Agent-driven power grid engineering project review method, system, device and storage medium

Abstract

This invention discloses an agent-driven method, system, device, and storage medium for reviewing power grid engineering projects. Applied to the field of power grid engineering project review technology, the method includes: acquiring the target power grid engineering project report and current review instructions, inputting them into a review model; the review model using a document parsing agent to split the report into paragraph text and table data; then, using a parallel pipeline composed of a text paragraph parsing agent and a table parsing agent to parse the paragraph text and table data respectively, obtaining corresponding textual evidence and table evidence; and finally, using an evidence fusion agent to align and fuse the textual and table evidence to form a joint context. This joint context fully preserves the argumentative information in the paragraphs and the structured data in the tables, providing a large language model for generating review results. This invention improves the accuracy and efficiency of evidence extraction and generation in long feasibility study documents through dual-channel parallel processing.

Description

Technical Field

[0001] This invention relates to the field of power grid engineering project review technology, and in particular to an agent-driven method, system, device and storage medium for power grid engineering project review. Background Technology

[0002] In engineering construction management, feasibility study review is a crucial step in project decision-making. The review process typically involves examining the necessity of the project, its construction content, technical solutions, and project cost. It encompasses multiple stages, including preliminary document review, feasibility study evaluation, and feedback and revisions. This process is broad in scope, involves numerous technical indicators, and is subject to strong regulatory constraints, demanding a high level of professional knowledge and experience from review experts. In scenarios supporting feasibility study reviews, existing publicly available technical approaches can be broadly categorized as rule-based automatic matching and verification, and key information extraction and content generation based on large-scale models.

[0003] Rule-based automatic matching and verification methods typically rely on manually preset rule templates and indicator systems. While they can effectively perform format checks, consistency verification, and matching of some normative clauses, their ability to understand complex semantic information is limited, making them unsuitable for the comprehensive review needs of feasibility study reports, which contain a large amount of unstructured text and cross-chapter related content. Key information extraction and content-assisted generation based on large models have strong capabilities in natural language understanding and text generation. However, feasibility study reports are lengthy, structurally complex, and informationally dispersed, making models susceptible to interference from redundant content. This can lead to insufficient evidence extraction or omission of key information, making it difficult to guarantee that the output meets actual business requirements in terms of professional standardization and logical consistency. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of the present invention provide an agent-driven method, system, device, and storage medium for reviewing power grid engineering projects. This addresses the issue that existing agent-driven power grid engineering project review methods suffer from inaccurate extraction of key information due to interference from redundant content in power grid project reports, resulting in low review accuracy.

[0005] A first aspect of this invention provides an agent-driven method for reviewing power grid engineering projects, comprising: Obtain the target power grid engineering project report and current review instructions; The target power grid project report and current review instructions are input into the review model to extract the joint context. The power grid project is then reviewed based on the joint context to obtain the review results. The extraction of the joint context is achieved through the following steps: The document parsing agent in the review model is used to split the target power grid project report into paragraph text and table data; Input the paragraph text and the current review instruction into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence; The table parsing agent in the review model is used to retrieve and filter table data and current review instructions to obtain table evidence; The evidence fusion agent in the review model is used to fuse textual evidence and tabular evidence to obtain a joint context.

[0006] Furthermore, the paragraph text and the current review instruction are input into the paragraph text parsing agent in the review model for retrieval and filtering to obtain textual evidence, including: The text paragraphs in the target power grid project report are segmented to obtain multiple sentence lists, where each sentence list corresponds to a sequential number; Perform dual-granularity semantic encoding on each list of sentences to obtain multiple paragraph-level vectors; Calculate the semantic similarity between each paragraph-level vector and the current review instruction to obtain the corresponding matching score for each paragraph-level vector; The paragraph-level vectors with matching scores greater than a preset matching threshold are used as candidate paragraphs to obtain a set of candidate paragraphs; Calculate the retention probability of each sentence in each candidate paragraph in the candidate paragraph set, process each candidate paragraph according to the retention probability of each sentence, and obtain the corresponding final candidate paragraph. Each final candidate paragraph was identified as textual evidence.

[0007] Furthermore, each candidate paragraph is processed based on the retention probability of each sentence to obtain the corresponding final candidate paragraphs, including: If the retention probability is greater than the preset probability threshold, the sentence will be retained. If the retention probability is less than the preset probability threshold, the sentence is deleted, and the corresponding final candidate paragraph is obtained.

[0008] Furthermore, the table parsing agent in the review model is used to retrieve and filter the table data and the current review instructions to obtain table evidence, including: Divide each row of data in each table to obtain the corresponding hierarchical path information and attribute information; Based on the path information and corresponding attribute information at each level, a feature tree is constructed for each table of data. Based on each feature tree, the intra-table relationship index of each table of data is obtained. Reasoning is performed on the current review instructions to obtain structured results. Based on the query terms in the structured results, the data in each table is searched to obtain an initial candidate table set. The query terms include at least entity terms and target attribute terms. The difficulty of the current review instruction is assessed based on the structured results. If the difficulty of the task is low complexity, it is selected from the initial candidate table set for screening to obtain tabular evidence. If the task difficulty is high complexity, a local graph structure is constructed based on the initial candidate table set. Table data related to each table data in the initial candidate table set is searched in the local graph structure to obtain an extended candidate set. Table evidence is obtained based on the extended candidate set.

[0009] Furthermore, based on the initial candidate table set, a local graph structure is constructed, including: Extract the two-dimensional matrix and merged cell information of each table data in the initial candidate table set; Based on the information of each two-dimensional matrix and merged cell, a local graph structure is constructed, which is used to describe the relationship between the data in each table in the initial candidate table set.

[0010] Furthermore, based on the expanded candidate set, tabular evidence is obtained, including: Calculate the consistency score between the tabular data and the structured results in the expanded candidate set; Calculate the semantic similarity between the data in each table in the expanded candidate set and the current review instruction to obtain a similarity score; Calculate the graph structure bonus score for each table data in the expanded candidate set and the current review instruction. Then, sum the graph structure bonus score, similarity score, and consistency score in a weighted manner to obtain the comprehensive score for each table data in the expanded candidate set. Table data with a comprehensive score greater than a preset score threshold will be used as target table data in the expanded candidate set; Based on the target table data, the final expanded candidate set is obtained, and the table data in the final expanded candidate set and the initial candidate table set are determined as table evidence.

[0011] Another embodiment of the present invention provides an agent-driven method system for reviewing power grid engineering projects, comprising: The acquisition module is used to acquire the target power grid engineering project report and current review instructions; The review module is used to input the target power grid engineering project report and the current review instructions into the review model to extract the joint context. Based on the joint context, the power grid engineering project is reviewed to obtain the review results. The extraction of the joint context is achieved through the following steps: The document parsing agent in the review model is used to split the target power grid project report into paragraph text and table data; Input the paragraph text and the current review instruction into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence; The table parsing agent in the review model is used to retrieve and filter table data and current review instructions to obtain table evidence; The evidence fusion agent in the review model is used to fuse textual evidence and tabular evidence to obtain a joint context.

[0012] A third aspect of the present invention provides a computer device, comprising: Memory, used to store computer programs; A processor is used to execute computer programs to implement steps of an agent-driven power grid engineering project review method, as described in the first aspect.

[0013] A fourth aspect of the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the agent-driven power grid engineering project review method of the first aspect.

[0014] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: By acquiring the target power grid project report and current review instructions, and inputting them into the review model, the model uses a document parsing agent to break down the report into paragraph text and table data. Then, a parallel pipeline consisting of a text paragraph parsing agent and a table parsing agent is constructed to parse the paragraph text and table data respectively, obtaining corresponding textual evidence and table evidence. Finally, an evidence fusion agent is used to align and fuse the textual and table evidence to form a joint context. This joint context fully preserves the argumentative information in the paragraphs and the structured data in the tables, providing a large language model for generating review results. This invention improves the accuracy and efficiency of extracting heterogeneous evidence from long feasibility study documents through a dual-channel parallel processing and precision-overhead balancing strategy. Attached Figure Description

[0015] Figure 1 This is a flowchart of an agent-driven power grid engineering project review method according to one embodiment of the present invention. Figure 2 This is a diagram illustrating the review model architecture of an agent-driven power grid engineering project review method according to one embodiment of the present invention. Figure 3 This is a flowchart illustrating the intelligent agent processing of an agent-driven power grid engineering project review method in one embodiment of the present invention. Figure 4 This is a flowchart of the table parsing intelligent agent processing of an agent-driven power grid engineering project review method in one embodiment of the present invention. Figure 5 This is a system block diagram of an agent-driven power grid engineering project review system according to one embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0018] One embodiment of the present invention provides an agent-driven method for reviewing power grid engineering projects. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The flowchart shown is a method for reviewing power grid engineering projects driven by an intelligent agent according to one embodiment of the present invention, including steps S101 to S102, the specific steps of which are as follows: S101. Obtain the target power grid project report and current review instructions.

[0019] The execution entity in this embodiment is an electronic device. This embodiment is not limited to a specific electronic device. For example, the electronic device in this embodiment can be a computer; or the electronic device in this embodiment can be a mobile phone; or the electronic device in this embodiment can also be a server.

[0020] In this embodiment, the target power grid project report refers to the documentation of the specific power grid project to be reviewed. It typically includes structured and unstructured content such as project overview, design drawings, bill of quantities, equipment and material details, construction plan, budget, contract terms, and acceptance records. The current review instruction refers to the specific review task or question raised by the user regarding the target power grid project report. For example: "Please identify the concrete-related items belonging to the foundation engineering and check if there are any abnormalities in their workload" or "Analyze the items with abnormal workloads in this project."

[0021] S102. Input the target power grid project report and the current review instruction into the review model to extract the joint context. Review the power grid project based on the joint context to obtain the review result. The extraction of the joint context is achieved through the following steps: use the document parsing agent in the review model to split the target power grid project report to obtain paragraph text and table data; input the paragraph text and the current review instruction into the paragraph text parsing agent in the review model for retrieval and filtering to obtain textual evidence; use the table parsing agent in the review model to retrieve and filter the table data and the current review instruction to obtain table evidence; use the evidence fusion agent in the review model to fuse the textual evidence and table evidence to obtain the joint context.

[0022] In this embodiment, the review model consists of multiple agents, such as Figure 2 As shown, this includes document parsing agents, paragraph text parsing agents, table parsing agents, and evidence fusion agents. Specifically, for the two types of heterogeneous information coexisting in a document—continuous text paragraphs and structured tables—separate paragraph parsing agents and table parsing agents are designed. Under the unified scheduling of the review model, text information and table information are processed separately through different channels. The two types of agents operate independently and synchronously during processing, thereby improving document parsing efficiency.

[0023] In this embodiment, textual evidence refers to relevant text fragments retrieved and filtered from the target power grid project report according to the current review instructions; tabular evidence refers to relevant table rows or cells retrieved and filtered from the target power grid project report according to the current review instructions; and the combined context refers to the comprehensive information formed by integrating the above textual and tabular evidence. This embodiment can simultaneously utilize unstructured text descriptions and structured tabular data to form a more complete and complementary evidentiary context, thereby improving the accuracy of power grid project assurance reviews.

[0024] In this embodiment, after inputting the target power grid project report and the current review instruction into the review model, the model supports importing the feasibility study report and traversing the paragraphs and tables in the original document order. The model then performs text formatting, such as removing leading and trailing whitespace, compressing consecutive spaces, and standardizing line breaks, to reduce the impact of document formatting differences on subsequent processing. For tables, it extracts two-dimensional matrix data and further converts it into row-level structured objects. The paragraph text parsing agent then refines and filters the paragraph content. If the review instruction is an input query, the model calculates the semantic matching degree in each paragraph of the feasibility study document, filters out key paragraphs highly relevant to the review focus, removes redundant or irrelevant content, and outputs refined textual evidence.

[0025] The system utilizes a table parsing agent to further parse and structure the table data, completing header normalization, duplicate field disambiguation, and row data expansion. Simultaneously, the system restores the hierarchical relationships of the tables and constructs a structured representation, forming queryable and scalable table evidence. Then, an evidence fusion agent is used to uniformly aggregate and integrate textual and table evidence, ultimately forming supporting content for review tasks.

[0026] In some embodiments, the paragraph text and the current review instruction are input into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence, including: The text paragraphs in the target power grid engineering project report are segmented to obtain multiple sentence lists; Perform dual-granularity semantic encoding on each list of sentences to obtain multiple paragraph-level vectors; Calculate the semantic similarity between each paragraph-level vector and the current review instruction to obtain the corresponding matching score for each paragraph-level vector; The paragraph-level vectors with matching scores greater than a preset matching threshold are used as candidate paragraphs to obtain a set of candidate paragraphs; Calculate the retention probability of each sentence in each candidate paragraph in the candidate paragraph set, process each candidate paragraph according to the retention probability of each sentence, and obtain the corresponding final candidate paragraph. Each final candidate paragraph was identified as textual evidence.

[0027] In this embodiment, the document parsing agent is first invoked to split the report, extracting the text paragraphs and table data. The extracted text paragraphs and the current review instructions are then input into the paragraph text parsing agent. For example... Figure 3 As shown, the paragraph text parsing agent first segments each paragraph into a list of sentences based on sentence-ending punctuation such as periods, question marks, exclamation marks, and semicolons, and assigns a sequential number to each sentence. The BGE-small-zh model is used to perform dual-granularity semantic encoding on each paragraph, generating both paragraph-level and sentence-level vectors. After encoding, the semantic similarity between each paragraph-level vector and the current review instruction is calculated, yielding a matching score for each vector. Paragraphs with matching scores greater than a threshold are marked as candidate paragraphs, forming a candidate paragraph set. For each paragraph in the candidate paragraph set, the system further uses a word-level content compression header to calculate the retention probability of each sentence within that paragraph. This retention probability represents the degree to which the sentence supports the current review instruction. Based on the retention probabilities of each sentence, candidate paragraphs are filtered to form final candidate paragraphs, which are then identified as textual evidence. This textual evidence, along with the tabular evidence output by the table parsing agent, is fed into the evidence fusion agent to construct a joint context, supporting the auxiliary generation of subsequent review content.

[0028] When training the text parsing agent for this paragraph, during the sample construction phase, sample data of R&D reports and review comments are acquired, and "review comments - candidate paragraphs" are defined as the basic processing unit. For each sample, information such as review comment identifier, paragraph identifier, review comment text, and the full text of the candidate paragraph are retained, and sentence-level segmentation is performed on the candidate paragraphs. Specifically, the initial review model splits the original paragraph into several sentence sequences based on Chinese sentence-ending punctuation such as periods, question marks, exclamation marks, and semicolons, and assigns sequential numbers to each sentence for subsequent annotation and supervised learning. In this way, the originally coarse-grained paragraph matching problem is further refined into the discrimination problem of "which sentences within a paragraph can directly support the current review comments," providing a more fine-grained modeling foundation for locating key evidence.

[0029] In the supervisory information construction phase, the system uses an automatic annotation method based on a large language model to generate silver labels, reducing the cost of manual annotation. Specifically, the system inputs review comments and numbered candidate sentences into the annotation model, guiding it to identify sentence numbers directly related to the review comments or providing supporting evidence; if no relevant content exists in a paragraph, an empty result is returned. Subsequently, the system parses the model output to obtain the set of relevant sentence indices corresponding to the current sample and saves it as a structured annotation result. In terms of model design, the Chinese semantic representation model BGE-small-zh is selected as the basic encoder, and a paragraph-level relevance scoring head and a word-level content compression head are added to it, constructing a dual-granularity joint modeling framework. Let the review comments be represented as q, and the candidate paragraphs as p. The concatenation of these two is input into the semantic encoder to obtain the context semantic representation matrix. in, L represents the length of the input sequence, and d represents the dimension of the hidden layer representation. Indicates the first The context semantic vector corresponding to each position.

[0030] This representation simultaneously encodes the interaction information between review comments and candidate paragraphs, providing a unified semantic foundation for subsequent paragraph-level matching judgments and intra-paragraph key content selection. Specifically, the paragraph-level relevance scoring head utilizes the overall semantic representation of the input sequence to model the degree of matching between review comments and candidate paragraphs, outputting a relevance score reflecting the overall support strength. Let... Let the overall representation vector of the input sequence be used, then the paragraph-level matching score can be expressed as: in, and These represent the weight parameters and bias terms of the paragraph-level relevance scoring head, respectively. (.) represents the activation function. This indicates the degree to which the candidate paragraph matches the current review comments. The larger the value, the more likely the paragraph as a whole is to serve as candidate evidence for the current review task.

[0031] The word-level content compression head performs binary classification judgments on each position in the input sequence to identify which content should be retained and which should be discarded, thereby realizing the location and screening of key evidence within the paragraph.

[0032] In some embodiments, each candidate paragraph is processed according to the retention probability of each sentence to obtain the corresponding final candidate paragraph, including: If the retention probability is greater than the preset probability threshold, the sentence will be retained. If the retention probability is less than the preset probability threshold, the sentence is deleted, and the corresponding final candidate paragraph is obtained.

[0033] In this embodiment, if the retention probability of a sentence exceeds a preset probability threshold, the sentence is retained; otherwise, it is deleted. Finally, the sentences retained from each candidate paragraph are recombined in their original order to form the final candidate paragraphs.

[0034] It should be noted that the preset matching threshold and preset probability threshold can be determined according to actual needs, and this embodiment does not impose specific limitations.

[0035] In some embodiments, a table parsing agent in the review model is used to retrieve and filter table data and current review instructions to obtain table evidence, including: Divide each row of data in each table to obtain the corresponding hierarchical path information and attribute information; Based on the path information and corresponding attribute information at each level, a feature tree is constructed for each table of data. Based on each feature tree, the intra-table relationship index of each table of data is obtained. Reasoning is performed on the current review instructions to obtain structured results. Based on the query terms in the structured results, the data in each table is searched to obtain an initial candidate table set. The query terms include at least entity terms and target attribute terms. The difficulty of the current review instruction is assessed based on the structured results. If the difficulty of the task is low complexity, it is selected from the initial candidate table set for screening to obtain tabular evidence. If the task difficulty is high complexity, a local graph structure is constructed based on the initial candidate table set. Table data related to each table data in the initial candidate table set is searched in the local graph structure to obtain an extended candidate set. Table evidence is obtained based on the extended candidate set.

[0036] In this embodiment, tabular data from the target power grid project report is obtained, such as the "Civil Engineering List" and "Equipment and Materials Summary Table," while receiving current review instructions, such as "Please analyze whether there are any abnormalities in the workload of concrete-related items under the foundation engineering item." Figure 4 As shown, the table parsing agent first divides each table's data into rows. For complex tables with hierarchical features, it identifies their hierarchical columns, such as first-level, second-level, and third-level projects, and splits each row into hierarchical path information (e.g., civil engineering / foundation engineering / concrete pouring) and attribute information. Subsequently, it constructs a feature tree for the table based on the hierarchical path information of each row, using column semantics as the index layer and cell values ​​as the content layer. An intra-table relationship index is then established on the feature tree to record attribute co-occurrence relationships within the same row, category affiliation relationships within the same column, and hierarchical relationships within the hierarchical path.

[0037] The current review instruction is input into the large language model, which performs one inference through a preset structured prompt template and outputs a structured result in a fixed JSON format. This result includes the task intent, entity words, target attribute words, search prompt words, expected headers, multi-hop relationship markers, and target evidence field features and keyword information. Based on the task intent and complexity criteria in this structured result, the review model assesses the task difficulty of the current review instruction. If it only involves reading a single entity attribute or extracting a simple one-hop correspondence, it is determined to be a low-complexity task; if it involves hierarchical attribution tracing, enumeration of peer projects, comparative analysis, or summary statistics, it is determined to be a high-complexity task. For example, in this embodiment, the current review instruction is to identify concrete projects under the foundation engineering item and check for workload anomalies. This instruction involves hierarchical attribution and comparative analysis, and is therefore determined to be a high-complexity task.

[0038] For highly complex tasks, the system first uses entity words, target attribute words, and related hints in the structured results to perform cross-table searches in the tabular data using feature trees and intra-table relationship indexes to construct an initial candidate table set. For example, it recalls row records containing "concrete" from the "Civil Engineering List" and "Equipment and Materials Summary Table" to obtain an initial candidate pool. Then, the system constructs a local graph structure based on the initial candidate table set: it extracts two-dimensional matrices and merged cell information from the tables to which the candidate records belong, constructing a local graph containing row relationships, column relationships, parent-child hierarchical relationships, sibling node relationships, and merged cell associations. Next, it uses a semantic retrieval model (such as the BGE-small-zh model) to calculate the semantic similarity between the current review requirements and the candidate records. Only when semantic consistency is high, it supplements the candidate records with relevant row content, column content, parent node information, and sibling node information according to the task type. The system adds these candidate records obtained through graph structure expansion to an expanded candidate set, and finally selects evidence rows with high matching and sufficient structural support from the expanded candidate set as tabular evidence output.

[0039] If the task is determined to be of low complexity, such as reading the workload values ​​of a concrete project, the review model can directly perform rapid filtering and rearrangement based on a tree structure in the initial candidate table set, outputting a small amount of highly relevant table evidence without the need to construct a local graph structure.

[0040] In some embodiments, a local graph structure is constructed based on an initial candidate table set, including: Extract the two-dimensional matrix and merged cell information of each table data in the initial candidate table set; Based on the information of each two-dimensional matrix and merged cell, a local graph structure is constructed, which is used to describe the relationship between the data in each table in the initial candidate table set.

[0041] In this embodiment, the review model has initially selected a candidate table set from multiple tables based on the entity words and target attribute words in the review instructions. For example, this set includes two tables: "Civil Engineering List" and "Equipment and Materials Summary Table". Then, based on the recall results, local graph structure enhancement is performed. First, the data in each table is preprocessed in a structured manner, converting the two-dimensional table into row-level records and identifying hierarchical columns to generate hierarchical fields such as path, parent, and child. At the same time, parent-child relationship edges and tree indexes are constructed based on the hierarchical paths. Subsequently, the two-dimensional matrix and merged cell information of the table to which the candidate records belong are extracted to construct a local graph structure that includes the relationships between rows, columns, parent and child levels, sibling nodes, and merged units.

[0042] In some embodiments, tabular evidence is obtained based on an expanded candidate set, including: Calculate the consistency score between the tabular data and the structured results in the expanded candidate set; Calculate the semantic similarity between the data in each table in the expanded candidate set and the current review instruction to obtain a similarity score; Calculate the graph structure bonus score for each table data in the expanded candidate set and the current review instruction. Then, sum the graph structure bonus score, similarity score, and consistency score in a weighted manner to obtain the comprehensive score for each table data in the expanded candidate set. Table data with a comprehensive score greater than a preset score threshold will be used as target table data in the expanded candidate set; Based on the target table data, the final expanded candidate set is obtained, and the table data in the final expanded candidate set and the initial candidate table set are determined as table evidence.

[0043] In this embodiment, for high-complexity tasks, an expanded candidate set has been obtained through graph structure enhancement. The table parsing agent further performs a comprehensive score on each table data in the expanded candidate set. First, a consistency score is calculated based on the structured result of the current review instruction: if a candidate record matches an entity word, target attribute column, or corresponding level field, a positive bonus is given. For example, "concrete pouring" matches the entity word "concrete" and its parent node is "foundation engineering," which matches the prompt that it belongs to foundation engineering, thus obtaining a high consistency score. Second, the semantic similarity between the current review instruction and the candidate record line text is calculated using the BGE-small-zh semantic retrieval model to obtain a similarity score. Third, for candidate records obtained from graph path expansion, the system extracts parent node information, sibling node information, row information, column information, and merged cell association information from its local graph neighborhood, and matches these relationship information with expected field names and expected keywords. A graph structure bonus score is calculated based on the number and type of hits. Then, the system weights and sums the consistency score, similarity score, and graph structure bonus score according to preset weights to obtain a comprehensive score for each candidate record. Candidate records with a comprehensive score greater than a preset score threshold are used as target table data. Finally, these target table data are added to the final expanded candidate set, and together with the table data in the initial candidate table set that were not expanded but also met the threshold conditions, they are identified as table evidence.

[0044] In some embodiments, calculating the graph structure bonus score for each tabular data in the extended candidate set and the current review instruction includes: Extract neighborhood relationship information of each table's data within the local graph structure data; The neighborhood relationship information is filtered according to the type of the current review instruction to obtain the filtered neighborhood relationship information; The filtered neighborhood relationship information is matched with the query terms in the current review instruction to obtain the matching results. The matching results include at least the number of fields and the number of keywords that are matched. The graph structure bonus score is calculated based on the number of fields and keywords hit.

[0045] In this embodiment, neighborhood relationship information is first extracted from its local graph structure, including parent node information, sibling node information, row information, column information, and association information formed by merged cells. Then, based on the task type of the current review instruction, the neighborhood relationship information is filtered. For example, when the task involves hierarchical attribution tracing and anomaly judgment, parent node information related to "basic engineering," sibling node information related to "concrete," and column information related to workload anomaly judgment are filtered out. The filtered neighborhood relationship information includes parent nodes, sibling nodes, and column workload. The filtered neighborhood relationship information is then matched with query terms in the current review instruction, such as entity words, target attribute words, expected field names, and expected keywords. If the matching results are as follows: the number of matched fields is 1, the number of matched keywords is 1, and the keyword for the "basic engineering" level path is also matched. According to preset rules, such as: In the formula, The number of fields hit. The number of keywords hit. These are the keywords for the hit hierarchical path. , and These are the corresponding weights.

[0046] This invention also provides an agent-driven power grid engineering project review method system, which is used to implement the agent-driven power grid engineering project review method of this embodiment. Figure 5 This is a structural block diagram of an agent-driven power grid engineering project review system 500 according to an embodiment of the present invention, including: Module 501 is used to obtain the target power grid engineering project report and current review instructions; The review module 502 is used to input the target power grid engineering project report and the current review instruction into the review model to extract the joint context, review the power grid engineering project based on the joint context, and obtain the review result. The extraction of the joint context is achieved through the following steps: using the document parsing agent in the review model to split the target power grid engineering project report to obtain paragraph text and table data; inputting the paragraph text and the current review instruction into the paragraph text parsing agent in the review model for retrieval and filtering to obtain textual evidence; using the table parsing agent in the review model to retrieve and filter table data and the current review instruction to obtain table evidence; and using the evidence fusion agent in the review model to fuse the textual evidence and table evidence to obtain the joint context.

[0047] The specific implementation method of the intelligent agent-driven power grid engineering project review method system is basically the same as the specific implementation method of the intelligent agent-driven power grid engineering project review method described above, and will not be repeated here.

[0048] In one embodiment, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the steps of the above-described method.

[0049] In one embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0050] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., SSD), etc.

[0051] 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 computer-readable storage medium, and when the computer program is executed, it can include the processes of the embodiments of the above methods.

[0052] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for reviewing power grid engineering projects driven by intelligent agents, characterized in that, include: Obtain the target power grid engineering project report and current review instructions; The target power grid project report and the current review instruction are input into the review model to extract the joint context. The power grid project is then reviewed based on the joint context to obtain the review result. The extraction of the joint context is achieved through the following steps: The document parsing agent in the review model is used to split the target power grid project report to obtain paragraph text and table data; The paragraph text and the current review instruction are input into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence; The table parsing agent in the review model is used to retrieve and filter the table data and the current review instruction to obtain table evidence; The evidence fusion agent in the review model is used to fuse the textual evidence and the tabular evidence to obtain a joint context.

2. The agent-driven power grid engineering project review method as described in claim 1, characterized in that, The process involves inputting the paragraph text and the current review instruction into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence, including: The text paragraphs in the target power grid project report are segmented to obtain multiple sentence lists; Perform dual-granularity semantic encoding on each of the sentence lists to obtain multiple paragraph-level vectors; Calculate the semantic similarity between each paragraph-level vector and the current review instruction to obtain the corresponding matching score for each paragraph-level vector; The paragraph-level vectors whose matching scores are greater than a preset matching threshold are used as candidate paragraphs to obtain a set of candidate paragraphs; Calculate the retention probability of each sentence in each candidate paragraph in the candidate paragraph set, and process each candidate paragraph according to the retention probability of each sentence to obtain the corresponding final candidate paragraph; Each of the aforementioned final candidate paragraphs was determined as textual evidence.

3. The agent-driven power grid engineering project review method as described in claim 2, characterized in that, The step of processing each candidate paragraph according to the retention probability of each sentence to obtain the corresponding final candidate paragraph includes: If the retention probability is greater than a preset probability threshold, then the sentence is retained; If the retention probability is less than the preset probability threshold, the sentence is deleted to obtain the corresponding final candidate paragraph.

4. The agent-driven power grid engineering project review method as described in claim 1, characterized in that, The process of using the table parsing agent in the review model to retrieve and filter the table data and the current review instruction to obtain table evidence includes: Each row of the data in each of the tables is divided to obtain the corresponding hierarchical path information and attribute information; Based on the hierarchical path information and corresponding attribute information, a feature tree for each table data is constructed, and based on the feature tree, an intra-table relation index for each table data is obtained. The current review instruction is reasoned to obtain a structured result. Based on the query terms in the structured result, a search is performed in each of the table data to obtain an initial candidate table set, wherein the query terms include at least entity terms and target attribute terms. The task difficulty of the current review instruction is assessed based on the structured results. If the task difficulty is low complexity, the candidate table is selected from the initial candidate table set for screening to obtain tabular evidence. If the task difficulty is high complexity, then a local graph structure is constructed based on the initial candidate table set. In the local graph structure, table data related to each table data in the initial candidate table set is searched to obtain an extended candidate set. Based on the extended candidate set, table evidence is obtained.

5. The agent-driven power grid engineering project review method as described in claim 4, characterized in that, The step of constructing a local graph structure based on the initial candidate table set includes: Extract the two-dimensional matrix and merged cell information of each table data in the initial candidate table set; Based on the two-dimensional matrices and the merged cell information, a local graph structure is constructed, wherein the local graph structure is used to describe the relationship between the table data in the initial candidate table set.

6. The agent-driven power grid engineering project review method as described in claim 4, characterized in that, The process of obtaining tabular evidence based on the expanded candidate set includes: Calculate the consistency score between each of the tabular data in the expanded candidate set and the structured result; Calculate the semantic similarity between each of the table data in the expanded candidate set and the current review instruction to obtain a similarity score; Calculate the graph structure bonus score for each of the table data in the expanded candidate set and the current review instruction, and then sum the graph structure bonus score, the similarity score, and the consistency score in a weighted manner to obtain the comprehensive score for each of the table data in the expanded candidate set. The table data whose overall score is greater than a preset score threshold is used as the target table data in the expanded candidate set; Based on the target table data, a final extended candidate set is obtained, and the table data in the final extended candidate set and the initial candidate table set are determined as table evidence.

7. The agent-driven power grid engineering project review method as described in claim 1, characterized in that, The calculation of the graph structure bonus score for each of the table data in the expanded candidate set and the current review instruction includes: Extract the neighborhood relationship information of each of the table data in the local graph structure data; The neighborhood relationship information is filtered according to the type of the current review instruction to obtain the filtered neighborhood relationship information; The filtered neighborhood relationship information is matched with the query terms in the current review instruction to obtain a matching result, wherein the matching result includes at least the number of fields and the number of keywords that are matched. The graph structure bonus score is calculated based on the number of fields and keywords hit.

8. A smart agent-driven power grid engineering project review system, characterized in that, include: The acquisition module is used to acquire the target power grid engineering project report and current review instructions; The review module is used to input the target power grid project report and the current review instruction into the review model to extract the joint context, review the power grid project based on the joint context, and obtain the review result. The extraction of the joint context is achieved through the following steps: The document parsing agent in the review model is used to split the target power grid project report to obtain paragraph text and table data; The paragraph text and the current review instruction are input into the paragraph text parsing agent in the review model for retrieval and filtering to obtain text evidence; The table parsing agent in the review model is used to retrieve and filter the table data and the current review instruction to obtain table evidence; The evidence fusion agent in the review model is used to fuse the textual evidence and the tabular evidence to obtain a joint context.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the agent-driven power grid engineering project review method as described in any one of claims 1 to 7 when executing the computer program.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the agent-driven power grid engineering project review method as described in any one of claims 1 to 7.

Images