A power constraint perception multi-agent collaboration method and system for power feasibility study report automatic generation

Abstract

The application relates to a power constraint perception multi-agent collaborative method and system for power feasibility report automatic generation, which comprises the following steps: acquiring a power feasibility report, constructing a task graph based on the power feasibility report, checking chapter nodes of the task graph, and obtaining a checking state; forming a candidate task set according to the checking state, screening an agent for each candidate task in the candidate task set, calculating a comprehensive dispatch score, and obtaining an optimal agent; calculating unified engineering index values of each optimal agent, determining an engineering conflict source, generating a supplementary certificate task, returning a task graph, and obtaining a final power feasibility report; performing quality scoring on the final power feasibility report, judging the quality scoring result through a hierarchical termination criterion, and outputting a current optimal version report, a low confidence content list and a to-be-supplemented list. The application organizes the power feasibility report into a task graph which is computable, traceable and closed-loop iterative.

Description

Technical Field

[0001] This invention relates to the fields of digital design of power engineering, intelligent document generation, knowledge enhancement generation and multi-agent collaborative decision-making technology, and in particular to a multi-agent collaborative method and system for power constraint perception in the automatic generation of power feasibility study reports. Background Technology

[0002] Public searches were conducted using keywords such as automatic generation of power feasibility study reports, automatic generation of power reports, and multi-agent report generation. Public clues indicate that related technologies can be broadly categorized into three types: The first type is automatic generation technology for power or distribution network feasibility study reports, such as automatic generation of distribution network feasibility study reports based on LDA probabilistic models, feasibility study report generation systems, and automatic generation methods for power reports; the second type is automatic generation technology for general project feasibility study reports, such as automatic generation of feasibility study reports based on large models or network systems; and the third type is general multi-agent report generation technology, such as multi-agent collaborative generation in areas such as property operation reports, budget reports, and investment research reports.

[0003] While the aforementioned technologies can improve report generation efficiency to some extent, they still have significant shortcomings: First, power feasibility study reports are not ordinary narrative documents; their chapters have strong engineering dependencies. For example, load forecasting directly affects access voltage levels, main equipment capacity, line cross-sections, engineering quantities, and investment estimates, and existing text generation methods cannot guarantee such cross-chapter linkage. Second, general multi-agent report generation typically divides tasks according to common roles such as data retrieval, writing, and review, without binding agent capabilities to the professional procedures of power feasibility studies. This can easily be perceived as simply applying existing multi-agent technologies to report writing. Third, existing methods lack explicit modeling of electrical constraints and regulatory clauses, which can easily lead to inconsistencies in line length, capacity, investment, access points, and conclusions / recommendations. Fourth, the generated report results lack a chain of evidence, making it difficult for reviewers to trace which data, regulation, or calculation a particular value, scheme, or conclusion originated from.

[0004] Therefore, it is imperative to focus the distinction on the power feasibility study task itself; among them, the multi-agent is not a generalized writing role, but a professional intelligent agent for power feasibility study; collaboration is not simply parallel generation, but driven by a coupled task graph of chapters, evidence, and constraints; and the review is not ordinary language proofreading, but cross-chapter electrical consistency arbitration and standard compliance verification. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide a multi-agent collaborative method and system for power constraint perception in the automatic generation of power feasibility study reports. By constructing a coupled task graph of power feasibility study chapters, evidence, and constraints, a constraint gap-driven assignment algorithm, a cross-chapter electrical consistency arbitration algorithm, and a quality closed-loop scoring algorithm, the generation of power feasibility study reports is transformed from general text generation to engineering document generation driven by professional evidence and electrical constraints.

[0006] To achieve the above objectives, the present invention provides the following solution: A multi-agent cooperative method for power constraint perception in the automatic generation of power feasibility study reports includes: Obtain the power feasibility study report, construct a task diagram based on the power feasibility study report, and verify the chapter nodes of the task diagram to obtain the verification status; Based on the verification status, a candidate task set is formed, an agent is selected for each candidate task in the candidate task set, and a comprehensive assignment score is calculated to obtain the optimal agent; Calculate the unified engineering index values ​​for each optimal agent, identify the sources of engineering conflicts, generate supplementary verification tasks and return the task graph to obtain the final power feasibility study report; The final power feasibility study report is scored for quality. The quality score results are judged by a hierarchical termination criterion, and the current best version report, a list of low-confidence content, and a list of supplementary certifications are output.

[0007] Optionally, the task graph includes: Extract the original evidence nodes from the power feasibility study report, and parse and normalize them to form power entity nodes and indicator nodes; The original evidence node, the power entity node, and the indicator node are introduced into the node set, and the chapter nodes, constraint nodes, agent capability nodes, and output nodes of the power feasibility study report are also included in the node set. The task graph is obtained by determining the edge relationship type and credibility of each node in the node set; the edge relationship type includes reference, dependency, generation, verification, backfilling, conflict and replacement.

[0008] Optionally, obtaining the verification status includes: The chapter type, dependency edges, and generation conditions of each chapter node in the task graph are obtained to verify whether the current chapter node has the necessary evidence nodes and constraint nodes. If the necessary evidence nodes and constraint nodes are present, the corresponding chapter of the current chapter node is split into data extraction, engineering calculation, text generation, and result verification subtasks, and added to the pending execution queue according to the chapter processing priority. The verification status is generated and written back to the corresponding chapter node and output node. If the necessary evidence nodes and constraint nodes are not present, a missing evidence task is generated, and the generation of the corresponding chapter text of the current chapter node is temporarily suspended.

[0009] Optionally, the processing priority of the chapters includes: ; in, Represents the k-th chapter node Processing priority; Chapter The proportion of missing necessary evidence; Chapter The intensity of existing constraint conflicts; Chapter Risk weights in the review rules; Chapter The degree to which it is relied upon by other chapters; This represents the weight coefficient of each evaluation factor.

[0010] Optionally, forming the candidate task set includes: Traverse the chapter nodes and output nodes in the task graph that are in a pending or reprocessing state, and read the missing original evidence, constraint violation, cross-chapter conflict indicators and downstream dependencies of the incoming edges; Based on the missing original evidence, a supplementary evidence task is generated; based on the violation of constraints, a recalculation or scheme adjustment task is generated; based on the cross-chapter conflict index, an index arbitration and chapter backfilling task is generated; based on the output that fails the verification, a review or rewriting task is generated; and for chapters that have not yet been generated but whose preconditions have been met, a chapter generation task is generated. The supplementary certification task, the recalculation or scheme adjustment task, the indicator arbitration and chapter backfilling task, the review or rewriting task, and the chapter generation task are coded according to task type, associated chapter, required capabilities, and priority, and then deduplicated to form the candidate task set.

[0011] Optionally, calculating the comprehensive dispatch score to obtain the optimal agent includes: ; ; in, Represents intelligent agents Execute the task The overall work assignment score; This indicates the degree of matching between the agent's capabilities and the task requirements; This indicates the expected reduction in the intensity of constraint conflicts; Indicates the expected amount of supplementary evidence; This indicates the agent's historical success rate or credibility in similar tasks; This refers to the cost of invoking, the time cost, or the cost of manual review; Indicates the weight of the capability matching item; Indicates the weight of the conflict reduction item; Indicates the weight of the evidence completion item; Indicates the weight of the historical credibility item; Indicates the weight of the cost penalty item; Indicates that the task has been selected to be executed. The optimal intelligent agent; Indicates the ability to handle tasks The set of candidate intelligent agents; This represents the m-th candidate agent; This represents the selection operation that maximizes the objective function.

[0012] Optionally, the method further includes: The optimal agent is established with the task, and the task status is set to dispatched. At the same time, the optimal agent reads the original evidence nodes, entity nodes, indicator nodes and constraint nodes associated with the task graph and performs supplementary evidence, data extraction, engineering calculation, chapter generation or result verification according to the task type. The execution results are written to the output node, and the satisfaction status, evidence completeness and credibility of the corresponding constraints are updated. If the verification fails, a recalculation, supplementary evidence or backfill sub-task is generated along the conflict edge to update the credibility. The candidate agent screening and dispatch are re-executed according to the updated credibility. During the assignment of candidate agents, if the conflict indicators can determine the workload and quota basis in the evidence node, a workload verification task is first established. If the verified workload and cost quota can calculate a new investment value, an investment estimation recalculation task is established. If the recalculated value passes the investment boundary constraint, a backfilling task for the general overview, investment estimation table, and economic evaluation is established. If the evidence is still insufficient or the recalculated value violates the investment boundary, a supplementary evidence task is established and the completion of relevant chapters is suspended. The task graph is updated after each subtask is completed. If the output of the current optimal agent leads to an increase in constraint conflicts or fails to improve the quality score for multiple consecutive rounds, the credibility of the current optimal agent on the corresponding task is reduced, and the assignment probability of other candidate agents is increased. At the same time, the credibility update result is written back to the corresponding agent capability node.

[0013] Optionally, calculating the uniform engineering performance index values ​​for each optimal agent includes: ; in, This represents the unified value of the p-th engineering indicator after arbitration; Indicates the candidate values; This represents the value of the p-th engineering metric output by the i-th chapter or the i-th agent; This indicates the number of outputs related to the indicator p; This represents the confidence weight of the i-th output; This represents the deviation loss between the candidate value and the original output; Indicates the power constraint penalty weight; This represents the penalty function for candidate values ​​violating power industry constraints. This penalty function is obtained by performing constraint penalty calculations after dimensional normalization of the voltage deviation constraint violation function, line current carrying capacity constraint violation function, and power supply radius constraint violation function.

[0014] Optionally, the output of the current best version report, the list of low-confidence content, and the list of content requiring supplementary verification includes: The hierarchical termination criterion is set to distinguish power constraints into hard constraints and soft indices. The remaining number of violations of the hard constraints must be zero to be considered as meeting the standard. At the same time, the soft indicators are used to score the quality of the final power feasibility study report, and the target threshold is used to determine whether to continue the next round of supplementary certification, recalculation, verification or rewriting tasks. After each round, the iteration is terminated based on a target termination determination; the target termination determination includes: Eligibility Exit: After each round of chapter assembly, engineering indicator arbitration, and full verification, if... If Q(k)≥τ, then this round of reports will be solidified into the final power feasibility study report and terminated; Resource Exit: If If Q(k) < τ, but Kmax has been reached, then stop adding new automatic tasks, retain the historical highest score snapshot, unmet constraints and their responsibility evidence / sections, and output a list of manual review; Convergence Exit: If If Q(k) < τ and ΔQ(k) < ε for consecutive preset rounds, it means that supplementary proof, recalculation and rewriting can no longer produce effective quality improvement. Roll back or keep the historical highest score snapshot and output the manual review list. Blocked Exit: If the remaining missing evidence satisfies If the credibility of all candidate agents for a critical pending task falls below the preset lower limit, then mark the cause of the blockage, related evidence, and affected chapters, terminate automatic task assignment, and transfer the task to manual verification or review. in, This indicates that all hard constraints have been met; Q(k) represents the quality score of the k-th round report, which is a comprehensive score obtained by weighting the power feasibility study report assembled in the k-th round according to the completeness of chapters, consistency of cross-chapter indicators, degree of satisfaction of key electrical constraints, degree of matching of normative clauses, degree of rationality of investment economy and traceability of evidence chain; τ represents the report quality threshold. This indicates that all hard constraints are not met; ΔQ(k) = Q(k) - Qbest(k-1) represents the increment of the current round score relative to the historical best snapshot score; Qbest(k-1) represents the historical highest score at the end of round (k-1); ε represents the minimum effective improvement threshold, calibrated by the upper bound of the inter-round noise of the score repeatedly generated with the same input; Kmax represents the maximum number of iteration rounds, calibrated by the number of rounds required for the vast majority of reports in historical projects to reach the quality threshold; M(k) represents the set of missing evidence in round k; Mext represents the set of evidence that needs to be supplemented externally and cannot be automatically obtained by the system. This indicates that the remaining missing evidence belongs to external evidence that the system cannot automatically obtain; Represents the empty set; ∧ represents a set inclusion relationship; ∧ represents a logical AND operation; ∨ represents a logical OR operation.

[0015] To achieve the above objectives, the present invention also provides a power constraint perception multi-agent cooperative system for automatic generation of power feasibility study reports, comprising: The task diagram construction module is used to obtain the power feasibility study report, construct a task diagram based on the power feasibility study report, and verify the chapter nodes of the task diagram to obtain the verification status. The multi-agent collaborative execution module is used to form a candidate task set according to the verification status, select agents for each candidate task in the candidate task set, calculate the comprehensive assignment score, and obtain the optimal agent. The verification and arbitration module is used to calculate the unified engineering index values ​​of each optimal agent, determine the source of engineering conflicts, generate supplementary verification tasks and return the task graph to obtain the final power feasibility study report. The report and audit package output module is used to score the quality of the final power feasibility study report, judge the quality score results through hierarchical termination criteria, and output the current best version report, the list of low confidence content, and the list of supplementary certifications.

[0016] The beneficial effects of this invention are as follows: This invention addresses the technical problems in the automatic generation of power feasibility study reports, such as the difficulty in matching input data with engineering objects, the easy conflict of engineering indicators across chapters, the lack of professional constraints on general intelligent agent assignment, and the difficulty in verifying the generated results. It employs four technical means: a unified task graph of "original evidence—power entity / indicator—chapter—constraint—intelligent agent—output," constraint gap-driven assignment, cross-chapter electrical consistency arbitration, and quality closed-loop. This ensures that each intelligent agent can only generate or backfill chapter content based on traceable evidence and engineering indicators that meet power constraints. It achieves consistency for the same engineering entity and the same indicator across different chapters, appendices, and conclusions, and outputs low-confidence content and a list of materials requiring supplementary verification, thereby improving the engineering completeness, verifiability, traceability, and engineering credibility of the report. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-agent collaborative method for power constraint perception in the automatic generation of power feasibility study reports, according to an embodiment of the present invention. Figure 2 This is a diagram illustrating the composition of a multi-agent collaborative system for power constraint perception, oriented towards automatic generation of power feasibility study reports, according to an embodiment of the present invention. Detailed Implementation

[0019] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1As shown in the figure, this embodiment discloses a multi-agent collaborative method for power constraint perception in the automatic generation of power feasibility study reports. The method includes: accessing original project data to form original evidence nodes; parsing and normalizing these to obtain power entity nodes and indicator nodes; constructing a task graph with chapter nodes, constraint nodes, agent capability nodes, and output nodes; forming a candidate task set based on the task graph state; selecting and dispatching the optimal agent for each task; the optimal agent performing supplementary verification, calculation, chapter generation, or verification on related entities and indicators, outputting chapter fragments, calculation results, and verification results; performing consistency arbitration, backfilling, and re-verification on cross-chapter engineering indicators to form the final power feasibility study report; performing a quality score on the final power feasibility study report, and outputting the current optimal version report, a list of low-confidence content, and a list of documents requiring supplementary verification through a hierarchical termination criterion.

[0022] Specifically, this embodiment discloses a multi-agent cooperative method for power constraint perception in the automatic generation of power feasibility study reports, including: S1 Project data access; S2 Power entity and indicator analysis; S3 Chapter-evidence-constraint coupling task graph construction; S4 Parallel generation of special intelligent agents and standard evidence intelligent agents; S5 Electrical constraint verification; S6 Cross-chapter consistency arbitration; S7 Quality closed-loop judgment; S8 Report and audit package output.

[0023] (I) Power Feasibility Study Chapter—Evidence—Constraint Coupling Task Graph Construction Algorithm: This algorithm transforms power feasibility study reports from a simple text task into an engineering constraint task. The system establishes chapter nodes, including general overview, project necessity, current power grid analysis, load forecasting, grid connection schemes, major engineering quantities, investment estimates, economic evaluation, environmental and safety considerations, and conclusions / recommendations. It also establishes original evidence nodes, including load curves, station topology, equipment ledgers, geographical routes, cost quotas, standard clauses, historical feasibility studies, and review opinions. Power entity nodes and indicator nodes are parsed from the original evidence nodes. Simultaneously, constraint nodes are established, including capacity adaptation, power supply radius, voltage deviation, line current carrying capacity, N-1 reliability, short-circuit current level, investment boundaries, and standard applicability. Agent capability nodes express the capability boundaries of different agents, and output nodes carry the generated chapter fragments, calculation results, verification results, and review opinions for each round.

[0024] S1 inputs the original data and forms the original evidence nodes. S2 parses and normalizes the original evidence nodes to form power entity nodes and indicator nodes. When constructing the task graph, S3 includes chapter nodes, original evidence nodes, power entity nodes, indicator nodes, constraint nodes, agent capability nodes, and output nodes into the node set. Among them, power entity nodes and indicator nodes are associated with the original evidence nodes through evidence pointer edges. The same input data is not repeatedly defined as two types of nodes.

[0025] ; Where G represents the coupled task graph of power feasibility study chapters, evidence, and constraints; V represents the set of all nodes; E represents the set of edges; Vc represents the set of nodes in the report chapter; Ve represents the set of original evidence nodes; Vn represents the set of power entity nodes obtained by parsing and normalizing the original evidence nodes; Vi represents the set of indicator nodes obtained by parsing and normalizing the original evidence nodes; Vk represents the set of power engineering constraint nodes; Va represents the set of agent capability nodes; Vo represents the set of agent output nodes; and ∪ represents the set union operation.

[0026] ; in, Indicates from node Pointing to node The directed edges; This represents the i-th graph node; This represents the j-th graph node; Represents the type of edge relationship, including reference, dependency, generation, verification, backfill, conflict, and replacement; Indicates relation weight or credibility.

[0027] To ensure the task graph guides actual generation, this invention further calculates chapter processing priorities. If a chapter lacks key evidence, has numerous conflicting constraints on related indicators, or is considered a high-risk chapter for review, it should be prioritized. For chapter nodes in the task graph with a status of "Pending Processing" or "Requires Reprocessing," the chapter type, dependent edges, and generation conditions of the node are read, and the associated original evidence nodes, entity nodes, indicator nodes, and constraint nodes are checked for completeness. If the conditions are met, the nodes are split into subtasks according to the dependency order of data extraction, engineering calculation, text generation, and result verification, and added to the execution queue. The output of the previous subtask serves as the input for the next subtask. The data extraction subtask supplements or updates entity / indicator nodes, the engineering calculation subtask generates constrained indicator values, the text generation subtask generates chapter fragments accordingly, and the result verification subtask forms the verified node status based on completeness, constraint satisfaction, and output credibility. The node status includes at least "Pending Processing," "Processing," "Completed," or "Requires Reprocessing." After each subtask is completed, the entity / indicator update, chapter fragment, and calculation result are written to the corresponding node or output node, and the verified node status is written back to the corresponding chapter node, indicator node, and output node for reuse in subsequent task generation, indicator arbitration, chapter backfilling, and quality scoring. If the conditions are not met, a missing certificate task is generated and the text generation of that chapter is temporarily suspended.

[0028] ; in, Represents the k-th chapter node Processing priority; Chapter The proportion of missing necessary evidence; Chapter The intensity of existing constraint conflicts; Chapter Risk weights in the review rules; Chapter The degree to which it is relied upon by other chapters; This represents the weight coefficient of each evaluation factor.

[0029] The specific function of this task graph construction algorithm is as follows: when the system generates the "Access System Scheme" section, it not only knows that this section requires textual descriptions, but also that it must reference the existing network structure, predicted load, available bays, path lengths, and standard clauses; when the system generates the "Investment Estimation" section, it must extract data from the engineering quantity nodes and cost quota nodes, and ensure consistency with the line lengths and equipment models in the scheme section. Thus, multi-agent collaboration is forcibly bound to the power feasibility study's professional objects and constraints.

[0030] (II) Constraint Gap-Driven Multi-Agent Dispatch Algorithm: This algorithm determines which agent handles which task in the current round. Unlike the general multi-agent approach of dividing tasks equally by chapter, this invention uses evidence gaps, constraint conflicts, capability matching, historical success rate, and invocation cost in the task graph as the basis for assignment. The system first generates a set of candidate tasks based on the task graph, then selects a set of agents with corresponding professional capabilities for each candidate task, and finally calculates the assignment reward.

[0031] The screening process for candidate agents is as follows: read the required capability tags, input evidence types, output formats, and constraint verification requirements of the candidate task, and match them with the professional fields, callable tools, processable evidence types, and output interfaces registered in the agent's capability nodes; retain only agents whose capability tags cover the task requirements, can access the evidence required by the task, and whose output interfaces meet the requirements of downstream nodes, and remove agents that are in a cooling-off period, have insufficient permissions, or whose credibility is below the preset lower limit.

[0032] The process of generating the candidate task set is as follows: traverse the chapter nodes, indicator nodes, and output nodes in the task graph that are in the state of "pending processing" or "re-processing required," and read the missing original evidence, entities / indicators to be updated, constraint violations, conflicting indicators, and downstream dependencies associated with their incoming edges; for missing evidence, generate supplementary evidence or data extraction tasks; for entities / indicators to be updated, generate entity normalization, indicator re-extraction, unit conversion, caliber correction, or indicator recalculation tasks based on their update reasons, and use the current value, source evidence, update reason, and target state of the entity / indicator to be updated as task inputs. After the task is completed, update the corresponding entity node or indicator node with the new value and evidence pointer for subsequent constraint verification, indicator arbitration, and chapter backfilling; for constraint violations, generate engineering recalculation or scheme adjustment tasks; for cross-chapter conflicting indicators, generate indicator arbitration and chapter backfilling tasks; for outputs that fail verification, generate review or rewriting tasks; for chapters that have not yet been generated but whose preconditions have been met, generate chapter generation tasks. The above tasks are coded according to task type, associated object, associated chapter, required capabilities, input nodes, expected output and priority, and then deduplicated to form a candidate task set.

[0033] ; in, Represents intelligent agents Execute the task The overall work assignment score; This indicates the degree of matching between the agent's capabilities and the task requirements; This indicates the expected reduction in the intensity of constraint conflicts; Indicates the expected amount of supplementary evidence; This indicates the agent's historical success rate or credibility in similar tasks; This refers to the cost of invoking, the time cost, or the cost of manual review; Indicates the weight of the capability matching item; Indicates the weight of the conflict reduction item; Indicates the weight of the evidence completion item; Indicates the weight of the historical credibility item; This indicates the weight of the cost penalty item.

[0034] ; in, Indicates that the task has been selected to be executed. The optimal intelligent agent; Indicates the ability to handle tasks The set of candidate intelligent agents; This represents the m-th candidate agent; This represents the selection operation that maximizes the objective function; This indicates the overall work assignment score.

[0035] After obtaining the optimal agent, the system establishes an "execution" edge between the agent and the task, and sets the task status to "assigned." The optimal agent reads the original evidence nodes, entity nodes, indicator nodes, and constraint nodes associated with the task graph, and performs supplementary evidence, data extraction, engineering calculations, chapter generation, or result verification according to the task type. Its execution results include newly added or updated evidence, entity / indicator values, constrained engineering calculation results, chapter fragments, and verification results, which are written to the corresponding nodes or output nodes, while updating the constraint satisfaction status, evidence completeness, and task credibility. If the output fails verification, a recalculation, supplementary evidence, backfilling, or rewriting subtask is generated along the "conflict" edge, and candidate agent screening and assignment are re-executed based on the updated credibility. Therefore, the assignment result of the optimal agent determines the executor of the current task, and its execution result is used to drive subsequent indicator arbitration, chapter backfilling, and final report formation, not just for recording or display.

[0036] For example, when the task diagram finds an inconsistency between the "total investment in the general overview" and the "summary investment in the investment estimate table," the system will not simply assign the task to the text editing agent. Instead, it will break the task down into three sub-tasks: quantity verification, cost quota recalculation, and chapter backfilling, which will be assigned to the quantity verification agent, the investment estimate agent, and the consistency arbitration agent, respectively. If insufficient evidence is found regarding the "maximum load in the long term," the system will prioritize calling the load forecasting agent to supplement the load forecasting basis, rather than directly requiring the text drafting agent to rewrite paragraphs.

[0037] The above-mentioned work assignments are executed according to the following judgment process: If the conflict indicators can be found in the evidence nodes with the basis for the quantity and quota, then a quantity verification task is first established; if the verified quantity and cost quota can be used to calculate a new investment value, then an investment estimate recalculation task is established; if the recalculated value passes the investment boundary constraints, then a task to fill in the general overview, investment estimate table, and economic evaluation is established; if the evidence is still insufficient or the recalculated value violates the investment boundary, then a supplementary evidence task is established and the completion of the relevant chapters is suspended. The task diagram is updated after each subtask is completed, and the next subtask is only executed when the status of its preceding output node is "valid".

[0038] To prevent a single agent from repeatedly generating invalid content, this invention also includes a reassignment mechanism. When an agent's output leads to increased constraint conflicts or fails to improve its quality score for multiple consecutive rounds, its credibility in the corresponding task is reduced, and the assignment probability of other candidate agents is increased.

[0039] The credibility update result is written back to the corresponding agent capability node and participates in the next round of candidate agent selection and task assignment score calculation; through this write-back, the verification result output in the previous round can affect the execution of subsequent tasks.

[0040] ; in, This indicates the updated agent's task credibility; This indicates the credibility of the agent's task before the update; This represents the historical reliability retention coefficient, with a value ranging from 0 to 1; This indicates whether the output of this round passes the constraint check. The value ranges from 0 to 1, with 1 indicating a complete pass and 0 indicating a failure. This represents the m-th agent; This represents the k-th task.

[0041] (III) Cross-Chapter Electrical Consistency Arbitration Algorithm: The professionalism of power feasibility study reports is largely reflected in cross-chapter consistency. For example, the long-term load given in the load forecasting section should be consistent with the main equipment capacity, line cross-section, equipment list in the investment estimate, and recommended scheme in the conclusions and recommendations. This algorithm performs entity alignment, index merging, credibility weighting, and constraint penalties on the outputs of each agent to obtain unified engineering index values, which are then backfilled into all relevant chapters.

[0042] ; in, This represents the unified value of the p-th engineering indicator after arbitration; Indicates the candidate values; This represents the value of the p-th engineering metric output by the i-th chapter or the i-th agent; This indicates the number of outputs related to the indicator p; This represents the confidence weight of the i-th output; This represents the deviation loss between the candidate value and the original output; Indicates the power constraint penalty weight; This represents the penalty function for engineering index x violating electrical engineering constraints.

[0043] ; in, This represents the confidence weight of the i-th output; represents the overall credibility score of the i-th output; exp represents the natural exponential function, and the credibility weights are obtained after normalizing all outputs; The data source quality score can be represented by the weights of the corresponding evidence-related edges in the task graph. Aggregation yields; Indicates the degree to which the output passes electrical verification; Indicates the credibility of similarity to historical projects; This indicates output uncertainty; Indicates the weight of the data source quality item; Indicates the weight of electrical verification items; Indicates the weight of historically similar items; This represents the weight of the uncertainty penalty term.

[0044] ; in, This represents the penalty for violating the constraint of the p-th engineering indicator; This represents the set of power constraints associated with index p; Indicates a specific constraint; Let r represent the function that violates the r-th constraint; (x)≤0 indicates that the constraint is satisfied. (x) > 0 indicates a violation of the constraint; This indicates that penalties will only be imposed on the portion of the violation.

[0045] To make the constraint violate the function In the power feasibility study scenario, the constraints can be directly calculated. This invention further provides the specific mathematical forms of typical electrical constraints. Taking the medium-voltage distribution network access scenario as an example, the voltage deviation constraint, line current carrying capacity constraint, and power supply radius constraint are constructed according to Equations (10), (11), and (12), respectively. The remaining constraints can be extended in the same way.

[0046] ; in, ΔU(x) represents the voltage deviation constraint violation function; ΔU(x) represents the percentage of voltage loss at the end of the line corresponding to candidate scheme x; P represents the active power transmitted by the line, in kilowatts; Q represents the reactive power transmitted by the line, in kilovars; R represents the line resistance, in ohms; X represents the line reactance, in ohms. This indicates the rated voltage of the line, in kilovolts. This indicates the permissible voltage drop limit according to the specifications.

[0047] ; in, Indicates the function that violates the line current carrying capacity constraint; This represents the maximum operating current of the line under the maximum load in the long term for candidate scheme x, in amperes. The value indicates the allowable current carrying capacity of the selected conductor or cable cross-section after correction for laying method and ambient temperature, in amperes; cosφ represents the load power factor.

[0048] ; in, L(x) represents the power supply radius constraint violation function; L(x) represents the power supply path length of candidate scheme x, in kilometers. This indicates the allowable power supply radius for the corresponding voltage level as determined by the technical guidelines for distribution network planning and design, in kilometers.

[0049] The constraint violation function, after dimensional normalization, is substituted into equation (9) for constraint penalty calculation. For constraints such as N-1 reliability and short-circuit current level, which are difficult to express as a single algebraic inequality, the system encapsulates them into a dedicated verification function and outputs them in numerical form as the degree of violation. This will be incorporated into a unified penalty scheme.

[0050] This algorithm not only unifies numerical values ​​but also outputs the sources of conflict. For example, if the general overview section uses an earlier calculated comprehensive absorption rate of 74.7%, while the electricity absorption analysis section obtains 74.23% by fitting the monthly load curves with the photovoltaic output curves, the arbitration algorithm will compare the evidence sources, data scope, and calculation time points of the two values. If the load output fitting evidence is of the highest quality, then 74.23% will be uniformly adopted, and the economic evaluation will be recalculated. If both values ​​lack reliable evidence, a supplementary evidence task will be generated and a task graph will be returned.

[0051] (iv) Power constraint verification and quality closed-loop algorithm: After each round of generation, the system scores the overall quality of the report. The report is assembled from chapter segments that have passed verification in this round, unified engineering indicator values ​​after consistency arbitration, and corresponding evidence citations: After multi-agent assignment completes supplementary certification, engineering calculations, and chapter generation, the chapter segments are written to the output node; the verification and arbitration module handles cross-chapter indicator conflicts and backfills the unified values ​​to the relevant chapters; when there are no confirmed conflicts to be backfilled in this round, the chapter segments that have passed verification are assembled into the current round's power feasibility study report according to the report directory. If a supplementary certification task is generated and the task graph is returned, the report is reassembled after completing the supplementary certification, recalculation, backfilling, and re-verification; the current round's report that meets hard constraints and whose quality score reaches the threshold is the final power feasibility study report. This scoring differs from ordinary text quality scoring; its core aspects are engineering completeness, indicator consistency, electrical constraint satisfaction, regulatory clause matching, investment economic rationality, and evidence traceability.

[0052] ; in, This indicates the overall quality score of the report; Indicates the chapter completeness score; Indicates the consistency score across chapter indicators; Indicates the score for meeting key electrical constraints; Indicates a score for matching the normative clauses; This indicates a score representing the reasonableness of the investment-economic boundary. Indicates a traceable chain of evidence score; Indicates the weight of chapter completeness; Indicates the consistency weight of the indicators; Indicates the electrical constraint weight; Indicates the standard matching weight; Indicates the economic weight of investment; Indicates the weight of evidence traceability.

[0053] ; in, This indicates the preset report quality threshold; Indicates the number of remaining constraint conflicts; This indicates that a set of key evidence is missing; This represents the set of outputs with low confidence. This indicates the next round of supplementary certification, recalculation, verification, or rewriting tasks; This represents a function that generates the next round of tasks based on the task graph and the problem set. This represents the coupled task diagram of the power feasibility study chapters, evidence, and constraints.

[0054] Based on equation (14), this invention further sets a hierarchical termination criterion. The system distinguishes between hard constraints and soft indicators: hard constraints include electrical safety constraints such as voltage deviation, line current carrying capacity, short-circuit current level, N-1 reliability, and power supply radius. The remaining number of violations must be zero to be considered as meeting the standard, and they do not participate in the weighted average; soft indicators are weighted and scored in equation (13) and judged by a threshold τ. After the k-th iteration, whether to terminate is determined according to equation (15).

[0055] ; Where Stop(k) represents the termination decision after the k-th round, with a value of 1 indicating termination of the iteration; H(k) represents the set of remaining hard constraint violations after the k-th round. This indicates that all hard constraints have been met; Q(k) represents the report quality score in the k-th round; τ represents the report quality threshold; ΔQ(k) = Q(k) - Qbest(k-1) represents the increment of the current round score relative to the historical best snapshot score; Qbest(k-1) represents the historical highest score at the end of the (k-1)-th round; ε represents the minimum effective improvement threshold, calibrated by the upper bound of the inter-round noise of the score generated repeatedly with the same input, or by taking 1% to 2% of the full score; Kmax represents the maximum number of iteration rounds, calibrated by the number of rounds required for the vast majority of reports in historical projects to reach the quality threshold; M(k) represents the set of missing evidence in the k-th round; Mext represents the set of evidence that needs to be supplemented externally and cannot be automatically obtained by the system. This indicates that the remaining missing evidence belongs to external evidence that the system cannot automatically obtain; Represents the empty set; ∧ represents a set inclusion relationship; ∧ represents a logical AND operation; ∨ represents a logical OR operation.

[0056] The four conditions of equation (15) correspond to the following specific process: First, exiting upon meeting the standard: After completing the chapter assembly, engineering index arbitration and full verification in the kth round, if And if Q(k)≥τ, then this round of reports will be solidified into the final power feasibility study report and terminated; secondly, resource withdrawal: if If Q(k) < τ, but Kmax has been reached, then stop adding new automatic tasks, retain the historical highest score snapshot, unmet constraints and their responsible evidence / sections, and output a list for manual review; third, convergence and exit: if Q(k) < τ and consecutive preset rounds This indicates that supplementary evidence, recalculation, and rewriting can no longer produce effective quality improvement. The system will either roll back or maintain a snapshot of the highest historical score and output a list of manual reviews; fourth, block exit: if the remaining missing evidence meets the requirements... If the credibility of all candidate agents for a critical pending task falls below a preset lower limit, the cause of the blockage, related evidence, and affected sections are marked, automatic task assignment is terminated, and the task is transferred to manual verification or review. When the score in each round is lower than the historical highest score snapshot, the system first rolls back to the snapshot and then reassigns the task to avoid the consistency arbitration oscillating between candidate values, which would cause a decline in quality. When the process terminates due to resource constraints, convergence, or blockage, the system outputs a report of the current best version, a list of low-confidence content, and a list of tasks requiring verification, and concludes the process through human-machine collaborative handover, rather than simply terminating with an error report.

[0057] Through the aforementioned closed loop, the system can be upgraded from "generate once and output" to an engineering process of "generation—verification—arbitration—backfilling—regeneration". This mechanism is particularly suitable for power feasibility study review scenarios, because the review focuses not only on whether the text is fluent, but also on whether the load, scheme, investment, specifications and conclusions support each other.

[0058] like Figure 2 As shown in the figure, this embodiment also provides a multi-agent collaborative system for power constraint perception in the automatic generation of power feasibility study reports, including: a task graph construction module, used to access original project data to form original evidence nodes, parse to obtain power entity nodes and indicator nodes, and construct a unified task graph with chapter nodes, constraint nodes, agent capability nodes and output nodes; a multi-agent collaborative execution module, used to form a candidate task set based on the task graph state, and select and dispatch the optimal agent for each task; a verification and arbitration module, used for the optimal agent to perform supplementary verification, calculation, chapter generation or verification on related entities and indicators, and to perform consistency arbitration, backfilling and re-verification on cross-chapter engineering indicators to form the final power feasibility study report; and a report and audit package output module, used to score the quality of the final power feasibility study report, and output the current optimal version report, a list of low-confidence content and a list of supplementary verification through hierarchical termination criteria.

[0059] Specifically, this embodiment also provides a power constraint perception multi-agent collaborative system for automatic generation of power feasibility study reports, including: The data access and evidence management capabilities are used to receive basic project information, current power grid data, historical load data, planned load data, station and equipment ledgers, geographical route data, typical cost databases, standard clause databases, historical feasibility study reports, and historical review opinions. This part is not a simple file upload module, but rather converts raw data into evidence objects that can be referenced in subsequent task diagrams, and records the source, time, scope of application, credibility, and traceability identifier for each evidence object.

[0060] The power entity and indicator analysis capability is used to identify key objects such as load points, substations, feeders, transformer areas, line sections, access points, equipment models, engineering quantities, investment items, and standard clauses from the input data corresponding to the original evidence nodes, and to extract key values ​​such as maximum load, power supply radius, line length, voltage level, equipment capacity, investment amount, and economic indicators. The analysis results form power entity nodes and indicator nodes, and are linked to the original evidence nodes through evidence pointer edges, so that the subsequent task graph, generation, and verification processes can be carried out around the power engineering objects.

[0061] This analytical capability employs a hybrid extraction pipeline categorized by evidence source type. For structured data such as equipment ledgers, cost quota databases, and geographical route data, values ​​are directly extracted through field mapping. For semi-structured content such as tables in reports and estimates, layout analysis is performed to restore the table structure, and then the semantics of the table headers are mapped to indicator types. For unstructured text such as report body, review comments, and specification clauses, ontology-constrained extraction is used. This involves pre-defining power entity types, indicator types, their units of measurement, and legal value ranges to constitute the power ontology. The extraction model outputs according to the structured pattern defined by this ontology. Strong pattern fields with fixed word formation methods, such as voltage levels, equipment models, and specification numbers, are identified using regular rules and a professional dictionary. Before entering the task graph, the extraction results undergo three checks: unit normalization, value range verification, and back-read verification. Each extracted value carries an original text anchor pointing to the source file, page number, or table cell location; this anchor is the physical carrier of the traceable evidence chain scoring.

[0062] To unify node definitions within the methodology and system, this invention strictly distinguishes between: "original evidence nodes" formed from unextracted external data such as original documents, original tables, original measurement records, and specification clauses; and "power entity nodes" and "indicator nodes" formed from load points, substations, feeders, line segments, equipment, engineering quantities, and indicators extracted and normalized from the original evidence nodes. The task graph is composed of chapter nodes, original evidence nodes, power entity nodes, indicator nodes, constraint nodes, agent capability nodes, and output nodes. Each entity node or indicator node is associated with its source original evidence node through an evidence pointer edge, carrying page number, table cell, time caliber, unit, and confidence level attributes. Chapter nodes are associated with the required entity / indicator nodes through reference or dependency edges, and constraint nodes are associated with constrained entity / indicator nodes through verification edges. Therefore, input data is not reused as both types of nodes simultaneously, but first enters the system as original evidence nodes, then is parsed to form entity / indicator nodes. The parsing results are traceable through evidence pointers and directly participate in chapter generation, engineering calculations, and constraint verification.

[0063] For different ways of writing the same project object in different materials, the parsing ability is used to unify entities by jointly judging the alias dictionary, string similarity and voltage level, geographical location and other attribute constraints; the same project indicator is used as the merging key by a quadruple consisting of entity identifier, indicator type, time caliber and scheme number. Multiple values ​​with the same quadruple are entered into the consistency arbitration of equation (7), and different values ​​of the quadruple are regarded as normal differences between multiple schemes or multiple periods and do not constitute a conflict. Key indicators are extracted by both the rule channel and the model channel. The results of the two channels are consistent and marked as high confidence. The results of the two channels are inconsistent and marked as low confidence and the uncertainty term in equation (8) is increased. The task is then transferred to a supplementary verification task or a manual review queue, thus enabling an operable mechanism for generating confidence levels.

[0064] The system of this invention includes the capability to construct a chapter-evidence-constraint coupled task graph, which is used to uniformly organize report chapters, evidence materials, electrical constraints, specification clauses, agent capabilities, and generated outputs into a heterogeneous graph. This task graph is the core of system scheduling; it determines whether a chapter meets the generation conditions, which evidence is still missing, which indicators conflict, which specialized agents need to be invoked, and which chapters or appendices the generated results need to be backfilled into.

[0065] The system of this invention includes multi-agent collaborative execution capabilities. The load forecasting agent is responsible for generating near- and long-term load forecasts based on historical load, industry planning, and parameters of similar projects. In the case of renewable energy grid connection, it is also responsible for generating power generation forecasts and power consumption analysis based on irradiance data, system efficiency, and load output fitting. The grid structure scheme agent is responsible for proposing grid connection schemes based on the current grid, access point conditions, and reliability requirements. The engineering quantity agent is responsible for generating a bill of quantities based on line routes, equipment configurations, and typical designs. The investment estimation agent is responsible for generating investment estimates by calling cost quotas and the bill of quantities. The economic evaluation agent is responsible for generating investment benefits, power supply reliability, and social benefits analyses. The compliance agent is responsible for matching national, industry, or enterprise standard clauses. The evidence traceability agent is responsible for establishing source indexes for key conclusions in the report. The consistency arbitration agent is responsible for handling cross-chapter conflicts and unified backfilling.

[0066] This invention's system includes electrical constraint verification and consistency arbitration capabilities. It is used to verify the outputs of each intelligent agent against constraints such as capacity, voltage drop, line current carrying capacity, short-circuit current, N-1 reliability, power supply radius, access conditions, and investment boundaries. It also performs reliable weighting and constraint penalties on different values ​​of the same engineering indicator across multiple chapters. Through this capability, the system can avoid common automatic generation errors, such as inconsistencies in line length between the scheme chapter and the investment estimation chapter, mismatches between load forecasting and equipment capacity, inconsistencies between the general investment and the appendix summary, and incorrect application scenarios for specification clauses.

[0067] The system of this invention includes the capability to generate and review reports, and is used to generate the main body of a power feasibility study report, appendices, evidence index, discrepancy list, and review comments after the quality closure loop is completed. The main body of the report is used for routine submission, the evidence index is used for review and traceability, the discrepancy list is used to explain which conflicting indicators the system modified during the arbitration process, and the review comments are used to indicate low-confidence content that still requires manual confirmation.

[0068] The six system capability modules and methodological steps correspond one-to-one. Data access and evidence management capability implements S1 and generates original evidence nodes; power entity and indicator analysis capability implements S2 and generates power entity nodes, indicator nodes, and evidence pointers; chapter-evidence-constraint coupling task graph construction capability implements S3; multi-agent collaborative execution capability implements S4 and writes the results of each agent into the output node; electrical constraint verification and consistency arbitration capability implements S5 and S6 sequentially, and backfills the unified indicators after arbitration into the relevant chapters; report drafting and review package output capability, combined with quality closed-loop judgment, implements S7 and S8. If S7 fails to meet the criteria, the task graph regenerates tasks S4 to S6 based on missing evidence, conflicts, and low confidence states until the termination criteria are met or the manual review exit is reached.

[0069] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-agent collaborative method for power constraint perception in the automatic generation of power feasibility study reports, characterized in that, include: Obtain the power feasibility study report, construct a task diagram based on the power feasibility study report, and verify the chapter nodes of the task diagram to obtain the verification status; Based on the verification status, a candidate task set is formed, an agent is selected for each candidate task in the candidate task set, and a comprehensive assignment score is calculated to obtain the optimal agent; Calculate the unified engineering index values ​​for each optimal agent, identify the sources of engineering conflicts, generate supplementary verification tasks and return the task graph to obtain the final power feasibility study report; The final power feasibility study report is scored for quality. The quality score results are judged by a hierarchical termination criterion, and the current best version report, a list of low-confidence content, and a list of supplementary certifications are output.

2. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, The task graph includes: Extract the original evidence nodes from the power feasibility study report, and parse and normalize them to form power entity nodes and indicator nodes; The original evidence node, the power entity node, and the indicator node are introduced into the node set, and the chapter nodes, constraint nodes, agent capability nodes, and output nodes of the power feasibility study report are also included in the node set. The task graph is obtained by determining the edge relationship type and credibility of each node in the node set; the edge relationship type includes reference, dependency, generation, verification, backfilling, conflict and replacement.

3. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, Obtaining the verification status includes: The chapter type, dependency edges, and generation conditions of each chapter node in the task graph are obtained to verify whether the current chapter node has the necessary evidence nodes and constraint nodes. If the necessary evidence nodes and constraint nodes are present, the corresponding chapter of the current chapter node is split into data extraction, engineering calculation, text generation, and result verification subtasks, and added to the pending execution queue according to the chapter processing priority. The verification status is generated and written back to the corresponding chapter node and output node. If the necessary evidence nodes and constraint nodes are not present, a missing evidence task is generated, and the generation of the corresponding chapter text of the current chapter node is temporarily suspended.

4. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 3, characterized in that, The processing priority of the aforementioned chapters includes: ； in, Represents the k-th chapter node Processing priority; Chapter The proportion of missing necessary evidence; Chapter The intensity of existing constraint conflicts; Chapter Risk weights in the review rules; Chapter The degree to which it is relied upon by other chapters; This represents the weight coefficient of each evaluation factor.

5. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, Forming the candidate task set includes: Traverse the chapter nodes and output nodes in the task graph that are in a pending or reprocessing state, and read the missing original evidence, constraint violation, cross-chapter conflict indicators and downstream dependencies of the incoming edges; Based on the missing original evidence, a supplementary evidence task is generated; based on the violation of constraints, a recalculation or scheme adjustment task is generated; based on the cross-chapter conflict index, an index arbitration and chapter backfilling task is generated; based on the output that fails the verification, a review or rewriting task is generated; and for chapters that have not yet been generated but whose preconditions have been met, a chapter generation task is generated. The supplementary certification task, the recalculation or scheme adjustment task, the indicator arbitration and chapter backfilling task, the review or rewriting task, and the chapter generation task are coded according to task type, associated chapter, required capabilities, and priority, and then deduplicated to form the candidate task set.

6. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, Calculating the comprehensive assignment score to obtain the optimal agent includes: ； ； in, Represents intelligent agents Execute the task The overall work assignment score; This indicates the degree of matching between the agent's capabilities and the task requirements; This indicates the expected reduction in the intensity of constraint conflicts; Indicates the expected amount of supplementary evidence; This indicates the agent's historical success rate or credibility in similar tasks; This refers to the cost of invoking, the time cost, or the cost of manual review; Indicates the weight of the capability matching item; Indicates the weight of the conflict reduction item; Indicates the weight of the evidence completion item; Indicates the weight of the historical credibility item; Indicates the weight of the cost penalty item; Indicates that the task has been selected to be executed. The optimal intelligent agent; Indicates the ability to handle tasks The set of candidate intelligent agents; This represents the m-th candidate agent; This represents the selection operation that maximizes the objective function.

7. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, The method also includes: The optimal agent is established with the task, and the task status is set to dispatched. At the same time, the optimal agent reads the original evidence nodes, entity nodes, indicator nodes and constraint nodes associated with the task graph and performs supplementary evidence, data extraction, engineering calculation, chapter generation or result verification according to the task type. The execution results are written to the output node, and the satisfaction status, evidence completeness and credibility of the corresponding constraints are updated. If the verification fails, a recalculation, supplementary evidence or backfill sub-task is generated along the conflict edge to update the credibility. The candidate agent screening and dispatch are re-executed according to the updated credibility. During the assignment of candidate agents, if the conflict indicators can determine the workload and quota basis in the evidence node, a workload verification task is first established. If the verified workload and cost quota can calculate a new investment value, an investment estimation recalculation task is established. If the recalculated value passes the investment boundary constraint, a backfilling task for the general overview, investment estimation table, and economic evaluation is established. If the evidence is still insufficient or the recalculated value violates the investment boundary, a supplementary evidence task is established and the completion of relevant chapters is suspended. The task graph is updated after each subtask is completed. If the output of the current optimal agent leads to an increase in constraint conflicts or fails to improve the quality score for multiple consecutive rounds, the credibility of the current optimal agent on the corresponding task is reduced, and the assignment probability of other candidate agents is increased. At the same time, the credibility update result is written back to the corresponding agent capability node.

8. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, The calculation of unified engineering performance values ​​for each optimal agent includes: ； in, This represents the unified value of the p-th engineering indicator after arbitration; Indicates the candidate values; This represents the value of the p-th engineering metric output by the i-th chapter or the i-th agent; This indicates the number of outputs related to the indicator p; This represents the confidence weight of the i-th output; This represents the deviation loss between the candidate value and the original output; Indicates the power constraint penalty weight; This represents the penalty function for candidate values ​​violating power industry constraints. This penalty function is obtained by performing constraint penalty calculations after dimensional normalization of the voltage deviation constraint violation function, line current carrying capacity constraint violation function, and power supply radius constraint violation function.

9. The multi-agent cooperative method for power constraint perception in automatic generation of power feasibility study reports according to claim 1, characterized in that, The output of the current best version report, the list of low-confidence content, and the list of content requiring supplementary verification includes: The hierarchical termination criterion is set to distinguish power constraints into hard constraints and soft indices. The remaining number of violations of the hard constraints must be zero to be considered as meeting the standard. At the same time, the soft indicators are used to score the quality of the final power feasibility study report, and the target threshold is used to determine whether to continue the next round of supplementary certification, recalculation, verification or rewriting tasks. After each round, the iteration is terminated based on a target termination determination; the target termination determination includes: Eligibility Exit: After each round of chapter assembly, engineering indicator arbitration, and full verification, if... If Q(k)≥τ, then this round of reports will be solidified into the final power feasibility study report and terminated; Resource Exit: If If Q(k) < τ, but Kmax has been reached, then stop adding new automatic tasks, retain the historical highest score snapshot, unmet constraints and their responsibility evidence / sections, and output a list of manual review; Convergence Exit: If If Q(k) < τ and ΔQ(k) < ε for consecutive preset rounds, it means that supplementary proof, recalculation and rewriting can no longer produce effective quality improvement. Roll back or keep the historical highest score snapshot and output the manual review list. Blocked Exit: If the remaining missing evidence satisfies If the credibility of all candidate agents for a critical pending task falls below the preset lower limit, then mark the cause of the blockage, related evidence, and affected chapters, terminate automatic task assignment, and transfer the task to manual verification or review. in, This indicates that all hard constraints have been met; Q(k) represents the quality score of the k-th round report, which is a comprehensive score obtained by weighting the power feasibility study report assembled in the k-th round according to the completeness of chapters, consistency of cross-chapter indicators, degree of satisfaction of key electrical constraints, degree of matching of normative clauses, degree of rationality of investment economy and traceability of evidence chain; τ represents the report quality threshold. This indicates that all hard constraints are not met; ΔQ(k) = Q(k) - Qbest(k-1) represents the increment of the current round score relative to the historical best snapshot score; Qbest(k-1) represents the historical highest score at the end of round (k-1); ε represents the minimum effective improvement threshold, calibrated by the upper bound of the inter-round noise of the score repeatedly generated with the same input; Kmax represents the maximum number of iteration rounds, calibrated by the number of rounds required for the vast majority of reports in historical projects to reach the quality threshold; M(k) represents the set of missing evidence in round k; Mext represents the set of evidence that needs to be supplemented externally and cannot be automatically obtained by the system. This indicates that the remaining missing evidence belongs to external evidence that the system cannot automatically obtain; Represents the empty set; ∧ represents a set inclusion relationship; ∧ represents a logical AND operation; ∨ represents a logical OR operation.

10. A multi-agent collaborative system for power constraint perception in the automatic generation of power feasibility study reports, implementing the method described in any one of claims 1-9, characterized in that, include: The task diagram construction module is used to obtain the power feasibility study report, construct a task diagram based on the power feasibility study report, and verify the chapter nodes of the task diagram to obtain the verification status. The multi-agent collaborative execution module is used to form a candidate task set according to the verification status, select agents for each candidate task in the candidate task set, calculate the comprehensive assignment score, and obtain the optimal agent. The verification and arbitration module is used to calculate the unified engineering index values ​​of each optimal agent, determine the source of engineering conflicts, generate supplementary verification tasks and return the task graph to obtain the final power feasibility study report. The report and audit package output module is used to score the quality of the final power feasibility study report, judge the quality score results through hierarchical termination criteria, and output the current best version report, the list of low confidence content, and the list of supplementary certifications.

Images