A method for real-time optimization of building performance for low-carbon goals

Abstract

This invention relates to the field of building performance optimization technology and discloses a real-time building performance optimization method for low-carbon goals. The method includes: loading a building performance database, a material property library, and a low-carbon design constraint library; initializing a lightweight simulation engine and a pre-trained surrogate model; acquiring and preprocessing building parameters; performing multi-physics performance prediction; combining simulation results with surrogate model output for fusion evaluation; constructing a multi-objective comprehensive objective function; generating candidate design adjustments; and outputting optimization suggestions. The method updates building parameters through closed-loop iteration, optimizes design schemes in real time, and ultimately outputs building schemes that meet low-carbon goals. This invention achieves efficient optimization of low-carbon building design through real-time feedback and multi-physics data fusion.

Description

Technical Field

[0001] This invention relates to the field of building performance optimization technology, and more specifically, to a method for real-time optimization of building performance aimed at low-carbon goals. Background Technology

[0002] As the global low-carbon transition accelerates, the construction industry, as a major source of energy consumption and carbon emissions, urgently needs to break through the limitations of traditional design models. In green building design practice, building performance optimization is a core element in achieving energy conservation and emission reduction goals, but the existing technological system contains deep-seated contradictions. Energy efficiency simulation and carbon emission calculation generally lag behind the design process, usually starting only in the scheme refinement or construction drawing stage, resulting in a lack of real-time performance feedback for designers in the conceptual and preliminary design phases. This delay mechanism deviates from quantitative basis for design decisions, making it impossible to dynamically adjust key parameters such as geometry and building envelope during the initial scheme development stage, leading to frequent rework and even making it difficult for the final scheme to meet stringent low-carbon performance requirements. The data fragmentation problem between multiphysics simulation tools further exacerbates the optimization difficulty. The evaluation of physical dimensions such as thermal performance, lighting, and ventilation relies on independent software platforms with incompatible data formats and a lack of interactive interfaces. Designers are forced to manually convert parameters between different tools, which is not only inefficient but also makes it difficult to capture parameter coupling effects, such as the two-way impact of window size adjustment on lighting gain and heat load, resulting in a one-sided performance evaluation and an inability to form a global optimization perspective. Meanwhile, low-carbon goals have failed to be internalized as a driving force in the design logic. Current methods treat carbon emission constraints as ex-post verification indicators rather than an integral part of the design process. Designers primarily rely on rules of thumb and regulatory provisions to advance their work, lacking mechanisms to embed operational carbon emissions and material-borne carbon targets into real-time parameter adjustments. This leads to a persistent disconnect between design intent and low-carbon performance goals, resulting in a broken feedback loop. Furthermore, existing systems cannot provide actionable optimization suggestions based on real-time data. Performance evaluation results are often generated only at the end of the design cycle, and the suggestions are vague and general, failing to specify parameter adjustment amounts, expected improvement effects, and constraint impacts. Designers struggle to verify adjustment schemes in real-time during interactive design, leading to a blind and inefficient optimization process. These shortcomings collectively hinder the evolution of building performance optimization towards real-time and integrated approaches, severely restricting the efficient achievement of low-carbon design goals. Summary of the Invention

[0003] The purpose of this invention is to provide a method for real-time optimization of building performance aimed at low-carbon goals, in order to solve the above-mentioned problems.

[0004] This invention provides a real-time building performance optimization method for low-carbon goals, comprising: Load the building performance database, material property database, material carbon emission factor database, location climate database, and low-carbon design constraint library; initialize the lightweight simulation engine and pre-trained proxy model; and establish the baseline scheme parameter set and the low-carbon target parameter set. Input building parameters to obtain a set of building parameters for the building scheme to be optimized. The set of building parameters includes geometric parameters, building envelope parameters, equipment and operation parameters, and location and climate parameters. Parameter preprocessing involves unifying units and structurally encoding the building parameter set to obtain a structured feature vector. Based on the low-carbon design constraint library, compliance detection is performed on the structured feature vector. When non-compliant items are detected, constraint repair is performed on the non-compliant items to obtain a feasible feature vector. Multiphysics performance prediction involves inputting the feasible feature vector into a lightweight simulation engine to obtain a first multiphysics performance index set, and inputting the feasible feature vector into a pre-trained surrogate model to obtain a second multiphysics performance index set. The fusion evaluation involves fusing and calibrating the first multiphysics performance index set with the second multiphysics performance index set to obtain the fusion evaluation result. The fusion evaluation result is then compared with the benchmark index and the target threshold to obtain the performance gap vector and index coupling metric. The objective is constructed by determining dynamic weights based on building type identifiers, design stage identifiers, the performance gap vector, and the index coupling metric, and then constructing a multi-objective comprehensive objective function and determining the constraint set based on the dynamic weights. Candidate design adjustment generation: Under the constraint set, a candidate design adjustment set is generated, and the expected performance improvement and constraint impact are calculated for each candidate design adjustment in the candidate design adjustment set. The optimization suggestions are output by adjusting the priority scores of each candidate design based on the comprehensive scoring rules and sorting them. The optimization suggestions include parameter adjustment amounts, expected performance improvement amounts, and explanations of the impact of constraints. Closed-loop iteration: receiving selection signals for the optimization suggestions, updating the building parameter set based on the selection signals to obtain an updated building parameter set, repeating the parameter preprocessing based on the updated building parameter set until the optimization suggestions are output, and recording the parameter change trajectory and the fusion evaluation results to form an optimization history library; The model update and termination determination involves incrementally updating the pre-trained proxy model based on the optimization history library, and outputting the optimized building scheme and the corresponding fusion evaluation result when the termination condition is met.

[0005] Furthermore, the low-carbon target parameter set includes an operational carbon emission target threshold, a material-implied carbon target threshold, and a comprehensive carbon emission target threshold. The comprehensive carbon emission target threshold is a weighted combination threshold or a summation combination threshold of the operational carbon emission target threshold and the material-implied carbon target threshold. The target threshold includes the operational carbon emission target threshold, the material-implied carbon target threshold, and the comprehensive carbon emission target threshold.

[0006] Furthermore, both the first and second multiphysics performance index sets include energy intensity index, operational carbon emission index, material implicit carbon index, lighting performance index, ventilation performance index, and thermal performance index; wherein, the material implicit carbon index is calculated from the material usage and the material carbon emission factor, and the material usage is determined by the geometric parameters and the enclosure structure parameters.

[0007] Furthermore, the structured encoding includes: standardizing or normalizing continuous parameters; performing one-hot encoding or embedding encoding on discrete parameters; and concatenating the processed parameters into a fixed-length structured feature vector according to a preset field order.

[0008] Furthermore, the compliance detection includes detection of geometric boundary constraints, material performance boundary constraints, equipment capacity boundary constraints, and specification threshold constraints; the constraint repair includes at least one of projection repair, truncation repair, and item-by-item repair based on constraint priority, such that the repaired feasible feature vector is located within the feasible domain corresponding to the constraint set.

[0009] Furthermore, the lightweight simulation engine uses a reduced-order model or a combination of pre-calculation lookup table and interpolation to calculate the first multiphysics performance index set, so that the calculation of the first multiphysics performance index set meets the preset response time limit.

[0010] Furthermore, the pre-trained proxy model outputs point estimates and uncertainty estimates of the second multiphysics performance index set; the comprehensive scoring rule includes a risk penalty term, which is determined based on the constraint violation probability determined by the uncertainty estimate.

[0011] Furthermore, the fusion calibration includes residual correction: obtaining the residual between the first multiphysics performance index set and the second multiphysics performance index set, constructing an error correction term based on historical residual statistics or residual recursion in the optimization history library, and applying the error correction term to the second multiphysics performance index set to obtain the fusion evaluation result.

[0012] Furthermore, the process of determining the dynamic weights includes: The carbon emission-related gap component in the performance gap vector is compared with a preset gap threshold. If the carbon emission-related gap component is less than the preset gap threshold, the dynamic weight of the carbon emission-related indicators remains unchanged. If the carbon emission-related gap component is greater than or equal to a preset gap threshold, then a gap ratio is determined based on the ratio of the carbon emission-related gap component to the preset gap threshold, and the dynamic weight of the carbon emission-related indicators is increased based on the gap ratio, and the increase in the dynamic weight of the carbon emission-related indicators is proportional to the gap ratio. The distance of the non-carbon performance constraint boundary corresponding to the coupled measurement of the index is compared with the preset safety margin; If the distance to the non-carbon performance constraint boundary is greater than or equal to the preset safety margin, the dynamic weight of the non-carbon performance index remains unchanged. If the distance to the non-carbon performance constraint boundary is less than a preset safety margin, a margin ratio is determined based on the ratio of the preset safety margin to the distance to the non-carbon performance constraint boundary. The dynamic weight of the non-carbon performance index is increased based on the margin ratio, and the increase in the dynamic weight of the non-carbon performance index is proportional to the margin ratio. Alternatively, the non-carbon performance index can be incorporated into the constraint set as a hard constraint.

[0013] Furthermore, generating the candidate design adjustment set includes: The sensitivity ranking of candidate parameter perturbations is calculated based on the index coupling measurement, and the sensitivity values ​​in the sensitivity ranking are compared with a preset sensitivity threshold. If the sensitivity value is less than the preset sensitivity threshold, a single-parameter candidate design adjustment is generated and the candidate design adjustment set is formed. If the sensitivity value is greater than or equal to the preset sensitivity threshold, then a multi-parameter linkage candidate design adjustment is generated and the candidate design adjustment set is formed. Establish a similarity index and cache table for the structured feature vectors, and associate and store the structured feature vectors in the cache table with the corresponding first multiphysics performance index set and the corresponding fusion evaluation results; The similarity between the newly generated structured feature vector and the structured feature vector in the cache table is calculated, and the similarity is compared with the preset similarity threshold. If the similarity is greater than or equal to the preset similarity threshold, the fusion evaluation result of the corresponding cached entry is reused, or the first multiphysics performance index set of the corresponding cached entry is reused. If the similarity is less than the preset similarity threshold, the lightweight simulation engine is invoked to recalculate the first multiphysics performance index set and perform the fusion calibration; wherein, the initial value of the solution of the lightweight simulation engine is determined based on the first multiphysics performance index set corresponding to the structured feature vector with the highest similarity in the cache table.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This method effectively addresses the issues of feedback lag, data fragmentation, and difficulty in embedding low-carbon goals in traditional building design. By real-time evaluation of building performance, integration of multi-physics data, dynamic adjustment of design weights, and generation of optimization suggestions based on real-time feedback, it ensures that building design remains consistent with low-carbon objectives. Furthermore, the method improves the efficiency and accuracy of design optimization through closed-loop iteration, historical optimization data updates, and caching acceleration techniques, thereby significantly enhancing low-carbon performance in the building design process and achieving precise attainment of low-carbon goals. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a flowchart illustrating a method for real-time optimization of building performance aimed at achieving low-carbon goals, as provided in an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] 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.

[0019] In traditional building performance optimization processes, energy efficiency simulations and carbon emission calculations are typically performed in the later stages of design, resulting in a lack of timely performance evaluation for design decisions. Specifically, multiphysics simulation tools operate independently with a lack of data interaction, making a comprehensive assessment of building performance impossible. Furthermore, low-carbon goals are not embedded in the design logic, leaving a lack of real-time feedback loop between design and performance objectives. Consequently, the low-carbon performance of building schemes is difficult to optimize in the early design phase, hindering the achievement of design goals.

[0020] For example, in a commercial complex design project, the design team failed to conduct real-time energy efficiency simulations after determining the building's form during the conceptual phase. When the design entered the detailed design phase, analyses of daylighting, ventilation, and thermal performance using independent software revealed that several indicators failed to meet standards, and the data output from different software programs could not be effectively integrated. Designers had to repeatedly adjust the design and resubmit simulations, leading to extended design cycles, and the final design still failed to meet the preset carbon emission targets. In this scenario, parameter adjustments lacked systematic guidance, and optimization suggestions relied on manual experience, failing to dynamically respond to changes in performance gaps.

[0021] If the above problems are not addressed, the lag in building performance assessment and the fragmentation of data will prevent design schemes from dynamically adapting to low-carbon requirements. Design decisions based on incomplete information may lead to excessive carbon emissions throughout the building's life cycle, affecting the achievement of industry emission reduction targets. In particular, the fragmentation of multiphysics data further weakens design accuracy, reduces the efficiency of implementing optimization measures, and continuously widens the gap between building performance and low-carbon goals.

[0022] See Figure 1 As shown, this embodiment of the invention provides a method for real-time optimization of building performance aimed at low-carbon goals, including: S1: Dataset establishment, loading building performance database, material property database, material carbon emission factor database, location climate database and low-carbon design constraint library, initializing lightweight simulation engine and pre-trained proxy model, and establishing benchmark scheme parameter set and low-carbon target parameter set; S2: Building parameter input, obtain the building parameter set of the building scheme to be optimized. The building parameter set includes geometric parameters, building envelope parameters, equipment and operation parameters, and location and climate parameters. S3: Parameter preprocessing: Unit unification and structured encoding of the building parameter set are performed to obtain structured feature vectors. Based on the low-carbon design constraint library, compliance detection is performed on the structured feature vectors. When non-compliant items are detected, constraint repair is performed on the non-compliant items to obtain feasible feature vectors. S4: Multiphysics performance prediction. Input the feasible feature vector into the lightweight simulation engine to obtain the first multiphysics performance index set, and input the feasible feature vector into the pre-trained surrogate model to obtain the second multiphysics performance index set. S5: Fusion evaluation, the first multiphysics performance index set and the second multiphysics performance index set are fused and calibrated to obtain the fusion evaluation result, and the fusion evaluation result is compared with the benchmark index and the target threshold respectively to obtain the performance gap vector and index coupling metric. S6: Target construction, based on building type identifier, design stage identifier, performance gap vector and index coupling metric, determine dynamic weights, and construct a multi-objective comprehensive objective function and determine the constraint set based on the dynamic weights; S7: Candidate design adjustment generation. Under this constraint set, a candidate design adjustment set is generated, and the expected performance improvement and constraint impact are calculated for each candidate design adjustment in the candidate design adjustment set. S8: Optimization suggestion output. Based on the comprehensive scoring rules, the priority score of each candidate design is adjusted, the ranking is determined, and optimization suggestions are output. These optimization suggestions include parameter adjustment amount, expected performance improvement amount, and constraint impact description. S9: Closed-loop iteration: Receive the selection signal for the optimization suggestion, update the building parameter set based on the selection signal to obtain the updated building parameter set, repeat the parameter preprocessing based on the updated building parameter set until the optimization suggestion is output, and record the parameter change trajectory and the fusion evaluation result to form an optimization history library; S10: Model update and termination determination. Based on the optimization history library, the pre-trained agent model is incrementally updated, and when the termination condition is met, the optimized building scheme and the corresponding fusion evaluation result are output.

[0023] For ease of understanding, the following explains some key terms in this embodiment: The building performance database stores performance data of historical building projects, such as energy consumption, carbon emissions, lighting, and ventilation. This data can be used as a benchmark for comparison or for model training.

[0024] The material property database is used to store detailed information on the physical, thermal, and optical properties of various building materials, such as thermal conductivity, density, specific heat capacity, and solar heat gain coefficient.

[0025] The Materials Carbon Emission Factor Database stores the carbon emission coefficients of different building materials throughout their life cycle, including production, transportation, construction, and demolition, in order to calculate the implicit carbon emissions of building materials.

[0026] Location-based climate databases are used to store climate data for specific geographical locations, including but not limited to temperature, humidity, solar radiation, wind speed, and rainfall. These data are important inputs for building performance simulation.

[0027] The low-carbon design constraint library stores specifications, standards, regulations, and user-defined constraints related to low-carbon building design, such as building energy consumption limits, carbon emission intensity requirements, and material usage restrictions.

[0028] The baseline parameter set represents the set of parameters for the current design or industry standard solution, used to compare the performance with the solution to be optimized.

[0029] The low-carbon target parameter set defines the desired low-carbon performance targets, such as the operating carbon emission target threshold, the material implicit carbon target threshold, and the comprehensive carbon emission target threshold.

[0030] The building parameter set is used to describe various characteristics of the building scheme to be optimized, including geometric parameters (such as the length, width, height, and number of floors of the building), envelope parameters (such as wall materials, window-to-wall ratio, and roof insulation thickness), equipment and operation parameters (such as air conditioning system type, lighting power density, and operating schedule), and location and climate parameters (such as the latitude and longitude, altitude, and local meteorological data of the project location).

[0031] Structured feature vectors are fixed-length numerical vectors formed by unifying and structuring the set of building parameters, making them easy to input into simulation engines and proxy models for processing.

[0032] A feasible feature vector is a structured feature vector that has undergone compliance testing and constraint repair, and it satisfies all preset design constraints.

[0033] The first multiphysics performance index set is a set of building performance indicators calculated by a lightweight simulation engine, such as energy intensity, operational carbon emissions, material-implied carbon, lighting performance, ventilation performance, and thermal performance.

[0034] The second set of multiphysics performance indicators is a set of building performance indicators predicted by a pre-trained surrogate model, which typically includes point estimates and uncertainty estimates.

[0035] A multi-objective integrated objective function is an optimization objective function constructed by combining multiple performance indicators and dynamic weights, used to guide the optimization process.

[0036] The constraint set consists of various technical, economic, and regulatory restrictions that need to be met during the optimization process, such as building height restrictions, total building area restrictions, and restrictions on the use of specific materials.

[0037] The candidate design adjustment set is a collection of potential improvement measures proposed for the current building scheme, such as adjusting the window-to-wall ratio, replacing insulation materials, and optimizing equipment operation strategies.

[0038] The expected performance improvement is the predicted effect of the candidate design adjustments on building performance, such as the expected reduction in energy consumption and carbon emissions.

[0039] Constraint impact is the predicted potential impact of candidate design adjustments on existing constraints, such as whether they will lead to violations of building height limits or cost budgets.

[0040] The comprehensive scoring rule is a criterion used to evaluate and rank candidate design adjustments. It can take into account factors such as expected performance improvement, constraint impact, and dynamic weights.

[0041] The selection signal is the user's feedback on whether to accept or reject optimization suggestions.

[0042] This embodiment provides a method for real-time optimization of building performance aimed at low-carbon goals. The method first loads multiple databases, including a building performance database, a material property database, a material carbon emission factor database, a location climate database, and a low-carbon design constraint library. These databases provide necessary data support for subsequent performance evaluation and optimization. Simultaneously, a lightweight simulation engine and a pre-trained surrogate model are initialized, and a baseline scheme parameter set and a low-carbon target parameter set are established. The lightweight simulation engine can use simplified physical models for rapid calculations, for example, estimating building energy consumption through a simplified form of the energy balance equation. The pre-trained surrogate model can be built based on a neural network model, quickly outputting predicted values ​​for various performance indicators by inputting building parameters. The baseline scheme parameter set can be manually input or selected from typical parameters in historical projects as a reference. The low-carbon target parameter set can be set according to national or industry standards, for example, setting the annual carbon emissions per unit area to not exceed a certain threshold.

[0043] There are several ways to obtain the set of building parameters for a proposed building design. For example, designers can manually input the building's geometric parameters, envelope parameters, equipment and operational parameters, and location climate parameters into the user interface. Alternatively, the system can interface with Building Information Modeling (BIM) software to automatically extract the required building parameters from the BIM model. Geometric parameters can include basic information such as the building's length, width, height, and number of stories. Envelope parameters can cover wall material types, window-to-wall ratios, and roof insulation thickness. Equipment and operational parameters can involve air conditioning system types, lighting power density, and operating schedules. Location climate parameters can be obtained from a location climate database, such as the project site's latitude, longitude, altitude, and local weather station data.

[0044] It should be noted that the BIM model described in this application can preferably be implemented using existing technologies in the field, and it is not the focus of the improvement claimed in this application. The focus of the improvement in this application lies in its process organization, data connection relationship, and constraint processing logic in low-carbon building design optimization. In some embodiments, the BIM model receives building design parameters as input, outputs building performance characteristic data, and interacts with the multiphysics simulation engine and low-carbon target assessment module through a standardized interface. Specifically, the BIM model can provide data on the building's geometry, envelope, equipment, and operation, which will be converted into structured feature vectors and further passed to the subsequent performance prediction module. The performance prediction module will calculate key performance indicators such as building energy consumption and carbon emissions based on this data, and compare them with target parameters stored in the low-carbon design constraint library to identify performance gaps and provide optimization suggestions.

[0045] Those skilled in the art can adapt, replace, or achieve equivalent implementations based on the input-output relationships, parameter configuration rules, and calling sequence disclosed in this application, combined with existing publicly available technologies or conventional engineering methods, without affecting the implementation of the technical solution of this application. In particular, the data extraction and processing process of the BIM model can be seamlessly integrated with existing building information modeling platforms, achieving efficient connection between building design and performance evaluation. In this way, the present invention can effectively utilize existing building information modeling technology while providing an innovative solution for low-carbon optimization tasks in the building design process.

[0046] In the parameter preprocessing stage, the building parameter set undergoes unit unification and structured encoding to obtain a structured feature vector. Unit unification ensures that all parameters use the same unit of measurement during calculation; for example, all length units are converted to meters, and temperature units are converted to degrees Celsius. Structured encoding converts different types of parameters into a unified numerical format. For example, continuous parameters can be directly used; discrete parameters, such as material types, can be converted into numerical representations using unique thermal encoding or embedded encoding. Subsequently, the structured feature vector is subjected to compliance checks based on the low-carbon design constraint library. Compliance checks verify whether parameters meet preset specifications or limitations; for example, checking whether the window-to-wall ratio exceeds the upper limit of the specification or whether the thermal conductivity of the material is within a reasonable range. When non-compliance items are detected, constraint repair is performed to obtain a feasible feature vector. Constraint repair can employ various strategies; for example, when the window-to-wall ratio exceeds the upper limit, it is adjusted to the upper limit value; when the thermal conductivity of the material is unreasonable, it is replaced with the default value or the most recent compliant value.

[0047] In the multiphysics performance prediction phase, the feasible feature vector is input into a lightweight simulation engine to obtain the first multiphysics performance index set, and then input into a pre-trained surrogate model to obtain the second multiphysics performance index set. The lightweight simulation engine can perform rapid calculations based on simplified physical models; for example, it can use a simplified form of the energy balance equation to estimate building energy consumption to meet preset response time limits. The pre-trained surrogate model can perform predictions based on neural network models; for example, it can input a structured feature vector and output predicted values ​​for indicators such as energy consumption, carbon emissions, daylighting, ventilation, and thermal performance. These two models run in parallel, each generating its own performance index set, thus providing multi-source data for subsequent fusion evaluation.

[0048] In the fusion evaluation phase, the first multiphysics performance index set and the second multiphysics performance index set are fused and calibrated to obtain the fusion evaluation result. Fusion calibration can be performed by merging the two index sets using a weighted average, or by correcting the surrogate model results using statistical methods, such as constructing an error correction term based on historical residual statistics and applying it to the surrogate model results. Subsequently, the fusion evaluation result is compared with the benchmark index and the target threshold, respectively, to obtain the performance gap vector and index coupling metric. Comparing the benchmark index can, for example, compare the energy consumption of the current scheme with the average energy consumption of similar buildings. Comparing the target threshold can, for example, compare the carbon emissions of the current scheme with a preset low-carbon target value. The performance gap vector records the specific numerical differences between each index and the benchmark or target. The index coupling metric can be calculated using sensitivity analysis or correlation analysis to determine the degree of influence of different parameters on different indices.

[0049] In the objective construction phase, dynamic weights are determined based on building type identifiers, design phase identifiers, the performance gap vector, and the coupling metric of the indicator. For example, for office buildings in the conceptual design phase, if the carbon emission gap is large, the dynamic weights of carbon emission-related indicators can be increased. Subsequently, a multi-objective comprehensive objective function and a constraint set are constructed based on these dynamic weights. The multi-objective comprehensive objective function can weight and sum indicators such as energy consumption, carbon emissions, and comfort to form a comprehensive optimization objective. The constraint set may include building height restrictions, total building area restrictions, and restrictions on the use of specific materials to ensure the practicality and compliance of the optimization results.

[0050] During the candidate design adjustment generation phase, a set of candidate design adjustments is generated under the given constraint set. The expected performance improvement and constraint impact are calculated for each candidate design adjustment in this set. The candidate design adjustment set can be generated based on preset parameter adjustment ranges and step sizes; for example, increasing the window-to-wall ratio by 5% or replacing the wall insulation material with another type. The expected performance improvement can be predicted by rerunning the lightweight simulation engine or surrogate model. The constraint impact assesses whether the adjustment will violate existing constraints; for example, will increasing the window-to-wall ratio exceed specifications or lead to cost overruns?

[0051] In the optimization suggestion output phase, priority scores are calculated and ranked for each candidate design adjustment based on the comprehensive scoring rules, and optimization suggestions are output. The comprehensive scoring rules can use weighted scoring based on expected performance improvement, constraint impact, and dynamic weights. Priority scores are calculated for each candidate adjustment according to the comprehensive scoring rules. Ranking arranges candidate adjustments from highest to lowest priority score. Optimization suggestions are presented in a clear format, for example, "It is recommended to adjust the window-to-wall ratio from 0.3 to 0.35, with an expected energy consumption reduction of 5% and no new constraint risks." This optimization suggestion includes the parameter adjustment amount, expected performance improvement, and constraint impact description, providing clear guidance for designers.

[0052] In the closed-loop iteration phase, a selection signal for the optimization suggestion is received. The user can manually choose to accept a specific optimization suggestion or reject all suggestions. Based on this selection signal, the building parameter set is updated. According to the user's selection, the corresponding parameter adjustments are applied to the current building parameter set. Subsequently, based on the updated building parameter set, the parameter preprocessing is repeated until the optimization suggestion is output, forming a closed-loop optimization process. Simultaneously, the parameter change trajectory and the fusion evaluation results are recorded to form an optimization history database for tracing the optimization process and updating the model.

[0053] During the model update and termination phase, the pre-trained proxy model is incrementally updated based on the optimization history database. New data from the database is used to train the model in small batches, improving its accuracy in specific projects or design phases. Upon meeting the termination conditions, the optimized building scheme and corresponding fusion evaluation results are output. Termination conditions can be set to reaching a preset low-carbon target threshold, reaching the maximum number of iterations, or insufficient optimization effect after multiple consecutive rounds. When the termination conditions are met, the system presents the final building scheme parameters and corresponding fusion evaluation results to the user.

[0054] Specifically, the lightweight simulation engine described in this application comprises the following hierarchical structure: (1) Engine scheduling layer: responsible for receiving structured feasible feature vectors, performing parameter distribution, multi-physics solution timing scheduling, computing resource allocation and result summarization; preset interactive design scenario response time limit ≤2s, and using thread pool to schedule the solution tasks of each physics field in parallel to ensure that multi-physics calculations are completed synchronously, solving the problems of data fragmentation and timing asynchrony in traditional tools. (2) Multi-physics reduced-order solution layer: the core of the engine, divided into thermal performance solution module, lighting performance solution module, and ventilation performance solution module. Each module adopts an adapted reduced-order model, solves independently and outputs the corresponding physics field index; the modules achieve data communication through a shared parameter pool, and the calculation results of the thermal module can be synchronously transmitted to the ventilation module to realize the linkage update of indoor and outdoor temperature difference and fully capture the parameter coupling effect. (3) Data interface layer: responsible for standardized data interaction with the upstream parameter preprocessing module and the downstream fusion evaluation module. The input is a feasible feature vector with fixed dimensions, and the output is a standardized first multiphysics performance index set. The data format adopts JSON structured output, which is compatible with the general interfaces of mainstream BIM software and building performance simulation software.

[0055] The core algorithm logic of each physics field solution module is as follows: (1) Thermal performance solution module: A reduced-order model (ROM) based on the response surface methodology is adopted. The core is to pre-generate a sample library covering the design parameter space through full-size simulation using EnergyPlus, and construct a second-order response surface model of the building envelope thermal parameters, meteorological parameters and building heating and cooling loads. The core calculation formula is: ; In the formula, Q load Let β0 be the hourly heating and cooling load of the building, and β be a constant term. i β is the coefficient of the linear term. ij x is the coefficient of the quadratic term. i x j The input parameters are normalized (including the heat transfer coefficient of the building envelope, the window-to-wall ratio, the building shape coefficient, the indoor set temperature, and the outdoor hourly temperature and humidity), and ε is the residual term.

[0056] The annual energy intensity index of the building is directly calculated based on this model, and then the operating carbon emission index is calculated by combining it with the local energy carbon emission factor.

[0057] (2) Lighting performance solution module: The method of pre-calculated radiance lookup table + bilinear interpolation is adopted. The indoor daylight coefficient under different window-to-wall ratio, window-to-floor ratio, glass transmittance, building orientation and sky model is pre-calculated using Radiance software. A parameterized lookup table database with a resolution of 0.05 is constructed. After inputting the geometric shape parameters and building envelope parameters in the feasible feature vector, the three closest pre-calculated samples are matched first. The indoor average daylight coefficient, the daylight compliance area ratio and other daylight performance indicators are calculated by bilinear interpolation.

[0058] (3) Ventilation performance solution module: A simplified network method based on the coupling of wind pressure and thermal pressure is adopted to reduce the order of the model. The building space is simplified into ventilation network nodes. Based on the building opening area, orientation, indoor and outdoor temperature difference, wind speed and wind direction, the indoor average air exchange rate is calculated by the following formula: ; Where: N is the average number of air changes per minute indoors, C i Let A be the flow coefficient of the i-th opening. i Let ΔP be the area of ​​the i-th opening. i Let ρ be the pressure difference between the inside and outside of the i-th opening (wind pressure + thermal pressure), ρ be the air density, and V be the volume of the building's interior space; based on this model, the ventilation performance index is calculated.

[0059] It should be noted that the solution time for the thermal performance module is ≤500ms, the solution time for the lighting performance module is ≤300ms, and the solution time for the ventilation performance module is ≤400ms. (1)-(3) The solution of each module adopts the offline pre-calculation + online fast solution mode. The total solution time for the full physics field of a single scheme is ≤1.5s, which meets the preset 2s interactive design response time limit. For inputs that exceed the range of pre-calculated parameters, the reduced-order model coefficients of the nearest neighbor sample are automatically called as initial values ​​to quickly complete the model correction and ensure the stability of the solution.

[0060] In addition, the source and composition of the training dataset for the pre-trained agent model in this application include: (1) Dataset sources: It is divided into three parts. First, the public dataset, including the ASHRAE global building energy consumption dataset and the China Academy of Building Research low-carbon building performance database, covering 12,000+ buildings of different types and different climate zones, with measured and simulated data. Second, the historical project dataset, including the full life cycle performance data of 3,000+ commercial, office and residential buildings. Third, the parametric simulation generated dataset, which generates 50,000+ sets of building parameter samples covering the entire design parameter space through Latin hypercube sampling, and completes the full physical field simulation through EnergyPlus, Radiance and Fluent to obtain the corresponding performance index data.

[0061] (2) Data set size: Total sample size ≥ 65,000 groups, divided into training set, validation set and test set in a ratio of 8:1:1.

[0062] (3) Dataset preprocessing: The input parameters are standardized / normalized and one-hot encoded in the same way as the original scheme. The output performance index is standardized by Z-score. Outliers exceeding 3 times the standard deviation are removed. The SMOTE algorithm is used to augment the data for small sample categories to ensure the balance of the dataset.

[0063] The model structure adopts a multi-output deep residual neural network (Multi-outputResNet) + Bayesian output layer architecture, which is suitable for multi-physics coupled prediction and uncertainty estimation requirements. The specific structure is as follows: Input layer: The dimension is consistent with the structured feature vector, fixed at 128 dimensions, corresponding to the encoding results of geometric shape parameters, enclosure structure parameters, equipment and operation parameters, and location climate parameters; Feature extraction layer: Contains 6 residual blocks. Each residual block consists of 2 fully connected layers, a batch normalization layer, a ReLU activation function, and a Dropout layer (dropout rate = 0.2). The residual connection solves the gradient vanishing problem in deep networks and extracts the nonlinear coupling features between building parameters and multiphysics performance. Multi-task branch layer: It is divided into two parallel branches. The first is the index prediction branch, which outputs the point estimates of the six core performance indicators. The second is the uncertainty estimation branch, which outputs the prediction variance (uncertainty estimate) of the corresponding six indicators. Output layer: Using a Gaussian distribution activation function, the output is a second set of multiphysics performance indicators, including point estimates and uncertainty estimates of six indicators: energy intensity, operational carbon emissions, material-implied carbon, lighting performance, ventilation performance, and thermal performance, for a total of 12 output dimensions.

[0064] The model training process and incremental update method of this model are as follows: ① Offline pre-training process: Hardware environment: Training can be completed with a single NVIDIA RTX 3090 GPU; Optimizer: AdamW optimizer is used, with an initial learning rate of 1e. -4 Weight decay = 1e -5 The learning rate decreases by 0.5 every 10 epochs; Loss function: Negative log-likelihood loss (NLLLoss) is used to simultaneously fit the predicted values ​​and uncertainties. The core formula is: ; In the formula: y i For the true value, These are the point estimates predicted by the model. The variance of the model predictions; Training process: batch size = 256, maximum training epochs = 200, early stopping mechanism is set to stop training if the validation set loss does not decrease for 20 consecutive epochs, and the final test set determination coefficient R² ≥ 0.95; Model saving: Saves the trained model weights, standardized parameters, and encoding rules for online prediction.

[0065] ② Incremental update method (corresponding to the model update in the closed-loop iteration of the original scheme): Based on the newly added samples in the optimization history database (the set of building parameters for each iteration and the corresponding fusion evaluation results), incremental training is performed on the model every 50 sets of newly added samples; the weights of the first 4 residual blocks are frozen during incremental training, and only the weights of the last 2 residual blocks and the output layer are fine-tuned, with an initial learning rate of 1e. -5 The training epoch is set to 20 to avoid model overfitting and improve the model's predictive adaptability to the current project.

[0066] It should be noted that the dynamic weights of this invention are determined using the following method: Based on the basic weight allocation rules in Table 1, determine the basic weights.

[0067] Building type Design phase Basic weights of carbon emission related indicators Basic weighting of non-carbon performance indicators (lighting / ventilation / thermal performance) office buildings Concept Design 0.6 0.4 (Equal score for all three items) office buildings Preliminary design 0.5 0.5 (Equal score for all three items) office buildings Construction drawing design 0.4 0.6 (average score of all three items) Residential buildings Concept Design 0.5 0.5 (Equal score for all three items) Residential buildings Preliminary design 0.4 0.6 (average score of all three items) Note: Carbon emission-related indicators include operational carbon emissions, material-embodied carbon, and overall carbon emissions, with the basic weights allocated among the three in a 1:1:1 ratio; the basic weights of non-carbon performance indicators can be adjusted according to the building's function.

[0068] (1) Calculation of dynamic weights for carbon emission-related indicators: Step 1: Calculate the carbon emission-related gap component and the gap ratio: ; In the formula: G carbon C represents the relative difference in carbon emissions. current C represents the overall carbon emissions of the current scheme. target To achieve the comprehensive carbon emission target threshold, R gap G represents the difference ratio. threshold The preset gap threshold is assumed to be 0.1 (i.e., 10%). Step 2: Calculate the dynamic weighting amplification factor for carbon emissions: λ carbon =max(1,R gap ); In the formula: λ carbon This is the carbon emission weighting factor, with a minimum value of 1 (i.e., no amplification), and is proportional to the gap ratio. Step 3: Calculate the final dynamic weights of carbon emission-related indicators: W carbon =W carbon_base ×λ carbon ; In the formula: W carbon_base Wcarbon represents the basic weights for carbon emission-related indicators, while Wcarbon represents the final dynamic weights.

[0069] Dynamic weighting calculation of non-carbon performance indicators: Step 1: Calculate the ratio of non-carbon performance constraint boundary distance to margin: ; ; In the formula: D non-carbon P represents the relative distance between the non-carbon performance constraint boundaries. current P represents the current non-carbon performance index value. bound R is the normative constraint boundary value of this indicator. margin M is the margin ratio. threshold To pre-determine the safety margin, we assume a value of 0.15 (i.e., 15%). Step 2: Calculate the dynamic weighting amplification factor for non-carbon performance: λ non-carbon =max(1,R margin In the formula: λ non-carbon This is the non-carbon performance weighting amplification factor, with a minimum value of 1, and is proportional to the margin ratio. Step 3: Hard constraint conversion rules for non-carbon performance indicators when D non-carbon When the value is ≤0.05 (i.e., less than 5% from the constraint boundary), the non-carbon performance index is directly incorporated into the constraint set as a hard constraint, and its specifications must be met during the optimization process; otherwise, a weight amplification calculation is performed: W non-carbon =W non-carbon_base ×λ non-carbon In the formula: W non-carbon_base W serves as the base weight for non-carbon performance indicators. non-carbon This represents the final dynamic weight.

[0070] Step 4: Weight Normalization After the dynamic weights of all indicators are calculated, normalization is performed to ensure that the sum of all weights is 1, thus avoiding weight imbalance.

[0071] 3. The complete mathematical expression of the multi-objective integrated objective function: Based on dynamic weights, a clear and computable objective function is constructed, which can be directly substituted into the calculation by those skilled in the art: ; In the formula: F is a multi-objective integrated objective function, and the optimization objective is to minimize the function value; n represents the total number of performance indicators, which is fixed at 6 items (operational carbon emissions, material-borne carbon, energy intensity, lighting performance, ventilation performance, and thermal performance). W i represents the normalized dynamic weight of the i-th indicator; X i This represents the current fusion evaluation result value of the i-th indicator; X i ,target is the target threshold of the i-th indicator; Note: For indicators such as lighting and ventilation, where a larger value is better, take their reciprocal and substitute it into the formula to ensure that the minimization direction of the objective function is consistent with the optimization direction.

[0072] The following example will provide a more detailed explanation of the above technical solution: Suppose User A is designing a new office building at location A, aiming for ultra-low energy consumption and near-zero carbon emissions. In traditional design processes, designers typically perform energy consumption simulations and carbon emission calculations only after the preliminary design is completed, resulting in delayed feedback and difficulty in effective optimization at the early stages of the design process. Furthermore, performance assessments for lighting, ventilation, and thermal performance often rely on different software, making data integration difficult and lacking real-time optimization suggestions.

[0073] The method in this embodiment first loads a climate database for location A, a database of properties and carbon emission factors for various building materials, and local low-carbon design specifications as a low-carbon design constraint library. Simultaneously, a lightweight simulation engine and a pre-trained proxy model are initialized, and a parameter set for a benchmark office building and a target parameter set for ultra-low energy consumption and near-zero carbon emissions are defined.

[0074] User A imported a preliminary architectural plan using BIM software. The system automatically acquired the building parameter set for the plan, including geometric parameters (such as volume and number of stories), envelope parameters (such as window-to-wall ratio and wall insulation materials), and equipment and operational parameters (such as air conditioning system type and lighting power density). The system standardized the units and structured the encoding of these parameters, forming a structured feature vector. Subsequently, the system performed compliance checks on this vector based on a low-carbon design constraint library. For example, it checked whether the window-to-wall ratio exceeded the maximum value specified in local regulations, or whether the wall insulation thickness was below the minimum value. If non-compliance was found, such as an excessively large window-to-wall ratio, the system automatically corrected it to the compliance range and generated a feasible feature vector.

[0075] Next, the feasible feature vector is simultaneously input into the lightweight simulation engine and the pre-trained surrogate model. The lightweight simulation engine quickly calculates the first set of multiphysics performance indicators, including energy intensity, operational carbon emissions, material-implied carbon, daylighting performance, ventilation performance, and thermal performance. The pre-trained surrogate model also predicts the second set of multiphysics performance indicators almost simultaneously and provides an estimate of the prediction uncertainty. The system fuses and calibrates these two sets of indicators; for example, it uses residual correction methods to correct the surrogate model's prediction results using historical residual statistics, resulting in a more accurate fusion evaluation. This fusion evaluation shows that the current scheme's energy consumption and carbon emissions are significantly higher than the baseline scheme and the low-carbon target threshold, while daylighting performance is also insufficient. The system thus generates a performance gap vector and indicator coupling metrics; for example, it finds that the window-to-wall ratio has a strong coupling effect on energy consumption and daylighting performance.

[0076] Based on the office building type, preliminary design stage, significant carbon emission gap, and the coupled influence of multiple indicators such as window-to-wall comparison, the system dynamically adjusted the weights of each performance indicator. For example, the weights of carbon emissions and daylighting performance were increased, and a multi-objective comprehensive objective function was constructed with the main objectives of reducing energy consumption and carbon emissions and improving daylighting performance. At the same time, the set of constraints such as building height and total building area were determined.

[0077] The system generated multiple candidate design adjustments under this set of constraints, such as "adjusting the window-to-wall ratio from 0.4 to 0.3", "replacing the wall insulation material with a higher-performance material", and "optimizing the lighting system control strategy". For each candidate adjustment, the system calculated its expected performance improvement (e.g., an expected energy consumption reduction of 10%) and constraint impact (e.g., replacing the insulation material may lead to a 5% increase in cost) by calling the lightweight simulation engine or proxy model again.

[0078] The system prioritizes these candidate adjustments based on a comprehensive scoring rule and outputs optimization suggestions. For example, the top-ranked suggestion might be: "Adjust the south-facing window-to-wall ratio from 0.4 to 0.3, which is expected to reduce energy consumption by 8%, reduce carbon emissions by 7%, slightly improve daylighting performance, and introduce no new constraints or risks."

[0079] After receiving the optimization suggestion, User A chose to adopt the suggestion to "adjust the south-facing window-to-wall ratio". Upon receiving the selection signal, the system immediately updated the building parameter set, adjusting the south-facing window-to-wall ratio to 0.3. Subsequently, based on the updated building parameter set, the system repeated the entire process of parameter preprocessing, multiphysics performance prediction, fusion evaluation, target construction, candidate design adjustment generation, and optimization suggestion output. The parameter change trajectory and fusion evaluation results of each iteration were recorded in the optimization history database.

[0080] After multiple rounds of closed-loop iterations, the building scheme's energy consumption and carbon emissions gradually approached the low-carbon target threshold, while its daylighting performance was also significantly improved. During this process, the system incrementally updated the pre-trained surrogate model using an optimization history database, making its predictive capabilities more accurate in specific project contexts. When both energy consumption and carbon emission indicators met the preset low-carbon target threshold, the system determined that optimization was terminated and output the final optimized building scheme parameters and the corresponding fusion evaluation results.

[0081] The method provided in this embodiment has significant technical contributions compared to traditional building performance optimization methods. Traditional methods generally suffer from problems such as lagging energy efficiency simulation and carbon emission calculation, fragmented multiphysics simulation software and data, difficulty in embedding low-carbon targets into design logic, and lack of real-time feedback and optimization suggestions.

[0082] In other embodiments, this application proposes a real-time optimization method for building performance oriented towards low-carbon goals. This method initializes a lightweight simulation engine and a pre-trained proxy model by loading a building performance database, a material property database, a material carbon emission factor database, a location climate database, and a low-carbon design constraint library, and establishes a baseline scheme parameter set and a low-carbon target parameter set. Through steps such as building parameter input, parameter preprocessing, multiphysics performance prediction, fusion evaluation, target construction, candidate design adjustment generation, optimization suggestion output, closed-loop iteration, and model update and termination determination, the method achieves real-time optimization of building schemes.

[0083] This application further proposes a set of low-carbon target parameters, including an operational carbon emission target threshold, a material-implied carbon target threshold, and a comprehensive carbon emission target threshold. The comprehensive carbon emission target threshold is a weighted combination threshold or a summation combination threshold of the operational carbon emission target threshold and the material-implied carbon target threshold. The target thresholds include the operational carbon emission target threshold, the material-implied carbon target threshold, and the comprehensive carbon emission target threshold.

[0084] The operational carbon emission target threshold refers to the upper limit set for carbon emissions generated during the building's operational phase. It measures the building's environmental impact during its use and can be set based on national or regional energy consumption standards, industry best practices, or carbon neutrality commitments for specific projects. Alternatively, it can be determined through analysis of historical building operational data combined with future energy structure forecasts. The material-implied carbon emission target threshold refers to the upper limit set for carbon emissions generated throughout the entire lifecycle of building materials, including production, transportation, construction, maintenance, and demolition. It measures the environmental impact of the building materials themselves. It can be set based on data from a materials carbon emission factor database, combined with estimates of building material usage, and reference to green building certification standards or low-carbon certification requirements for specific materials. Alternatively, it can be set through carbon footprint analysis of different material supply chains to establish challenging emission reduction targets. The comprehensive carbon emission target threshold refers to the upper limit of total carbon emissions set after comprehensively considering operational carbon emissions and the carbon contained in materials. It provides a more comprehensive perspective on low-carbon goals, avoiding the "carbon transfer" problem that may result from focusing solely on a single carbon emission source. This ensures the low-carbon performance of buildings throughout their entire life cycle. The threshold can be obtained by simply summing the operational carbon emission target threshold and the carbon contained in materials target threshold, or by weighting the two based on project characteristics, design phase focus, or policy guidance. The weighted combination threshold or summation combination threshold refers to the specific calculation method of the comprehensive carbon emission target threshold. It can be a simple addition of the operational carbon emission target threshold and the carbon contained in materials target threshold, or a weighted summation based on preset weights, providing flexibility and allowing adjustment of the importance of different carbon emission sources in the overall target according to actual needs and optimization strategies. The target threshold is a set of preset performance standards used in the fusion evaluation phase to compare with the fusion evaluation results. It serves as a benchmark for measuring the performance of building schemes, guiding the optimization direction, and determining whether the current scheme meets low-carbon requirements. It can be manually entered by the user when the project starts, or the system can automatically load recommended values ​​from the preset low-carbon design constraint library based on information such as building type, geographical location, and design stage.

[0085] As a specific implementation method, during the initialization phase, the low-carbon target parameter set can be set as follows: the operational carbon emission target threshold is 15 kg CO2 equivalent per square meter per year, and the material-implied carbon target threshold is 50 kg CO2 equivalent per square meter. In this case, the comprehensive carbon emission target threshold can be calculated using a summation method, resulting in 65 kg CO2 equivalent per square meter per year. Alternatively, considering the emphasis on material selection in the early stages of the project, a weighted combination method can be used. For example, assigning a weight of 0.6 to the material-implied carbon target threshold and a weight of 0.4 to the operational carbon emission target threshold, the comprehensive carbon emission target threshold is calculated as (15 * 0.4) + (50 * 0.6) = 6 + 30 = 36 kg CO2 equivalent. In subsequent fusion evaluation steps, the system will compare the actual operational carbon emissions, material-implied carbon, and comprehensive carbon emissions of the current building scheme with the aforementioned set operational carbon emission target thresholds, material-implied carbon target thresholds, and comprehensive carbon emission target thresholds to quantify the performance gaps.

[0086] In other implementations, multiphysics performance prediction is proposed using lightweight simulation engines and pre-trained proxy models. However, if the specific performance indicators to be predicted are not clearly defined during the implementation process, the evaluation results may not be comprehensive or accurate enough. Especially under the goal of low carbon, how to comprehensively and accurately quantify the various performance indicators of buildings and ensure that these indicators can effectively support subsequent integrated evaluation and optimization decisions is the key to the effectiveness of the method.

[0087] In this regard, this application further proposes that both the first multiphysics performance index set and the second multiphysics performance index set include energy intensity index, operational carbon emission index, material implicit carbon index, lighting performance index, ventilation performance index, and thermal performance index; wherein, the material implicit carbon index is calculated from the material usage and the material carbon emission factor, and the material usage is determined from the geometric parameters and the enclosure structure parameters.

[0088] Among these, the calculation method of the implicit carbon index of materials is particularly crucial. This implicit carbon index quantifies the carbon emissions generated by building materials throughout their entire life cycle, from production, transportation, construction, and maintenance to demolition and recycling. Its concept lies in assessing the impact of building material selection on the overall carbon footprint. Possible implementation methods include: consulting a material carbon emission factor database and cumulatively calculating the amount of material used in each building component; or using Life Cycle Assessment (LCA) software, inputting a list of building materials and relevant parameters for detailed analysis. By directly linking the calculation of the implicit carbon index of materials to the building's geometric parameters and envelope parameters, it achieves a comprehensive assessment from the initial selection of building form and structure in the design phase, to the accurate estimation of material usage in the later stages, and finally to the quantification of carbon emissions. Specifically, geometric parameters refer to parameters related to the building's geometry, such as its external outline, volume, orientation, and window-to-wall ratio, which affect the building's lighting, ventilation, thermal performance, and material usage. Possible parameters include: building length, width, height, number of stories, facade window area, and roof type. Envelope parameters refer to the material composition, thickness, and thermal performance of building envelope components such as exterior walls, roof, ground, doors, and windows. Their role is to influence the building's thermal insulation performance, material usage, and operational energy consumption. Possible parameters include: wall material type, insulation layer thickness, window glass type, and frame material. These two parameters together determine the material usage of each part of the building. For example, the building's volume and shape determine the area of ​​the exterior walls and roof, while the envelope parameters determine the specific types and thicknesses of materials used in these areas, thus affecting the overall material usage. Once the material usage is determined, it refers to the actual amount of various materials used in the building, such as the volume of concrete, the mass of steel, and the area of ​​glass. Its function is to serve as the basis for calculating the implicit carbon content of materials. Possible methods of determination include: directly extracting the volume and area information of each component from the building design model (such as a BIM model) and calculating it in conjunction with material density or thickness; or statistically analyzing it through a bill of quantities based on architectural drawings and construction specifications. Combining material carbon emission factors, which refer to the carbon emissions generated per unit mass or unit volume of a material throughout its entire life cycle, quantifies the material's contribution to carbon emissions. Possible methods of obtaining this information include consulting national or industry-published standard databases, such as the "Standard for Carbon Emission Accounting of Building Materials," or obtaining Product Environmental Declarations (EPD) data from suppliers. This allows for accurate calculation of the material's implicit carbon index. This clear chain from source parameters to final carbon emission indicators enables more effective identification and adjustment of key design elements affecting low-carbon goals during the optimization process. It provides comprehensive and detailed performance evaluation data during the multiphysics performance prediction stage, laying a solid foundation for subsequent integrated evaluation, target construction, and optimization recommendation output, ensuring the scientific validity and effectiveness of optimization decisions.

[0089] For example, in multiphysics performance prediction, when feasible feature vectors are input into the lightweight simulation engine and the pre-trained surrogate model, both modules output a performance metric set containing six core metrics. For instance, the energy intensity metric can be specifically expressed as the electricity consumption per square meter of building per year (kWh / m²). 2 a) and gas volume (m³) 3 / m 2 •a) This indicator quantifies the energy consumption per unit area or volume of a building during operation. Its concept is to reflect building energy efficiency and serves as the basis for assessing carbon emissions during building operation. Possible implementation methods include: calculating the building's average annual energy consumption under typical climatic conditions through simulation and dividing by the building area or volume; or standardizing historical operational data through statistical analysis, combined with building type and usage function. The operational carbon emission indicator can then be based on this energy consumption data, combined with the local power grid's carbon emission factor (e.g., 0.58 kgCO2 / kWh) and the gas emission factor (e.g., 2.0 kgCO2 / m³). 3 This indicator, calculated to quantify the carbon dioxide emissions generated by a building's energy consumption during operation, directly reflects the impact of building operations on climate change. Possible implementation methods include: calculation based on energy intensity indicators and local power grid or fuel carbon emission factors; or dynamic estimation through real-time monitoring of building energy consumption data combined with relevant carbon emission factors.

[0090] The calculation process for the implicit carbon index of materials is as follows: First, geometric parameters are extracted from the building parameter set, such as the total building area, floor height, exterior wall area, and roof area, as well as envelope parameters, such as the type and thickness of the insulation material for the exterior walls (e.g., extruded polystyrene board), the type of window glass (e.g., double-glazed Low-E glass), and the frame material (e.g., thermally broken aluminum alloy). Based on these parameters, the material usage of each part of the building, including the main structure, envelope, and interior decoration, can be accurately estimated, such as the cubic meters of concrete, the tons of steel reinforcement, the square meters of insulation board, and the square meters of glass. Then, by querying the material carbon emission factor database, the carbon emission factor corresponding to each unit usage of each material is obtained, such as the implicit carbon emissions per cubic meter of concrete and per ton of steel reinforcement. Multiplying the usage of each material by its corresponding carbon emission factor and summing the results yields the implicit carbon index of the entire building.

[0091] In addition, daylighting performance indicators can be specifically expressed as the average and minimum values ​​of the solar radiation coefficient or daylight factor of the main functional spaces indoors. This indicator is used to quantify the sufficiency and uniformity of natural lighting in the building's indoor spaces, and its role is to assess the building's efficiency in utilizing natural light, affecting living comfort and operational energy consumption. Possible implementation methods include: simulating and calculating parameters such as indoor average illuminance, uniformity, and glare probability; or evaluating using traditional indicators such as solar radiation coefficient and daylight factor. Ventilation performance indicators can be expressed as the average number of air changes or air age under typical summer conditions. This indicator is used to quantify the efficiency and quality of indoor air circulation in buildings, and its role is to assess the impact of the building's natural or mechanical ventilation systems on the indoor air environment, affecting living comfort and health. Possible implementation methods include: calculating parameters such as air change rate, air age, and pollutant concentration distribution; or evaluating by simulating airflow organization and wind pressure distribution. Thermal performance indicators can be expressed as the average heat transfer coefficient (U-value) of the building envelope, such as exterior walls and roofs, or the range of indoor temperature fluctuations. These indicators quantify the building envelope's ability to impede heat transfer between the interior and exterior, thus assessing the building's thermal insulation performance and directly impacting energy consumption and indoor thermal comfort. Possible implementation methods include calculating parameters such as the heat transfer coefficient (U-value), thermal inertia index, and condensation risk; or simulating the range of indoor temperature fluctuations under different seasonal and climatic conditions. This approach ensures a comprehensive quantification and evaluation of building performance.

[0092] In other implementations, structured coding is proposed, including: standardizing or normalizing continuous parameters; performing one-hot coding or embedding coding on discrete parameters; and concatenating the processed parameters into a fixed-length structured feature vector according to a preset field order.

[0093] Structured coding refers to converting raw, heterogeneous building parameters into a unified, machine-readable, fixed-format numerical representation. Its purpose is to eliminate differences in parameter types and dimensions, enabling these parameters to be efficiently processed and analyzed by lightweight simulation engines and pre-trained proxy models. Structured coding ensures the consistency and standardization of input data, thereby improving the accuracy and stability of subsequent calculations. Continuous parameters are those that can take any value within a certain range, such as building area, window-to-wall ratio, and U-value. Standardization typically transforms data into a distribution with a mean of 0 and a standard deviation of 1, for example, by subtracting the mean and dividing by the standard deviation. This method eliminates the influence of dimensions, making features at different scales comparable and helping to accelerate model convergence. Normalization typically scales data to a fixed interval, such as [0, 1] or [-1, 1], for example, by scaling using maximum and minimum values. This method avoids certain features being too large or too small, thus allowing them to dominate model training and helping to maintain the model's sensitivity to all features. Discrete parameters are parameters that can only take a finite number of specific values, such as building type, material type, and equipment model. One-hot encoding converts discrete features into binary vectors. For a discrete feature with N possible values, one-hot encoding creates an N-dimensional binary vector where only one position is 1 and the rest are 0, representing the current value of the feature. This method avoids the model misinterpreting the order relationship between discrete features and effectively handles categorical data. Embedding encoding maps discrete features to a low-dimensional continuous vector space. This approach is typically used in scenarios with a large number of discrete categories or where it's necessary to capture semantic relationships between categories, such as learning the embedding vector for each category through a neural network. Embedding encoding can represent discrete features more compactly and may capture similarities between categories. Concatenation combines all parameters processed by standardization / normalization or one-hot / embedding encoding into a single, fixed-length vector in a predefined order. Predefined field order ensures that parameters of the same type are always in the same position in the vector with each input, thus guaranteeing the consistency of the input data. Fixed-length structured feature vectors are a common input requirement for many machine learning models. They enable models to handle parameters of different architectural schemes in a uniform way, simplifying model design and training processes.

[0094] In other embodiments, this application proposes a real-time building performance optimization method for low-carbon goals. In the parameter preprocessing stage, this method unifies the units and structures the building parameter set to obtain structured feature vectors. However, these original building parameters may not conform to various constraints, specifications, or physical boundaries in actual engineering design. Directly using non-compliant feature vectors for subsequent multiphysics performance prediction and optimization may lead to distorted prediction results, incorrect optimization directions, or even the generation of building schemes that are impractical, thus affecting the effectiveness and reliability of the optimization method.

[0095] In response, this application further proposes parameter preprocessing steps including: compliance detection including detection of geometric boundary constraints, material performance boundary constraints, equipment capacity boundary constraints and specification threshold constraints; constraint repair including at least one of projection repair, truncation repair and item-by-item repair based on constraint priority, so that the repaired feasible feature vector is located in the feasible domain corresponding to the constraint set.

[0096] Compliance testing refers to checking the parameters in a structured feature vector to ensure they conform to pre-defined rules, standards, or restrictions. Its function is to identify parameter values ​​that do not comply with design specifications, physical limitations, or engineering realities. Compliance testing can be achieved in various ways, such as checking each parameter item by using pre-defined logical judgment rules, or comparing parameter values ​​with legal ranges stored in a database. Geometric boundary constraints are restrictions imposed on the geometric parameters of a building design. These constraints typically stem from building design codes, structural stability requirements, or site conditions, such as building height, volume proportions, and window-to-wall ratios. Material performance boundary constraints are restrictions imposed on the material properties involved in the building envelope parameters. These constraints are typically based on the physical properties, durability, fire resistance rating, or environmental requirements of materials, such as the thermal conductivity range of insulation materials and the heat transfer coefficient of glass. Equipment capacity boundary constraints are restrictions imposed on the equipment selection and operating parameters involved in building equipment and operating parameters. These constraints typically stem from the rated power, efficiency range, operating mode, or system integration requirements of equipment, such as the cooling / heating range of air conditioning systems and the airflow of ventilation equipment. Normative threshold constraints refer to mandatory or recommended restrictions on building performance or parameters stipulated by national, industry, or local standards, such as the upper limit of the heat transfer coefficient of the building envelope specified in building energy-saving design standards.

[0097] Constraint remediation refers to the process of adjusting non-compliant parameter values ​​to meet all constraints after non-compliance checks are detected. Its purpose is to ensure that the generated feature vectors are valid and feasible, providing reliable input for subsequent performance prediction and optimization. Constraint remediation can be achieved through various strategies, such as directly adjusting out-of-range parameter values ​​to boundary values, or adjusting them in a coordinated manner based on dependencies between parameters. Projection remediation is a method that maps non-compliant parameter values ​​to the nearest compliant region boundary. For example, if a parameter value exceeds its maximum allowed value, it is directly set to that maximum value. Truncation remediation typically refers to directly truncating parameter values ​​exceeding a preset range to the boundary value of that range. For example, when a parameter value is below the minimum value, it is set to the minimum value; when it is above the maximum value, it is set to the maximum value. Constraint priority-based item-by-item remediation refers to remediating parameters sequentially according to a preset priority order when multiple conflicting or mutually influential constraints exist. For example, mandatory specification constraints are remediated first, followed by recommended design constraints, to ensure that key constraints are met. The feasible region refers to the multi-dimensional space formed by all parameter values ​​under all constraints. Any eigenvector located within the feasible region represents a compliant and feasible building solution.

[0098] In other embodiments, this application proposes a real-time optimization method for building performance oriented towards low-carbon goals. This method includes: loading a building performance database, a material property database, a material carbon emission factor database, a location climate database, and a low-carbon design constraint library; initializing a lightweight simulation engine and a pre-trained proxy model; and establishing a baseline scheme parameter set and a low-carbon target parameter set. Building parameter input involves obtaining the building parameter set of the building scheme to be optimized, including geometric parameters, envelope parameters, equipment and operation parameters, and location climate parameters. Parameter preprocessing involves unifying units and structurally encoding the building parameter set to obtain a structured feature vector. Based on the low-carbon design constraint library, compliance checks are performed on the structured feature vector. When non-compliant items are detected, constraint repair is performed on the non-compliant items to obtain a feasible feature vector. Multiphysics performance prediction involves inputting the feasible feature vector into the lightweight simulation engine to obtain a first multiphysics performance index set, and inputting the feasible feature vector into the pre-trained proxy model to obtain a second multiphysics performance index set. Fusion evaluation involves fusing and calibrating the first and second multiphysics performance index sets to obtain a fusion evaluation result. The fusion evaluation results are compared with benchmark indicators and target thresholds to obtain performance gap vectors and indicator coupling measures. Target construction involves determining dynamic weights based on building type identifiers, design stage identifiers, performance gap vectors, and indicator coupling measures, and then constructing a multi-objective comprehensive objective function and determining the constraint set based on these dynamic weights. Candidate design adjustment generation involves generating a candidate design adjustment set under the constraint set, and calculating the expected performance improvement and constraint impact for each candidate design adjustment. Optimization suggestion output involves calculating and ranking the priority scores of each candidate design adjustment based on comprehensive scoring rules, and outputting optimization suggestions, including parameter adjustment amounts, expected performance improvement amounts, and constraint impact descriptions. Closed-loop iteration involves receiving selection signals for optimization suggestions, updating the building parameter set based on these signals, and repeating parameter preprocessing until optimization suggestions are output, recording parameter change trajectories and fusion evaluation results to form an optimization history database. Model update and termination determination involve incrementally updating the pre-trained surrogate model based on the optimization history database, and outputting the optimized building scheme and corresponding fusion evaluation results when termination conditions are met.

[0099] In some of the embodiments described above in this application, a lightweight simulation engine is proposed to use a reduced-order model or a combination of pre-calculation lookup table and interpolation to calculate the first multiphysics performance index set, so that the calculation of the first multiphysics performance index set meets the preset response time limit.

[0100] Reduced-order models are techniques that reduce the degrees of freedom or dimensionality of complex system models using mathematical methods. They significantly reduce computational load by preserving the system's primary dynamic characteristics while ignoring secondary or high-frequency dynamics. For example, reduced-order models can be constructed using intrinsic orthogonal decomposition (POD), modal analysis, or machine learning-based dimensionality reduction techniques. These models capture key characteristics of building performance while avoiding the computational burden of full-scale simulation models.

[0101] The pre-calculation lookup table combined with interpolation method generates data points covering the design parameter space by running a large number of full-scale simulations in advance, and stores these data points and their corresponding performance indicators in a lookup table. When it is necessary to predict the performance of a specific design scheme, the system does not rerun the simulation. Instead, it queries the lookup table and uses an interpolation algorithm (such as linear interpolation, polynomial interpolation, or radial basis function interpolation) to estimate the target performance indicators based on the neighboring known data points. This method shifts the computationally intensive task to the offline stage, thereby achieving extremely fast response speeds during online prediction.

[0102] The first multiphysics performance index set is a collection of building performance data calculated by the lightweight simulation engine. It characterizes a building's performance across multiple physical dimensions, including energy consumption, carbon emissions, lighting, ventilation, and thermal performance. It serves as a crucial basis for evaluating building designs and is fused and calibrated with the second multiphysics performance index set output by the pre-trained proxy model.

[0103] A preset response time limit refers to the time constraint within which a system must complete performance prediction and provide results within a specific time window after receiving input parameters. This time limit is usually determined by the requirements of the application scenario. For example, interactive design tools may require feedback within seconds; real-time optimization control may require calculations to be completed in milliseconds. Meeting the preset response time limit is crucial for achieving real-time optimization and rapid iteration.

[0104] In this application's approach to real-time building performance optimization for low-carbon goals, a lightweight simulation engine is introduced in the multiphysics performance prediction stage to overcome the limitations of traditional full-physics simulation models, which are computationally time-consuming. This engine does not directly perform a complete physical simulation; instead, it calculates the first set of multiphysics performance indicators using either a reduced-order model or a combination of pre-calculation lookup tables and interpolation. Specifically, upon receiving feasible feature vectors obtained through parameter preprocessing, the lightweight simulation engine quickly processes these inputs. If a reduced-order model is used, it utilizes a pre-built simplified mathematical model to quickly solve or approximate the physical behavior of the building, thereby obtaining various performance indicators. If a combination of pre-calculation lookup tables and interpolation is used, the engine quickly searches through a large amount of pre-stored simulation result data based on the parameters in the feasible feature vectors and calculates the performance indicators of the current scheme using an interpolation algorithm. Both methods aim to significantly reduce the computational load, enabling the calculation of the first set of multiphysics performance indicators to be completed within a preset response time limit. In this way, the method can quickly obtain preliminary performance evaluation of building schemes, providing timely data support for subsequent fusion evaluation, target construction and optimization suggestion generation, thereby significantly improving the real-time performance and efficiency of the entire optimization process.

[0105] In some of the embodiments described above in this application, multiphysics performance prediction is performed using a pre-trained surrogate model. However, the pre-trained surrogate model may have certain errors or uncertainties during the prediction process. If these are not taken into account, the risk of the optimization suggestions may increase, affecting the reliability of the optimization results.

[0106] The pre-trained surrogate model outputs point estimates and uncertainty estimates for the second set of multiphysics performance indicators; the comprehensive scoring rules include a risk penalty term, which is determined based on the constraint violation probability determined by the uncertainty estimate.

[0107] The point estimate refers to the most probable prediction given by the pre-trained surrogate model for the second multiphysics performance index set, usually a single numerical value. This point estimate can be directly output by various regression models, for example, the output value obtained by forward propagating the input data through a trained neural network model. The uncertainty estimate is a quantification of the reliability of the prediction results of the pre-trained surrogate model, reflecting the model's confidence in its prediction results or the prediction range. The uncertainty estimate can be obtained in various ways, for example, by using a Bayesian neural network to simultaneously output the prediction mean (point estimate) and prediction variance (uncertainty estimate) during prediction; or by using ensemble learning methods, such as using Dropout Monte Carlo techniques, to perform multiple forward propagations on the same input and then calculate the statistics (such as standard deviation or confidence interval width) of the multiple prediction results as the uncertainty estimate; or by using a Gaussian process regression model, which naturally outputs the prediction mean and variance.

[0108] The comprehensive scoring rule is a criterion used to calculate the priority score for each candidate design adjustment in the candidate design adjustment set. The purpose of introducing a risk penalty term is to quantify and consider the potential risks of candidate design adjustments in addition to their expected performance improvement when evaluating them, in order to avoid recommending high-risk optimization schemes. The risk penalty term is a component of the comprehensive scoring rule, and its function is to adjust the comprehensive score of candidate design adjustments based on the uncertainty of the prediction. For example, when the prediction result of a candidate design adjustment has high uncertainty, the risk penalty term will be increased accordingly, thereby reducing the priority score of that candidate design adjustment.

[0109] The generation strategy for the candidate design adjustment set can be implemented through the following process: ① Determining the type of parameter adjustment based on sensitivity: Step 1: Calculate the sensitivity value of the parameter perturbation using the local sensitivity analysis method. The core formula is: ; In the formula: S k Let Δx be the sensitivity value of the k-th building parameter, ΔF be the change in the multi-objective integrated objective function, and Δx be the value of the k-th building parameter. k x is a small perturbation of the k-th parameter (recommended to be 1% of the parameter's value range). k This is the current value of the k-th parameter; Step 2: Sensitivity threshold determination, preset sensitivity threshold S threshold =0.03; If S k threshold ​If a parameter is identified as low-sensitivity, a single-parameter candidate design is generated for adjustment. The adjustment step size is 2% to 5% of the parameter value range. Three gradient adjustment schemes (increase, decrease, and optimal step size) are generated for each parameter. If S k ≥S threshold The parameters are identified as high-sensitivity parameters, and multi-parameter linkage candidate design adjustments are generated. Based on the index coupling metric, parameters with strong coupling relationships (Pearson correlation coefficient |r|≥0.7) are grouped into linkage parameter groups. Each linkage parameter group generates 5 collaborative adjustment schemes to ensure that the parameter adjustment conforms to the physical coupling logic (such as adjusting the glass heat transfer coefficient synchronously when the window-to-wall ratio increases to avoid a surge in heat load).

[0110] ② Computation acceleration strategy based on similarity caching: Step 1: Construct a cache table and a similarity index. The Locality Sensitive Hash (LSH) algorithm is used to construct a similarity index for structured feature vectors. The cache table stores the following: structured feature vectors, the corresponding first multiphysics performance index set, fusion evaluation results, and parameter adjustment records. Step 2: Similarity calculation and reuse rules. The cosine similarity is used to calculate the similarity between the newly generated structured feature vector and the vector in the cache table. The preset similarity threshold is Simthreshold=0.99. If the similarity is ≥0.99, the fusion evaluation result of the corresponding cache entry and the first multiphysics performance index set can be directly reused without repeating the simulation calculation. If the similarity is less than 0.99, the lightweight simulation engine is invoked to recalculate, and the first multiphysics performance index set corresponding to the vector with the highest similarity in the cache table is used as the initial value for solving, thereby accelerating the simulation convergence.

[0111] ③ The selection rules for the candidate design adjustment set are to remove candidate solutions with a constraint violation probability > 20% and invalid solutions with an expected performance improvement of < 0.5%. The final size of the candidate design adjustment set is controlled at 10 to 20 sets to ensure optimization efficiency.

[0112] 2. Specific quantitative implementation of the comprehensive scoring rules: The complete formula for calculating priority scores is given, and the ranking rules are clearly defined, so that those skilled in the art can calculate it directly: ① The core formula for calculating priority scores is Score=α×ΔFnorm-β×Risknorm-γ×Constraintnorm; In the formula: The score represents the priority score for adjusting the candidate design; the higher the score, the higher the priority. ΔFnorm is the normalized expected performance improvement, i.e., the decrease in the multi-objective comprehensive objective function after the candidate solution is implemented, normalized to the [0,1] interval; α is the improvement weight, and a value of 0.6 is recommended. Risknorm is the normalized risk penalty term, which calculates the probability of constraint violation based on the uncertainty estimate output by the pre-trained surrogate model and normalizes it to the [0,1] interval; β is the risk weight, and a value of 0.25 is recommended. Constraintnorm is the normalized constraint influence, which is the degree of influence of the candidate solution on the constraint boundary. The closer to the constraint boundary, the larger this value is, and it is normalized to the [0,1] interval; γ is the constraint influence weight, and a value of 0.15 is recommended. ② Specific calculation method for risk penalty items: ; In the formula: P violation To constrain the probability of violation, Φ is the cumulative distribution function of the standard normal distribution, and P... bound For the constraint boundary values ​​of the index, the formula here... σ represents the point estimate predicted by the surrogate model, and σ is the prediction standard deviation (uncertainty estimate). ③ The sorting and output rules sort the candidate design adjustments from high to low priority scores, and take the top 5 to 10 groups as optimization suggestions. Each group of suggestions is clearly marked with: parameter adjustment amount, expected performance improvement amount (carbon emission reduction ratio, energy consumption reduction ratio, etc.), constraint violation probability, and constraint impact description.

[0113] The risk penalty term is determined based on the constraint violation probability, which is based on the uncertainty estimate. The constraint violation probability refers to the likelihood that a set of second multiphysics performance indicators will violate predefined constraints, considering the uncertainty of the pre-trained surrogate model's predictions. This probability can be calculated based on the uncertainty estimate. For example, if the uncertainty estimate is given in the form of standard deviation, and assuming the prediction error follows a normal distribution, the constraint violation probability can be determined by calculating the cumulative probability that the predicted value falls outside the constraint range. Alternatively, the constraint violation probability can be estimated by performing multiple random samplings from the uncertainty distribution using Monte Carlo simulation methods and statistically analyzing the proportion of samples violating the constraints. Or, methods such as quantile regression can be used to directly output the probability of satisfying or violating the constraints. The value of the risk penalty term can be proportional to the constraint violation probability, or it can be mapped through a nonlinear function such that the penalty term increases sharply when the constraint violation probability exceeds a certain threshold.

[0114] As a specific implementation, in the multiphysics performance prediction step, when the feasible feature vector is input into the pre-trained surrogate model, this pre-trained surrogate model can be a trained Bayesian neural network. This Bayesian neural network outputs point estimates of the second set of multiphysics performance indicators (e.g., operating carbon emission indicators) along with the corresponding prediction variance, which is the uncertainty estimate. Assuming that for a candidate design adjustment, the pre-trained surrogate model predicts its operating carbon emission indicator as a point estimate of 45 kgCO2 / m³. 2 / year, the uncertainty estimate (standard deviation) is 5 kg CO2 / m 2 / year. If the low-carbon target is set at a carbon emission target threshold of 50 kgCO2 / m³, the annual target is [not specified]. 2 If the annual carbon emission rate exceeds 50 kgCO2 / m³, then based on the point estimate and uncertainty estimate, and assuming the prediction error follows a normal distribution, the actual operating carbon emission index can be calculated. 2 The probability of constraint violation per year. For example, if this probability is calculated to be 15%. During the optimization recommendation output phase, the comprehensive scoring rule will calculate a risk penalty term based on this 15% constraint violation probability. This risk penalty term will be incorporated into the priority score calculation for the candidate design adjustment, for example, deducted from its initial score calculated based on the expected performance improvement. In this way, even if the point estimate of a candidate design adjustment looks attractive, if its high uncertainty leads to a high probability of constraint violation, its final priority score will be reduced by the risk penalty term, thereby avoiding recommending high-risk optimization solutions.

[0115] In some other embodiments, the fusion calibration of this application includes residual correction: obtaining the residual between the first multiphysics performance index set and the second multiphysics performance index set, constructing an error correction term based on historical residual statistics or residual recursion based on an optimized historical library, and applying the error correction term to the second multiphysics performance index set to obtain the fusion evaluation result.

[0116] Residual correction is a method to correct prediction results by analyzing the differences between different prediction sources. Its core idea is to identify and quantify the systematic bias or random error of one prediction source relative to another, and then construct a correction model to eliminate or reduce these biases, thereby improving the accuracy and reliability of the fusion results. For example, statistical regression methods can be used to model the residuals, or machine learning algorithms can be used to learn the patterns of the residuals. Residuals refer to the difference between two predicted values; in this application, it specifically refers to the difference between the corresponding index values ​​of the first multiphysics performance index set output by the lightweight simulation engine and the second multiphysics performance index set output by the pre-trained surrogate model under the same input conditions. By calculating these residuals, the inconsistency between the two prediction sources can be quantified. For example, the difference between corresponding indices can be calculated directly, or their relative difference can be calculated. The construction of the error correction term is crucial for residual correction. On one hand, it can be constructed based on historical residual statistics. For example, the mean, variance, or distribution characteristics of the residuals between the first and second multiphysics performance index sets over a past period can be statistically analyzed, and a fixed or condition-varying correction factor can be determined accordingly. On the other hand, it can be constructed based on the residual recursion of an optimized historical database. The optimized historical database records the parameter change trajectory and fusion evaluation results for each iteration, containing a large amount of historical residual data. This historical data can be used to dynamically learn and update the residual patterns through recursive algorithms (e.g., Kalman filtering, exponential smoothing), thereby constructing an adaptive error correction term. The error correction term is typically applied by adding or subtracting a correction term from the second multiphysics performance index set to make it closer to the first multiphysics performance index set or more realistic. For example, if the error correction term indicates a systematic underestimation of the second multiphysics performance index set relative to the first multiphysics performance index set, then this correction term can be added to the second multiphysics performance index set. This application method aims to utilize the relative accuracy of the lightweight simulation engine to correct the predictions of the pre-trained surrogate model, thereby obtaining a more reliable fusion evaluation result.

[0117] The following example illustrates how, after obtaining the residuals between the first and second multiphysics performance index sets, an error correction term can be constructed and applied. For instance, for a specific performance index (such as energy intensity), the average residual between the predicted values ​​from the lightweight simulation engine and the pre-trained surrogate model over the past N optimization iterations can be calculated. This average residual serves as the error correction term for that index. In the current iteration, this average residual is directly added to the energy intensity index predicted by the pre-trained surrogate model, resulting in a corrected energy intensity index as part of the fusion evaluation result. Alternatively, a more dynamic approach can be used. For example, a simple linear regression model can be trained using residual data recorded in the optimization history database. This model takes some features of the current building parameter set as input, predicts the current residual value, and applies this predicted residual value as an error correction term to the second multiphysics performance index set. In this way, the correction amount can be adaptively adjusted based on historical data and the current context, further improving the accuracy of the fusion evaluation.

[0118] In some of the aforementioned implementations, a real-time building performance optimization method oriented towards low-carbon goals was proposed. This method obtains a performance gap vector and coupled metrics through fusion evaluation, and determines dynamic weights based on this to construct a multi-objective comprehensive objective function. However, in the actual optimization process, how to intelligently adjust these dynamic weights according to the current performance of the building scheme to ensure that while actively pursuing low-carbon goals, the deterioration or violation of other key performance indicators is effectively avoided is a problem that requires careful handling. Simply adopting fixed weights or insensitive dynamic adjustment strategies may lead to low optimization efficiency or unacceptable compromises in certain performance dimensions.

[0119] To this end, this application further proposes a process for determining the aforementioned dynamic weights, including: comparing the carbon emission-related gap component of the performance gap vector with a preset gap threshold; if the carbon emission-related gap component is less than the preset gap threshold, maintaining the dynamic weights of the carbon emission-related indicators unchanged; if the carbon emission-related gap component is greater than or equal to the preset gap threshold, determining a gap ratio based on the ratio of the carbon emission-related gap component to the preset gap threshold, and increasing the dynamic weights of the carbon emission-related indicators based on the gap ratio, wherein the increase in the dynamic weights of the carbon emission-related indicators is proportional to the gap ratio; and so on. The distance between the non-carbon performance constraint boundary corresponding to the coupled metric is compared with a preset safety margin. If the distance between the non-carbon performance constraint boundary is greater than or equal to the preset safety margin, the dynamic weight of the non-carbon performance index remains unchanged. If the distance between the non-carbon performance constraint boundary is less than the preset safety margin, a margin ratio is determined based on the ratio of the preset safety margin to the distance between the non-carbon performance constraint boundary. The dynamic weight of the non-carbon performance index is increased based on the margin ratio, and the increase in the dynamic weight of the non-carbon performance index is proportional to the margin ratio. Alternatively, the non-carbon performance index may be incorporated into the constraint set as a hard constraint.

[0120] The carbon emission-related gap component of the aforementioned performance gap vector refers to the carbon emission-related part extracted from the performance gap vector obtained by comparing the fusion evaluation results with the target threshold. It quantifies the gap between the current building scheme's carbon emission performance and the preset low-carbon target. Possible implementation methods include treating the gap in total carbon emissions (e.g., operational carbon emissions, carbon contained in materials) as a single component, or considering the gaps in each carbon emission indicator as independent components. The aforementioned preset gap threshold is a pre-set reference value used to determine whether the carbon emission performance gap is significant. When the carbon emission-related gap component exceeds this threshold, it indicates that the current scheme has significant room for improvement or risk in terms of carbon emissions, requiring stronger optimization guidance. Possible implementation methods include setting a fixed value based on industry standards, design phase requirements, or user-defined preferences; or dynamically calculating an adaptive threshold based on historical optimization data and the performance distribution of benchmark schemes. The dynamic weights of the aforementioned carbon emission-related indicators are the weights assigned to the carbon emission-related performance indicators in the multi-objective comprehensive objective function. Its dynamic adjustment aims to flexibly enhance or weaken the influence of carbon emission targets in the optimization process based on the current performance gap. Possible implementation methods include: calculating weight values ​​using linear, exponential, or piecewise functions; or employing adaptive adjustments using methods such as fuzzy logic or reinforcement learning.

[0121] The non-carbon performance constraint boundary distance corresponding to the aforementioned index coupling metric refers to the index coupling metric reflecting the mutual influence relationship between different performance indicators. The non-carbon performance constraint boundary distance indicates how close the current building scheme is to its preset constraint boundary in terms of non-carbon performance (e.g., lighting, ventilation, thermal performance). It is used to assess the risk or potential violation level of non-carbon performance. Possible implementation methods include: calculating the absolute or relative difference between the current performance value and the constraint boundary value; or unifying the distances of different indicators to the same scale through normalization. The aforementioned preset safety margin is a pre-set reference value used to determine whether non-carbon performance is close to the constraint boundary. When the non-carbon performance constraint boundary distance is less than this margin, it indicates that the current scheme has potential risks in non-carbon performance and requires special attention. Possible implementation methods include: setting a fixed percentage or absolute value based on building codes, user comfort requirements, or design experience; or determining a statistically safe range based on historical data analysis. The dynamic weights of the aforementioned non-carbon performance indicators are the weights assigned to the non-carbon performance indicators in the multi-objective comprehensive objective function. Its dynamic adjustment aims to flexibly enhance or weaken the influence of non-carbon performance objectives in the optimization process based on their distance from the constraint boundary. Possible implementation methods include: calculating weight values ​​using linear, exponential, or piecewise functions; or employing fuzzy logic or reinforcement learning for adaptive adjustment. Incorporating the aforementioned non-carbon performance indicators into the constraint set as hard constraints means that these hard constraints must be strictly satisfied during the optimization process; any solution that violates the hard constraints will be considered infeasible. Incorporating non-carbon performance indicators into the constraint set means that these performance indicators will transform from being part of the objective function into strict restrictions that must be met. Possible implementation methods include: significantly reducing the fitness of solutions that violate hard constraints through a penalty function in the optimization algorithm; or employing a feasible region search algorithm to ensure that all generated candidate solutions satisfy the hard constraints.

[0122] In other implementations, generating a candidate design adjustment set includes: calculating the sensitivity ranking of candidate parameter perturbations based on index coupling metrics, and comparing the sensitivity values ​​in the sensitivity ranking with a preset sensitivity threshold; if the sensitivity value is less than the preset sensitivity threshold, generating a single-parameter candidate design adjustment and forming a candidate design adjustment set; if the sensitivity value is greater than or equal to the preset sensitivity threshold, generating a multi-parameter linked candidate design adjustment and forming a candidate design adjustment set; establishing a similarity index and cache table for structured feature vectors, and combining the structured feature vectors in the cache table with the corresponding first multiphysics performance index set and the corresponding fusion... The evaluation results are stored in association; the similarity between the newly generated structured feature vector and the structured feature vector in the cache table is calculated, and the similarity is compared with a preset similarity threshold; if the similarity is greater than or equal to the preset similarity threshold, the fusion evaluation result of the corresponding cache entry is reused, or the first multiphysics performance index set of the corresponding cache entry is reused; if the similarity is less than the preset similarity threshold, the lightweight simulation engine is called to recalculate the first multiphysics performance index set and perform fusion calibration; wherein, the initial value of the solution of the lightweight simulation engine is determined based on the first multiphysics performance index set corresponding to the structured feature vector with the highest similarity in the cache table.

[0123] Specifically, the sensitivity ranking of candidate parameter perturbations based on index coupling measurement, and the comparison of the sensitivity values ​​in the sensitivity ranking with preset sensitivity thresholds, refers to assessing the magnitude and direction of the impact of small changes in one or a group of building parameters on overall performance indicators (such as energy consumption, carbon emissions, daylighting, etc.). This can be achieved through gradient analysis, finite difference method, or sensitivity analysis methods based on surrogate models. Comparing the calculated sensitivity values ​​with preset sensitivity thresholds aims to identify parameters with significant performance impacts, so as to generate targeted design adjustment schemes. If the sensitivity value is less than the preset sensitivity threshold, a single-parameter candidate design adjustment is generated and forms a candidate design adjustment set. When the sensitivity value of a parameter is low, i.e., its change has no significant impact on overall performance, the system can generate candidate design schemes involving only the adjustment of that single parameter. This single-parameter adjustment can be a small increase or decrease in the parameter, for example, fine-tuning the window-to-wall ratio by 0.05, or increasing the insulation material thickness by 5mm. If the sensitivity value is greater than or equal to the preset sensitivity threshold, multi-parameter linked candidate design adjustments are generated and forms a candidate design adjustment set. When the sensitivity value of a parameter or a set of parameters is high, indicating a significant impact on performance, the system can generate candidate design schemes involving the coordinated adjustment of multiple related parameters. For example, when the sensitivity of external wall insulation performance is high, the type and thickness of the external wall material, as well as the heat transfer coefficient of the windows, may be adjusted simultaneously. This multi-parameter coordinated adjustment can be achieved through predefined parameter combination rules, expert knowledge-based coordination strategies, or by exploring the parameter space through optimization algorithms.

[0124] A similarity index and cache table for structured feature vectors are established. The structured feature vectors in the cache table are associated with and stored along with the corresponding first multiphysics performance index set and the corresponding fusion evaluation results. This aims to store calculated design schemes and their corresponding performance evaluation results. The similarity index can employ data structures such as Kd-trees and Locality Sensitive Hash (LSH) for quick retrieval of historical schemes similar to the new scheme. The cache table stores the structured feature vectors of these historical schemes, the first multiphysics performance index set calculated by the lightweight simulation engine, and the fusion evaluation results after fusion calibration. The similarity between newly generated structured feature vectors and the structured feature vectors in the cache table is calculated and compared with a preset similarity threshold to determine if the new scheme is sufficiently close to historical schemes. Similarity calculation can use metrics such as Euclidean distance, cosine similarity, or Manhattan distance. If the similarity is greater than or equal to the preset similarity threshold, the fusion evaluation result of the corresponding cached entry or the first multiphysics performance index set of the corresponding cached entry is reused. When the similarity between a newly generated structured feature vector and a historical scheme in the cache table reaches or exceeds a preset threshold, the system can directly retrieve the pre-calculated fusion evaluation result of that historical scheme from the cache table, or simply reuse the first multiphysics performance index set. If the similarity is less than the preset similarity threshold, the lightweight simulation engine is invoked to recalculate the first multiphysics performance index set and perform fusion calibration. If the similarity between the newly generated structured feature vector and any historical scheme in the cache table is less than the preset threshold, the system needs to invoke the lightweight simulation engine to predict the performance of the new scheme, calculate its first multiphysics performance index set, and then perform fusion calibration. The initial solution value of the lightweight simulation engine is determined based on the first multiphysics performance index set corresponding to the structured feature vector with the highest similarity in the cache table. To improve the computational efficiency of the lightweight simulation engine, when it is necessary to recalculate the first multiphysics performance index set of a new scheme, the calculation result of the historical scheme with the highest similarity to the new scheme in the cache table can be used as the initial solution value.

[0125] The proposed solution, when generating a candidate design adjustment set, first ranks the candidate parameter perturbations based on the sensitivity of the index coupling metric. By comparing the sensitivity values ​​with preset sensitivity thresholds, it intelligently determines whether to generate single-parameter adjustments or multi-parameter linkage adjustments. When a parameter has a small impact on performance, a single-parameter adjustment is generated for fine-tuning; when the parameter has a significant impact, a multi-parameter linkage adjustment is generated to achieve greater optimization. This sensitivity-based dynamic adjustment strategy makes the generation of candidate design adjustments more targeted and efficient, avoiding blind exploration. Furthermore, to further improve computational efficiency, this application establishes a similarity index and cache table for structured feature vectors. This cache table stores the structured feature vectors of historical design schemes and their corresponding first multiphysics performance index sets and fusion evaluation results. When generating a new candidate design adjustment, the system calculates the similarity between the newly generated structured feature vector and the existing vectors in the cache table. If the similarity reaches a preset threshold, the performance evaluation results in the cache are directly reused, thus avoiding repeated simulation calculations and fusion calibration processes. If the similarity is insufficient, the lightweight simulation engine is invoked for recalculation. However, in this case, the initial values ​​of the lightweight simulation engine can be determined based on the performance metrics set of the historical solutions with the highest similarity in the cache table, which significantly accelerates the simulation convergence process. In this way, when generating candidate design adjustments, the proposed solution can not only intelligently select adjustment strategies based on the degree of parameter influence, but also effectively utilize historical calculation results, greatly reducing the computational resources and time required in real-time optimization. This solves the problems of low generation efficiency, high computational resource consumption, and redundant calculations in real-time optimization scenarios, ensuring rapid response and high efficiency in the optimization process.

[0126] The following is a concrete example to illustrate this. When generating a set of candidate design adjustments, the system first performs sensitivity analysis on various parameters of the current building scheme. For example, for the window-to-wall ratio parameter, its sensitivity can be calculated by perturbing this parameter and observing changes in energy intensity indicators, operational carbon emission indicators, etc. If the sensitivity value of the window-to-wall ratio is lower than a preset threshold (e.g., 0.01), the system may generate a single-parameter adjustment scheme that increases the window-to-wall ratio by 0.02. If the sensitivity value of the thickness of the external wall insulation material is higher than a preset threshold (e.g., 0.05), the system may generate a multi-parameter linkage adjustment scheme, for example, simultaneously increasing the thickness of the external wall insulation material by 10mm and adjusting the window glass type to low-emissivity glass. Simultaneously, the system maintains a cache table that stores the structured feature vectors of thousands of previously evaluated building schemes, their corresponding first multiphysics performance index set (calculated by the lightweight simulation engine), and the fusion evaluation results. When a new candidate design adjustment scheme is generated (e.g., a scheme that increases the window-to-wall ratio by 0.02), the system calculates the similarity between the structured feature vector of the new scheme and all historical schemes in the cache table. This can be achieved by calculating the Euclidean distance between two vectors. If the Euclidean distance between the structured feature vector of a historical solution and the new solution is less than a preset similarity threshold (e.g., 0.001), the system will directly extract the fusion evaluation result of that historical solution from the cache table and use it as the evaluation result of the new solution, without having to run the simulation again. Conversely, if the similarity of all historical solutions does not reach the threshold, the system will call the lightweight simulation engine to perform simulation calculations on the new solution. In this case, to accelerate the simulation, the system will find the historical solution most similar to the new solution from the cache table and use its corresponding first multiphysics performance index set as the initial value for the lightweight simulation engine, thereby speeding up the simulation convergence.

[0127] 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 by those skilled in the art to the technical solutions of the present invention 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 method for real-time optimization of building performance aimed at low-carbon goals, characterized in that, include: Load the building performance database, material property database, material carbon emission factor database, location climate database, and low-carbon design constraint library; initialize the lightweight simulation engine and pre-trained proxy model; and establish the baseline scheme parameter set and the low-carbon target parameter set. Input building parameters to obtain a set of building parameters for the building scheme to be optimized. The set of building parameters includes geometric parameters, building envelope parameters, equipment and operation parameters, and location and climate parameters. Parameter preprocessing involves unifying units and structurally encoding the building parameter set to obtain a structured feature vector. Based on the low-carbon design constraint library, compliance detection is performed on the structured feature vector. When non-compliant items are detected, constraint repair is performed on the non-compliant items to obtain a feasible feature vector. Multiphysics performance prediction involves inputting the feasible feature vector into a lightweight simulation engine to obtain a first multiphysics performance index set, and inputting the feasible feature vector into a pre-trained surrogate model to obtain a second multiphysics performance index set. The fusion evaluation involves fusing and calibrating the first multiphysics performance index set with the second multiphysics performance index set to obtain the fusion evaluation result. The fusion evaluation result is then compared with the benchmark index and the target threshold to obtain the performance gap vector and index coupling metric. The objective is constructed by determining dynamic weights based on building type identifiers, design stage identifiers, the performance gap vector, and the index coupling metric, and then constructing a multi-objective comprehensive objective function and determining the constraint set based on the dynamic weights. Candidate design adjustment generation: Under the constraint set, a candidate design adjustment set is generated, and the expected performance improvement and constraint impact are calculated for each candidate design adjustment in the candidate design adjustment set. The optimization suggestions are output by adjusting the priority scores of each candidate design based on the comprehensive scoring rules and sorting them. The optimization suggestions include parameter adjustment amounts, expected performance improvement amounts, and explanations of the impact of constraints. Closed-loop iteration: receiving selection signals for the optimization suggestions, updating the building parameter set based on the selection signals to obtain an updated building parameter set, repeating the parameter preprocessing based on the updated building parameter set until the optimization suggestions are output, and recording the parameter change trajectory and the fusion evaluation results to form an optimization history library; The model update and termination determination involves incrementally updating the pre-trained proxy model based on the optimization history library, and outputting the optimized building scheme and the corresponding fusion evaluation result when the termination condition is met.

2. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The low-carbon target parameter set includes an operational carbon emission target threshold, a material-implied carbon target threshold, and a comprehensive carbon emission target threshold. The comprehensive carbon emission target threshold is a weighted combination threshold or a summation combination threshold of the operational carbon emission target threshold and the material-implied carbon target threshold. The target thresholds include the operational carbon emission target threshold, the material-implied carbon target threshold, and the comprehensive carbon emission target threshold.

3. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, Both the first and second multiphysics performance index sets include energy intensity index, operational carbon emission index, material implicit carbon index, lighting performance index, ventilation performance index, and thermal performance index; wherein, the material implicit carbon index is calculated from the material usage and the material carbon emission factor, and the material usage is determined by the geometric parameters and the enclosure structure parameters.

4. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The structured encoding includes: standardizing or normalizing continuous parameters; performing one-hot encoding or embedding encoding on discrete parameters; and concatenating the processed parameters into a fixed-length structured feature vector according to a preset field order.

5. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The compliance testing includes testing for geometric boundary constraints, material performance boundary constraints, equipment capacity boundary constraints, and specification threshold constraints; the constraint repair includes at least one of projection repair, truncation repair, and item-by-item repair based on constraint priority, such that the repaired feasible feature vector is located within the feasible domain corresponding to the constraint set.

6. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The lightweight simulation engine uses a reduced-order model or a combination of pre-calculation lookup and interpolation to calculate the first multiphysics performance index set, so that the calculation of the first multiphysics performance index set meets the preset response time limit.

7. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The pre-trained proxy model outputs point estimates and uncertainty estimates for the second multiphysics performance index set; the comprehensive scoring rule includes a risk penalty term, which is determined based on the constraint violation probability determined by the uncertainty estimate.

8. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The fusion calibration includes residual correction: obtaining the residual between the first multiphysics performance index set and the second multiphysics performance index set, constructing an error correction term based on historical residual statistics or residual recursion in the optimization history library, and applying the error correction term to the second multiphysics performance index set to obtain the fusion evaluation result.

9. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, The process of determining the dynamic weights includes: The carbon emission-related gap component in the performance gap vector is compared with a preset gap threshold. If the carbon emission-related gap component is less than the preset gap threshold, the dynamic weight of the carbon emission-related indicators remains unchanged. If the carbon emission-related gap component is greater than or equal to a preset gap threshold, then a gap ratio is determined based on the ratio of the carbon emission-related gap component to the preset gap threshold, and the dynamic weight of the carbon emission-related indicators is increased based on the gap ratio, and the increase in the dynamic weight of the carbon emission-related indicators is proportional to the gap ratio. The distance of the non-carbon performance constraint boundary corresponding to the coupled measurement of the index is compared with the preset safety margin; If the distance to the non-carbon performance constraint boundary is greater than or equal to the preset safety margin, the dynamic weight of the non-carbon performance index remains unchanged. If the distance to the non-carbon performance constraint boundary is less than a preset safety margin, a margin ratio is determined based on the ratio of the preset safety margin to the distance to the non-carbon performance constraint boundary. The dynamic weight of the non-carbon performance index is increased based on the margin ratio, and the increase in the dynamic weight of the non-carbon performance index is proportional to the margin ratio. Alternatively, the non-carbon performance index can be incorporated into the constraint set as a hard constraint.

10. The method for real-time optimization of building performance for low-carbon goals according to claim 1, characterized in that, Generating the candidate design adjustment set includes: The sensitivity ranking of candidate parameter perturbations is calculated based on the index coupling measurement, and the sensitivity values ​​in the sensitivity ranking are compared with a preset sensitivity threshold. If the sensitivity value is less than the preset sensitivity threshold, a single-parameter candidate design adjustment is generated and the candidate design adjustment set is formed. If the sensitivity value is greater than or equal to the preset sensitivity threshold, then a multi-parameter linkage candidate design adjustment is generated and the candidate design adjustment set is formed. Establish a similarity index and cache table for the structured feature vectors, and associate and store the structured feature vectors in the cache table with the corresponding first multiphysics performance index set and the corresponding fusion evaluation results; The similarity between the newly generated structured feature vector and the structured feature vector in the cache table is calculated, and the similarity is compared with the preset similarity threshold. If the similarity is greater than or equal to the preset similarity threshold, the fusion evaluation result of the corresponding cached entry is reused, or the first multiphysics performance index set of the corresponding cached entry is reused. If the similarity is less than the preset similarity threshold, the lightweight simulation engine is invoked to recalculate the first multiphysics performance index set and perform the fusion calibration; wherein, the initial value of the solution of the lightweight simulation engine is determined based on the first multiphysics performance index set corresponding to the structured feature vector with the highest similarity in the cache table.

Images