Power grid project cost automatic examination method, system, equipment and medium

Abstract

The invention discloses a power grid project cost automatic review method, system, device and medium, and belongs to the technical field of power grid project cost review. The power grid project cost automatic review method designs a set of complete multi-agent workflow for complex project data in power grid cost data; related data files can be intelligently indexed according to given rules, and reports and key information can be automatically extracted from multi-source input documents by utilizing an OCR technology and a Python tool library. And performing data analysis, verification and review report summarization by using an agent collaboration framework based on a multi-modal large model. According to the method, accurate identification and analysis of multi-source data can be realized, and a relatively comprehensive review conclusion can be quickly output. According to the method, the labor cost is reduced, the review efficiency and the data processing quality are greatly improved, and efficient, reliable and traceable technical support is provided for power grid project cost review.

Description

Technical Field

[0001] This application belongs to the technical field of power grid engineering cost review, and in particular relates to an automated method, system, equipment and medium for power grid engineering cost review. Background Technology

[0002] Cost review of power grid projects is a crucial link in power grid construction management. With the continuous expansion of power grid project scale, the volume of data involved in cost review is also growing exponentially. This data encompasses everything from design drawings and construction plans to material lists and final settlement data, exhibiting complex data types and diverse sources. Currently, the cost data review work at Tianjin Company still relies primarily on manual labor. This traditional review method has several problems: First, cost data review mainly relies on manual verification of project quantities, quota applications, equipment and material prices, and fee standards, resulting in a massive workload and low overall efficiency. Second, it depends on a large amount of professional manpower, leading to high costs. Third, the types of standards and documents used for review are numerous, including national and industry standards, internal company regulations, pricing standards, design plans, and construction drawings, making manual processing difficult. Furthermore, differences in the experience and professional level of reviewers can easily lead to inconsistent results, affecting review quality and decision-making accuracy.

[0003] Some existing technologies attempt to improve review efficiency by introducing information technology and intelligent methods. For example, by establishing a review rule base, using OCR and NLP technologies to recognize and parse the text in the tender documents, converting the text information into structured data, and comparing it item by item according to preset review rules, the final review results and bid evaluation information are generated. However, current methods mainly target directly readable documents or table files, as well as PDF files containing simple tables. Engineering drawings in power grid engineering cost documents often contain complex multi-level header structures, and the table content may span multiple pages. Traditional technologies are prone to problems such as misaligned headers, data loss, and confused row and column correspondences when processing such complex tables. At the same time, power grid drawings contain a large number of professional graphic symbols. These graphics have specific industry standards and complex geometric features, and text annotations are closely intertwined with graphic elements. Text may be located inside graphics, overlap with lines, or use special fonts and layouts. The accuracy of existing OCR technology is clearly insufficient and cannot meet the needs of power grid engineering cost review.

[0004] In recent years, with the rapid development of large language models in natural language processing, text understanding, and multimodal information analysis, their applications in automated document parsing, intelligent data analysis, and decision support have received widespread attention. Large language models can effectively handle complex tasks that traditional rule engines struggle to cover through deep semantic understanding and contextual reasoning, demonstrating significant advantages, especially in the analysis of unstructured data and cross-modal information. However, when faced with tasks like power grid engineering cost review, which involve cross-document and cross-format data and include complex reports and graphical information, a single model often cannot simultaneously handle the multiple needs of document parsing, data analysis, and report generation. It lacks a complete end-to-end workflow and a specialized division of labor and collaboration mechanism, resulting in limited overall performance and difficulty in meeting the data processing needs of large-scale engineering projects.

[0005] Therefore, it is necessary to provide a new method, system, equipment, and medium for automated review of power grid engineering costs to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this disclosure is to provide an automated method, system, device, and medium for reviewing the cost of power grid projects in order to solve the above-mentioned problems.

[0007] This disclosure achieves the above objectives through the following technical solutions: An automated method for reviewing the cost of power grid projects includes the following steps: Construct document parsing agent, task allocation agent, data analysis agent, result verification agent, and report summary agent; The document parsing agent performs unified parsing of multi-source engineering documents to generate structured data. The task allocation agent automatically breaks down the review rules and assigns different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution. The structured data is analyzed collaboratively from multiple specialized intelligent agents within the data analysis intelligent agent in a multi-dimensional manner. The result verification agent performs a secondary review of the output results of all the professional agents. The report summarizes how the intelligent agent integrates the analysis and review results to generate the final engineering review report.

[0008] As a further optimization of this disclosure, a document parsing intelligent agent is used to uniformly parse multi-source engineering documents to generate structured data, including: The document parsing agent receives the review objectives and rule descriptions input by the user, identifies the required file name or path according to the rules, and then reads the corresponding engineering cost document from the specified file server or archive. For files of different formats, the document parsing agent adopts a customized parsing process to transform the original file content into a unified structured representation while preserving document hierarchy and location information; For complex tables and scanned PDF drawings, a training set of OCR models specifically for power grid engineering is constructed, and the OCR models are optimized and adjusted.

[0009] As a further optimization of this disclosure, a training set for a dedicated OCR model for power grid engineering is constructed. By collecting construction drawings and material lists from actual engineering projects, the original data source is formed; Combine manual annotation with annotations to mark the table structure, text content, and formulas; By introducing review experts to assist in quality verification, the completeness, semantic consistency, and format specifications of the labeled data are automatically checked and optimized.

[0010] As a further optimization of this disclosure, the review rules are automatically broken down by a task allocation agent, and different types of sub-tasks are assigned to the corresponding professional agents in the data analysis agent for execution, including: The task allocation agent receives the review rules, transforms the review rules into a structured task description, and obtains a set of subtasks. Based on the set of subtasks, the task allocation planning agent assigns the subtasks to corresponding specialized agents for execution according to their attributes and data requirements; when there are dependencies between subtasks, the task allocation agent plans and determines the execution order of the subtasks.

[0011] As a further optimization of this disclosure, the structured data is subjected to multi-dimensional collaborative analysis by multiple specialized intelligent agents within the data analysis intelligent agent, including: The data analysis agent includes a numerical calculation agent, a data review agent, and a standardization review agent; The numerical computation agent parses structured tabular data, identifies the calculation data and corresponding calculation formulas in the bill of quantities, and executes the calculation tasks. The data review agent performs consistency and rationality analysis on various types of structured data in engineering projects; The standardized review agent performs automated and standardized checks on engineering project documents based on industry standards, national regulations, and internal company procedures.

[0012] As a further optimization of this disclosure, a result verification agent performs a secondary review of the output results of all the aforementioned specialized agents, including: After the professional intelligent agent completes its execution, the result verification intelligent agent receives the review results, original engineering documents, and supporting materials from the professional intelligent agent, and performs a second verification on the review results of the professional intelligent agent, including data integrity, calculation correctness, and logical consistency. When abnormal errors are detected in the review results, the cause of the error is analyzed, the analysis strategy is adjusted, and the task allocation intelligent agent and related professional intelligent agents are automatically triggered to re-analyze, or a list of issues requiring manual review is generated.

[0013] As a further optimization of this disclosure, the report summarizing agent integrates the analysis and review results to generate a final engineering review report, including: After completing the professional intelligent agent analysis and secondary review, the report summarizes and organizes all review results to form the final engineering review report. The final engineering review report categorizes and prioritizes various issues, marks risk points and sources of abnormal data, and locates the relevant document locations, specific fields, and values. The final engineering review report may include tables, charts, or highlighted original images to present issues intuitively. The final engineering review report generates targeted modification suggestions, proposing improvement plans for calculation errors, data inconsistencies, or standardization deficiencies. The final engineering review report supports multi-level output formats, including detailed technical reports and management summary reports.

[0014] An automated cost review system for power grid projects includes: The agent building module is used to build document parsing agents, task allocation agents, data analysis agents, result verification agents, and report summarization agents; The document parsing module is used to uniformly parse multi-source engineering documents through the document parsing intelligent agent to generate structured data; The task allocation module is used to automatically break down the review rules through the task allocation agent and assign different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution. The data analysis module performs multi-dimensional collaborative analysis of the structured data through multiple specialized intelligent agents within the data analysis intelligent agent. The result verification module is used to perform a secondary review of the output results of all the professional agents through the result verification agent. The report summary module is used to integrate the analysis and review results through the report summary agent to generate the final engineering review report.

[0015] As a further optimization of this disclosure, the document parsing module performs unified parsing of multi-source engineering documents through the document parsing intelligent agent to generate structured data, including: The document parsing agent receives the review objectives and rule descriptions input by the user, identifies the required file name or path according to the rules, and then reads the corresponding engineering cost document from the specified file server or archive. For files of different formats, the document parsing agent adopts a customized parsing process to transform the original file content into a unified structured representation while preserving document hierarchy and location information; For complex tables and scanned PDF drawings, a training set of OCR models specifically for power grid engineering is constructed, and the OCR models are optimized and adjusted.

[0016] As a further optimization of this disclosure, a training set for a dedicated OCR model for power grid engineering is constructed, including: By collecting construction drawings and material lists from actual engineering projects, the original data source is formed; Combine manual annotation with annotations to mark the table structure, text content, and formulas; By introducing review experts to assist in quality verification, the completeness, semantic consistency, and format specifications of the labeled data are automatically checked and optimized.

[0017] As a further optimization of this disclosure, the task allocation module automatically decomposes the review rules through the task allocation agent, and assigns different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution, including: The task allocation agent receives the review rules, transforms the review rules into a structured task description, and obtains a set of subtasks. Based on the set of subtasks, the task allocation planning agent assigns the subtasks to corresponding specialized agents for execution according to their attributes and data requirements; when there are dependencies between subtasks, the task allocation agent plans and determines the execution order of the subtasks.

[0018] As a further optimization of this disclosure, the data analysis module performs multi-dimensional collaborative analysis of the structured data through multiple specialized agents within the data analysis agent, including: The data analysis agent includes a numerical calculation agent, a data review agent, and a standardization review agent; The numerical computation agent parses structured tabular data, identifies the calculation data and corresponding calculation formulas in the bill of quantities, and executes the calculation tasks. The data review agent performs consistency and rationality analysis on various types of structured data in engineering projects; The standardized review agent performs automated and standardized checks on engineering project documents based on industry standards, national regulations, and internal company procedures.

[0019] As a further optimization of this disclosure, the result verification module performs a secondary review of the output results of all the professional agents through the result verification agent, including: After the professional intelligent agent completes its execution, the result verification intelligent agent receives the review results, original engineering documents, and supporting materials from the professional intelligent agent, and performs a second verification on the review results of the professional intelligent agent, including data integrity, calculation correctness, and logical consistency. When abnormal errors are detected in the review results, the cause of the error is analyzed, the analysis strategy is adjusted, and the task allocation intelligent agent and related professional intelligent agents are automatically triggered to re-analyze, or a list of issues requiring manual review is generated.

[0020] As a further optimization of this disclosure, the report summary module integrates the analysis and review results through the report summary agent to generate a final engineering review report, including: After completing the professional intelligent agent analysis and secondary review, the report summarizes and organizes all review results to form the final engineering review report. The final engineering review report categorizes and prioritizes various issues, marks risk points and sources of abnormal data, and locates the relevant document locations, specific fields, and values. The final engineering review report may include tables, charts, or highlighted original images to present issues intuitively. The final engineering review report generates targeted modification suggestions, proposing improvement plans for calculation errors, data inconsistencies, or standardization deficiencies. The final engineering review report supports multi-level output formats, including detailed technical reports and management summary reports.

[0021] An electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor is used to execute the program stored in the memory to implement the automated review method for power grid engineering costs.

[0022] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the automated review method for power grid engineering costs.

[0023] The beneficial effects of this disclosure are as follows: This disclosure introduces a task decomposition and intelligent agent division of labor mechanism to break down complex review rules into multiple types of sub-tasks, such as numerical calculation, data consistency verification, and standardization specification inspection, which are then executed by corresponding professional intelligent agents. This significantly improves the overall efficiency of power grid engineering cost review, reduces human intervention, and increases work accuracy. This disclosure utilizes a multimodal large model to jointly analyze structured data and engineering document images, and fine-tunes the OCR model using a dedicated training set built in the field of power grid engineering, significantly improving its recognition capabilities in scenarios such as multi-page tables, multi-level headers, and complex reports. Through this collaborative approach of multimodal and domain-specific OCR fine-tuning, the system can accurately parse scanned PDFs and image / text information, effectively overcoming the misidentification and missed detection problems that are prone to occur in traditional OCR and single rule engines in complex report processing, ensuring the integrity and accuracy of key data, and providing reliable data support for review. This disclosure uses a result verification agent to perform secondary verification on the output results of various professional agents, automatically discovering potential misjudgments or omissions, and can trigger re-analysis or generate a manual review checklist, thereby enhancing the reliability, traceability and completeness of the review results, and ensuring the accuracy and credibility of the final report. This disclosure integrates analysis results into a well-organized engineering review report through a report summarizing intelligent agent, enabling visualization of anomalies and facilitating rapid understanding of issues and development of improvement plans by managers, engineers, and auditors. It also supports multi-level output to meet the diverse needs of technical and management levels, achieving high readability and usability of the review report. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a method in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the document parsing intelligent agent large model instructions in an embodiment of this disclosure; Figure 3 This is a schematic diagram of the task allocation agent large model instruction in an embodiment of this disclosure; Figure 4 This is a schematic diagram of the numerical computation agent large model instruction in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the large model instructions for the data review agent in an embodiment of this disclosure; Figure 6 This is a schematic diagram of the standardized review agent large model instructions in the embodiments of this disclosure; Figure 7 This is a schematic diagram of the large model instructions for result verification intelligent agent in an embodiment of this disclosure; Figure 8This is a schematic diagram of the large model instructions for the report generation agent in an embodiment of this disclosure; Figure 9 This is a system framework diagram of an embodiment of this disclosure; Figure 10 This is a block diagram of the device structure in an embodiment of this disclosure. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] like Figure 1 As shown, an automated method for reviewing the cost of power grid projects includes the following steps: S1. Construct document parsing agent, task allocation agent, data analysis agent, result verification agent, and report summary agent; Each agent employs a tailored construction strategy to adapt to the power grid engineering cost review scenario. All agents are built upon the Qwen2.5-VL-32B-AWQ multimodal large model, achieving expert behavior through prompting engineering and tool scheduling control modules. The models used in this disclosure are deployed using the vLLM inference framework and support further acceleration of inference through ONNXRuntime or TensorRT in computationally limited scenarios.

[0028] The document parsing agent serves as the underlying data entry point for this disclosure. Based on a large model, it parses required documents according to rules. It utilizes stable and reliable parsing tools and a finely tuned OCR model for the power grid field to achieve unified extraction and processing of multi-source engineering documents. Upon receiving the user's input description of review rules and document directory, the document parsing agent first parses the review rules based on the large model's natural language understanding capabilities, automatically identifying the document types and file paths required for this round of review. The large model's instructions are as follows: Figure 2As shown. Subsequently, a professional toolchain is used for document parsing. For example, for Excel documents, the Pandas library is used to read the table data row by row and column by column, parsing cell attribute information. For Doc documents, the Python-docx library is used for document extraction. For PDF documents, the PyMuPDF library is used to convert PDF pages into image format and perform image preprocessing, including noise reduction, enhancement, and skew correction. For PDF and image files, an OCR model is used for parsing. The OCR model used in this invention is based on the Dots-OCR platform and fine-tuned on an NVIDIA A100 processor using a domain-specific OCR training set annotated by experts in the field of power grid engineering. After parsing is complete, the document parsing agent outputs the parsing results in JSON format for subsequent processes.

[0029] The task allocation agent leverages the natural language understanding and task planning capabilities of the large model. Through rule parsing (Prompt), it transforms review rules into structured key points, identifies required data and verification objectives, and breaks them down into executable subtasks. Each subtask contains standardized attributes such as `task_type`, `required_fields`, `input_documents`, and `upstream_tasks`, describing its task type, field requirements, document input, and dependencies. Subsequently, based on the subtask's attributes and data requirements, the agent automatically routes it to a numerical computation agent, a data review agent, or a standardization review agent, forming a schedulable execution sequence. This sequence is then dynamically triggered by the task scheduler according to dependencies, enabling collaborative execution by multiple agents. The instructions used by the task allocation agent are as follows: Figure 3 As shown. Furthermore, if inconsistencies or misjudgments occur in subsequent processes, the task allocation agent can receive a rollback signal from the result verification agent, and re-plan and generate relevant sub-tasks based on task dependencies.

[0030] The data analysis agent consists of multiple specialized agents, including a numerical computation agent, a data review agent, and a standardization review agent, all built on Qwen2.5-VL-32B-AWQ. The numerical computation agent uses instructions such as... Figure 4 As shown, the system identifies calculation formulas, variables, and units, and converts the calculation process into executable Python code for precise evaluation. The data review agent leverages the model's multimodal understanding capabilities to perform cross-document consistency checks on multi-source data, effectively identifying issues such as numerical inconsistencies, cross-table contradictions, and omissions. The instructions used by the data review agent are as follows: Figure 5 As shown. The main objective of the standardization review agent is to check the format and content of engineering cost documents, bill of quantities data, normative texts, and drawings for compliance with standards. The instructions used by the standardization review agent are as follows: Figure 6As shown.

[0031] After the professional intelligent agent completes the initial review, the result verification intelligent agent reflects and reasons based on the large model, performing a secondary verification of the aforementioned review output, original documents, OCR results, and supporting materials. The verification focuses on data integrity, calculation correctness, and cross-document logical consistency. The instructions used by the result verification intelligent agent are as follows: Figure 7 As shown, when a misjudgment or missed detection is detected by the preceding agent, the result verification agent can output a rollback signal to automatically trigger the task allocation agent and the corresponding professional agent to perform backtracking analysis.

[0032] The report summary agent is responsible for integrating all review results, verification information, and document references to automatically generate a structured, well-organized engineering cost review report suitable for auditors, managers, and technical personnel. The instructions used include... Figure 8 As shown.

[0033] S2. A document parsing agent performs unified parsing of multi-source engineering documents to generate structured data, specifically including: Step 2.1 Reading relevant documents The document parsing agent receives the review objectives and rule descriptions input by the user (such as checking whether the budget list is complete, checking whether the material quantity in the construction material list is consistent with the construction drawings, etc.), identifies the required file name or path (budget, construction drawings, contract, settlement statement, material list, technical specifications, etc.) according to the rules, and then reads the corresponding project cost document from the specified file server or archive.

[0034] Step 2.2 Multi-source document parsing For files of different formats, the document parsing agent adopts a customized parsing process to convert the original file content into a unified structured representation (such as JSON or Markdown) while preserving the document hierarchy and position information.

[0035] Excel-type documents are read row by row and column by column by calling toolkits (such as Python's Pandas library) to parse cell attribute information (merged cells, formulas, comments, formatting information).

[0036] Documents of type Doc can use toolkits (such as Python's Python-docx library) to extract text and table content from paragraphs and tables, while preserving chapter levels and heading numbers.

[0037] PDF documents can be converted to image format using toolkits (such as Python's PyMuPDF library) and preprocessed, including denoising, enhancement, and skew correction. The images are then input into an OCR model for parsing. Image documents do not require conversion and are directly preprocessed before being input into the OCR model.

[0038] Step 2.3 Fine-tuning of the OCR model Power grid cost engineering drawings contain numerous complex tables, such as merged headers, oversized tables, and tables spanning multiple pages. They also include many specialized graphics with specific industry standards and complex geometric features, with text annotations intertwined with graphic elements; text may be located inside graphics or overlap with lines. Existing OCR technology struggles to accurately recognize and parse this type of data. To improve the accuracy of OCR model recognition, this invention constructs a dedicated OCR training set for the power grid engineering field.

[0039] The training set covers various scenarios in power grid cost engineering, such as complex tables (including multi-page tables, multi-level headers, nested tables, etc.) and various graphic structures and technical terms. The construction process of the training set includes the following steps: First, raw data sources are formed by collecting construction drawings and material lists from actual engineering projects; then, combined with manual annotation, the table structure, text content, and formulas are precisely annotated to ensure the accuracy of the training samples; finally, expert review is introduced to assist in quality verification, automatically checking and optimizing the completeness, semantic consistency, and format specifications of the annotated data. By fine-tuning the OCR model on this dataset, it can accurately parse complex tables in power grid engineering drawings and documents, while enhancing the OCR model's ability to recognize power grid technical terms, engineering formulas, and standard numbers, ensuring table hierarchy and data integrity, and providing accurate basic data for subsequent intelligent agent analysis.

[0040] For OCR model base selection, open-source weighted models such as PaddleOCR, OCR-Flux, and Dots-OCR can be used. The initial checkpoints of the model can be selected based on task characteristics; for example, OCR-Flux can be used for documents containing complex tables, while PaddleOCR can be used for text-intensive scenarios. Hardware configuration, computing resources, and latency requirements also need to be considered to select a model with appropriate parameter sizes. To improve the execution efficiency of the OCR model, it is deployed on a vLLM inference framework, supporting batch inference and concurrent pipeline scheduling. In resource-constrained scenarios, ONNX Runtime or TensorRT can also be used for acceleration.

[0041] S3. The review rules are automatically broken down by the task allocation agent, and different types of sub-tasks are assigned to the corresponding professional agents in the data analysis agent for execution. Specifically, this includes: A single project review rule may involve a large amount of data or multiple steps of analysis, affecting the analytical accuracy of the large model. Furthermore, the analytical methods and data types required for different sub-tasks may differ significantly, such as formula calculations, tabular data verification, and document format validation; a single model cannot accommodate multiple professional analytical capabilities simultaneously. In this invention, the task allocation agent, as the task planning and assignment unit in multi-agent collaboration, automatically breaks down complex review rules into executable sub-tasks and assigns them to the corresponding professional agents according to task type, achieving accuracy and efficiency throughout the entire review process.

[0042] Step 3.1 Rule parsing and task breakdown The task allocation agent first receives the review rules and, through the natural language parsing capabilities of the large model, transforms them into a structured task description. Specifically, the agent identifies the required engineering document categories, such as budgets, drawings, and material lists, and extracts key verification points from the rules. Based on this, the agent further divides the overall review objective into sub-tasks, providing clear inputs and task boundaries for subsequent execution by specialized agents.

[0043] Step 3.2 Subtask Assignment After obtaining the set of subtasks, the task allocation and planning agent will assign them to appropriate specialized agents based on the task's attributes and data requirements. For example, tasks related to numerical calculations will be handled by the numerical calculation agent, tasks involving cross-table data consistency will be handled by the data review agent, and tasks related to format and standardization will be handled by the standardization review agent. When there are dependencies between subtasks, such as when cost calculations must be completed before subsequent data reviews can be carried out, the agent can plan and determine the reasonable execution order of the subtasks, thereby ensuring the orderly connection of the overall process.

[0044] S4. The structured data is subjected to multi-dimensional collaborative analysis by multiple specialized intelligent agents within the data analysis intelligent agent, specifically including: The parsed structured data and original documents are analyzed from multiple dimensions by various specialized intelligent agents to achieve automated review. During the analysis, each agent uses the original engineering document image as visual input and the text information recognized by OCR as text input. These are then jointly processed by a multimodal large-scale model to improve the accuracy of table data recognition. This processing effectively addresses issues that may occur when using OCR alone, such as table misalignment, missing content, or loss of information across pages. The multimodal large-scale model comprehensively analyzes the table layout and text information in the image, correcting and supplementing the OCR results, thereby ensuring the integrity and accuracy of key data.

[0045] The data analysis intelligent agent comprises multiple specialized agents, designed according to different review rule types. It is built upon an open-source multimodal large model, combining prompt word engineering and Python-based engineering functions. Specifically, it includes: (1) Numerical Computation Agent: The numerical computation agent is specifically designed to handle the calculation of basic engineering quantities. By parsing tabular data output from structured Excel or OCR, it identifies the calculation data and corresponding calculation formulas in the bill of quantities and executes the calculation tasks. This agent supports converting complex formulas in the calculation into executable code and performing calculations in a Python execution environment, ensuring the accuracy of numerical calculations. At the same time, this agent can record each calculation step and the source cell, achieving data traceability.

[0046] (2) Data Review Agent: The data review agent is responsible for analyzing the consistency and rationality of various structured data in the project. It detects outliers, data inconsistencies, and potential errors by comparing budget, contract terms, construction drawing annotations, and material list data. For example, when the quantity of materials in the budget list does not match the annotations on the construction drawings, the agent automatically generates an error message and marks the location of the relevant documents. The agent can also perform logical checks, such as checking whether the total of the sub-items equals the total price and whether the unit price matches the total price, thereby discovering possible calculation errors or data entry problems.

[0047] (3) Standardization Review Agent: The standardization review agent performs automated standardization checks on engineering project documents based on industry standards, national regulations, and internal company procedures. For engineering documents to be reviewed, the standardization review agent can check whether the documents meet the standards (e.g., whether the project characteristics are complete, whether the project name description is accurate, whether there are any missing or omitted items in the bill of quantities, whether the terminology is accurate, etc.) according to the supporting materials involved in the review rules. The agent will also check the title format, paragraph layout, and table style to ensure that the document as a whole meets the company or industry standards.

[0048] S5. The result verification agent performs a secondary review of the output results of all the aforementioned professional agents, specifically including: After the specialized AI agent completes its execution, the result verification AI agent receives the review results, original engineering documents, and supporting materials from the specialized AI agent. It then performs a secondary verification of the review results, including checks for data integrity, calculation correctness, and logical consistency. Data integrity refers to comparing the original document with the OCR parsing results to check for missing or truncated data. Specifically, this includes verifying the completeness of table rows and columns; checking for missing or replaced characters due to OCR errors; and ensuring correct recognition of technical terms. Calculation correctness involves mathematically verifying the output of the numerical calculation AI agent. Logical consistency verification checks the consistency of data or rules across different documents. By comparing the outputs of each specialized AI agent with the rule requirements, the result verification AI agent can identify potential misjudgments or omissions in the previous execution process. When an abnormal error is detected in the review results, requiring a re-review, the result verification AI agent can analyze the cause of the error, adjust the analysis strategy, and automatically trigger the task allocation AI agent and related specialized AI agents to re-analyze, or generate a list of issues requiring manual review to prevent error propagation and improve the completeness and accuracy of the review results.

[0049] S6. The report summarizes the intelligent agent's integration of analysis and review results to generate the final engineering review report, which includes: After the professional intelligent agent analysis and result verification are completed, the report summary intelligent agent summarizes and organizes all review results to form the final engineering review report. The core task of the report summary intelligent agent is to transform scattered and complex data and analysis results into a structured, clear, and highly readable comprehensive review report, making it easy for managers, engineers, and auditors to understand and use.

[0050] During report generation, the AI ​​agent categorizes and prioritizes various issues, clearly identifies risk points and sources of abnormal data, and precisely locates the relevant document positions, specific fields, and values. The report can also include tables, charts, or highlighted original images to present issues intuitively, allowing users to quickly understand the anomalies and their context. In the report summary, the AI ​​agent generates targeted modification suggestions, proposing improvement plans for calculation errors, data inconsistencies, or standardization deficiencies to help relevant personnel efficiently correct problems. Furthermore, the report supports multi-level output formats, including detailed technical reports and management summary reports, to meet the needs of different use cases.

[0051] In summary, after the above-mentioned multi-agent automated analysis and review process, the final review report can be clearly summarized according to the review rules, forming an accurate, complete, and traceable engineering cost review report.

[0052] This disclosure introduces a multi-agent collaborative system for the first time in the cost review of power grid projects. Through multi-source file reading, automatic task decomposition and dynamic division of labor, it enables the parallel and efficient execution of multiple sub-tasks such as numerical calculation, data consistency comparison and standardization review, which significantly improves the automation level and overall efficiency of the review.

[0053] This disclosure utilizes multimodal large-scale power grid engineering cost documents (including scanned PDFs, multi-page tables, complex reports, and mixed text and image data) for joint text and image modeling, which can overcome the limitations of traditional single OCR parsing engines in processing complex documents, such as misidentification and missed detection.

[0054] In response to the characteristics of professional documents and drawings in power grid engineering scenarios (complex tables, technical terms, graphic symbols, etc.), this disclosure constructs a special training set for the power grid field and performs domain-specific fine-tuning on the OCR model, which effectively improves the OCR model's ability to identify and detect false positives and false negatives during the recognition process, and ensures the integrity and accuracy of key data.

[0055] This disclosure constructs a self-reflection and backtracking mechanism based on a result verification agent. After the professional agent completes the initial review, it performs a second verification of data integrity, calculation correctness, and logical consistency. When an anomaly is detected, it automatically locates the source and triggers a rerun or generates a manual review list, thereby ensuring the consistency and traceability of the conclusions.

[0056] like Figure 9 As shown, embodiments of this disclosure provide an automated cost review system for power grid engineering projects, including: The agent building module is used to build document parsing agents, task allocation agents, data analysis agents, result verification agents, and report summarization agents; The document parsing module is used to uniformly parse multi-source engineering documents through the document parsing intelligent agent to generate structured data; The task allocation module is used to automatically break down the review rules through the task allocation agent and assign different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution. The data analysis module performs multi-dimensional collaborative analysis of the structured data through multiple specialized intelligent agents within the data analysis intelligent agent. The result verification module is used to perform a secondary review of the output results of all the professional agents through the result verification agent. The report summary module is used to integrate the analysis and review results through the report summary agent to generate the final engineering review report.

[0057] See Figure 10The electronic device provided in the embodiments of this disclosure includes a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120 and the memory 1130 communicate with each other through the communication bus 1140. Memory 1130 is used to store computer programs; The processor 1110, when executing the program stored in the memory 1130, implements the above-described automated review method for power grid engineering costs. The aforementioned communication bus 1140 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0058] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0059] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1130 may also be at least one storage device located remotely from the aforementioned processor 1110.

[0060] Embodiments of this disclosure also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the automated cost review method for power grid engineering projects as described above.

[0061] The embodiments described above are merely examples of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these modifications and improvements all fall within the protection scope of this disclosure.

Claims

1. An automated method for reviewing the cost of power grid projects, characterized in that, Includes the following steps: Construct document parsing agent, task allocation agent, data analysis agent, result verification agent, and report summary agent; The document parsing agent performs unified parsing of multi-source engineering documents to generate structured data. The task allocation agent automatically breaks down the review rules and assigns different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution. The structured data is analyzed collaboratively from multiple specialized intelligent agents within the data analysis intelligent agent in a multi-dimensional manner. The result verification agent performs a secondary review of the output results of all the professional agents. The report summarizes how the intelligent agent integrates the analysis and review results to generate the final engineering review report.

2. The method for automated review of power grid engineering costs according to claim 1, characterized in that, A document parsing agent is used to uniformly parse multi-source engineering documents, generating structured data, including: The document parsing agent receives the review objectives and rule descriptions input by the user, identifies the required file name or path according to the rules, and then reads the corresponding engineering cost document from the specified file server or archive. For files of different formats, the document parsing agent adopts a customized parsing process to transform the original file content into a unified structured representation while preserving document hierarchy and location information; For complex tables and scanned PDF drawings, a training set of OCR models specifically for power grid engineering is constructed, and the OCR models are optimized and adjusted.

3. The automated cost review method for power grid projects according to claim 2, characterized in that, Construct a training set for a dedicated OCR model for power grid engineering, including: By collecting construction drawings and material lists from actual engineering projects, the original data source is formed; Combine manual annotation with annotations to mark the table structure, text content, and formulas; By introducing review experts to assist in quality verification, the completeness, semantic consistency, and format specifications of the labeled data are automatically checked and optimized.

4. The automated cost review method for power grid projects according to claim 1, characterized in that, The review rules are automatically broken down by a task allocation agent, and different types of sub-tasks are assigned to the corresponding specialized agents in the data analysis agent for execution, including: The task allocation agent receives the review rules, transforms the review rules into a structured task description, and obtains a set of subtasks. Based on the set of subtasks, the task allocation planning agent assigns the subtasks to corresponding specialized agents for execution according to their attributes and data requirements; when there are dependencies between subtasks, the task allocation agent plans and determines the execution order of the subtasks.

5. The automated cost review method for power grid projects according to claim 1, characterized in that, The structured data is analyzed collaboratively from multiple specialized agents within the data analysis agent, including: The data analysis agent includes a numerical calculation agent, a data review agent, and a standardization review agent; The numerical computation agent parses structured tabular data, identifies the calculation data and corresponding calculation formulas in the bill of quantities, and executes the calculation tasks. The data review agent performs consistency and rationality analysis on various types of structured data in engineering projects; The standardized review agent performs automated and standardized checks on engineering project documents based on industry standards, national regulations, and internal company procedures.

6. The method for automated review of power grid engineering costs according to claim 1, characterized in that, The result verification agent performs a secondary review of the output results of all the aforementioned specialized agents, including: After the professional intelligent agent completes its execution, the result verification intelligent agent receives the review results, original engineering documents, and supporting materials from the professional intelligent agent, and performs a second verification on the review results of the professional intelligent agent, including data integrity, calculation correctness, and logical consistency. When abnormal errors are detected in the review results, the cause of the error is analyzed, the analysis strategy is adjusted, and the task allocation intelligent agent and related professional intelligent agents are automatically triggered to re-analyze, or a list of issues requiring manual review is generated.

7. The automated cost review method for power grid projects according to claim 1, characterized in that, The report summarizes the intelligent agent's integration of analysis and review results to generate the final engineering review report, including: After completing the professional intelligent agent analysis and secondary review, the report summarizes and organizes all review results to form the final engineering review report. The final engineering review report categorizes and prioritizes various issues, marks risk points and sources of abnormal data, and locates the relevant document locations, specific fields, and values. The final engineering review report may include tables, charts, or highlighted original images to present issues intuitively. The final engineering review report generates targeted modification suggestions, proposing improvement plans for calculation errors, data inconsistencies, or standardization deficiencies. The final engineering review report supports multi-level output formats, including detailed technical reports and management summary reports.

8. An automated cost review system for power grid projects, characterized in that, include: The agent building module is used to build document parsing agents, task allocation agents, data analysis agents, result verification agents, and report summarization agents; The document parsing module is used to uniformly parse multi-source engineering documents through the document parsing intelligent agent to generate structured data; The task allocation module is used to automatically break down the review rules through the task allocation agent and assign different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution. The data analysis module performs multi-dimensional collaborative analysis of the structured data through multiple specialized intelligent agents within the data analysis intelligent agent. The result verification module is used to perform a secondary review of the output results of all the professional agents through the result verification agent. The report summary module is used to integrate the analysis and review results through the report summary agent to generate the final engineering review report.

9. The automated cost review system for power grid engineering projects according to claim 8, characterized in that, The document parsing module uses the document parsing intelligent agent to uniformly parse multi-source engineering documents, generating structured data, including: The document parsing agent receives the review objectives and rule descriptions input by the user, identifies the required file name or path according to the rules, and then reads the corresponding engineering cost document from the specified file server or archive. For files of different formats, the document parsing agent adopts a customized parsing process to transform the original file content into a unified structured representation while preserving document hierarchy and location information; For complex tables and scanned PDF drawings, a training set of OCR models specifically for power grid engineering is constructed, and the OCR models are optimized and adjusted.

10. The automated cost review system for power grid engineering projects according to claim 9, characterized in that, Construct a training set for a dedicated OCR model for power grid engineering, including: By collecting construction drawings and material lists from actual engineering projects, the original data source is formed; Combine manual annotation with annotations to mark the table structure, text content, and formulas; By introducing review experts to assist in quality verification, the completeness, semantic consistency, and format specifications of the labeled data are automatically checked and optimized.

11. The automated cost review system for power grid engineering according to claim 8, characterized in that, The task allocation module automatically breaks down the review rules through the task allocation agent, and assigns different types of sub-tasks to the corresponding professional agents in the data analysis agent for execution, including: The task allocation agent receives the review rules, transforms the review rules into a structured task description, and obtains a set of subtasks. Based on the set of subtasks, the task allocation planning agent assigns the subtasks to corresponding specialized agents for execution according to their attributes and data requirements; when there are dependencies between subtasks, the task allocation agent plans and determines the execution order of the subtasks.

12. The automated cost review system for power grid engineering according to claim 8, characterized in that, The data analysis module performs multi-dimensional collaborative analysis of the structured data through multiple specialized agents within the data analysis intelligent agent, including: The data analysis agent includes a numerical calculation agent, a data review agent, and a standardization review agent; The numerical computation agent parses structured tabular data, identifies the calculation data and corresponding calculation formulas in the bill of quantities, and executes the calculation tasks. The data review agent performs consistency and rationality analysis on various types of structured data in engineering projects; The standardized review agent performs automated and standardized checks on engineering project documents based on industry standards, national regulations, and internal company procedures.

13. The automated cost review system for power grid engineering according to claim 8, characterized in that, The result verification module performs a secondary review of the output results of all the professional agents through the result verification agent, including: After the professional intelligent agent completes its execution, the result verification intelligent agent receives the review results, original engineering documents, and supporting materials from the professional intelligent agent, and performs a second verification on the review results of the professional intelligent agent, including data integrity, calculation correctness, and logical consistency. When abnormal errors are detected in the review results, the cause of the error is analyzed, the analysis strategy is adjusted, and the task allocation intelligent agent and related professional intelligent agents are automatically triggered to re-analyze, or a list of issues requiring manual review is generated.

14. The automated cost review system for power grid engineering projects according to claim 8, characterized in that, The report summary module integrates the analysis and review results through the report summary agent to generate the final engineering review report, including: After completing the professional intelligent agent analysis and secondary review, the report summarizes and organizes all review results to form the final engineering review report. The final engineering review report categorizes and prioritizes various issues, marks risk points and sources of abnormal data, and locates the relevant document locations, specific fields, and values. The final engineering review report may include tables, charts, or highlighted original images to present issues intuitively. The final engineering review report generates targeted modification suggestions, proposing improvement plans for calculation errors, data inconsistencies, or standardization deficiencies. The final engineering review report supports multi-level output formats, including detailed technical reports and management summary reports.

15. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor is used to execute a program stored in a memory to implement the power grid engineering cost automation review method according to any one of claims 1-7.

16. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the automated review method for power grid engineering cost as described in any one of claims 1-7.

Images