion results according to the specified structure, and adds the visualization images and three types of structured parameters to the report as appendices, ultimately generating a 13,748-word structured report.

[0129] The end user reviews the report and provides specific modification suggestions. In this embodiment, the three types of structured parameters are automatically optimized and a new DAG structure instance is created to regenerate the report, ensuring an improvement in the quality of the final generated report.

[0130] It should also be noted that the greenhouse gas emission analysis method based on multimodal intelligent agents in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a greenhouse gas emission analysis system based on multimodal intelligent agents, corresponding to the greenhouse gas emission analysis method based on multimodal intelligent agents provided in the above embodiments, such as... Figure 5 As shown, it includes:

[0131] The data acquisition module is used to acquire natural language questions related to greenhouse gas emissions input by the user;

[0132] The report generation module is used to process the natural language questions acquired by the data acquisition module according to the greenhouse gas emission analysis method based on multimodal intelligent agents described in the above embodiments, and generate and visualize a greenhouse gas emission analysis report that meets the user's needs.

[0133] It is understood that the greenhouse gas emission analysis method based on multimodal intelligent agents described in S1-S3 above can essentially be implemented by a computer program. Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the greenhouse gas emission analysis method based on multimodal intelligent agents provided in the above embodiments, such as... Figure 6 As shown, it includes a memory and a processor;

[0134] The memory is used to store computer programs;

[0135] The processor is configured to implement the greenhouse gas emission analysis method based on multimodal intelligent agents in the above embodiments when executing the computer program.

[0136] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0137] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be 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, or discrete hardware components.

[0138] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0139] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for analyzing greenhouse gas emissions based on multimodal intelligent agents, characterized in that, Includes the following steps: S1: Acquire report documents related to greenhouse gas emissions, time-series data of greenhouse gas emissions, and geospatial data within the target area. Process each of the three types of data to form the first data, the second data, and the third data. Based on the first data, construct a dedicated knowledge base in the field of greenhouse gas emissions, and based on the third data, construct a standardized dataset that supports multimodal data visualization. S2: Based on the logical dependencies and parallel processing requirements of the greenhouse gas emission analysis report generation, a hybrid directed acyclic graph is constructed. A multi-agent collaborative architecture is built using the hybrid directed acyclic graph, and a state transition mechanism between agents is established. The multi-agent collaborative architecture includes six agents: semantic parsing, knowledge retrieval, data visualization, image understanding, content fusion, and report generation. The data visualization agent generates visualization images based on standardized datasets and / or secondary data. The knowledge retrieval agent performs knowledge retrieval based on a dedicated knowledge base. By repeatedly calling the knowledge retrieval agent and the content fusion agent, the tasks of each sub-chapter of the greenhouse gas emission analysis report are processed in parallel. S3: Obtain the user's input of a natural language question related to greenhouse gas emissions, call the multi-agent collaborative architecture to generate a greenhouse gas emissions analysis report based on the input natural language question, and feed the report back to the user in a preset manner. When the user's input of modification suggestions is received, a new multi-agent collaborative architecture is created to regenerate the greenhouse gas emissions analysis report until the user's needs are met. In the S2 multi-agent collaborative architecture, the level value of an agent is equal to the maximum level value among all its predecessor agents plus 1; predecessor agents and successor agents define the execution order constraints between agents: a predecessor agent refers to an agent that, for a given agent, has a higher level value than its predecessor agents. If there exists a directed edge in the set of directed edges Then the other agent connected by the directed edge Then it is The precursor intelligent agent, representing It needs to rely on when performing tasks The output result; the successor agent refers to the agent's .... If there exists a directed edge in the set of directed edges ,but for The successor intelligent agent, representing Will use The output is used as input; This represents the set of all directed edges in a mixed directed acyclic graph. The specific layered structure of the multi-agent collaborative architecture is as follows: the first layer is the semantic parsing agent, the second layer is the data visualization agent and the knowledge retrieval agent, the third layer is the image understanding agent, the fourth layer is the content fusion agent, and the fifth layer is the report generation agent.

2. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 1, characterized in that, In S1, the data processing procedure is as follows: text is extracted from the report document data using optical character recognition technology and cleaned using regular expressions to generate standardized corpus. The generated standardized corpus is used as the first data. The standardized corpus is then segmented into text blocks. Entity recognition and relation extraction are performed from the text blocks using a large language model to construct an entity relation graph, thereby forming a dedicated knowledge base. For greenhouse gas emission time series data, missing values ​​are filled using the average of the previous and subsequent data, and the annual change gradient is calculated. The processed greenhouse gas emission time series data is used as the second data. Spatial indexing technology is applied to geospatial data to establish a mapping relationship between target area codes and geographic coordinates, and the processed geospatial data is used as third-party data.

3. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 1, characterized in that, In the S2 multi-agent collaborative architecture, the specific functions of each agent are implemented as follows: The semantic parsing agent serves as the starting agent. It employs hierarchical intent recognition and, based on a pre-designed semantic parsing function prompt word template, first extracts key entities from the input user question. Then, it combines the extracted key entities to determine the user's analysis needs through intent classification. The task of generating the greenhouse gas emission analysis report is decomposed into three types of structured parameters: visualization requirement parameters, sub-chapter task parameters, and report structure parameters. The data visualization intelligent agent receives visualization requirement parameters based on pre-designed visualization function prompt word templates, determines specific visualization tasks, and calls the corresponding image generation tools to automatically extract secondary data and / or standardized datasets to generate statistical images that meet professional standards. The knowledge retrieval agent receives task parameters from each sub-chapter based on pre-designed retrieval function prompt word templates, executes the retrieval tasks of each sub-chapter in parallel, and uses a multi-source parallel retrieval strategy to obtain relevant background knowledge from the constructed dedicated knowledge base and real-time web page information that has been evaluated for credibility, and outputs the knowledge retrieval results of each sub-chapter. The image understanding agent utilizes the visual-text cross-modal understanding capability of a multimodal large language model to perform in-depth analysis on generated statistical images based on pre-designed image understanding function prompt word templates. It identifies the statistical image type, extracts data trends from the statistical images, analyzes data change characteristics, forms analysis results, and converts them into structured text format. This text, along with the statistical image file path and statistical image type, serves as the output information of the image understanding agent. The content fusion agent employs entity alignment and semantic relationship reasoning. Based on a pre-designed content fusion function prompt word template, it fuses the knowledge retrieval results of each sub-chapter with the output information of the image understanding agent. Through entity recognition and standardization processing, it unifies the entity representations from different agents. By calling multiple times to process the fusion tasks of different sub-chaps in parallel, it generates the structured content of each sub-chapter. As a terminal agent, the report generation agent receives report structure parameters based on a pre-designed report generation function prompt template to determine the overall structure of the report. It then organizes the structured content of each sub-chapter output by the content fusion agent according to the specified chapter logical order to generate a logically coherent and complete greenhouse gas emission analysis report.

4. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 3, characterized in that, The visualization requirements parameters include data dimensions, image type, data range, and visualization objectives; the task parameters for each sub-chapter include chapter theme, key search terms, analysis dimensions, and output requirements; the report structure parameters include a list of chapter titles, logical order of chapters, overall report framework, and format requirements; and the functional prompt templates for each agent include role definition, task description, input format, output requirements, and example display.

5. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 3, characterized in that, When evaluating the credibility of the i-th real-time webpage information, a preset authority weight is used to weight the authority score of the i-th real-time webpage information, a preset timeliness weight is used to weight the timeliness score of the i-th real-time webpage information, and a preset content relevance weight is used to weight the content relevance score of the i-th real-time webpage information. The weighted authority score, timeliness score, and content relevance score are added together to obtain the credibility score of the i-th real-time webpage information. When the credibility score of the i-th real-time webpage information is greater than the preset credibility score threshold, it means that it passes the credibility evaluation. Among them, the timeliness score is obtained by multiplying the initial timeliness score by an exponential decay term. The exponential decay term is a power operation with the natural constant as the base, and the exponent is the negative of the time decay coefficient multiplied by the time interval.

6. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 3, characterized in that, In the S2 multi-agent collaborative architecture, each agent maintains four execution states: waiting state, running state, completed state, and failed state. The waiting state indicates that an agent is waiting for the predecessor agent to complete; the running state indicates that the agent is executing a task; the completed state indicates that the agent has successfully completed the task and can pass the result to the successor agent; and the failed state indicates that the agent encountered an error during execution.

7. The greenhouse gas emission analysis method based on multimodal intelligent agents as described in claim 6, characterized in that, In the S2 multi-agent collaborative architecture, the state transition mechanism between agents is as follows: when all predecessor agents of an agent have reached the completion state, the agent automatically transitions from the waiting state to the running state; when the agent successfully completes the task, it automatically notifies all its successor agents to check the pre-built start conditions; when the agent fails to complete the task, it initiates the error propagation mechanism, sets all its successor agents to the waiting state, and triggers a retry or degradation processing strategy.

8. A greenhouse gas emission analysis system based on multimodal intelligent agents, characterized in that, include: The data acquisition module is used to acquire natural language questions related to greenhouse gas emissions input by the user; The report generation module is used to process the natural language questions acquired by the data acquisition module according to the greenhouse gas emission analysis method based on multimodal intelligent agents as described in any one of claims 1 to 7, and generate and visualize a greenhouse gas emission analysis report that meets the user's needs.

9. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the greenhouse gas emission analysis method based on multimodal intelligent agents as described in any one of claims 1 to 7 when executing the computer program.

Images