A multi-agent large language model-based urban building low-carbon renovation decision system and method

Abstract

The application discloses a kind of urban building low carbon reconstruction decision system and method based on multi-agent large language model, comprising: obtaining city building stock, energy consumption simulation and budget policy data;By semantic processing agent, call large language model to modify the semantics of building attributes and generate semantic confidence, aggregate buildings into semantic decision units and calculate comprehensive confidence;By knowledge agent, construct reconstruction strategy library;By scene agent, generate unit-strategy candidate matrix, and use building type, age, climate zone, policy access and confidence threshold for gate control screening;Under the constraint of total budget and single-strategy budget proportion, generate reconstruction combination scheme through multi-objective optimization;Reviewed and output audit record by audit agent.The application converts building energy consumption diagnosis into executable and traceable reconstruction list, improves the reliability of urban scale low carbon reconstruction decision and budget utilization efficiency.

Description

Technical Field

[0001] This invention relates to urban planning, building energy conservation, urban building energy consumption simulation, low-carbon retrofitting combination optimization, and artificial intelligence decision-making systems and methods, and particularly to an urban building low-carbon retrofitting decision-making system and method based on a multi-agent large language model. Background Technology

[0002] Building energy consumption and building-related carbon emissions are key targets in urban low-carbon transformation. Urban building energy consumption modeling can estimate the energy consumption levels and energy-saving potential of a large number of buildings at the city scale. However, in actual low-carbon retrofitting work, simply knowing which buildings have high energy consumption is insufficient to formulate an actionable plan. Governments or park management agencies usually need to determine which buildings to prioritize for retrofitting, what measures to take for each type of building, and how to allocate the budget among various measures, under constraints such as fixed budgets, policy access, building type suitability, construction feasibility, and strategy diversity.

[0003] Existing publicly available technologies include methods for calculating carbon emissions from existing building renovations, comprehensively optimizing energy consumption and microclimate in urban building clusters, managing building energy efficiency, and optimizing and controlling building equipment. For example, existing methods can evaluate existing building renovation schemes based on life-cycle carbon emission calculations; others can optimize energy consumption and comfort in urban building cluster design schemes by incorporating local microclimates; and still others can predict, monitor, or schedule building energy consumption. While these technologies can address specific problems in building design, operation control, or energy consumption assessment, they typically remain at the level of assessing energy-saving potential, optimizing individual or partial equipment, and comparing design schemes.

[0004] From the perspective of urban governance and renovation implementation, low-carbon building retrofitting is more like a combination allocation problem with policy and budget constraints. On the one hand, the large-scale building stock is significantly heterogeneous, with different buildings having different functions, construction years, thermal formwork, microclimate exposure levels, and roof usability. On the other hand, retrofitting funds are limited, and specific measures are also restricted by local policies, implementation order, property rights conditions, and construction feasibility. Therefore, prioritizing retrofitting plans solely based on energy intensity or single emission reduction potential can easily lead to a concentration of retrofitting schemes on a few low-cost measures, making it difficult to meet the requirements of multi-strategy coordination and policy implementation.

[0005] Meanwhile, urban building data often suffers from incomplete or inconsistent semantic attributes. For example, building function, construction year, and thermal performance specifications may come from different data sources, leading to conflicts between them. Points of interest and building names can provide important semantic clues, but their reliability is affected by spatial matching accuracy and information update timeliness. Existing optimization methods typically treat these building attributes as deterministic inputs, rarely explicitly passing semantic uncertainty to the applicability judgment and combination optimization process of renovation strategies. This may result in the misallocation of measures such as deep renovation, which heavily rely on building semantic attributes.

[0006] Large language models and agent workflows can be used for semantic understanding, rule extraction, task decomposition, and result interpretation. However, if numerical modification schemes are directly generated by large language models, problems such as unverifiable results or inconsistencies with physical calculation results can easily arise. A more prudent approach is to limit the large language model agent to semantic correction, rule organization, candidate construction, and audit interpretation, while delegating energy consumption calculation, cost calculation, and combinatorial optimization to deterministic models or optimization algorithms.

[0007] Therefore, it is necessary to propose a method that combines microclimate energy consumption evidence, building semantic confidence, multi-agent large language model workflow, policy budget constraints, and combinatorial optimization algorithms, so that urban building energy consumption diagnosis results can be further transformed into executable, interpretable, and auditable low-carbon retrofitting schemes. Summary of the Invention

[0008] Purpose of the invention: In order to overcome the shortcomings of existing methods for low-carbon retrofitting of urban buildings in terms of urban-scale combination and allocation, transmission of semantic uncertainty of buildings, coupling of budget policy constraints and auditability of results, the purpose of this invention is to propose a decision-making system and method for low-carbon retrofitting of urban buildings based on a multi-agent large language model.

[0009] The method of this invention includes four core technical components: first, a physical evidence layer consisting of urban building energy consumption models, microclimate meteorological boundary conditions, building stock data, and carbon emission factors; second, an intelligent agent decision-making layer consisting of data intelligent agents, knowledge intelligent agents, scenario intelligent agents, optimization intelligent agents, and auditing intelligent agents; third, a candidate generation mechanism consisting of semantic decision-making units, unit-policy candidate matrices, confidence gating, and policy access control; and fourth, a combined decision-making mechanism consisting of multi-objective optimization, budget constraints, policy diversity constraints, and audit interpretation.

[0010] The core innovation of this method lies in: inputting building interest points, building names, land use, and building energy consumption template information into a large language model agent; semantically correcting the building function type, construction year, and thermal template; and transforming the corrected results into semantic decision units with confidence; setting semantic confidence gating, building type gating, construction year gating, local climate zone gating, and policy access gating in the unit-strategy candidate matrix to exclude uncertain or unsuitable candidates before optimization; using total budget, budget ratio of single-strategy, candidate mutual exclusion, and policy access as combined optimization constraints; generating city-scale low-carbon transformation combinations through multi-objective optimization algorithms and / or mixed-integer linear programming algorithms; and verifying the budget, policy, applicability, confidence, and evidence chain through an auditing agent.

[0011] Technical solution: The decision-making method for low-carbon retrofitting of urban buildings based on a multi-agent large language model of the present invention includes the following steps: Step 1: Obtain building stock data, microclimate energy consumption simulation results, renovation strategy library, and budget and policy constraint data for the target city; Step 2: The semantic processing agent calls the large language model to perform semantic correction on the building function type, construction year and thermal template based on building name, point of interest data and land use data, and generates the semantic confidence of each building; buildings in the same spatial grid with the same or similar function type, construction year and thermal template are aggregated into semantic decision units, and the comprehensive confidence is calculated by weighting by building area. Step 3: The knowledge intelligence agent extracts the applicable conditions, cost parameters, energy-saving parameters, carbon emission reduction parameters, policy access conditions, and minimum confidence requirements of the transformation strategies from transformation literature, policy documents, and engineering rules to form a transformation strategy library; Step 4: The scene intelligence agent combines each semantic decision unit with each strategy in the transformation strategy library to generate a unit-strategy candidate matrix. The matrix is ​​then gating and screening based on building type, year of construction, local climate zone, policy access conditions, and a comprehensive confidence level that is not lower than the minimum confidence level requirement corresponding to the strategy. Candidates that pass all conditions are then entered into the optimization set. Step 5: For each candidate option entering the optimization set, calculate the investment cost, energy saving per representative cycle, carbon emission reduction per representative cycle, and annualized carbon emission reduction. Step 6: Under the constraints that the total budget does not exceed the budget limit, the total investment of any transformation strategy does not exceed a preset proportion of the total budget, and each semantic decision unit can select at most one transformation strategy, a multi-objective optimization algorithm is used to generate a transformation combination scheme. Step 7: The auditing agent reviews the budget compliance, policy applicability, strategy feasibility, and evidence sources of the proposed transformation and combination scheme, and generates audit records. Step 8: Output a list of urban building low-carbon retrofit combinations that includes the retrofit targets, retrofit strategies, investment costs, annualized carbon emission reductions, and audit results.

[0012] Furthermore, the overall confidence level of the semantic decision-making unit: ; in, Represents a semantic decision-making unit. This refers to buildings that belong to this semantic decision-making unit. Represents architecture The building area, Represents architecture The semantic confidence.

[0013] Furthermore, the admission indication value I(u,s) for the gating screening is determined according to the following rules: ; in, Represents a semantic decision-making unit. Indicates the transformation strategy. and These represent the building function, age category, local climate zone, and overall confidence level of the semantic decision-making unit, respectively. , and Representing strategies The set of applicable functions, the set of applicable eras, and the set of applicable local climate zones. Indicates policy access criteria, Representation strategy Minimum requirements for semantic confidence.

[0014] Furthermore, in step five, the annualized carbon emission reduction is calculated based on the carbon emission reduction of the representative period and the annual correction coefficient, and the annual correction coefficient is determined by the correspondence between the representative period baseline energy consumption and the annual baseline energy consumption.

[0015] Furthermore, in step six, the multi-objective optimization algorithm aims to maximize one or more of the following: annualized carbon emission reduction, maximize the number of covered buildings, and minimize investment cost. The constraints are that the total investment cost does not exceed the budget limit, the total investment amount of any type of transformation strategy does not exceed the budget ratio limit corresponding to that type of strategy, and each semantic decision unit can select at most one transformation strategy. The 0-1 variables represent whether each candidate option is selected.

[0016] Furthermore, the multi-objective optimization algorithm employs NSGA-II and / or a mixed-integer linear programming algorithm.

[0017] Furthermore, in step seven, the auditing agent outputs an audit label for each selected candidate; if there is a budget overrun, inapplicable policy, insufficient building semantic confidence, mismatch between strategy and building type, or missing cost parameters, an error item or warning item is generated, and the corresponding candidate is returned for reprocessing.

[0018] Furthermore, in step one, the microclimate energy consumption simulation results are generated by the urban building energy consumption model driven by microclimate meteorological boundary conditions.

[0019] Furthermore, it includes a data intelligent agent module, a knowledge intelligent agent module, a scenario intelligent agent module, an optimization intelligent agent module, an auditing intelligent agent module, and an output interaction module, which respectively execute the functions of each intelligent agent; it also includes a processor, a memory, and a computer program stored in the memory that can be executed by the processor, wherein the program, when executed, implements the method described above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves a closed loop from urban building energy consumption diagnosis to the allocation of low-carbon retrofit combinations. Compared with urban building energy consumption models that only output energy-saving potential or energy consumption hotspots, this invention further transforms energy consumption evidence, retrofit strategies, and policy budget constraints into an actionable retrofit combination list.

[0021] 2. This invention explicitly incorporates architectural semantic uncertainty into the decision-making process. By using semantic confidence and policy confidence thresholds, the system can avoid implementing renovation strategies that strongly depend on building type and age for semantically uncertain building units, thereby improving the robustness of the combined scheme.

[0022] 3. This invention improves policy consistency for low-carbon retrofitting at the city scale. By setting budget caps, budget proportions for single-strategy approaches, and policy access controls, the system avoids excessive concentration of optimization results on a single low-cost measure, making retrofitting combinations more aligned with multi-strategy synergy and practical implementation requirements.

[0023] 4. This invention has good interpretability and auditability. The auditing agent can review the budget, policy, applicability, and source of evidence for each selected candidate, and can explain the reasons for selection and non-selection of a specified building or region through a question-and-answer format.

[0024] 5. In the Shanghai urban building stock example, the present invention can process approximately 592,795 buildings and form approximately 38,338 semantic decision units; under the constraints of a budget of 1 billion yuan and the proportion of budget for a single strategy, it can form a renovation combination covering approximately 30,075 buildings and achieve an annualized carbon emission reduction of approximately 126.2 ktCO2 / yr, indicating that the method can form a combinatorial optimization result with engineering feasibility on a large-scale urban building stock. Attached Figure Description

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a roadmap for constructing the semantic decision unit of this invention; Figure 3 This is a roadmap for the coordination and execution of the five multi-agent systems in this invention; Figure 4 This is a diagram showing the allocation of emission reduction measures and dominant strategies in Shanghai as described in this invention; Figure 5 This is a performance comparison chart of the present invention tested using ten large language models; Figure 6 This is a ranking of the modification measures selected by multiple agents for different building types according to the present invention; Figure 7 This is a schematic diagram of the digital twin platform that accompanies this invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0028] This embodiment focuses on the optimization of low-carbon retrofitting of buildings at the urban scale. The goal is to select priority retrofit targets and corresponding retrofitting measures from a large pool of urban buildings, given budget, policy constraints, and a retrofitting strategy library. Instead of directly generating final numerical conclusions from a large language model, the method utilizes a large language model agent for semantic correction, rule organization, candidate generation, and audit interpretation, while a deterministic computation module handles energy consumption, cost, carbon emission reduction, and optimization solutions, thus ensuring verifiable results.

[0029] The decision-making system and method for low-carbon retrofitting of urban buildings based on a multi-agent large language model of the present invention includes the following steps: S1. Construct the physical evidence layer. Obtain the building outline, building height, building use, year of construction, building area, local climate zone, administrative division, spatial grid, microclimate energy consumption simulation results, and electricity carbon emission factor of the target city; among which, the microclimate energy consumption simulation results can be generated by the city building energy consumption model and may include representative cycle energy consumption, annual baseline energy consumption, cooling energy consumption, electricity intensity, and carbon emission data per building or per grid.

[0030] S2. Construct a multi-agent decision-making layer. Set up a data agent, a knowledge agent, a scenario agent, an optimization agent, and an audit agent; among them, the data agent is responsible for correcting building attributes and organizing data, the knowledge agent is responsible for transforming documents and policies into policy rules, the scenario agent is responsible for constructing the candidate matrix, the optimization agent is responsible for calling optimization algorithms, and the audit agent is responsible for compliance and evidence chain verification.

[0031] S3. Construct semantic decision units. The data agent reads points of interest, building names, land use, and building spatial context, corrects the building function type, construction year, and thermal template, and generates the semantic confidence score for each building. Then, buildings located in the same spatial grid and with the same or similar function type, construction year, and thermal template are aggregated into semantic decision units, and the area-weighted comprehensive confidence score is calculated.

[0032] S4. Construct a retrofit strategy library. The knowledge agent extracts the applicable conditions and parameters of retrofit measures from retrofit literature, policy texts, and engineering experience to form a strategy library. The strategy library may include strategies such as reflective roofs, operation control, roof photovoltaics, external shading, building envelope retrofitting, and deep energy-saving retrofitting, and records the applicable building type, applicable era, applicable local climate zone, unit cost, energy-saving parameters, carbon emission reduction parameters, and policy access conditions for each type of strategy.

[0033] S5. Generate and filter the candidate matrix. The scene agent combines each semantic decision unit with each type of transformation strategy to generate a unit-strategy candidate matrix; for each candidate, gating is performed based on building type, era category, local climate zone, policy access, and semantic confidence. Only candidates that simultaneously meet all gating conditions are included in the optimization set.

[0034] S6. Calculate candidate indexes. For each candidate index entering the optimization set, calculate the investment cost, energy savings in the representative period, carbon emission reduction in the representative period, annualized carbon emission reduction, number of covered buildings, and unit carbon emission reduction cost. The annualized carbon emission reduction can be calculated using the representative period results and the annual correction coefficient to anchor the representative period simulation results to the annual scale.

[0035] S7. Perform combinatorial optimization. Under constraints such as the total budget not exceeding the budget limit, the budget ratio of a single policy not exceeding a preset proportion, each semantic decision unit selecting at most one transformation policy, and candidate policies meeting the admission indicator value, the optimization agent calls the NSGA-II multi-objective optimization algorithm and / or mixed integer linear programming algorithm to obtain one or more non-dominated transformation combination schemes, and selects the recommended scheme according to the policy objective.

[0036] S8. Execution Review and Traceability. The review agent verifies the recommended solution, checking whether the total budget, strategy budget ratio, policy applicability, candidate mutual exclusivity, semantic confidence, cost parameters, and carbon emission reduction calculation meet the preset rules; for candidates that do not meet the rules, it outputs errors or warnings and generates audit records.

[0037] S9. Output executable results. The system output includes a list of renovations with semantic decision unit number, spatial grid, administrative region, local climate zone, number of buildings, building area, recommended renovation strategy, investment cost, annualized carbon emission reduction, unit emission reduction cost, semantic confidence, and audit tags. It also generates a spatial distribution map, strategy proportion map, sensitivity analysis results, and question and answer explanation results.

[0038] like Figure 1 As shown, in one specific embodiment, The urban building low-carbon retrofitting combined optimization method based on a multi-agent large language model adopted in this embodiment includes the following steps and main technical processes: S1. Obtain building and energy consumption data for the target city. This includes: S11. Obtain data on building outlines, building heights, building areas, building uses, land use, building names, points of interest, administrative divisions, and local climate zone classifications for the target city. S12. Assign buildings to spatial grids. Preferably, the spatial grids can be 500m×500m in scale, but other scales can be used depending on the accuracy of the urban data. S13. Read the microclimate-driven urban building energy consumption simulation results, which include representative cycle power consumption, annual baseline power consumption, cooling energy consumption, carbon emissions, and energy intensity for each building or each grid. S14. Read the electricity carbon emission factor, retrofit cost parameters, budget parameters, and policy constraint parameters.

[0039] S2. Set up a multi-agent workflow. Specifically, this includes: setting up planning logic to break down tasks into five sub-tasks: data processing, knowledge rule construction, candidate generation, optimization solution, and review and traceability; setting up a data agent responsible for correcting building semantic attributes, calculating confidence levels, and constructing semantic decision units; setting up a knowledge agent responsible for converting renovation documents, policy texts, engineering rules, and cost parameters into a structured strategy library; setting up a scenario agent responsible for combining semantic decision units with renovation strategies and establishing a candidate matrix; setting up an optimization agent responsible for calling the NSGA-II and / or mixed-integer linear programming solver; and setting up a review agent responsible for checking budget, policy, applicability, confidence levels, and evidence chains.

[0040] S3. Constructing Semantic Decision Units. The data agent reads the original use, building name, point of interest tags, and land use category of each building; it semantically corrects the building function type through a large language model and maps the results to residential, office, educational, public, commercial, or other preset types; it infers or corrects the construction year category based on the building name, land use, spatial location, and existing age information; it maps the corrected building function and construction year to a thermal template; it generates a semantic confidence score for each building, which can be determined by evidence consistency, point of interest density, name matching degree, land use consistency, and rule verification results; and it aggregates buildings located in the same spatial grid with the same function type, age category, and thermal template into semantic decision units.

[0041] S4. Construct a renovation strategy library. The knowledge agent reads renovation measure literature, policy documents, engineering experience tables, and cost databases; for each type of renovation strategy, it sets the applicable building type set, applicable era set, applicable local climate zone set, minimum confidence threshold, unit cost, energy saving rate, or emission reduction coefficient; renovation strategies may include reflective roofs, operation control, roof photovoltaics, external shading, building envelope renovation, and deep energy-saving renovation; it configures policy access conditions for each type of strategy, such as whether it is allowed for a certain type of public building, whether it is applicable to buildings of a certain era, and whether it is restricted by roof area or property rights conditions; the strategy library is saved as a structured table so that subsequent candidate matrices can be directly called.

[0042] S5. Generate unit-policy candidate matrix. The scenario agent generates candidate options for each semantic decision unit and each type of transformation policy; performs function type gating, era gating, local climate zone gating, policy gating, and confidence gating on the candidate options; if any gating condition is not met, the candidate option is excluded; all candidate options with full gating are entered into the optimization set, and the admission reason for each candidate option is recorded.

[0043] S6. Calculate candidate costs and emission reduction targets. Calculate the investment cost of candidate options based on the building area, roof area, number of buildings, or other scale factors of the semantic decision unit; calculate the energy saving for the representative period based on the microclimate energy consumption simulation results and strategy energy saving parameters; calculate the carbon emission reduction for the representative period based on the electricity carbon emission factor; calculate the annualized carbon emission reduction based on the annual correction coefficient, which can be determined by the proportional relationship between the annual baseline simulation and the representative period simulation; calculate the unit carbon emission reduction cost, the number of covered buildings, and the strategy budget ratio; and write the above indicators into the candidate matrix.

[0044] S7. Perform budget constraint combinatorial optimization. Set decision variables. If the candidate If selected =1, otherwise =0; Set total budget constraints, single-strategy budget ratio constraints, candidate mutual exclusion constraints, and candidate admission constraints; With the goal of maximizing annualized carbon emission reduction, maximizing the number of covered buildings, minimizing investment costs, or minimizing unit carbon emission reduction costs, call the NSGA-II algorithm to generate a non-dominated solution set, and can further call the mixed integer linear programming algorithm to perform baseline or exact solutions for the specified objectives.

[0045] S8. Execute the auditing agent for review. The auditing agent reads the recommended combination, candidate matrix, strategy library, and constraint parameters. It checks whether the total investment cost exceeds the budget, whether the budget percentage for each strategy exceeds the corresponding upper limit, whether each selected candidate option meets the functional type, era category, local climate zone, policy access, and confidence gating, and whether the annualized carbon emission reduction and unit carbon emission reduction cost are calculated from the parameters in the candidate matrix. If errors are found, an error label is generated and the process is returned to the corresponding agent for reprocessing; if only insufficient data or parameters need to be reviewed, a warning label is generated.

[0046] S9. Output Transformation Combinations and Interpretation Results. The system outputs a list of urban building low-carbon transformation combinations, with fields including semantic decision unit number, spatial grid number, administrative region, local climate zone, building function, era category, number of buildings, building area, recommended strategy, investment cost, annualized carbon emission reduction, unit carbon emission reduction cost, semantic confidence, and review label. It also outputs spatial distribution results, strategy proportion and budget proportion results, as well as Q&A interpretation results, to answer the reasons for selection, non-selection, and evidence sources for a specific building, grid, administrative region, or strategy.

[0047] In one specific embodiment, the target city is Shanghai. The system reads building stock data from approximately 592,795 buildings and microclimate-driven urban building energy consumption simulation results, aggregating the buildings into approximately 38,338 semantic decision units. Retrofit strategies include reflective roofs, operational control, rooftop photovoltaics, external shading, building envelope modifications, and deep energy-saving retrofits. Under a budget of 1 billion yuan and a 35% single-strategy budget cap, the system generates a recommended combination covering approximately 30,075 buildings, yielding an annualized carbon emission reduction of approximately 126.2 ktCO2 / yr. Compared to ranking solely by energy consumption or emission reduction potential, this embodiment simultaneously satisfies budget, strategy diversity, semantic confidence, and policy access criteria, and each selected candidate has a traceable chain of evidence.

[0048] In another alternative embodiment, if the target city lacks annual microclimate energy consumption simulation results, the candidate emission reduction estimates can be formed first using the representative period simulation results, and then annualized using the annual baseline correction coefficient; when the annual simulation results are obtained later, the candidate emission reduction parameters can be recalibrated and the combined optimization can be run again.

[0049] In another alternative embodiment, if there is insufficient semantic evidence of buildings in the target city, the confidence threshold of the deep renovation strategy can be increased, or uncertain building units can be allowed to enter only the low-sensitivity strategy candidate set, such as reflective roof or operation control strategy, thereby reducing the impact of incorrect semantic classification on the combination scheme.

[0050] This invention is not limited to the embodiments described above. For different climate zones, different city scales, different budget sizes, and different policy objectives, only the building stock data, microclimate energy consumption results, strategy library, and constraint parameters need to be replaced to generate corresponding urban building low-carbon retrofitting schemes.

[0051] In the description of this specification, references to terms such as "embodiment," "example," "by way of," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0052] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A decision-making method for low-carbon retrofitting of urban buildings based on a multi-agent large language model, characterized in that, Includes the following steps: Step 1: Obtain building stock data, microclimate energy consumption simulation results, renovation strategy library, and budget and policy constraint data for the target city; Step 2: The semantic processing agent calls the large language model to perform semantic correction on the building function type, construction year and thermal template based on building name, point of interest data and land use data, and generates the semantic confidence of each building; buildings in the same spatial grid with the same or similar function type, construction year and thermal template are aggregated into semantic decision units, and the comprehensive confidence is calculated by weighting by building area. Step 3: The knowledge intelligence agent extracts the applicable conditions, cost parameters, energy-saving parameters, carbon emission reduction parameters, policy access conditions, and minimum confidence requirements of the transformation strategies from transformation literature, policy documents, and engineering rules to form a transformation strategy library; Step 4: The scene intelligence agent combines each semantic decision unit with each strategy in the transformation strategy library to generate a unit-strategy candidate matrix. The matrix is ​​then gating and screening based on building type, year of construction, local climate zone, policy access conditions, and a comprehensive confidence level that is not lower than the minimum confidence level requirement corresponding to the strategy. Candidates that pass all conditions are then entered into the optimization set. Step 5: For each candidate option entering the optimization set, calculate the investment cost, energy saving per representative cycle, carbon emission reduction per representative cycle, and annualized carbon emission reduction. Step 6: Under the constraints that the total budget does not exceed the budget limit, the total investment of any transformation strategy does not exceed a preset proportion of the total budget, and each semantic decision unit can select at most one transformation strategy, a multi-objective optimization algorithm is used to generate a transformation combination scheme. Step 7: The auditing agent reviews the budget compliance, policy applicability, strategy feasibility, and evidence sources of the proposed transformation and combination scheme, and generates audit records. Step 8: Output a list of urban building low-carbon retrofit combinations that includes the retrofit targets, retrofit strategies, investment costs, annualized carbon emission reductions, and audit results.

2. The method according to claim 1, characterized in that, In step two, the overall confidence level of the semantic decision-making unit. cu Calculated using the following formula: ； in, Represents a semantic decision-making unit. This refers to buildings that belong to this semantic decision-making unit. Represents architecture The building area, Represents architecture The semantic confidence.

3. The method according to claim 1, characterized in that, Step four, the access indication value for the gating screening. I(u, s) Determined by the following rules: ； in, Represents a semantic decision-making unit. Indicates the transformation strategy. and These represent the building function, age category, local climate zone, and overall confidence level of the semantic decision-making unit, respectively. , and Representing strategies The set of applicable functions, the set of applicable eras, and the set of applicable local climate zones. Indicates policy access criteria, Representation Strategy Minimum requirements for semantic confidence.

4. The method according to claim 1, characterized in that, In step five, the annualized carbon emission reduction is calculated based on the carbon emission reduction of the representative period and the annual correction coefficient, which is determined by the correspondence between the representative period baseline energy consumption and the annual baseline energy consumption.

5. The method according to claim 1, characterized in that, In step six, the multi-objective optimization algorithm aims to maximize one or more of the following: annualized carbon emission reduction, maximize the number of covered buildings, and minimize investment cost. The constraints are that the total investment cost does not exceed the budget limit, the total investment amount of any type of transformation strategy does not exceed the budget ratio limit of that type of strategy, and each semantic decision unit can select at most one transformation strategy. The 0-1 variables represent whether each candidate option is selected.

6. The method according to claim 5, characterized in that, The multi-objective optimization algorithm employs NSGA-II and / or mixed integer linear programming algorithms.

7. The method according to claim 1, characterized in that, In step seven, the auditing agent outputs an audit label for each selected candidate; if there is a budget overrun, inapplicable policy, insufficient building semantic confidence, mismatch between strategy and building type, or missing cost parameters, an error item or warning item is generated, and the corresponding candidate is returned for reprocessing.

8. The method according to claim 1, characterized in that, In step one, the microclimate energy consumption simulation results are generated by the urban building energy consumption model driven by microclimate meteorological boundary conditions.

9. A decision-making system for low-carbon retrofitting of urban buildings based on a multi-agent large language model, characterized in that, It includes a data intelligent agent module, a knowledge intelligent agent module, a scenario intelligent agent module, an optimization intelligent agent module, an auditing intelligent agent module, and an output interaction module, which respectively perform the functions of each intelligent agent as described in any one of claims 1 to 8; it also includes a processor, a memory, and a computer program stored in the memory that can be executed by the processor, wherein when the program is executed, it implements the method described in any one of claims 1 to 8.

Images