Method, device and storage medium for processing double-carbon business

Abstract

The application relates to a double-carbon business processing method, device, equipment and storage medium. The method comprises the following steps: receiving a double-carbon business analysis request input by a user, calling a preset large language model to perform semantic analysis on the double-carbon business analysis request, and generating a task set containing multiple sub-tasks and weight values corresponding to each sub-task; based on a preset business logic dependency relationship and the task set, a layered scheduling logic is generated for the double-carbon business analysis request according to a weight self-adaptive ant colony optimization algorithm; according to the layered scheduling logic, the large language model and a pre-constructed double-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed, and the reasoning results of each sub-task are fused to generate a structured decision report. The method can improve the analysis efficiency and decision quality in the double-carbon business scenario.

Description

Technical Field

[0001] This application relates to the field of dual-carbon business management technology, and in particular to a method, apparatus, equipment and storage medium for processing dual-carbon business. Background Technology

[0002] With the development of artificial intelligence technology, large language models have demonstrated powerful capabilities in natural language understanding and reasoning tasks, and are therefore widely used in knowledge-based question answering and decision support systems across various professional fields. Meanwhile, retrieval-enhanced generation technology, by injecting supporting text from external knowledge bases into the reasoning context of large language models, effectively alleviates the problems of insufficient knowledge coverage and lack of real-time information in general-purpose large models, and has become an important technical means to improve the analytical accuracy of intelligent systems in professional fields.

[0003] Currently, in the field of dual-carbon business management, related analytical tasks are typically handled using a question-and-answer system based on a general large language model. Specifically, users input business questions in natural language, and the question-and-answer system then directly generates answers by calling the large language model.

[0004] However, this method cannot decompose the comprehensive business quality of user input, resulting in low analysis efficiency and unreliable reasoning quality of the question-answering system in complex business scenarios. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, equipment, and storage medium for processing dual-carbon data to address the aforementioned technical problems.

[0006] Firstly, this application provides a method for processing dual-carbon transactions, including:

[0007] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0008] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0009] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0010] Secondly, this application also provides a processing apparatus for dual-carbon transactions, comprising:

[0011] The receiving module is used to receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request to generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0012] The scheduling module is used to generate hierarchical scheduling logic for dual-carbon business analysis requests based on preset business logic dependencies and task sets, and according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0013] The fusion module is used to call the large language model and the pre-built dual-carbon knowledge base layer by layer according to the hierarchical scheduling logic to perform reasoning until all sub-tasks are completed, and then fuse the reasoning results of each sub-task to generate a structured decision report; the dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0014] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0015] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0016] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0017] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0019] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0020] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0021] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0022] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0023] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0024] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0025] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0026] The aforementioned dual-carbon service processing method, apparatus, equipment, and storage medium receive dual-carbon service analysis requests input by users and call a preset large language model to perform semantic analysis on the dual-carbon service analysis requests, generating a task set containing multiple sub-tasks and their corresponding weight values. Based on preset business logic dependencies and the task set, a hierarchical scheduling logic is generated for the dual-carbon service analysis requests according to a weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponding to a set of sub-tasks, and each execution layer is executed serially according to business logic dependencies, while the sub-tasks within the same execution layer are executed concurrently. According to the hierarchical scheduling logic, the large language model and a pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed, and the reasoning results of each sub-task are fused to generate a structured decision report. The dual-carbon knowledge base is used to store semantic vector indexes of dual-carbon domain documents. The above method enables the automatic decomposition of complex business instructions, and the integration of hierarchical scheduling logic with large language model reasoning effectively solves the technical problems of low analysis efficiency and unreliable reasoning quality in traditional question-answering systems. It realizes end-to-end automated processing from user instructions to structured decision reports, and significantly improves the analysis efficiency and decision quality in dual-carbon business scenarios. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is an application environment diagram of a dual-carbon service processing method in one embodiment;

[0029] Figure 2 This is one of the flowcharts illustrating a dual-carbon service processing method in one embodiment;

[0030] Figure 3 This is a second flowchart illustrating a dual-carbon business processing method in one embodiment;

[0031] Figure 4 This is the third flowchart illustrating a dual-carbon business processing method in one embodiment;

[0032] Figure 5 This is a flowchart of a dual-carbon business processing method in one embodiment (fourth example).

[0033] Figure 6 This is the fifth flowchart illustrating a dual-carbon service processing method in one embodiment;

[0034] Figure 7 This is a flowchart of a dual-carbon business processing method in one embodiment, number six.

[0035] Figure 8 This is a structural block diagram of a dual-carbon service processing apparatus in one embodiment. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0038] With the development of artificial intelligence technology, large language models have demonstrated powerful capabilities in natural language understanding and reasoning tasks, and are therefore widely used in knowledge-based question answering and decision support systems across various professional fields. Meanwhile, retrieval-enhanced generation technology, by injecting supporting text from external knowledge bases into the reasoning context of large language models, effectively alleviates the problems of insufficient knowledge coverage and lack of real-time information in general-purpose large models, and has become an important technical means to improve the analytical accuracy of intelligent systems in professional fields.

[0039] Currently, in the field of dual-carbon business management, related analytical tasks are typically handled using a question-and-answer system based on a general large language model. Specifically, users input business questions in natural language, and the question-and-answer system then directly generates answers by calling the large language model.

[0040] However, this method cannot decompose the comprehensive business quality of user input, resulting in low analysis efficiency and unreliable reasoning quality of the question-answering system in complex business scenarios.

[0041] In view of the above-mentioned technical problems, this application provides a method for processing dual-carbon services. The following embodiments will specifically illustrate the method for processing dual-carbon services.

[0042] The dual-carbon business processing method provided in this application embodiment can be applied to, for example, Figure 1The illustrated computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a dual-carbon business processing method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0043] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0044] In one exemplary embodiment, such as Figure 2 As shown, a method for processing dual-carbon business is provided, which is applied to... Figure 1 The following explanation uses computer equipment as an example, including:

[0045] S201: Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, generating a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0046] In this embodiment, the computer device receives a dual-carbon business analysis request input by a user. This request includes at least one of text instructions and local documents. Text instructions are business instructions input by the user using natural language, including at least one of carbon verification requests, compliance inquiries, and policy consultations. Local documents are at least one of carbon emission quota allocation schemes, industry carbon verification reports, carbon trading market policy documents, and historical emission data ledgers. Subsequently, the computer device parses the input format of the received dual-carbon business analysis request. Specifically, if it is a text instruction, the computer device can directly perform encoding preprocessing; if it is a local document, it supports parsing PDF and Word documents. The computer device can uniformly convert original documents of different formats into plain text sequences to eliminate interference from format differences in subsequent processing, thereby achieving input standardization.

[0047] Subsequently, the computer device can invoke a pre-defined large language model to perform semantic parsing on the normalized input, extracting the core objective and constraints. The core objective is the user's ultimate business goal, including completing the carbon quota compliance assessment; the constraints include at least one of time range, industry category, and geographical scope. Based on the semantic parsing results, the computer device can construct a hierarchical analysis model, which includes an objective layer, a criterion layer, and a solution layer. Specifically, the objective layer contains the core objective, the criterion layer contains multiple analytical dimensions required to achieve the core objective, and the solution layer contains multiple sub-tasks under each analytical dimension. Next, the computer device invokes the large language model based on pre-defined dual-carbon business rules to perform pairwise importance comparisons on each sub-task in the solution layer, and constructs a judgment matrix based on the comparison results. This judgment matrix is ​​assigned values ​​using a 1-9 scale and satisfies the positive reciprocity condition. Then, the computer device solves for the judgment matrix to obtain the largest eigenvalue and its corresponding eigenvector, and normalizes the eigenvector to obtain the weight values ​​corresponding to each sub-task. This weight value is used to characterize the degree of contribution of each subtask to the core objective. The larger the weight value, the higher the degree of contribution of the corresponding subtask to the final decision objective.

[0048] Finally, the computer calculates a dynamic filtering threshold based on the weight values ​​of each subtask. Subtasks with weight values ​​lower than this threshold are removed from the solution layer, while subtasks with weight values ​​not lower than the threshold are retained, generating a task set. This dynamic filtering threshold is used to adaptively filter each subtask to reduce the computational complexity of subsequent path planning.

[0049] S202, based on the preset business logic dependencies and task set, generates hierarchical scheduling logic for dual-carbon business analysis requests according to the weighted adaptive ant colony optimization algorithm.

[0050] The hierarchical scheduling logic includes at least one execution layer, each execution layer corresponds to a set of subtasks, and each execution layer is executed serially according to the business logic dependency relationship, while the subtasks in the same execution layer are executed concurrently.

[0051] In this embodiment, after generating a task set containing each subtask and its weight value, the computer device can generate hierarchical scheduling logic for the dual-carbon business analysis request based on a weighted adaptive ant colony optimization algorithm according to a preset business logic dependency relationship and the task set. This hierarchical scheduling logic is used to characterize the execution order and concurrency relationship of each subtask.

[0052] Specifically, the computer device first maps each subtask in the task set to a graph node, and determines the directed edges between the graph nodes based on business logic dependencies, constructing a directed acyclic graph (DAG). Each graph node corresponds to a subtask, and the attributes of the graph node include the weight value and execution cost of the subtask. Simultaneously, the computer device sets a start node and a stop node in this DAG, transforming the path optimization problem into an optimal path search problem from the start node to the stop node. Subsequently, the computer device performs path search on this DAG based on a weighted adaptive ant colony optimization algorithm. This algorithm simulates the foraging behavior of an ant colony and incorporates weighted adaptive improvements to address the uneven weight distribution and significant differences in execution costs among the subtasks, prioritizing nodes with high task weights and low execution costs during the path search process. The computer device iteratively searches using this algorithm until a preset convergence condition is met, obtaining the optimal path from the start node to the stop node. After obtaining the optimal path, the computer device performs hierarchical partitioning based on the business logic dependencies of each graph node within that path. Graph nodes with the same dependency depth are grouped into the same execution layer, while those with different dependency depths are grouped into different execution layers. The execution order of each execution layer is then determined according to the order of dependency depth from shallowest to deepest. Subtasks within the same execution layer have no data dependencies and can be executed concurrently; subtasks between different execution layers are executed sequentially according to the order of dependency depth from shallowest to deepest. Thus, the computer device generates hierarchical scheduling logic, which serves as the basis for subsequent steps involving layer-by-layer inference using the large language model and the pre-built dual-carbon knowledge base.

[0053] S203, according to the hierarchical scheduling logic, calls the large language model and the pre-built dual-carbon knowledge base layer by layer for reasoning until all sub-tasks are completed, and then merges the reasoning results of each sub-task to generate a structured decision report.

[0054] The dual-carbon knowledge base is used to store semantic vector indexes of documents in the dual-carbon domain.

[0055] In this embodiment, the computer device adds each subtask in the ready state in the current execution layer to the ready queue according to the order of the execution layers in the hierarchical scheduling logic. Then, it retrieves the subtask from the ready queue, generates a retrieval query instruction based on the semantic information of the subtask, and selects multiple candidate texts related to the subtask from the dual-carbon knowledge base according to the retrieval query instruction. The computer device sorts the candidate texts based on their relevance to the retrieval query instruction and selects the top-ranked candidate text as the target text. The computer device assembles the target text with the subtask to generate a reasoning context, and performs reasoning on the reasoning context based on a large language model to obtain the reasoning result of the subtask. When all subtasks in the current execution layer have been completed, the computer device adds the subtasks in the next execution layer that meet the prerequisite dependency conditions to the ready queue, repeating the above steps until all subtasks in all execution layers have been completed. Then, the computer device merges the inference results of each subtask and, based on a preset prompt template, assembles the dual-carbon business analysis request, preset industry macro-constraints, the inference results of each subtask, and the calculation results of the dual-carbon core indicators to generate a fusion prompt. The calculation results of the dual-carbon core indicators are obtained by a large language model based on industry standards and verified inference results. The computer device then calls the large language model to perform global inference based on the fusion prompt to generate a structured decision report.

[0056] The dual-carbon knowledge base is pre-built to provide retrieval support for sub-task reasoning. This knowledge base stores documents in the dual-carbon field, including but not limited to carbon emission quota allocation schemes, industry carbon verification reports, carbon trading market policy documents, and historical emission data ledgers. These documents were imported and the knowledge base was built during the deployment of the computer equipment. When using the knowledge base, users can directly input dual-carbon business analysis requests, and the computer equipment will call the pre-built knowledge base for retrieval and reasoning. Simultaneously, it supports users uploading local documents for incremental updates; that is, users can upload new documents along with their dual-carbon business analysis requests, and the computer equipment will append these new documents to the knowledge base, achieving dynamic knowledge supplementation. When users do not upload new documents, the computer equipment directly calls the already built dual-carbon knowledge base for subsequent processing.

[0057] Specifically, the pre-construction process of the dual-carbon knowledge base is as follows: The computer device first performs text segmentation processing on the acquired dual-carbon domain documents. Since the context window of the large language model has a length limit, and the granularity of vector retrieval requires text blocks to have relatively independent semantic integrity, it is necessary to segment the dual-carbon domain documents. The computer device uses a recursive character text segmentation algorithm to segment each dual-carbon domain document. This algorithm recursively segments according to a pre-defined list of delimiters (such as paragraph marks, line breaks, periods, commas, etc.), prioritizing sentence breaks at semantic boundaries to preserve the contextual coherence within the text block. The computer device sets the target upper limit for the text block length to be [missing information]. (Unit: number of characters or number of tokens), set the overlap length between adjacent text blocks to [value]. ( This is used to preserve contextual transition information between adjacent blocks, preventing semantic breaks at block boundaries. Through the above block segmentation process, the computer device divides each document into multiple text blocks, obtaining a set of text blocks. .in, The total number of blocks, and each text block satisfy Adjacent text blocks and There exists a length of The overlapping area, i.e. .

[0058] Subsequently, the computer device performs semantic vectorization processing on each text block. The computer device invokes a pre-defined embedding model to vectorize each text block, mapping each text block into a semantic vector. This embedding model, based on a general language model and continuously pre-trained on corpora from fields such as carbon, carbon trading, and carbon verification, possesses the semantic encoding capability for specialized concepts such as carbon quotas, carbon intensity, carbon sinks, and compliance costs. Specifically, the computer device uses an embedding function... ,in For vocabulary size, The dimension of the embedding vector. The computer device considers a set of text blocks. Each text block in Perform embedding mapping, that is, for each text block Mapped to high-dimensional semantic vectors After the computer device performs mapping operations on all text blocks, it can obtain a set of semantic vectors. To ensure the numerical stability of subsequent cosine similarity calculations, the computer equipment performs [a certain procedure] on each vector. The normalization process can be represented by relation (1), which is shown below:

[0059] (1);

[0060] After normalization, the cosine similarity between any two vectors is equivalent to their inner product, i.e. As a result, computer equipment further reduces the computational complexity in subsequent retrieval processes.

[0061] Then, the computer device builds an index for the vector database. The computer device uses the normalized set of semantic vectors. and its corresponding original text block The vector database stores data in key-value pairs, forming a two-carbon knowledge base. The vector database employs an approximate nearest neighbor retrieval structure based on inverted file indexing and product quantization. Specifically, in terms of clustering, the computer device processes the vector set... The k-means algorithm was used to divide the data into... Let the i-th cluster be the i-th cluster. The centroids of each cluster are The optimization objective of clustering is Therefore, each vector Assigned to the nearest cluster It is stored in an inverted list format, reducing the retrieval complexity from Reduce to In terms of product quantization compression, computer equipment will... The dimensional vector space is divided into There are 1 subspace, and the dimension of each subspace is 1. K-means quantization is performed independently on a subset of vectors in each subspace to generate the codebook for each subspace. ( Each vector is quantized by a quantization code in each subspace. This means that the original floating-point vector is compressed into a compact quantized code sequence. As the vectors are entered into the dual-carbon knowledge base, the computer equipment adds structured metadata to each record, including the document source, document type, original text block content, and the text block's position offset in the original text. The metadata is linked to the vector index through a unique key, ensuring that the search results can be traced back to the original document.

[0062] In addition, the computer equipment supports an incremental update mechanism for the dual-carbon knowledge base. When a user uploads a new document, the computer equipment only performs chunking and vectorization operations on that new document and appends the new vector to the existing index, without needing to rebuild the index for the entire database vector. If the set of the dual-carbon knowledge base before the update is... The vector set generated by the newly added document is The updated dual-carbon knowledge base is as follows: .

[0063] For expired policy documents that need to be revoked, the computer equipment locates and marks the failure vector based on the chunk_id, and filters them during retrieval to ensure the timeliness and accuracy of the knowledge base content.

[0064] Using the above method, a dual-carbon knowledge base is pre-built by computer equipment.

[0065] In the above-mentioned dual-carbon business processing method, a dual-carbon business analysis request is received from the user, and a preset large language model is invoked to perform semantic analysis on the dual-carbon business analysis request, generating a task set containing multiple sub-tasks and their corresponding weight values. Based on the preset business logic dependencies and the task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently. According to the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are invoked layer by layer for reasoning until all sub-tasks are completed, and the reasoning results of each sub-task are fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents. The above method enables the automatic decomposition of complex business instructions, and the integration of hierarchical scheduling logic with large language model reasoning effectively solves the technical problems of low analysis efficiency and unreliable reasoning quality in traditional question-answering systems. It realizes end-to-end automated processing from user instructions to structured decision reports, and significantly improves the analysis efficiency and decision quality in dual-carbon business scenarios.

[0066] In an exemplary embodiment, the above-mentioned step S201, "receiving a dual-carbon service analysis request input by the user, and calling a preset large language model to perform semantic analysis on the dual-carbon service analysis request, generating a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task," is as follows: Figure 3 As shown, it includes:

[0067] S301 performs semantic parsing on dual-carbon business analysis requests and determines the hierarchical analysis module based on the parsing results.

[0068] The hierarchical analysis module includes an objective layer, a criterion layer, and a solution layer; the solution layer contains multiple sub-tasks.

[0069] In this embodiment, user instructions in dual-carbon business scenarios are often highly comprehensive. For example, the instruction "Assess the carbon quota compliance gap for this quarter and provide the optimal trading suggestion" implicitly includes multiple interrelated sub-tasks such as carbon emission accounting, quota surplus / deficit prediction, market price analysis, and compliance cost optimization. If such instructions are directly submitted to a large language model for processing, the large language model is prone to logical jumps or omissions of key analysis dimensions. Therefore, the computer device calls a preset large language model to perform semantic parsing on the dual-carbon business analysis request, extracts the core target entities and constraints in the instructions, summarizes them into target layer nodes, and denotes the target layer nodes as G. Subsequently, the computer device expands the target layer into several criterion layer nodes through the large language model and dual-carbon business logic, denoted as G. This represents several analytical dimensions required to achieve the goal, such as carbon emission assessment, carbon trend forecasting, and carbon financial accounting. Below the criteria layer, the computer equipment further refines this into a solution layer using a large language model; this layer comprises a set of specific, single-point executable subtasks, denoted as... Examples include "querying the current CCER price" and "calculating the current month's direct emissions." The aforementioned target layer, criterion layer, and scenario layer constitute a complete hierarchical analysis model, formally represented as triples. .

[0070] For example, a user inputs a dual-carbon business analysis request as "assess the carbon quota compliance gap for this quarter and provide the optimal trading recommendation." The computer device calls a large language model to perform semantic parsing on this dual-carbon business analysis request, extracting the core objective as "complete the carbon quota compliance assessment and trading decision," and extracting the constraint condition as "the time range is this quarter." The computer device expands the objective layer into a criterion layer based on the dual-carbon business logic using the large language model, including the dimensions of "carbon emission accounting" and "market price analysis." Further, under the criterion layer, it refines this into multiple sub-tasks at the scheme layer, including "querying the current period's free quota," "querying the current period's actual emissions," and "querying the current period's carbon market price."

[0071] It is worth noting that this large language model possesses the ability to understand dual-carbon business logic. By inputting prompt words containing dual-carbon business rules into the large language model, the computer device enables the model to identify specialized terms in the dual-carbon field and their logical relationships during semantic parsing. Therefore, when performing semantic parsing on user commands, the large language model can determine the analytical dimensions of the criterion layer based on the aforementioned rules, and further refine them into sub-tasks of the scheme layer, thereby generating a hierarchical analysis model consistent with the dual-carbon business logic.

[0072] S302, based on the preset dual-carbon business rules, calls the large language model to compare the importance of each subtask in the solution layer pairwise, determines the judgment matrix based on the comparison results, and solves the judgment matrix to obtain the weight value corresponding to each subtask.

[0073] In this embodiment, the traditional analytic hierarchy process (AHP) relies heavily on domain experts manually filling in the judgment matrix, which is time-consuming and highly subjective. Instead, the computer device uses a large language model based on pre-set dual-carbon industry rules and historical decision-making experience (which may include historical compliance cases, transaction strategy execution records, and industry best practices in the dual-carbon field) to automatically compare the pairwise importance of each subtask in the solution layer. Specifically, the computer device evaluates any two subtasks using the large language model. and The relative importance of the higher-level objectives is assigned using the Saaty1~9 scaling method to construct... Order judgment matrix ,in Subtasks Compared to The importance of this. At the same time, this judgment matrix satisfies the positive reciprocity condition, which can be expressed by relation (2), as shown below:

[0074] (2);

[0075] For example, the computer equipment uses a large language model to determine, based on the dual-carbon business rules, that querying the current period's free allowance and querying the current period's actual emissions are the basis for calculating the allowance gap, and their importance is higher than querying the current period's carbon market price. Therefore, in the judgment matrix, the relative importance of the first two is 1 (equally important), while the importance of both relative to querying the current period's carbon market price is 5 (significantly important).

[0076] Subsequently, the computer equipment performs a judgment matrix. Solving for the largest eigenvalue and its corresponding eigenvectors Satisfying the characteristic equation Computer equipment for feature vectors After normalization, the weight allocation vectors for each subtask are obtained, as shown in equation (3):

[0077] (3);

[0078] Final weight vector satisfy , ,in The size reflects the subtask The degree of relative contribution under the current business objectives.

[0079] Because large language models are prone to inconsistencies in logic before and after evaluation during the automatic construction of the judgment matrix, the computer device needs to perform a consistency check on the judgment matrix after solving it. Specifically, the computer device calculates the consistency index CI, which can be represented by the relation (4), as shown below:

[0080] (4);

[0081] Simultaneously, the average random consistency index RI of random judgment matrices of the same order is introduced (its value is determined by looking up a table as the matrix order n increases), and the consistency ratio is calculated. When CR < 0.1, the computer device determines that the judgment matrix has satisfactory consistency, and the weight vector... Effective. When CR ≥ 0.1, the computer device triggers a matrix correction mechanism. Specifically, it locates inconsistent element pairs based on the source of the largest eigenvalue deviation using a large language model. That is, for the element pairs that satisfy relation (5), the assignment at that position is iteratively adjusted, and the eigenvalue solution and consistency check are re-executed until CR < 0.1. Relation (5) is shown below:

[0082] (5);

[0083] This demonstrates that the correction mechanism can effectively overcome the limitations of the traditional analytic hierarchy process, which relies on repeated manual verification.

[0084] S303: Calculate the dynamic screening threshold based on the weight value of each subtask, compare the dynamic screening threshold with each weight value, and generate a task set based on the comparison result.

[0085] In this embodiment of the application, after calculating the weights of each subtask, the set of subtasks is... Perform dimensionality reduction filtering. The computer equipment uses dynamic filtering thresholds. This replaces the fixed threshold to address the issue of varying subtask numbers across different complex business scenarios. Dynamic threshold selection. It can be represented by relation (6), which is shown below:

[0086] (6);

[0087] That is, the weighted average with weighted standard deviation The difference. Regarding the weights subtasks If a computer device determines that its contribution to the current business objective is below the mean and the deviation exceeds one standard deviation, it is considered an inefficient and redundant task and is removed; the computer device retains weights that meet the requirements. The subtasks constitute the streamlined core task set. .set up The task size after dimensionality reduction is reduced from the original Dimensional compression to dimension( This effectively reduces the computational complexity of subsequent path planning and agent scheduling.

[0088] For example, consider the three subtasks: "Query the current period's free allowance," "Query the current period's actual emissions," and "Query the current period's carbon market price." Their weights are 0.45, 0.45, and 0.10, respectively, with a mean of 0.333, a standard deviation of 0.202, and a dynamic filtering threshold of 0.131. The two subtasks with a weight of 0.45, both greater than or equal to 0.131, are retained; the subtask with a weight of 0.10, less than 0.131, is removed. Thus, the original task set containing three subtasks is simplified to two.

[0089] Furthermore, the computer device sorts the retained core task set in descending order of weight, resulting in an ordered task sequence with priority labels. ,in In order to make The established order. This priority information will serve as the initial input for subsequent path planning steps, guiding the algorithm to prioritize the execution paths of high-contribution tasks, thereby maximizing analysis quality within computational constraints. Finally, the computer outputs a concise and ordered set of core tasks. And the weight vectors corresponding to each task.

[0090] In an exemplary embodiment, the "generating hierarchical scheduling logic for dual-carbon business analysis requests based on preset business logic dependencies and task sets, according to a weighted adaptive ant colony optimization algorithm" in S202 above, such as Figure 4 As shown, it includes:

[0091] S401 maps each subtask in the task set to a graph node, determines the directed edges between graph nodes based on business logic dependencies, and determines a directed acyclic graph.

[0092] A directed acyclic graph includes a start node and an end node.

[0093] In this embodiment of the application, the computer device sets core tasks. Mapped to a set of graph nodes Based on this, the computer device defines a set of directed edges according to the data flow and prerequisite dependencies between each subtask. If subtask Execution requires subtasks The output of the task is used as the input condition, and then the computer device performs the subtask. sub-tasks Establish a directed edge between them , indicating subtask It is a subtask The direct predecessor node. To prevent circular dependencies from causing the algorithm to deadlock, the computer performs a topological sorting verification on the graph structure after the edge set is constructed to confirm... It satisfies the definition of a directed acyclic graph.

[0094] Subsequently, the computer device sets up a virtual start node s and a virtual end node e in G. s points to all task nodes without predecessor nodes (i.e., graph nodes corresponding to subtasks that do not depend on any other subtasks and can be executed directly), and all task nodes without successor nodes (i.e., graph nodes corresponding to subtasks that are not depended on by any other subtasks) point to e. Through this method, the computer device transforms the path optimization problem into a unified optimal path search problem from s to e.

[0095] For each task node in the graph Computer devices define three types of attributes. Specifically, the first attribute is task weight. The first attribute, derived from the calculation results of the analytic hierarchy process (AHP), represents the degree to which the task contributes to the final decision objective; the second attribute, namely the execution cost. The first attribute represents the computing resources consumed in executing the subtask. The computer equipment sets a baseline initial value for the execution cost of each type of subtask based on its historical computation time and resource consumption. This execution cost is then dynamically adjusted in subsequent iterations based on the historical execution records of the subtask. The third attribute is the execute status flag. Initially, all tasks without predecessor nodes are set to 1, and the rest are set to 0. During execution, the settings are dynamically updated according to the completion status of the predecessor nodes.

[0096] For example, suppose the current task set contains three sub-tasks: querying the current period's free quota (node ​​A, weight 0.45), querying the current period's actual emissions (node ​​B, weight 0.45), and calculating the quota gap (node ​​C, weight 0.10). The calculation of the quota gap depends on the execution results of querying the current period's free quota and querying the current period's actual emissions, while there is no mutual dependency between querying the current period's free quota and querying the current period's actual emissions. The computer device establishes directed edges based on the above business logic dependencies, i.e., from node A to node C, and from node B to node C. Further, the computer device sets a virtual starting node s pointing to nodes A and B, and node C pointing to a virtual ending node e, thus constructing a directed acyclic graph. Simultaneously, the computer device defines attributes for each node: node A has a weight of 0.45, node B has a weight of 0.45, and node C has a weight of 0.10; nodes A and B are nodes without predecessors, and their executable status flag "ready" is set to 1, while node C's "ready" flag is set to 0. With the above setup, the business problem of assessing the carbon quota compliance gap for this quarter is transformed into an optimal path search problem from s to e.

[0097] S402 uses a weighted adaptive ant colony optimization algorithm to perform path search on a directed acyclic graph and obtain the optimal path from the starting node to the ending node.

[0098] In this embodiment, after constructing a directed acyclic graph (DAG), the computer device performs path search on the DAG based on a weighted adaptive ant colony optimization algorithm to find an optimal path from the starting node to the ending node. Specifically, this weighted adaptive ant colony optimization algorithm is a heuristic search algorithm that simulates the foraging behavior of ant colonies, achieving path optimization through the accumulation and volatilization mechanism of pheromones. The computer device has improved the algorithm with weight adaptation, making it suitable for dual-carbon business scheduling scenarios where subtasks have uneven weight distribution and significant differences in execution costs.

[0099] In the path search initiation phase, the computer device sets a non-uniform initial pheromone concentration for all paths in the directed acyclic graph (DAG), ensuring that paths containing nodes with higher task weights receive higher initial pheromone concentrations. This allows the weighted adaptive ant colony optimization algorithm to prioritize exploring paths with high contribution tasks from the outset. During path search, the computer device controls multiple ants to move across the DAG. Each ant determines its direction of movement based on the pheromone concentration between its current node and the next node, along with heuristic information. This heuristic information is dynamically calculated based on the task weight and execution cost of the subtask corresponding to the next node, causing the ant to prioritize nodes with high task weights and low execution costs. After each ant moves from its starting node to its ending node, multiple candidate execution paths are generated, with each ant's movement trajectory constituting a candidate path. The computer device then evaluates the quality of each candidate execution path. Path quality is determined by the sum of the task weights of the nodes contained within the path and their execution costs. Subsequently, the computer determines the current pheromone evaporation rate based on the current iteration count. This evaporation rate decreases as the iteration count increases, causing the algorithm to tend towards broad exploration in the early stages of iteration and fine convergence in the later stages. The computer updates the pheromone concentration in the graph based on the path quality of each candidate execution path and the current pheromone evaporation rate. For candidate execution paths whose quality meets preset conditions, the computer performs pheromone enhancement, further increasing the pheromone concentration on high-quality paths and guiding ants in subsequent iterations to prioritize these paths.

[0100] The computer device repeatedly executes the above path search and pheromone update process until the preset convergence condition is met, and then determines the candidate execution path with the highest pheromone concentration as the optimal path. Through the above path search process, the computer device finds an optimal execution path with high decision contribution and low execution cost, while satisfying the pre-dependency constraints of each subtask.

[0101] S403: Based on the business logic dependencies of each graph node in the optimal path, hierarchical division is performed to generate hierarchical scheduling logic.

[0102] In this embodiment, the computer device determines the execution layer to which each graph node belongs based on its dependency depth in the optimal path. Graph nodes with the same dependency depth are assigned to the same execution layer, while graph nodes with different dependency depths are assigned to different execution layers. The computer device determines the execution order of each execution layer according to the order of dependency depth from shallow to deep, generating hierarchical scheduling logic containing at least one execution layer. Subtasks within the same execution layer have no data dependencies and can be executed concurrently; subtasks between different execution layers are executed serially according to the order of dependency depth from shallow to deep.

[0103] For example, the optimal execution path for the three subtasks (i.e., querying the current free quota (node ​​A, weight 0.45), querying the current actual emissions (node ​​B, weight 0.45), and calculating the quota gap (node ​​C, weight 0.10)) is s to A to C to e (or s to B to C to e). Here, the dependency depth of nodes A and B is both 1 (directly dependent on the starting node s), and the dependency depth of node C is 2 (dependent on both A and B). The computer device performs hierarchical partitioning based on dependency depth: nodes A and B are assigned to the first execution layer. Node C is assigned to the second execution layer. The hierarchical scheduling logic is represented as Flow={ , },in ={A, B} ={C}. During execution, nodes A and B in the first execution layer execute concurrently. After both have completed their execution, node C in the second execution layer begins execution.

[0104] If the optimal path includes all three nodes (s to A to B to C to e), then the dependency depths are: node A depth 1, node B depth 2, and node C depth 3, and the hierarchy is divided as follows: ={A}, ={B}, ={C}, each layer executes sequentially. By dividing the optimal path into layers, the computer device ultimately generates hierarchical scheduling logic that enables concurrent execution within the same layer and sequential execution between layers. Where M is the total number of layers.

[0105] In an exemplary embodiment, the "path search of the directed acyclic graph based on the weighted adaptive ant colony optimization algorithm to obtain the optimal path from the starting node to the ending node" in S402 above, such as Figure 5 As shown, it includes:

[0106] S501, in the current iteration, control multiple ants to move on the directed acyclic graph based on the state transition probability. Each ant moves from the starting node to the ending node, generating candidate execution paths.

[0107] The state transition probability is determined based on the pheromone concentration between the current node and the next node on the candidate execution path and the heuristic information. The heuristic information is used to guide the ant to prioritize nodes with high task weight and low execution cost.

[0108] In this embodiment of the application, the computer device processes all directed edges in the directed acyclic graph. The initial pheromone concentration is non-uniformly initialized to replace the uniform initialization strategy of the standard ant colony algorithm. The computer device assigns differentiated values ​​to the initial pheromone based on the task weight, which can be represented by relation (7), as shown below:

[0109] (7);

[0110] in, This is the basic pheromone concentration parameter. This initialization strategy ensures that the path containing high-weight tasks has a higher pheromone concentration at the algorithm's startup, accelerating the algorithm's early convergence towards the optimal path while retaining some room for exploration to avoid prematurely getting trapped in local optima.

[0111] In the In the next iteration, the ants Located at node At that time, from the set of candidate nodes that satisfy the executable state Select the next execution node The computer device makes a selection based on the state transition probability, which can be represented by the relation (8), as shown below:

[0112] (8);

[0113] in, As a pheromone heuristic factor, it controls the intensity of the influence of historical path experience on decision-making; The expected heuristic factor controls the strength of the influence of heuristic information on the current decision. For ants At the node This refers to the set of all executable nodes that satisfy the prerequisite dependency constraints and have not yet been visited. ,in For ants The set of nodes visited in the current iteration. This is for cases where a next node cannot be selected from the candidate set (i.e., ...). However, the termination point has not yet been reached. The computer device triggers a dependency waiting mechanism, marking the current ant as suspended, waiting for the predecessor node to complete and update. Status and restore selection.

[0114] Heuristic information in standard ant colony algorithm Typically, it's only linked to the reciprocal of the path length, failing to effectively utilize the semantic attributes of the task itself. Therefore, computer devices rely heavily on heuristic information. The design was reworked to take both task weights and execution costs into account, and the path was adjusted accordingly. The comprehensive heuristic information above is redefined and can be represented by relation (9), which is shown below:

[0115] (9);

[0116] in For the target node Task weights The cost of its execution, and These are the weight contribution index and the cost penalty index, used to adjust the influence of execution cost on heuristic information. This definition gives high-contribution, low-cost task nodes higher guiding priority in path selection, thereby guiding the algorithm to prioritize planning high-quality analysis paths under computational constraints.

[0117] For example, an ant at the starting node s faces two choices, node A and node B. Since their heuristic information values ​​are equal, the ant chooses node A or node B with similar probabilities. After each ant completes its movement from s to e, multiple candidate execution paths are generated.

[0118] S502, for each candidate execution path, calculate the path quality of the candidate execution path based on the sum of the task weights of the graph nodes contained in the candidate execution path and the execution cost, and perform pheromone enhancement on the candidate execution paths that meet the preset path quality conditions.

[0119] In this embodiment of the application, each ant completes one cycle from... arrive After a complete path search, an execution sequence is generated. ,in To satisfy the node access arrangement of all dependency constraints, the computer device performs a comprehensive quality evaluation on each candidate execution path and defines a path quality function, which can be represented by relation (10), as shown below:

[0120] (10);

[0121] Where the numerator is the sum of the task weights of all nodes on the path, reflecting the total decision contribution of the path; in the denominator... The total execution cost of the path. The critical path length corresponding to this path (i.e., from) arrive The longest dependency chain time). This is the critical path penalty coefficient, used to constrain the overall execution latency. The larger the value, the higher the decision contribution of the path per unit cost, and the better the path quality.

[0122] For example, consider two candidate paths, s to A to C to e and s to B to C to e. Both paths contain two task nodes, and their weights are the same as their execution costs. Therefore, the two paths... Equal. Computer equipment for path quality. Candidate execution paths that meet preset conditions (such as ranking in the top N) are enhanced with pheromones. This means that additional pheromone increments are applied to elite paths to further increase the pheromone concentration on high-quality paths, guiding ants in subsequent iterations to prioritize these paths.

[0123] S503, determine the current pheromone evaporation rate based on the current iteration number.

[0124] Among them, the current pheromone evaporation rate decreases as the number of iterations increases.

[0125] In this embodiment, after completing path quality evaluation and elite enhancement for each candidate execution path, the computer device further determines the pheromone evaporation rate for this iteration. Unlike the standard ant colony algorithm, which uses a fixed evaporation rate leading to a decline in exploration capability in the later stages of iteration, the computer device designs a dynamic evaporation rate. The dynamic evaporation rate can be adaptively adjusted with the number of iterations. It can be represented by relation (11), which is shown below:

[0126] (11);

[0127] in, and These are the upper and lower bound parameters for the volatile rate; This represents the maximum number of iterations. This dynamic evaporation rate allows the algorithm to have a high evaporation rate in the early stages of iteration, accelerating the fading of old path information and facilitating the expansion of the search range; while in the later stages of iteration, it has a low evaporation rate, preserving the pheromones accumulated on high-quality paths and promoting the algorithm to converge to the global optimum.

[0128] S504 Updates the pheromone concentration in the directed acyclic graph based on the path quality of each candidate execution path and the current pheromone evaporation rate.

[0129] In this embodiment of the application, after completing one iteration, the computer device performs pheromone updates on all edges of the directed acyclic graph, i.e. .in, This indicates the pheromones remaining after evaporation. This represents the increase in pheromone deposited by all ants on the new path during this iteration.

[0130] Among them, pheromone increment It consists of two superimposed parts, namely The first part represents the contribution of ordinary ants, and the second part represents the enhancement amount of elite paths.

[0131] For ordinary ants Its border The pheromone contribution can be represented by relation (12), which is shown below:

[0132] (12);

[0133] in, The number of edges contained in the path is used to normalize the effect of path length on pheromone distribution.

[0134] For Elite Path Enhancement Quantity In other words, it mainly refers to the path with the best quality in the current iteration. The path applies an additional pheromone reward, and the enhancement amount of this elite path can be represented by relation (13), which is shown below:

[0135] (13);

[0136] in, For the first Elite Pathway Elite enhancement coefficient, To reduce the weight of ranking, this ensures that higher-ranking elite paths receive stronger pheromone boosts. As an indicator function, when the edge The value is 1 if it belongs to the elite path, and 0 otherwise.

[0137] Through the aforementioned global pheromone update mechanism, the computer device realizes the complete process of pheromone evaporation and deposition in one iteration. In the next iteration, the ants will make decisions on path selection based on the updated pheromone concentration, thereby gradually improving path quality.

[0138] S505, repeatedly execute the steps of controlling multiple ants to move on the directed acyclic graph based on the state transition probability until the preset convergence condition is met, and determine the candidate execution path with the highest pheromone concentration as the optimal path.

[0139] In this embodiment, the computer device repeatedly executes the iterative process from S501 to S504 until a preset convergence condition is met. This convergence condition may be that the number of iterations reaches a preset maximum number of iterations. Alternatively, the optimal path may remain unchanged throughout multiple iterations. After the iterations are complete, the computer determines the candidate execution path with the highest pheromone concentration as the optimal path.

[0140] After the algorithm converges, the computer device moves from the globally optimal path. Extract and execute parallel execution logic. Specifically, computer devices... Execution hierarchy partitioning involves grouping all tasks with the same in-degree and whose predecessor nodes have all been completed into the same execution layer. ,Right now ;in This indicates the layer number. Tasks belonging to the same execution layer have no data dependencies and can be concurrently scheduled and executed by computer devices; different execution layers proceed sequentially according to their layer numbers, and the next layer can only start after the previous layer is completed. This ultimately forms a structured, layered parallel execution logic. .

[0141] In an exemplary embodiment, the above-mentioned S203, "according to the hierarchical scheduling logic, calling the large language model and the pre-built dual-carbon knowledge base layer by layer for reasoning until all sub-tasks are completed, and fusing the reasoning results of each sub-task to generate a structured decision report," is as follows: Figure 6 As shown, it includes:

[0142] S601, according to the order of each execution layer in the hierarchical scheduling logic, add each subtask in the ready state in the current execution layer to the ready queue.

[0143] In this embodiment, the hierarchical scheduling logic clarifies the execution order and concurrency relationship of each subtask. Therefore, when the computer device receives this hierarchical scheduling logic... Then, check the global task status table. Perform initialization for all subtask nodes. Perform status registration. Each task node maintains the following fields in the status table: Execution Status Precursor completion counter (Records the number of completed predecessor nodes), result cache slots (Used to store the result data returned after the node completes execution) and a retry counter. (Initialized to 0). Subsequently, the computer device performs scheduling according to the order of execution layers in the hierarchical scheduling logic. Specifically, after initialization, the computer device scans the first execution layer of the logic flow. Set the state of all task nodes to [state]. and add it to the ready queue. The task is awaiting scheduling and execution. The ready state indicates that all prerequisites for this subtask have been completed and it can be executed immediately. For subsequent tasks, their initial state is always "pending execution." Once all predecessor nodes have completed their tasks, the state transition will be triggered by the computer equipment.

[0144] For example, the hierarchical scheduling logic is Flow={ , },in ={node A, node B}, ={Node C}. The computer device first scans... Nodes A and B in the list have no prerequisites and are in a ready state, so nodes A and B are added to the ready queue. Node C in the diagram depends on nodes A and B, and its initial state is pending execution. It is not added to the ready queue.

[0145] S602, retrieve the subtask from the ready queue, generate a retrieval query instruction based on the semantic information of the subtask, and select multiple candidate texts related to the subtask from the dual-carbon knowledge base according to the retrieval query instruction.

[0146] In this embodiment of the application, the computer device is removed from the ready queue. Take out the ready state Subtask nodes of state Transform it into standardized search commands The encoding process of this retrieval instruction comprehensively considers three types of information sources. Specifically, the first type of information source is the semantics of the subtask itself, that is, the query intent and business meaning determined during the decomposition process of the subtask, such as "querying the free quota allocation of controlled emission enterprises within the current compliance period"; the second type of information source is the constraints in the user's dual-carbon business analysis request, including limiting parameters such as time range, industry category, and geographical range; the third type of information source is the output result of the predecessor node, that is, if the task There exists a set of predecessor nodes. The computer device then caches the results of all predecessor nodes. The concatenation adds contextual information and, together with the semantics of the subtask ontology, constitutes a composite query instruction to support cross-task data transfer and conditional acceptance.

[0147] Computer devices can generate standard retrieval instructions in natural language form by assembling the three types of information using templates. It serves as a unified input for knowledge base retrieval and large language model invocation.

[0148] Subsequently, the computer device invokes a preset dual-carbon domain embedding model. Search instructions for this standard Perform mapping to convert it into a query semantic vector. and execute on the query vector The normalization process can be represented by relation (14), which is shown below:

[0149] (14);

[0150] The normalized query vector and the normalized text block vector stored in the knowledge base are in the same metric space, and semantic similarity can be directly calculated through inner product operation to ensure the consistency of retrieval calculation.

[0151] The computer device normalizes the query vector. Submitted to a vector database, an approximate nearest neighbor search is performed using the IVF-PQ index structure. The search process first determines the candidate cluster set to which the query vector belongs (taking the cluster closest to the query vector). (15) clusters), and then calculate the cosine similarity between the query vector and each knowledge base vector within the candidate clusters. It can be represented by relation (15), which is shown below:

[0152] (15);

[0153] Computer devices are sorted from highest to lowest cosine similarity, and the devices with the highest similarity are extracted. A set of text blocks as candidate texts, i.e. Meanwhile, to further ensure the quality of search results, computer equipment optimizes the candidate support set. Apply a similarity lower bound filter to remove noisy text blocks with insufficient semantic relevance, i.e. ,in This is a preset similarity filtering threshold. When When the set is empty, it means that there is no supporting text related to the current subtask in the knowledge base. The computer device records this situation and marks the knowledge base coverage of the subtask as zero in the final report. The large language model makes inferences based on its own parameterized knowledge, and at the same time issues a knowledge blind spot warning to the user.

[0154] For example, the computer device retrieves the subtask corresponding to node A from the ready queue, namely, querying the current period's free quota. Since this subtask has no predecessor node, the computer device generates a retrieval query instruction (i.e., what is the free quota for this quarter) based on its ontology semantics (querying the free quota) and the constraints in the user's dual-carbon business analysis request (the time range is this quarter). The computer device then invokes the dual-carbon domain embedding model. The query command is converted into a query vector and retrieved from the dual-carbon knowledge base to obtain multiple candidate texts related to the free quota. For the subtask corresponding to node B, which queries the actual emissions for the current period, the computer generates a corresponding retrieval query command and retrieves candidate texts related to the actual emissions from the knowledge base.

[0155] S603, sort the candidate texts based on their relevance to the search query command, and select the candidate text with the highest ranking as the target text.

[0156] In this embodiment of the application, after the computer device completes the lower similarity limit filtering, it can obtain multiple candidate texts (i.e., candidate support sets). Since the search results may contain semantically repetitive text blocks (such as adjacent blocks of the same policy document), directly injecting them all into the context of a large language model would result in redundant information accumulation, affecting the inference quality of the large language model. Therefore, the computer device uses a candidate support set... The maximum marginal relevance (MMR) iterative selection strategy is implemented, and each time a text block satisfying the following criteria is selected from the remaining candidate set and added to the final result set. ,Right now ,in This is a balance coefficient between relevance and diversity; the first term measures the semantic relevance between the candidate text block and the query; the second term measures the overlap between the candidate text block and the most similar text block in the selected set; and the weights of the first and second terms are determined by... Control. The computer device repeats the above selection process until a selection is made. text blocks ( This yields a final search result set that balances relevance and diversity. By employing an iterative selection strategy based on maximum marginal relevance, the computer device effectively enhances the information diversity within the result set while ensuring the relevance of candidate texts to the retrieval query. The computer device will then select... Each text block serves as the target text and is used to construct the reasoning context in subsequent steps.

[0157] S604 assembles the target text with the subtask to generate a reasoning context, and performs reasoning on the reasoning context based on the large language model to obtain the reasoning result of the subtask.

[0158] In this embodiment of the application, the computer device will retrieve the final search result set. The text blocks within serve as external knowledge support, combined with standardized instructions for subtasks. Construct a complete reasoning context according to the preset prompt word template. ,Right now .in, The system prompts for the dual-carbon domain include role definitions, answer guidelines, numerical accuracy requirements, and output format constraints. This is the query command for the current subtask; Supporting text for knowledge base retrieval; This is the result data for the preceding task. After assembling the above four types of information according to a fixed template, the computer device can call the large language model to perform inference based on this inference context and return the structured inference result of the current subtask. .

[0159] For example, the computer device assembles the target text (the rules for allocating free quotas) with the subtask (querying the current period's free quotas) to generate complete prompts containing role settings, query instructions, and supporting documents. After calling a large language model, it obtains structured quota data. Thus, the computer device completes the entire execution process of a subtask from retrieval to inference.

[0160] S605, if all subtasks in the current execution layer have been completed, add the subtasks in the next execution layer that meet the prerequisite dependency conditions to the ready queue, and repeat the step of retrieving subtasks from the ready queue until all subtasks in all execution layers have been completed.

[0161] In this embodiment, after a single subtask completes execution and its inference result is cached, the computer device continues to retrieve the next subtask from the ready queue for execution, until all subtasks in the current execution layer have been completed. After all subtasks in the current execution layer have been completed, the computer device performs hierarchical scheduling according to the hierarchical scheduling logic. The hierarchical structure summarizes and confirms the execution status within each execution layer, and then performs synchronization control between execution layers. Specifically, when the first... The execution state of all subtasks in the execution layer is transferred to... Then, the computer device continues scanning the next execution layer (the...). Within each subtask in the layer, the computer checks whether all its direct predecessor nodes have been completed. For subtasks that meet the prerequisites, the computer updates their status to ready. The task is entered into a ready state and added to the ready queue; for subtasks that do not meet the prerequisite dependency conditions, they remain in a waiting state. The inter-layer synchronization condition is determined if and only if the task... When all direct predecessor nodes have been completed, i.e. At this time, the task It can enter the ready state to participate in scheduling. Different execution layers proceed sequentially according to their layer numbers; the next execution layer can only start after the previous one has completed. Simultaneously, for multiple ready tasks within the same execution layer, the computer device employs a concurrent scheduling strategy, simultaneously initiating multiple retrieval and inference requests to the knowledge base and the large language model, fully utilizing the system's concurrent processing capabilities.

[0162] For example, Nodes A and B in the system are executed simultaneously using a concurrent scheduling strategy. After both have completed their execution, the computer device will then scan. Node C in the process checks that its dependent nodes A and B have both completed, and then adds node C to the ready queue. The computer device retrieves node C from the ready queue, performs the corresponding retrieval and inference operations, and completes the subtask of calculating the quota gap. At this point, all subtasks in all execution layers have been completed.

[0163] Furthermore, the computer equipment is equipped with a tiered exception handling mechanism during task execution. When a large language model call or knowledge base retrieval fails due to network timeouts or non-compliant output formats of the large language model, the computer equipment first determines the execution status of the subtask. Specifically, if the subtask... If the first execution fails, the computer device will increment the retry counter by 1 and rebuild the inference context. A retry is then initiated, with format validation constraints added to the prompt during the retry to correct format deviations in the model output. The maximum number of retries is reached. If this still fails, the computer device will terminate the subtask. Set the status to failure The impact range is assessed based on the subtask's position in the hierarchical scheduling logic. If the subtask is a non-critical path node, it is skipped, and the result cache slot is filled with a missing marker. Continue to advance the execution of subsequent tasks; if this subtask If it is a critical path node, a global alarm will be triggered, the execution of the current logical flow will be suspended, and the user will be notified to intervene manually.

[0164] Finally, after all execution layers of the hierarchical scheduling logic have completed, the computer device retrieves the data from the global task state table. The results of each subtask are aggregated and cached to form a complete set of subtask results. Each result It includes the subtask number, execution status, returned data content, and source information of the referenced knowledge base text block.

[0165] S606 integrates the reasoning results of each subtask to generate a structured decision report.

[0166] In this embodiment of the application, the computer device obtains the subtask result set. Subsequently, since the results of each subtask still exist in discrete form at this stage and lack logical connections between them, they cannot directly meet the user's need for a comprehensive decision report. Therefore, the computer equipment calls a large language model to perform global knowledge fusion on the aforementioned discrete results, calculates the dual-carbon core indicators, and ultimately generates a structured decision report for the user.

[0167] In the subtask result set Before being fed into the large language model for fusion, the computer equipment first performs preprocessing and quality checks on the results of each subtask. Specifically, for numerical results, the computer equipment performs a dimensional consistency check (i.e., dimensional unification check), converting the values ​​returned by each subtask to a standard unit system (e.g., carbon emissions are uniformly measured in tons of equivalent, and monetary indicators are uniformly measured in RMB). For result entries with missing labels (i.e., placeholder results from failed tasks), the computer equipment marks the data confidence as low during fusion and notes the reason for the missing data in the corresponding analysis dimension of the final report. The computer equipment defines the set of valid results after preprocessing as... Its scale meets Subsequently, the computer device, based on the set of valid results... By leveraging a large language model and invoking internal reasoning capabilities, combined with industry-standard calculation methods, the system outputs core quantitative indicators for dual-carbon processes. The computer equipment then processes dual-carbon business analysis requests. Industry macroeconomic constraints (Including current compliance policy requirements, industry benchmark emission intensity, market price upper and lower limits, etc.), set of valid results The calculation results of the dual-carbon core quantitative indicators are then assembled into a complete reasoning context according to the fusion prompt word template. Submit the large language model to perform global inference. The prompt word template explicitly constrains the output of the large language model to cover four dimensions: carbon emission status analysis, compliance situation assessment, trading strategy suggestions, and risk warnings, and to be returned in a structured format so that the subsequent report rendering module can directly parse it.

[0168] Large language models based on fusion context After completing the reasoning, output a structured comprehensive decision report. The report should cover four parts: the first part is a carbon emission status analysis, including total emissions, emission intensity, and a comparative analysis with historical data for the same period within the statistical period; the second part is a compliance situation assessment, based on the quota compliance gap. The report is divided into four parts: 1) The first part provides an assessment of the current compliance pressure level, along with the compliance deadline and an estimated quota shortfall and required replenishment funds; 2) The second part offers carbon trading strategy recommendations, providing optimal purchase timing and quantity suggestions for entities with compliance shortfalls based on market price trends; and 3) For entities with quota surpluses, providing strategies for holding or selling surplus quotas and estimated expected returns; and 4) The third part outlines risk warnings, listing key uncertainties affecting the final decision, including carbon price volatility risk, policy adjustment risk, and explanations of the confidence level of the analysis due to data gaps. The final report, along with a list of sources cited in the dual-carbon knowledge base for each sub-task, is output. Users can trace the original policy documents or verified data upon which each conclusion in the report is based using the citation tracing function, ensuring the interpretability and verifiability of the report content.

[0169] In summary, based on all the above embodiments, a method for processing dual-carbon business is also provided, such as... Figure 7 As shown, the method includes:

[0170] S701 receives user input requests for dual-carbon service analysis;

[0171] S702 performs semantic parsing on dual-carbon business analysis requests and determines the hierarchical analysis module based on the parsing results;

[0172] S703, based on the preset dual-carbon business rules, calls the large language model to compare the importance of each subtask in the solution layer pairwise, determines the judgment matrix based on the comparison results, and solves the judgment matrix to obtain the weight value corresponding to each subtask.

[0173] S704, calculate the dynamic filtering threshold based on the weight value of each sub-task, compare the dynamic filtering threshold with each weight value, and generate a task set based on the comparison result;

[0174] S705 maps each subtask in the task set to a graph node, determines the directed edges between graph nodes based on business logic dependencies, and determines a directed acyclic graph.

[0175] S706 controls multiple ants to move on a directed acyclic graph based on state transition probabilities. Each ant moves from the starting node to the ending node, generating candidate execution paths.

[0176] S707, For each candidate execution path, calculate the path quality of the candidate execution path based on the sum of the task weights of the graph nodes contained in the candidate execution path and the execution cost, and perform pheromone enhancement on the candidate execution paths that meet the preset path quality conditions.

[0177] S708, determine the current pheromone evaporation rate based on the current iteration number;

[0178] S709, Update the pheromone concentration in the directed acyclic graph based on the path quality of each candidate execution path and the current pheromone evaporation rate;

[0179] S710, repeat S706-S709 until the preset convergence condition is met, and determine the candidate execution path with the highest pheromone concentration as the optimal path;

[0180] S711, based on the business logic dependencies of each graph node in the optimal path, performs hierarchical division and generates hierarchical scheduling logic;

[0181] S712, according to the order of each execution layer in the hierarchical scheduling logic, add each subtask in the current execution layer that is in the ready state to the ready queue;

[0182] S713, retrieve the subtask from the ready queue, generate a retrieval query instruction based on the semantic information of the subtask, and select multiple candidate texts related to the subtask from the dual-carbon knowledge base according to the retrieval query instruction; execute S720-S722 to construct the dual-carbon knowledge base;

[0183] S714, sort the candidate texts based on their relevance to the search query command, and select the candidate texts with the highest ranking as the target text;

[0184] S715 assembles the target text with the subtask to generate a reasoning context, and performs reasoning on the reasoning context based on the large language model to obtain the reasoning result of the subtask.

[0185] S716, if all subtasks in the current execution layer have been completed, add the subtasks in the next execution layer that meet the prerequisite dependency conditions to the ready queue, and repeat S713-S715 until all subtasks in all execution layers have been completed.

[0186] S717: Obtain the reasoning results of each subtask, perform dimensional uniformity verification on each reasoning result, and mark the reasoning results that fail the verification.

[0187] S718, based on the preset prompt template, assembles the dual-carbon business analysis request, the preset industry macro constraints, the verified inference results, and the calculation results of the dual-carbon core indicators into a fusion prompt word;

[0188] S719 calls upon a large language model to perform global reasoning based on fused prompt words, generating a structured decision report;

[0189] S720: Obtain documents in the dual-carbon domain and perform text segmentation on the documents to generate multiple text blocks;

[0190] S721, based on a preset embedding model, vectorizes each text block to generate a semantic vector index corresponding to each text block;

[0191] S722 stores each semantic vector and its corresponding text block into a vector database to construct a dual-carbon knowledge base.

[0192] The methods described in each of the above steps have been described in the foregoing embodiments. For details, please refer to the foregoing descriptions. They will not be repeated here.

[0193] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0194] Based on the same inventive concept, this application also provides a dual-carbon service processing apparatus for implementing the dual-carbon service processing method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more dual-carbon service processing apparatus embodiments provided below can be found in the limitations of the dual-carbon service processing method described above, and will not be repeated here.

[0195] In one exemplary embodiment, such as Figure 8 As shown, a dual-carbon business processing apparatus is provided, comprising:

[0196] The receiving module 11 is used to receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request to generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0197] The scheduling module 12 is used to generate hierarchical scheduling logic for dual-carbon business analysis requests based on preset business logic dependencies and task sets, and according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0198] The fusion module 13 is used to perform reasoning by calling the large language model and the pre-built dual-carbon knowledge base layer by layer according to the hierarchical scheduling logic until all sub-tasks are completed. It then fuses the reasoning results of each sub-task to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector indexes of documents in the dual-carbon domain. Each module in the above-mentioned dual-carbon business processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0199] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0200] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0201] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0202] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0203] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0204] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0205] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0206] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0207] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0208] Receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request, and generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task.

[0209] Based on the preset business logic dependencies and task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm. The hierarchical scheduling logic contains at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, while the sub-tasks in the same execution layer are executed concurrently.

[0210] Based on the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

[0211] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0212] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0213] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for processing dual-carbon transactions, characterized in that, The method includes: Receive a dual-carbon business analysis request input by the user, and call a preset large language model to perform semantic analysis on the dual-carbon business analysis request to generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task. Based on the preset business logic dependencies and the task set, a hierarchical scheduling logic is generated for the dual-carbon business analysis request according to the weighted adaptive ant colony optimization algorithm; the hierarchical scheduling logic includes at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, and each sub-task in the same execution layer is executed concurrently. According to the hierarchical scheduling logic, the large language model and the pre-built dual-carbon knowledge base are called layer by layer for reasoning until all the sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report. The dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

2. The method according to claim 1, characterized in that, The process of receiving a dual-carbon service analysis request from a user and then using a pre-defined large language model to perform semantic analysis on the request generates a task set containing multiple sub-tasks and their corresponding weight values, including: The dual-carbon business analysis request is semantically parsed, and a hierarchical analysis module is determined based on the parsing results; the hierarchical analysis module includes an objective layer, a criterion layer, and a solution layer; the solution layer contains the multiple sub-tasks. Based on the preset dual-carbon business rules, the large language model is invoked to compare the importance of each subtask in the scheme layer pairwise. The judgment matrix is ​​determined according to the comparison results, and the judgment matrix is ​​solved to obtain the weight value corresponding to each subtask. A dynamic filtering threshold is calculated based on the weight value of each subtask, and the dynamic filtering threshold is compared with each weight value. The task set is generated based on the comparison result.

3. The method according to claim 1, characterized in that, The process of generating hierarchical scheduling logic for the dual-carbon business analysis request based on preset business logic dependencies and the task set, using a weighted adaptive ant colony optimization algorithm, includes: Each subtask in the task set is mapped to a graph node, and the directed edges between the graph nodes are determined according to the business logic dependencies to form a directed acyclic graph; the directed acyclic graph includes a start node and an end node; The weighted adaptive ant colony optimization algorithm is used to perform path search on the directed acyclic graph to obtain the optimal path from the starting node to the ending node. The hierarchical scheduling logic is generated by dividing the graph nodes in the optimal path into layers based on their business logic dependencies.

4. The method according to claim 3, characterized in that, The step of performing path search on the directed acyclic graph based on the weighted adaptive ant colony optimization algorithm to obtain the optimal path from the starting node to the ending node includes: In the current iteration, multiple ants are controlled to move on the directed acyclic graph based on state transition probabilities. Each ant moves from the starting node to the ending node, generating candidate execution paths. The state transition probabilities are determined based on the pheromone concentration between the current node and the next node on the candidate execution path and heuristic information. The heuristic information is used to guide the ants to prioritize nodes with high task weights and low execution costs. For each candidate execution path, the path quality of the candidate execution path is calculated based on the sum of the task weights of the graph nodes contained in the candidate execution path and the execution cost, and pheromone enhancement is performed on the candidate execution paths that meet the preset path quality conditions. The current pheromone evaporation rate is determined based on the current iteration number; the current pheromone evaporation rate decreases as the iteration number increases; The pheromone concentration in the directed acyclic graph is updated based on the path quality of each candidate execution path and the current pheromone evaporation rate. The process of controlling multiple ants to move on the directed acyclic graph based on state transition probabilities is repeated until a preset convergence condition is met, and the candidate execution path with the highest pheromone concentration is determined as the optimal path.

5. The method according to claim 1, characterized in that, The process involves sequentially calling the large language model and the pre-built dual-carbon knowledge base for reasoning according to the hierarchical scheduling logic, until all sub-tasks are completed. The reasoning results of each sub-task are then fused to generate a structured decision report, including: According to the order of each execution layer in the hierarchical scheduling logic, each subtask in the current execution layer that is in the ready state is added to the ready queue; The subtask is retrieved from the ready queue, a retrieval query instruction is generated based on the semantic information of the subtask, and multiple candidate texts related to the subtask are selected from the dual-carbon knowledge base based on the retrieval query instruction. The candidate texts are sorted according to their relevance to the search query command, and the candidate texts with the highest ranking are selected as the target texts. The target text is assembled with the subtask to generate a reasoning context, and reasoning is performed on the reasoning context based on the large language model to obtain the reasoning result of the subtask. If all the subtasks in the current execution layer have been completed, the subtasks in the next execution layer that meet the prerequisite dependency conditions are added to the ready queue. The step of retrieving the subtasks from the ready queue is repeated until all subtasks in all execution layers have been completed. The reasoning results of each of the sub-tasks are merged to generate the structured decision report.

6. The method according to claim 5, characterized in that, The step of fusing the reasoning results of each of the sub-tasks to generate a structured decision report includes: Obtain the reasoning results of each subtask, perform a dimensional uniformity check on each reasoning result, and mark the reasoning results that fail the check. According to the preset prompt template, the dual-carbon business analysis request, the preset industry macro constraints, the verified inference results, and the dual-carbon core indicator calculation results are assembled into a fusion prompt word; the dual-carbon core indicator calculation results are obtained by the large language model based on the verified inference results of the verified inference results. The large language model is invoked to perform global reasoning based on the fused prompt words, generating a structured decision report; the structured decision report includes an analysis of the current status of carbon emissions, an assessment of compliance, suggestions for carbon trading strategies, and risk warnings.

7. The method according to any one of claims 1-6, characterized in that, The construction of the dual-carbon knowledge base includes: Obtain the dual-carbon domain document and perform text block processing on the dual-carbon domain document to generate multiple text blocks; Based on a preset embedding model, each text block is vectorized to generate a semantic vector index corresponding to each text block. Each semantic vector and its corresponding text block are stored in a vector database to construct the dual-carbon knowledge base.

8. A processing apparatus for dual-carbon business, characterized in that, The device includes: The receiving module is used to receive the dual-carbon business analysis request input by the user, and call the preset large language model to perform semantic analysis on the dual-carbon business analysis request to generate a task set containing multiple sub-tasks and the weight values ​​corresponding to each sub-task. The scheduling module is used to generate hierarchical scheduling logic for the dual-carbon business analysis request based on the preset business logic dependencies and the task set, according to the weighted adaptive ant colony optimization algorithm; the hierarchical scheduling logic includes at least one execution layer, each execution layer corresponds to a set of sub-tasks, and each execution layer is executed serially according to the business logic dependencies, and each sub-task in the same execution layer is executed concurrently. The fusion module is used to call the large language model and the pre-built dual-carbon knowledge base layer by layer according to the hierarchical scheduling logic to perform reasoning until all the sub-tasks are completed, and to fuse the reasoning results of each sub-task to generate a structured decision report; the dual-carbon knowledge base is used to store the semantic vector index of dual-carbon domain documents.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images