A green building energy-saving carbon efficiency evaluation method and system based on BIM and energy consumption data fusion

Abstract

The application belongs to the technical field of building carbon efficiency evaluation, and discloses a green building energy-saving carbon efficiency evaluation method and system based on BIM and energy consumption data fusion. The standard IFC data model is extended with carbon efficiency evaluation special semantic ontology, a carbon efficiency attribute extension entity is derived under the IfcBuildingElement entity subclass, carbon efficiency related extension attribute fields are supplemented for various building components, and carbon efficiency semantic standardization predefinition on the BIM side is independently completed. Four-level granularity of building as a whole, floor partition, room unit and component monomer is adopted to regularly divide the BIM space, generate a hierarchical space index grid, and support a three-segment space coding rule containing IFC entity type identification, space level identification and unique sequence identification. The corresponding BIM component attribute set is bound for each grid unit, and a unified space reference for cross-domain data fusion is constructed.

Description

Technical Field

[0001] This invention belongs to the field of building carbon efficiency evaluation technology, specifically a green building energy-saving carbon efficiency evaluation method and system based on the fusion of BIM and energy consumption data. Background Technology

[0002] Building Information Modeling (BIM), based on its ability to integrate information throughout the entire building lifecycle and combined with operational data from building energy consumption monitoring systems, is widely used in green building performance assessment. Current technologies generally employ a technical approach of BIM modeling, energy consumption simulation software verification, and scoring of individual indicators. However, these existing technologies suffer from the following technical problems: BIM models are static spatial geometry and attribute information collections, while energy consumption data are time-series dynamic data. The two belong to different information dimensions. Current integration is only a simple data hooking and label mapping, lacking a unified spatiotemporal indexing mechanism. Spatial positioning accuracy and time resolution are mutually constrained, making it impossible to achieve synchronous and accurate matching of component-level spatial granularity and minute-level time granularity. The evaluation results can only cover the overall building or system sub-items.

[0003] Existing carbon efficiency assessments mostly use simulated values ​​from the design phase or fixed annual baselines as the basis for evaluation. They cannot respond to baseline deviations caused by time-varying factors such as personnel behavior drift, equipment performance degradation, and interannual changes in meteorological parameters. The evaluation conclusions lack dynamic confidence indicators, and the reference value of the evaluation results drops significantly when the operating conditions deviate from the design state.

[0004] Existing data is integrated into a one-way path, with energy consumption simulation driven by parameters derived from the BIM model in a forward direction. However, actual operational energy consumption data cannot be used to calibrate the embedded parameters of the BIM model in reverse. As the years of operation increase, the information discrepancy between the BIM model and the real building continues to accumulate, and the model's fidelity gradually decreases.

[0005] The existing evaluation system can only output the overall carbon efficiency score of the building or its sub-systems. It cannot pinpoint the carbon efficiency anomalies of specific spatial units or even the component level, nor can it quantify the contribution of each spatial area and energy-consuming equipment to the overall carbon efficiency, making it difficult to directly guide precise energy-saving renovations. Summary of the Invention

[0006] The purpose of this invention is to provide a green building energy-saving carbon efficiency evaluation method and system based on the fusion of BIM and energy consumption data, so as to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a green building energy-saving carbon efficiency evaluation method based on the fusion of BIM and energy consumption data, comprising the following specific steps: Furthermore, a dedicated semantic ontology for carbon efficiency evaluation is extended to the standard IFC data model. A carbon efficiency attribute extension entity is derived under the IfcBuildingElement entity subclass to supplement extended attribute fields for building components. The BIM space is systematically divided into four levels of granularity: overall building, floor zoning, room unit, and individual component, generating a hierarchical spatial index grid. A three-segment spatial coding rule is established, with the code sequentially including IFC entity type identifier, spatial level identifier, and unique sequence identifier. Each grid unit is bound to a unique spatial code and the corresponding BIM component attribute set, achieving a strong correlation mapping between the grid and the IFC entity through coding. The standardization of spatial information and the predefinition of carbon efficiency semantics on the BIM side are completed, forming a unified spatial benchmark for cross-domain data fusion.

[0008] Furthermore, energy consumption data from building sub-meters, cold and heat source systems, environmental sensors, and meteorological data from weather stations are collected. Based on a unified time reference, interpolation alignment and outlier repair of time-series data are completed. The spatial index grid is matched with the spatial coordinates of the energy consumption collection points, and a spatial attribution code is marked for each energy consumption data point. The original energy consumption sequence is decoupled into three types of energy consumption feature components through a feature decoupling algorithm: climate-related energy consumption component, base load energy consumption component, and operating condition-driven energy consumption component. These correspond to three types of driving factors: outdoor meteorological influence, building thermal performance, and energy consumption behavior of personnel and equipment, respectively. A mapping relationship between energy consumption data and spatial index grid is established to form a standardized input for cross-domain mapping.

[0009] Furthermore, by combining the geometric properties, thermal parameters, topological adjacency relationships, and decoupled energy consumption feature components of the spatial index grid, a bidirectional mapping model of the graph neural network is constructed; the forward branch uses the geometric properties and thermal parameters of the spatial index grid as graph node features and the topological adjacency relationships as graph edge features to output the time-period predicted energy consumption distribution of the corresponding spatial unit. A gradient constraint calibration mechanism is set up in the reverse branch. Taking the measured energy consumption time series characteristics as input, the embedded parameters in the BIM model that cannot be directly observed are calibrated in reverse along the spatial topology relationship. Physical boundary constraints of the parameters are set during the calibration process. The interactive correction of BIM spatial information and energy consumption time series information is realized through the graph neural network bidirectional mapping model.

[0010] Furthermore, by combining historical operating datasets with meteorological parameters from the same period, a dual sliding window incremental learning framework is adopted, setting a baseline update window and a confidence calibration window to generate a carbon efficiency baseline; a working condition drift detection mechanism is set up, which triggers the reconstruction of the carbon efficiency baseline when the similarity between the current operating condition and the historical baseline operating condition is lower than a set threshold. A baseline confidence assessment model is established. Based on factors including the similarity between current and historical operating conditions, data sample size, and data quality, the confidence weights of the carbon efficiency baseline for each time period are calculated. These confidence weights serve as quantitative values ​​of the confidence level. When structural changes occur in the building's operating conditions, the carbon efficiency baseline automatically shifts accordingly. The evaluation conclusion is accompanied by a confidence level indicator, which directly participates in the interval determination of the carbon efficiency level.

[0011] Furthermore, by combining the output results of the carbon efficiency baseline and the graph neural network bidirectional mapping model, the deviation of the overall building carbon efficiency from the carbon efficiency baseline and the corresponding confidence interval is determined. The carbon efficiency data of the corresponding floor zone is matched to complete the anomaly screening. The carbon efficiency anomaly points of the room unit and the individual component are located downwards. The spatial transmission path of the carbon efficiency anomaly is tracked simultaneously to clarify the transmission link of the anomaly signal from the component to the whole building. A dual-dimensional contribution decomposition algorithm based on structure and operation is used to quantify the contribution ratio of spatial units and energy consumption items to the overall carbon efficiency, distinguishing between structural carbon efficiency differences and operational carbon efficiency differences; the output includes targeted localization results containing abnormal location, abnormality degree, contribution ratio, and difference type.

[0012] Furthermore, by matching the weak links in carbon efficiency identified through carbon efficiency anomaly localization and path tracing, energy-saving renovation scenario combinations are constructed, including categories such as building envelope renovation, equipment system performance optimization, and operation strategy optimization; simulation calculations are performed based on the calibrated graph neural network bidirectional mapping model to obtain carbon efficiency improvement curves under different renovation intensities, and the marginal diminishing effect of each measure is quantified. With the dual objectives of improving renovation costs and carbon efficiency, the Pareto optimal solution set is obtained, and the optimal renovation combination scheme under different budget constraints is output. Simultaneously, a BIM parameter adjustment list corresponding to the renovation scheme is generated, clarifying the changes in BIM component parameters corresponding to each renovation measure.

[0013] Furthermore, a mapping rule is established between carbon efficiency information and IFC semantic entities. The carbon efficiency evaluation results, anomaly markers, and transformation scheme parameters are back-mounted to the spatial index grid and component entities corresponding to the BIM model in the form of extended semantic attributes, so as to realize the spatial display and interactive query of carbon efficiency distribution in the BIM visualization environment. Set up a model fidelity attenuation early warning mechanism. After each data iteration, calculate the degree of deviation between the BIM model parameters and the actual working conditions. When the deviation exceeds the threshold, trigger a full calibration of the embedded parameters of the BIM model. The modified expected carbon efficiency parameters are fed back into the dedicated semantic ontology for carbon efficiency evaluation to update the model baseline parameters.

[0014] This invention also provides a green building energy-saving carbon efficiency evaluation system based on the fusion of BIM and energy consumption data, used to implement the above method, comprising the following modules: The semantic extension and spatial partitioning module extends the standard IFC data model with a dedicated semantic ontology for carbon efficiency evaluation, derives carbon efficiency attribute extension entities, partitions BIM space according to four-level granularity rules to generate a hierarchical spatial index grid, establishes three-segment spatial coding rules and binds them to the corresponding BIM component attribute sets. The data registration and decoupling module collects energy consumption data, completes time-series interpolation alignment and outlier repair, matches spatial index grid encoding to mark the spatial affiliation of energy consumption data, decouples to obtain three types of energy consumption feature components, and establishes a mapping relationship between energy consumption data and spatial index grid. The bidirectional mapping network module constructs a graph neural network bidirectional mapping model. The forward branch outputs the predicted value of energy consumption distribution of each spatial unit in time period, and the reverse branch iteratively calibrates the embedded parameters of the BIM model along the spatial topology relationship. The calibration process applies physical boundary constraints to the parameters. The benchmark generation module uses a dual sliding window incremental learning framework to generate a carbon efficiency benchmark, sets a working condition drift detection mechanism to trigger benchmark reconstruction, and combines multi-dimensional factors to calculate the benchmark confidence weight and incorporate it into the carbon efficiency level determination. The anomaly tracing and attribution module locates the anomaly points of carbon efficiency at a multi-level granularity and tracks the spatial transmission path. It uses a two-dimensional contribution decomposition algorithm to quantify the carbon efficiency contribution ratio of each unit, distinguishes between structural carbon efficiency differences and operational carbon efficiency differences, and outputs targeted location results. The renovation scheme optimization module constructs energy-saving renovation scenario combinations, calculates the carbon efficiency improvement curve based on the calibrated graph neural network bidirectional mapping model, solves the Pareto optimal solution set of the dual objectives of renovation cost and carbon efficiency improvement, and generates a corresponding BIM parameter adjustment list. The semantic mounting closed-loop module establishes mapping rules between carbon efficiency information and IFC semantic entities, reverse-mounts carbon efficiency evaluation results, anomaly markers and retrofit scheme parameters to the BIM model, sets up a model fidelity attenuation early warning mechanism, feeds back retrofit parameters to the dedicated semantic ontology for carbon efficiency evaluation and updates the model baseline parameters.

[0015] The beneficial effects of this invention are as follows: 1. This invention extends the standard IFC data model with a dedicated semantic ontology for carbon efficiency evaluation. It derives carbon efficiency attribute extension entities under the IfcBuildingElement entity subclass, supplementing various building components with carbon efficiency-related extension attribute fields, and independently completing the standardized pre-definition of carbon efficiency semantics on the BIM side. It employs a four-level granularity approach—building as a whole, floor zoning, room unit, and individual component—to perform regularized partitioning of the BIM space, generating a hierarchical spatial index grid. This is accompanied by a three-segment spatial coding rule containing IFC entity type identifiers, spatial hierarchy identifiers, and unique sequence identifiers. Each grid unit is bound to a corresponding BIM component attribute set, constructing a unified spatial benchmark for cross-domain data fusion. Combining temporal interpolation alignment, outlier repair, and spatial attribution coding of energy consumption data, a precise mapping is established with the spatial index grid after decoupling from three types of energy consumption feature components, achieving synchronous matching between component-level spatial granularity and high temporal resolution.

[0016] 2. This invention constructs a bidirectional mapping model using a graph neural network. The forward branch takes the geometric properties, thermal parameters, and topological adjacency relationships of spatial units as input, and outputs the predicted energy consumption distribution of each spatial unit over time. The reverse branch takes the measured energy consumption time-series characteristics as input and iterates backward along the spatial topological relationships under the constraints of the physical boundaries of the parameters. This calibrates embedded parameters in the BIM model that cannot be directly observed, reversing the problem of model deviation accumulating year by year caused by traditional one-way data fusion, and continuously maintaining the consistency between the BIM model and the actual building conditions. A matching dual-sliding-window incremental learning framework generates a dynamic carbon efficiency baseline. Combined with a condition drift detection mechanism, it triggers baseline reconstruction. The baseline confidence weight is calculated through multi-dimensional factors and directly incorporated into the carbon efficiency level determination, which can adapt to time-varying factors such as personnel behavior, equipment degradation, and interannual meteorological variations.

[0017] 3. This invention achieves step-by-step anomaly localization from the overall building to individual components through a four-level spatial granularity system, simultaneously tracking the spatial transmission path of carbon efficiency anomalies and clarifying the transmission link of anomaly signals from components to the overall building; combined with a dual-dimensional contribution decomposition algorithm of structure and operation, it quantifies the carbon efficiency contribution ratio of each spatial unit and energy consumption item, clearly distinguishes between structural carbon efficiency differences and operational carbon efficiency differences, and outputs targeted localization results with clear location, degree of anomaly, and cause type; constructs multiple transformation scenario combinations for weak carbon efficiency links, calculates the carbon efficiency improvement curve based on a calibrated graph neural network bidirectional mapping model, solves the dual-objective Pareto optimal solution set of transformation cost and carbon efficiency improvement, and generates a BIM parameter adjustment list; finally, it completes model updates through semantic reverse mounting and fidelity early warning mechanisms. Attached Figure Description

[0018] Figure 1 This is the overall flowchart for the green building energy-saving carbon efficiency evaluation of this invention; Figure 2This is an internal flowchart of the bidirectional mapping model of the graph neural network of the present invention; Figure 3 This is a flowchart of the dynamic generation and confidence calibration process of the carbon efficiency baseline in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1 to 3 As shown, this embodiment of the invention provides a green building energy-saving carbon efficiency evaluation method based on the fusion of BIM and energy consumption data; This embodiment uses a 12-story Class A office building in a hot-summer, cold-winter region as the application scenario. The building has a total floor area of ​​18,600 square meters and has completed basic BIM modeling and output standard IFC4 format files. The standard IFC data model is extended with a dedicated semantic ontology for carbon efficiency evaluation. Based on the IFC standard attribute set extension mechanism, a custom attribute set `CarbonEfficiency_Pset` is defined and associated with the `IfcBuildingElement` entity subclass. Extended attribute fields such as unit energy consumption carbon emission factor, design heat transfer coefficient, and rated energy consumption power are added to components such as walls, windows, and heating / cooling source equipment. The BIM space is systematically divided into three levels of granularity: overall building, floor zoning, and room units, generating a hierarchical spatial index grid. A one-to-one mapping association is simultaneously established between the spatial grid and two types of component entities: internal envelope structure and equipment terminals. A three-segment spatial coding rule is established, with the code sequentially including the IFC entity type identifier, spatial level identifier, and unique sequence identifier. Each grid unit is bound to a unique spatial code and the corresponding BIM component attribute set, achieving a strong association mapping between the grid and the IFC entity through coding. The IfcCarbonEfficiencyExtension entity is bound to the original IFC component entity using an attribute set association method. The attribute set is named CarbonEfficiency_Pset, and it does not change the original attributes and topological relationships of the standard IFC entity. The extension entity is divided into three attribute groups according to function: carbon emission basic attribute group, thermal performance attribute group, and operation evaluation attribute group. The carbon emission basic attribute group stores basic parameters such as the energy type, carbon emission factor per unit energy consumption, and carbon emission coefficient throughout the life cycle of the corresponding component. In this project, the carbon emission factor for electricity is 0.5810 kgCO2 / kWh, and the carbon emission factor for natural gas is 2.1622 kgCO2 / m³. 3 All of these are stored in the corresponding equipment component's carbon emission basic attribute group.

[0021] The thermal performance attribute group stores parameters such as heat transfer coefficient, heat storage coefficient, and solar radiation absorption coefficient of the building envelope components, serving as the source of node features for the graph neural network model. In this project, the design heat transfer coefficient of the exterior wall is 0.78 W / (m²). 2 ·K), exterior window design heat transfer coefficient 2.7W / (m²) 2 ·K) are all stored in the thermal performance attribute group of the corresponding component; the operation evaluation attribute group is used to store dynamically generated evaluation information such as carbon efficiency evaluation results, anomaly markers, and modification parameters. All dynamic evaluation data are written to this attribute group and do not occupy the IFC native attribute fields.

[0022] After the semantic ontology extension is completed, all extended attribute groups are included in the unified management of CarbonEfficiency_Pset, which is independent of the standard IFC native attributes. They can be synchronously called through the association relationship of the entity's globally unique identifier, and the extended attributes can be read directly in mainstream BIM software.

[0023] The standardized meshing adopts a top-down hierarchical approach. The overall building level corresponds to the complete building space boundary and serves as the top-level benchmark for meshing; this project corresponds to one overall building grid unit. The floor zoning level is divided according to the building's natural floors, with each natural floor corresponding to one zoning unit, preserving the floor's elevation and boundary information. This project has a total of 12 floor zoning units. The room unit level extracts boundaries from the room partition components in the BIM model; each independent room corresponds to one grid unit. This project's standard floor has a total of 42 room units. The component unit level decomposes into units for two core component types: the building envelope and equipment terminals. A single wall, a single window, and a single piece of equipment each correspond to an independent grid unit. This project's standard floor has a total of 96 component unit units; the boundaries of each level of grid unit completely coincide with the geometric boundaries of the corresponding BIM entity.

[0024] The three-segment spatial coding uses a 16-bit fixed-length character encoding format. The first two bits are the IFC entity type identifier segment, using fixed characters to distinguish the entity category: BL corresponds to the entire building, FL to the floor zone, RM to the room unit, and EL to the individual component. The middle eight bits are the spatial level identifier segment, using a numerical sequence to identify the building, floor, and room level to which the grid belongs, reflecting the hierarchical nesting relationship of the grid. For example, 00030008 represents room 08 on the 3rd floor. The last six bits are the unique sequence identifier segment, using sequential numbering to uniquely encode all grid units under the same level. A one-to-one correspondence is established between the spatial code of each grid unit and the globally unique identifier of the corresponding IFC entity. The attribute binding process is automatically completed through code matching. For example, the complete code for the south-facing window of room 08 on the 3rd floor is EL00030008000012.

[0025] This part of the work does not rely on external energy consumption data. It independently completes the standardization of spatial information and the pre-definition of carbon efficiency semantics on the BIM side, forming a unified spatial benchmark for cross-domain fusion of BIM geometric space and energy consumption time series data, providing a standardized spatial coordinate system for subsequent data mapping and model training.

[0026] Hourly energy consumption and environmental data were collected from 144 sub-meters, 4 chiller / heater units, and 432 indoor temperature and humidity sensors in this project. Hourly dry-bulb temperature and total horizontal radiation meteorological data from the local meteorological station were simultaneously accessed. Using Beijing time at the top of the hour as the unified time reference, linear interpolation was used to complete missing data, and the Laida criterion was used to identify and repair outliers, resulting in a complete time-series dataset with a time resolution of 1 hour. The spatial index grid was matched to the installation coordinates of the energy consumption collection points, and a spatial attribution code was assigned to each energy consumption data point. Spatial attribution matching primarily uses point coordinates as the determining factor, supplemented by system logical attribution. For room-level meters and indoor sensors with precise installation coordinates, the device coordinates are compared with the three-dimensional boundaries of each spatial index grid. When the coordinates fall completely within a grid, the spatial code of the corresponding grid is directly marked. When multiple similar collection points exist within the same grid, the corresponding time-series data are summed and aggregated before being uniformly bound to that grid code.

[0027] For main line equipment such as floor main electricity meters and cold and heat source main pipes covering multiple spatial units, the grid code of the corresponding level is matched according to the metering coverage of the equipment. Combined with the design load ratio and historical energy consumption characteristics of each room, the data is split into lower-level sub-grids. For example, the main electricity meter data of the 3rd floor is split into each room grid according to the building area ratio of the 42 rooms on that floor.

[0028] For data collection points without precise coordinates, spatial codes of the corresponding level are matched according to the logical affiliation of the system loop to which the data belongs and the installation location. After all energy consumption data is affixed and marked, the correspondence between the original data collection point identifier and the grid code is retained, and the complete data spatial mapping link can be traced.

[0029] By employing a feature decoupling algorithm combining variational mode decomposition and multiple linear regression constraints, the original energy consumption sequence is decoupled into three types of energy consumption characteristic components: climate-related energy consumption component, base load energy consumption component, and operating condition-driven energy consumption component. These correspond to three driving factors: outdoor meteorological influence, building thermal performance, and energy consumption behavior of personnel and equipment, respectively. A mapping relationship between energy consumption data and a spatial index grid is established, forming a standardized input for cross-domain mapping. The specific execution steps are as follows: The first step is to use the hourly energy consumption time-series data, which has been time-aligned and anomaly-corrected, as the original input sequence, and simultaneously input the outdoor dry-bulb temperature and total horizontal radiation meteorological parameters of the same time period as constraint variables.

[0030] The second step is to set the number of modal components to 3, initialize the center frequency and bandwidth constraints of each component, and solve the variational constraint problem iteratively by the alternating direction multiplier method to obtain three sets of initial modal components.

[0031] The third step is to perform regression fitting on the high-frequency modal components using meteorological parameters as independent variables, and the goodness of fit R is calculated. 2 A value higher than 0.7 is considered a climate-related energy consumption component, corresponding to fluctuations in building heating and cooling loads caused by changes in outdoor weather.

[0032] The fourth step is to determine the medium-frequency steady-state mode as the base load energy consumption component, which corresponds to the fixed baseline energy consumption determined by the thermal performance of the building itself; and to determine the low-frequency fluctuation mode as the operating condition-driven energy consumption component, which corresponds to the energy consumption changes brought about by personnel activities, equipment start-up and shutdown, and operating strategies.

[0033] The fifth step is to sum and verify the three types of components to ensure that the relative deviation between the component superposition result and the original energy consumption sequence is less than 2%, and output the three types of energy consumption characteristic component sequences. The temporal resolution and spatial attribution label of the three types of energy consumption characteristic component sequences are consistent with the original energy consumption data. In this project, the proportions of the three types of components in room 08 on the 3rd floor during summer working days are 38%, 22%, and 40%, respectively.

[0034] All energy consumption data are converted into carbon efficiency values ​​according to the carbon emission factors of the corresponding energy types. The conversion process is completed simultaneously with the decoupling of energy consumption characteristics, and three types of carbon efficiency components are generated synchronously for the three types of energy consumption characteristic components. The carbon emission factors are taken from the carbon emission basic attribute group of the corresponding component in the dedicated semantic ontology for carbon efficiency evaluation. Different energy consumption types adopt corresponding exclusive carbon emission factors. In this project, the conversion factor for electricity is 0.5810 kgCO2 / kWh, and the conversion factor for natural gas is 2.1622 kgCO2 / m³. 3 .

[0035] The conversion is performed on a time-period basis according to spatial grid units. The time-period energy consumption data of each spatial unit is matched with the carbon emission factor of the corresponding energy type to calculate the carbon emission of the corresponding time period. This is used as the basic quantitative value for carbon efficiency evaluation. For example, the hourly carbon efficiency value of room 08 on the 3rd floor during summer workdays is about 1.2-1.8 kgCO2 / h. After the conversion is completed, all subsequent baseline generation, abnormal deviation judgment, and carbon efficiency improvement calculation are based on the converted carbon efficiency value.

[0036] By combining the geometric properties, thermal parameters, topological adjacency relationships, and decoupled energy consumption feature components of the spatial index grid, a bidirectional mapping model based on a graph attention network is constructed. The forward branch uses the geometric properties and thermal parameters of the spatial index grid as static graph node features, and topological adjacency relationships as graph edge features. It simultaneously incorporates time-period outdoor meteorological parameters and personnel / equipment operating condition parameters as dynamic input features, outputting the time-period predicted energy consumption distribution value for the corresponding spatial unit. The network structure and execution flow of the forward branch are as follows: The network structure consists of four layers: an input layer, a first-layer graph attention convolutional layer, a second-layer graph attention convolutional layer, and a fully connected output layer. The input layer receives the node features and adjacency matrix of all spatial index grids. The feature dimension of a single node is 6, corresponding to three geometric attributes: grid projected area, net floor height, and proportion of the outer envelope structure, as well as three thermal parameters: heat transfer coefficient of the outer envelope structure, heat storage coefficient of the wall, and solar radiation absorption coefficient of the outer surface. For example, the six feature values ​​of node 08 on the 3rd floor are 32m. 2 3.6m, 0.28, 2.7W / (m 2 ·K), 10.2kJ / (m 2 ·K), 0.7. The adjacency matrix is ​​constructed based on the physical adjacency of spatial units. Two units are marked as adjacent if they have a common enclosure. The edge weight is the product of the area of ​​the common enclosure and the heat transfer coefficient. For example, the area of ​​the common partition wall between room 08 on the 3rd floor and the adjacent room 09 is 12㎡, and the corresponding edge weight is 12×0.78=9.36.

[0037] The first layer, a graph attention convolutional layer, uses eight attention heads to upscale the 6-dimensional node features to 64-dimensional features using linear transformation. It calculates the weight coefficients of neighboring nodes through an attention mechanism and fuses the features of neighboring nodes to output the first-level features. The second layer, a graph attention convolutional layer, uses four attention heads to perform secondary neighborhood fusion on the first-level features, outputting 128-dimensional high-order spatial correlation features. The fully connected output layer performs dimensionality reduction mapping on the high-order features and outputs the hourly energy consumption prediction value of the corresponding spatial unit. This model can simultaneously output the hourly energy consumption prediction results of 42 room units in the standard layer at a single time step.

[0038] The training process for the forward branch used 8760 hours of hourly measured energy consumption data from the previous year as labels, with mean squared error as the loss function, and the Adam optimizer was used for iterative training. The training batch size was set to 16, and the initial learning rate was set to 0.001. Training was stopped when the loss function decreased by less than 1e-5 for 10 consecutive iterations. The training process of this project converged after 126 iterations, completing the parameter solidification of the forward branch, and the average prediction bias of the test set was controlled within 7.2%.

[0039] A gradient constraint calibration mechanism is set up in the reverse branch. Taking the measured energy consumption time series characteristics as input, the embedded parameters in the BIM model, such as thermal performance parameters and equipment efficiency parameters, which cannot be directly observed, are calibrated in reverse along the spatial topology. During the calibration process, physical boundary constraints of the parameters are set to make the parameter values ​​conform to the actual laws of engineering. The interactive correction of BIM spatial information and energy consumption time series information is realized through a graph neural network bidirectional mapping model.

[0040] The reverse calibration process fixes all network weights and topology of the forward branches, uses energy consumption prediction deviation as the loss function, and employs a gradient constraint calibration mechanism to perform reverse iterative optimization of BIM input parameters. The specific process is as follows: Input the current 30-day hourly measured energy consumption data into the model, calculate the deviation between the positive prediction value and the measured value, and propagate the gradient backward along the network starting from the deviation value. Map the gradient to the thermal parameters of the BIM components and the equipment efficiency parameters corresponding to each node to generate parameter update quantities.

[0041] The parameter update process applies physical boundary constraints, such as limiting the range of the external wall heat transfer coefficient to 0.4-1.2 W / (m²). 2 The heat transfer coefficient of the external window is limited to 1.8-3.5 W / (m²·K). 2 (K), the COP of the chiller unit is limited to 4.2-6.1, and the update amount exceeding the range is directly truncated to the boundary value.

[0042] A successive iterative calibration method was adopted. After each iteration, the positive prediction deviation was recalculated. Calibration was considered converged when the deviation change rate was less than 0.5% for three consecutive iterations, or when the number of iterations reached the upper limit of 50. All embedded parameters of the calibrated BIM model were then output, completing one model parameter update. This was the first calibration after one year of operation in this project, and the external wall heat transfer coefficient was calibrated from the design value of 0.78 to the actual value of 0.86 W / (m²). 2 ·K).

[0043] Inverse parameter calibration formula with physical boundary constraints: This represents the current value of the embedded parameter of the m-th BIM model at the t-th iteration, covering thermal and equipment parameters that cannot be directly observed, such as the heat transfer coefficient of the building envelope, the heat storage coefficient of the wall, and equipment efficiency parameters. This represents the parameter value after the (t+1)th iteration update; The learning rate represents the gradient update and controls the step size of parameter updates in each iteration. This represents the gradient value of the loss function L with respect to the m-th parameter, which is calculated through backpropagation of a graph neural network; This represents a boundary truncation function, used to implement the "parameter physical boundary constraint" described in the document, ensuring that the parameter values ​​always conform to the actual range of the project. This represents the lower limit of the physical value of the m-th parameter, corresponding to the lower limit of the conventional engineering value for the component type; This represents the upper limit of the physical value of the m-th parameter, corresponding to the upper limit of the conventional engineering value for the component type; The mean squared error loss function represents the positive prediction and is used to quantify the deviation between the predicted and measured energy consumption values. Indicates the total number of cells in the spatial index grid; The time-period energy consumption prediction value of the i-th spatial unit is output by the positive branch of the graph neural network; This represents the time-period measured energy consumption value of the i-th spatial unit.

[0044] The three decoupled energy consumption characteristic components correspond to different functional links of the bidirectional mapping model: the climate-related energy consumption component is linked with the meteorological parameters of the same period and is used to verify the influence weight of meteorological factors on the energy consumption of spatial units in the model, thereby improving the prediction accuracy under different meteorological conditions; the basic load energy consumption component corresponds to the steady-state output of the thermal performance of the building body. In the reverse calibration process, the prediction deviation generated by this component is mainly used to calibrate the embedded parameters of the building envelope, such as the heat transfer coefficient and heat storage coefficient. For example, in this project, the deviation of the basic load component is mainly used to calibrate the actual heat transfer coefficient of the south-facing exterior wall and windows on the third floor.

[0045] The operating condition-driven energy consumption component corresponds to the dynamic energy consumption behavior of personnel and equipment. During reverse calibration, the prediction deviation generated by this component is mainly used to calibrate embedded parameters of equipment such as actual operating efficiency and load rate. For example, in this project, the deviation of the operating condition-driven component is mainly used to calibrate the actual operating efficiency of the three-layer air conditioning terminal fans. By layering the three types of components, the mutual interference of deviations from different driving factors is avoided, improving the accuracy of reverse parameter calibration and ensuring that the calibration results of structural and equipment parameters conform to physical reality.

[0046] By combining historical operating datasets with meteorological parameters from the same period, a dual sliding window incremental learning framework is adopted, setting a baseline update window and a confidence calibration window to generate a dynamic carbon efficiency baseline. A working condition drift detection mechanism is set up, which triggers the reconstruction of the carbon efficiency baseline when the similarity between the current operating condition and the historical baseline operating condition is lower than a set threshold.

[0047] The dual sliding window incremental learning framework uses an online ridge regression algorithm with a forgetting factor to generate a carbon efficiency baseline. The specific settings and execution process are as follows: Two parallel sliding windows are set up: a baseline update window and a confidence calibration window. The baseline update window has a duration of 90 days and a sliding step of 7 days, which is used to carry out incremental updates of the baseline. The confidence calibration window has a duration of 30 days, which overlaps with the tail end of the baseline update window, and is used to carry out calibration calculations for the baseline confidence.

[0048] The baseline generation process uses the baseline values ​​of outdoor dry-bulb temperature, solar radiation, and personnel and equipment load under standard working hours as inputs, and carbon emissions per unit area as output. A ridge regression model with a forgetting factor is used for incremental fitting; the forgetting factor is set to 0.95, and the regularization coefficient α is 0.01. Newly entered data receives higher weight, gradually weakening the influence of long-term data, and outputting a time-period carbon efficiency baseline sequence. The baseline value of carbon efficiency per unit area on summer working days for this project is approximately 0.022 kgCO2 / (m²). 2 ·h).

[0049] Incremental ridge regression formula with forgetting factor: Let represent the ridge regression loss function with a forgetting factor, which is the optimization objective for fitting the baseline; This represents the regression coefficient vector to be fitted. After the fitting is completed, the carbon efficiency benchmark value for that period can be calculated by inputting the features of the corresponding time period. This represents the forgetting factor, which is set to 0.95 in the document. It is used to give more weight to recent data and gradually weaken the influence of older data, thereby achieving incremental updates to the baseline. This indicates the total number of time periods within the baseline update window. The duration of the baseline update window in the document is 90 days. The carbon emissions per unit area in time period t (actual carbon efficiency) are used as the label variable for regression fitting. This represents the input feature vector for time period t, which includes outdoor meteorological parameters and basic load energy consumption components during the same period. Representing the eigenvector Transpose of; This represents the ridge regression regularization coefficient, used to apply L2 regularization constraints and prevent the model from overfitting. This represents the squared L2 norm of the regression coefficient vector, i.e., the regularization term.

[0050] The operating condition drift detection uses cosine similarity calculation to extract four features in the current window: the proportion of climate-related energy consumption, the proportion of operating condition driving energy consumption, the average outdoor temperature, and the average operating condition driving energy consumption per unit area. These features form the current operating condition feature vector. The corresponding features in the baseline window are extracted to form the baseline operating condition feature vector. The cosine similarity between the two feature vectors is calculated, and the similarity threshold is set to 0.75.

[0051] When the similarity is below the threshold for three consecutive testing cycles, it is determined that a working condition drift has occurred, triggering a full reconstruction of the carbon efficiency baseline. The baseline is regenerated using the current window data as the initial sample. For example, after the occupancy rate of this project increased from 65% to 92%, the working condition similarity was below 0.7 for three consecutive weeks, triggering a full reconstruction of the baseline.

[0052] A baseline confidence assessment model is established. Based on factors such as the similarity between the current operating conditions and historical operating conditions, the data sample size, and the data quality, the confidence weights of the carbon efficiency baselines for each time period are calculated. These confidence weights are used as the quantitative values ​​of the confidence metric.

[0053] The calculation of the baseline confidence weight includes three influencing factors: the first is the operating condition similarity coefficient, which is consistent with the calculated cosine similarity; the second is the data sample size coefficient, which is calculated based on the ratio of the number of valid data days in the calibration window to the total number of days in the window; and the third is the data quality coefficient, which is calculated as "data integrity rate - 2 × outlier percentage", with a lower limit of 0.5.

[0054] The final baseline confidence weight is obtained by multiplying the three coefficients. The weight value ranges from 0 to 1. For example, if the similarity of working conditions in a certain period is 0.82, the effective data is 28 days, and the data completeness rate is 96%, the confidence weight is calculated to be 0.82×(28 / 30)×(0.96-2×0.02)≈0.71.

[0055] Carbon efficiency is graded into five levels: Excellent, Good, Satisfactory, Poor, and Very Poor, corresponding to baseline deviation ranges of ≤-15%, -15%~-5%, -5%~+5%, +5%~+15%, and >+15%, respectively. During the assessment process, the carbon efficiency baseline is used as the benchmark value. The deviation threshold of the baseline level range is divided by the confidence weight of the corresponding time period to obtain the upper and lower limits of the corrected level assessment range. The lower the confidence weight, the wider the level assessment range, and the stronger the conservatism of the evaluation result. The actual carbon efficiency value is compared and matched with the corrected range to output the corresponding carbon efficiency level.

[0056] When structural changes occur in the building's operating conditions, the carbon efficiency baseline automatically shifts accordingly, and the evaluation conclusion is accompanied by a confidence level indicator. The confidence level directly participates in the determination of the carbon efficiency level range. For example, in August, the overall carbon efficiency per unit area of ​​this project was 12% higher than the baseline value, corresponding to a confidence weight of 0.85. After correction, the upper limit of the qualified range is 5% / 0.85≈5.88%, so it is judged as a poor level.

[0057] By combining the output of the carbon efficiency baseline and the bidirectional mapping model of the graph neural network, the deviation of the overall carbon efficiency of the building from the carbon efficiency baseline and the corresponding confidence interval is determined. Anomaly screening is completed by matching the carbon efficiency data of the corresponding floor zones, locating the carbon efficiency anomaly points down to room units and individual components, and simultaneously tracing the spatial transmission path of the carbon efficiency anomaly to clarify the transmission link of the anomaly signal from the component to the whole building. For example, after the overall carbon efficiency of this project was determined to be poor in August, step-by-step screening revealed that the carbon efficiency deviation was highest in the 3rd floor zone. Further location was made in room 08 on the southwest side of the 3rd floor, ultimately identifying the south-facing window of this room as the core anomaly component. Spatial transmission path tracing is achieved through a combination of spatial hierarchical inclusion relationships and grid topology adjacency relationships. Starting with identified component-level carbon efficiency anomalies, the tracing proceeds hierarchically upwards along the spatial hierarchy, simultaneously identifying the lateral transmission impact of units at the same level using an adjacency matrix. During the backtracking process, the carbon efficiency deviation percentage of each adjacent upper-level unit is calculated, and units with a deviation percentage higher than 15% are marked as transmission path nodes. This hierarchical backtracking is completed sequentially from individual components to room units, from room units to floor zones, and from floor zones to the entire building, connecting all marked transmission nodes to form a complete spatial transmission path. In this project, the complete transmission path of this anomaly is: south-facing window of room 08 on the 3rd floor → room 08 on the 3rd floor → floor zone on the 3rd floor → the entire building, with the window anomaly contributing 62% of the room's carbon efficiency deviation.

[0058] Anomaly detection employs a tiered judgment rule combining relative deviation thresholds and confidence weights. The initial deviation thresholds for the four granularity levels are: 10% for the overall building, 15% for floor zones, 20% for room units, and 25% for individual components. Using the carbon efficiency baseline value for the corresponding time period as the benchmark, the relative deviation between the actual carbon efficiency value and the benchmark value is calculated. The deviation threshold is then adjusted based on the confidence weight for the corresponding time period; a lower confidence weight results in a more lenient judgment threshold. When the deviation at the overall building level exceeds the threshold, anomaly screening at the next level is triggered.

[0059] The carbon efficiency deviation value of each floor zone is calculated one by one, and zones with deviations exceeding the corresponding thresholds are screened out and marked as abnormal zones. For abnormal zones, all room units under them are further screened to identify room-level abnormal points. For room-level abnormal points, they are further broken down to internal building envelope components and energy-consuming equipment components, and the carbon efficiency contribution deviation of each individual component is calculated to finally locate the component-level abnormal points. The deviation magnitude and the proportion exceeding the threshold are recorded simultaneously for each level of abnormal points. For example, the carbon efficiency deviation of the south-facing window of room 08 on the 3rd floor of this project is 32%, which exceeds the component-level threshold by 7 percentage points.

[0060] For scenarios with multiple anomaly points, each anomaly point is independently backtracked, and overlapping path nodes are merged to form a conduction path network. The conduction direction and contribution percentage of each node on the path are simultaneously marked, distinguishing between the anomaly source node and the nodes affecting the conduction, thus clarifying the transmission logic of carbon efficiency anomalies from local components to the entire building. For example, in this project, when there are two anomalies simultaneously—heat transfer anomalies at the 3rd floor exterior windows and chiller unit efficiency degradation at the 5th floor—the two paths are backtracked separately, and overlapping nodes from the floors to the entire building are merged to form a complete anomaly conduction network.

[0061] A dual-dimensional contribution decomposition algorithm based on structure and operation is used to quantify the contribution ratio of spatial units and energy consumption items to the overall carbon efficiency, distinguishing between structural carbon efficiency differences and operational carbon efficiency differences; the output includes targeted localization results containing abnormal location, abnormality degree, contribution ratio, and difference type.

[0062] The structure and operation dual-dimensional contribution decomposition algorithm is implemented using the log-mean Dichroic exponential decomposition method, and the specific execution steps are as follows: The first step is to determine the decomposition dimensions. The structural dimensions include three types of inherent parameters: heat transfer coefficient of the building envelope, rated efficiency of equipment, and shape coefficient of space unit. The operational dimensions include four types of dynamic parameters: outdoor meteorological parameters, personnel density, equipment operating time, and system control strategy. The parameters corresponding to the carbon efficiency baseline are used as the benchmark values, and the current actual parameters are used as the comparison values.

[0063] The second step involves using a logarithmic average weighting method to calculate the change in carbon efficiency caused by changes in each type of structural parameter, and summing them up to obtain the total value of structural carbon efficiency difference, which corresponds to the carbon efficiency deviation caused by the building's inherent attributes; the same step is to calculate the change in carbon efficiency caused by changes in each type of operational parameter, and sum them up to obtain the total value of operational carbon efficiency difference, which corresponds to the carbon efficiency deviation caused by dynamic factors during the operational phase.

[0064] The third step involves breaking down the two types of differences into spatial units and energy consumption items, calculating the contribution percentage of each unit and item to the overall carbon efficiency deviation, and, combined with the anomaly location results, marking the dominant difference type at each anomaly location, and outputting the complete contribution decomposition results. In the total carbon efficiency deviation of this project in August, structural differences accounted for 31%, and operational differences accounted for 69%; among these, the higher heat transfer coefficient of external windows contributed 64% to the structural differences, and the increase in personnel density contributed 43% to the operational differences.

[0065] Log-mean Dijkstra's index (LMDI) decomposition formula: This represents the total carbon efficiency difference, which is the difference between the current actual carbon efficiency and the benchmark carbon efficiency for the same period. This represents the carbon efficiency value under baseline conditions, corresponding to the same period value of the carbon efficiency baseline. This indicates the carbon efficiency value under the current actual operating conditions; The total number of factors affecting carbon efficiency is divided into two main categories: structural factors and operational factors. This represents the contribution of the i-th influencing factor to the overall carbon efficiency difference. A positive value indicates that the factor increases the overall carbon efficiency, while a negative value indicates that the factor decreases the overall carbon efficiency. This represents the baseline value of the i-th influencing factor; This represents the current actual value of the i-th influencing factor; This represents the logarithmic average weighting function, which is the core calculation function of the LMDI decomposition method and can achieve decomposition results without residuals. Structural factors include: heat transfer coefficient of building envelope, rated efficiency of equipment, and building shape coefficient; operational factors include: outdoor meteorological parameters, population density, equipment operating time, and system control strategy.

[0066] By matching the carbon efficiency anomaly location and path tracing to identify weak links in carbon efficiency, three major categories of energy-saving renovation scenarios are constructed: building envelope modification, equipment system performance optimization, and operation strategy optimization. Simulation calculations are performed based on a calibrated graph neural network bidirectional mapping model to obtain carbon efficiency improvement curves under different renovation intensities, quantifying the diminishing marginal returns of each measure. For example, for the structural anomaly of the south-facing windows on the third floor, three renovation intensities are set: Intensity I is upgrading the double-glazed windows to Low-E double-glazed windows; Intensity II is adding fixed external shading; and Intensity III is replacing the window frames with thermally broken aluminum frames and Low-E double-glazed windows. The corresponding window heat transfer coefficients are reduced to 2.2, 2.2, and 1.6 W / (m²), respectively. 2 (·K), corresponding to an external window solar heat gain coefficient of 0.65, 0.45, and 0.35, respectively; The calculation process is based on the calibrated graph neural network bidirectional mapping model. For each renovation measure, the thermal parameters or equipment efficiency parameters of the corresponding BIM components are adjusted according to the set renovation intensity. The adjusted parameters are updated to the node features of the model. The forward prediction branch is run to obtain the time-period energy consumption distribution after renovation. The carbon efficiency level after renovation is obtained by carbon efficiency conversion.

[0067] The carbon efficiency value after the renovation is compared with the baseline carbon efficiency value before the renovation on a unit-by-unit basis. The carbon emission reduction per unit time and the cumulative carbon emission reduction are calculated, corresponding to the carbon efficiency improvement of a single renovation measure. When multiple measures are combined for renovation, the parameters of all corresponding components are adjusted simultaneously and a unified positive prediction is run to obtain the overall carbon efficiency improvement result of the combined renovation. By gradually increasing the renovation intensity and repeating the calculation, the carbon efficiency improvement curve of the corresponding renovation measures is obtained, and the marginal effect change law in the process of intensity improvement is identified simultaneously. For example, when the heat transfer coefficient of the external window in this project decreases from 2.7 to 2.2, the annual CO2 emission reduction is 11.8 tons; when it decreases from 2.2 to 1.9, the annual emission reduction is 7.3 tons; and when it decreases from 1.9 to 1.6, the annual emission reduction is 4.5 tons, showing a clear marginal reduction effect.

[0068] The scenario combination construction adopts a matching mechanism of anomaly type and transformation measures: for weak links with structural carbon efficiency anomalies as the main focus, priority is given to matching measures such as building envelope transformation and equipment system performance optimization, and specific measures such as wall insulation upgrade, window performance improvement, and replacement of cold and heat source equipment are selected according to the type of abnormal component; for weak links with operational carbon efficiency anomalies as the main focus, priority is given to matching measures such as operation strategy optimization, and specific measures such as air conditioning operation period adjustment, equipment start and stop logic optimization, and indoor temperature and humidity setpoint adjustment are selected accordingly.

[0069] When two types of anomalies exist simultaneously in the same weak link, corresponding renovation measures are matched synchronously to form a combined renovation scenario. Three renovation intensities are set for each measure, with different intensities corresponding to different parameter adjustment ranges and renovation costs. A complete set of renovation scenarios is formed by combining multiple intensities. For example, the unit cost of the three renovation levels for exterior windows is 120 yuan / m². 2 180 yuan / m 2 360 yuan / m 2 .

[0070] With the dual objectives of minimizing the total life-cycle renovation cost over 10 years and maximizing annual carbon emission reduction, the Pareto optimal solution set is obtained, and the optimal renovation combination scheme under different budget constraints is output. Simultaneously, a BIM parameter adjustment list corresponding to the renovation scheme is generated, clarifying the changes to the BIM component parameters corresponding to each renovation measure.

[0071] The bi-objective Pareto optimal solution is implemented using a non-dominated sorting genetic algorithm with an elitist strategy. The specific execution flow is as follows: The first step is to define two optimization objectives: the first objective is to minimize the 10-year life cycle cost of the renovation project, including three cost items: material cost, construction cost, and operation and maintenance cost; the second objective is to maximize the annual carbon emission reduction after the renovation, which is calculated based on a calibrated graph neural network bidirectional mapping model.

[0072] The second step is to set constraints, including ensuring that the heat transfer coefficient of the building envelope is not less than 1.2 W / (m²). 2 The engineering lower limit of K), the performance upper limit of the COP of the chiller unit not exceeding 6.1, the proportion of exterior window renovation of a single building not exceeding 100%, and the minimum requirements for energy-saving design of public buildings after renovation.

[0073] The third step is to initialize the population, setting the population size to 100, and using combinations of different modification measures and modification intensities as individual codes; to classify the individuals in the population using non-dominated sorting, calculate the crowding degree to retain elite individuals, and generate the next generation population through selection, crossover, and mutation operations.

[0074] The fourth step involves setting the iteration count to 200 generations. After the iterations are complete, the Pareto optimal solution set, composed of all non-dominated solutions, is output. For different budget constraint thresholds, the renovation combination scheme that satisfies the budget conditions and offers the highest carbon efficiency improvement is selected from the solution set. Simultaneously, the component parameter changes corresponding to the schemes are mapped to a BIM parameter adjustment list, clearly defining the component code and parameter value corresponding to each adjustment. For example, under a budget of 500,000 yuan, the optimal solution is to upgrade all exterior windows to Low-E glass and optimize the air conditioning operation strategy, resulting in an annual CO2 reduction of approximately 86 tons.

[0075] The parameter adjustment list is generated using an itemized structure. Each item contains six core components: component spatial code, globally unique IFC entity identifier, parameter name, current parameter value, target parameter value, and corresponding modification measure. The list generation process is based on the modification combination scheme corresponding to the Pareto optimal solution. The performance requirements of each modification measure are broken down into specific component parameter change values. The spatial code is matched with the corresponding BIM component's IFC entity to complete the one-to-one mapping between parameters and entities.

[0076] After the bill of quantities is generated, parameter validity is automatically checked. The system compares the target parameter values ​​against the physical boundary ranges corresponding to the component type, verifying whether they conform to the standard engineering range. Items exceeding the boundaries are automatically marked and prompted for adjustment. Bill of quantities that pass the check can be directly integrated with mainstream BIM modeling software, updating component parameters in batches through code matching. For example, a single bill of quantities item might be: Component spatial code EL00030008000012, IFC global ID 2O2Bkf8mzDfQe7bF9aX8sZ, parameter name heat transfer coefficient, current parameter value 2.7W / (m³). 2 • K), Target value of parameter 2.2W / (m 2 •K), corresponding renovation measures include upgrading the insulated glass to Low-E insulated glass.

[0077] Establish mapping rules between carbon efficiency information and IFC semantic entities, and back-mount carbon efficiency evaluation results, anomaly markers and remediation scheme parameters to the corresponding spatial index grid and component entities of the BIM model in the form of extended semantic attributes, and write them into the CarbonEfficiency_Pset attribute set in a unified manner; realize the spatial display of carbon efficiency distribution through color gradient rendering in the BIM visualization environment, and support interactive query of detailed evaluation data and remediation scheme by clicking on the component; The mapping rules match information types with entity levels: carbon efficiency evaluation results are mapped to the corresponding IFC entities at four spatial granularities; the overall annual carbon efficiency and carbon efficiency level of the building are mapped to the IfcBuilding entity; the monthly carbon efficiency and deviation rate of each floor are mapped to the IfcBuildingStorey entity; and the hourly carbon efficiency and contribution ratio of each room and component are mapped to the corresponding IfcSpace and IfcBuildingElement entities. Anomaly marker information is directly mapped to the spatial grid and component entities of the corresponding anomaly points, with attribute fields for anomaly degree and anomaly type. The parameters of the renovation plan are mapped to the corresponding component entities to be renovated, including three types of attribute fields: parameters before renovation, expected parameters after renovation, and parameter adjustment amount.

[0078] All the extended attributes are incorporated into the CarbonEfficiency_Pset attribute set of the dedicated semantic ontology for carbon efficiency evaluation, and remain independent of the original IFC standard attributes. During the attachment process, the target entity is automatically located by matching the spatial encoding with the global identifier of the IFC entity, and the entity association relationship of the semantic ontology is updated synchronously after the attribute assignment is completed.

[0079] A model fidelity decay early warning mechanism is set up. After each monthly data iteration, the deviation between the BIM model parameters and the actual working conditions is calculated. When the deviation exceeds the threshold, the embedded parameter calibration of the BIM model is triggered.

[0080] The model fidelity deviation calculation encompasses two independent dimensions: the first dimension is energy consumption prediction deviation, calculated based on the positive prediction results of the graph neural network bidirectional mapping model and the measured energy consumption data from the same period, to determine the average relative deviation between the overall building and each zone; the second dimension is parameter deviation, calculated by comparing the embedded parameters of the current BIM model with the initial design parameter values ​​to determine the average offset of core thermal parameters and equipment parameters. The deviation results for both dimensions are compared with corresponding thresholds: an 8% threshold for energy consumption prediction deviation and a 10% threshold for parameter deviation. Exceeding either threshold triggers an alert.

[0081] The early warning system operates in two levels: Level 1 warnings occur when a single-dimensional deviation exceeds a threshold, triggering targeted calibration of the corresponding dimension's parameters and updating only the parameter category for the deviation exceeding the limit; Level 2 warnings occur when deviations in two dimensions simultaneously exceed the threshold, triggering full calibration of all embedded parameters in the BIM model and performing a complete reverse iterative calibration process on all calibrable parameters. After each calibration, the fidelity deviation is recalculated, and the warning is lifted once the deviation falls back to within the threshold range. For example, after 24 months of operation in this project, the energy consumption prediction deviation rose to 9.1%, triggering a Level 1 warning. After targeted calibration of the building envelope's thermal parameters, the deviation fell back to 6.8%, and the warning was lifted.

[0082] The modified expected carbon efficiency parameters are fed back into the dedicated semantic ontology for carbon efficiency evaluation to update the model's baseline parameters.

[0083] The semantic ontology's baseline parameter updates cover three core categories: The first category consists of component-level baseline carbon emission factors and thermal parameter baseline values, which are updated based on the embedded parameters of the BIM model after reverse calibration, replacing the default baseline values ​​for the corresponding component types in the semantic ontology. For example, after calibration, the baseline value of the heat transfer coefficient of a 200mm aerated concrete exterior wall is updated from 0.78 to 0.85 W / (m²). 2The first category is the carbon efficiency level determination benchmark parameters, which are updated synchronously with the iteration results of the dynamic carbon efficiency benchmark line to adjust the carbon efficiency range thresholds corresponding to each level; the second category is the carbon efficiency improvement benchmark value of the modification measures, which is updated according to the calculation results of the modification scheme to enrich the performance reference database of modification measures in the semantic ontology, such as supplementing the unit area emission reduction benchmark value of Low-E window modification.

[0084] The parameter update trigger scenarios include three categories: completing a full model calibration, completing the reconstruction of the carbon efficiency baseline, and completing the calculation of the renovation plan. Each update retains the baseline parameter records of the historical version, supports the retrospective comparison of baseline values ​​at different times, and the updated baseline parameters are synchronized to all associated spatial index grids and component entities.

[0085] This embodiment also provides a green building energy-saving carbon efficiency evaluation system based on the fusion of BIM and energy consumption data, used to realize the full-process carbon efficiency evaluation of the above-mentioned office building scenario, including the following modules: The semantic extension and spatial subdivision module extends the input IFC4 format office building model with a dedicated semantic ontology for carbon efficiency evaluation, derives the IfcCarbonEfficiencyExtension extended entity, and subdivides the BIM space according to a four-level granularity rule of building as a whole - floor zoning - room unit - component individual, generates a hierarchical spatial index grid, establishes a three-segment fixed-length spatial coding rule and binds it to the corresponding BIM component attribute set. The data registration and decoupling module accesses multi-source data from sub-meters, heat and cold sources, sensors and weather stations, completes hourly interpolation alignment and outlier repair, matches spatial index grid encoding to mark the spatial affiliation of energy consumption data, decouples to obtain three types of energy consumption feature components, and establishes a precise mapping relationship between energy consumption data and spatial index grid. The bidirectional mapping network module constructs a graph attention network bidirectional mapping model. The forward branch outputs the predicted value of energy consumption distribution of each spatial unit in time period, and the reverse branch iteratively calibrates the embedded parameters of thermal and equipment classes of the BIM model along the spatial topology relationship. The calibration process applies engineering physical boundary constraints. The benchmark generation module uses a dual sliding incremental learning framework with a 90-day update window and a 30-day calibration window to generate a dynamic carbon efficiency benchmark. A cosine similarity working condition drift detection mechanism is set to trigger benchmark reconstruction. The benchmark confidence weight is calculated by combining working condition similarity, sample size, and data quality and incorporated into the carbon efficiency level determination. The anomaly tracing and attribution module locates the carbon efficiency anomaly points at four levels of granularity and tracks the spatial transmission path. It uses the logarithmic average Dichroic exponential decomposition algorithm to quantify the carbon efficiency contribution ratio of each unit, distinguishes between structural and operational carbon efficiency differences, and outputs targeted location results. The renovation scheme optimization module matches the anomaly type to construct energy-saving renovation scenario combinations with multiple intensity levels, calculates the carbon efficiency improvement curve based on the calibrated bidirectional mapping model, solves the Pareto optimal solution set for the dual objectives of life cycle cost and carbon emission reduction, and generates a parameter adjustment list that can be directly connected to BIM software. The semantic mounting closed-loop module establishes hierarchical mapping rules between carbon efficiency information and IFC semantic entities, and reverse-mounts evaluation results, anomaly markers and transformation parameters to the corresponding entities in the BIM model. It sets up a two-level model fidelity decay early warning mechanism, feeds back transformation parameters to the dedicated semantic ontology for carbon efficiency evaluation and dynamically updates the benchmark parameters.

[0086] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A green building energy-saving carbon efficiency evaluation method based on the fusion of BIM and energy consumption data, characterized in that, The specific steps include the following: Extend the standard IFC data model with a dedicated semantic ontology for carbon efficiency evaluation, segment the BIM space to generate a spatial index grid and bind the corresponding BIM component attributes; Collect energy consumption data, complete time-series preprocessing and spatial attribution labeling, decouple to obtain multiple types of energy consumption feature components, and establish a mapping relationship between the energy consumption data and the spatial index grid; Combining the attributes and topological adjacency of the spatial index grid with the decoupled multi-type energy consumption feature components, a graph neural network bidirectional mapping model is constructed; the forward branch outputs the time-period predicted energy consumption distribution value of the corresponding spatial unit, and the reverse branch reversely calibrates the embedded parameters of the BIM model under parameter constraints. Combining historical operational datasets with meteorological parameters from the same period, an incremental learning framework is used to generate a carbon efficiency baseline. A working condition drift detection mechanism is set up to trigger the reconstruction of the carbon efficiency baseline. The baseline confidence weight is calculated and incorporated into the carbon efficiency level determination process. Combining the carbon efficiency baseline with the output of the graph neural network bidirectional mapping model, the abnormal carbon efficiency points are located level by level according to building level and the spatial transmission path is tracked. The carbon efficiency contribution ratio of each unit is quantified, the structural and operational carbon efficiency differences are distinguished, and the location results are output. Energy-saving renovation scenario combinations are constructed to address weak links in carbon efficiency. The carbon efficiency improvement curve is calculated based on the calibrated graph neural network bidirectional mapping model. The optimal solution set for the dual objectives of renovation cost and carbon efficiency improvement is solved, and a corresponding BIM parameter adjustment list is generated. Establish mapping rules between carbon efficiency information and IFC semantic entities, reverse-mount carbon efficiency evaluation results and transformation parameters to the BIM model, set a model fidelity attenuation early warning mechanism, and update the baseline parameters of the carbon efficiency evaluation-specific semantic ontology.

2. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 1, characterized in that, Derive a carbon efficiency attribute extension entity from the IfcBuildingElement entity subclass to supplement building components with extended attribute fields; The BIM space is divided according to a four-level granularity rule of building as a whole, floor zoning, room unit, and individual component, and the spatial index grid is generated. Each level of spatial index grid follows the boundary division rules of the corresponding level, and the spatial coding rules are kept consistent. Establish a three-segment spatial coding rule that includes IFC entity type identifier, spatial level identifier, and unique sequence identifier. Each grid cell is bound to a unique spatial code and the corresponding BIM component attribute set to achieve a strong association mapping between the grid and the IFC entity.

3. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 2, characterized in that, Collect energy consumption data from building sub-meters, cold and heat source systems, environmental sensors, and meteorological data from weather stations. Based on a unified time reference, complete the interpolation and alignment of time-series data and outlier repair. Match the encoding of the spatial index grid with the spatial coordinates of the energy consumption collection points and mark the spatial affiliation. The original energy consumption sequence is decoupled into three types of energy consumption characteristic components by feature decoupling algorithm: climate correlation, base load, and operating condition drive, which correspond to three types of energy consumption driving factors: outdoor meteorological influence, building thermal performance, and energy consumption behavior of people and equipment.

4. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 3, characterized in that, The forward branch uses the geometric properties and thermal parameters of the spatial index grid as graph node features and the topological adjacency relationship as graph edge features to output the time-period energy consumption distribution prediction value of the corresponding spatial unit. A gradient constraint calibration mechanism is set up in the reverse branch. The measured energy consumption time series characteristics are used as input to iteratively calibrate the embedded parameters of the BIM model in reverse along the spatial topology relationship. The physical boundary constraints of the parameters are set during the calibration process.

5. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 4, characterized in that, A dual sliding window incremental learning framework is adopted, and a benchmark update window and a confidence calibration window are set to generate the carbon efficiency baseline; when the similarity between the current operating condition and the historical benchmark operating condition is lower than a set threshold, the carbon efficiency baseline reconstruction is triggered. The confidence weights of the baselines for each time period are calculated based on the similarity of operating conditions, the amount of data sample, and the data quality. These confidence weights are directly used to determine the carbon efficiency level range.

6. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 5, characterized in that, Determine the deviation of the overall carbon efficiency of the building from the carbon efficiency baseline and the corresponding confidence weight correction level range, and locate the carbon efficiency anomaly points of floor zones, room units and individual components step by step, and simultaneously track the spatial transmission path of carbon efficiency anomalies. A dual-dimensional contribution decomposition algorithm based on structure and operation is used to quantify the carbon efficiency contribution ratio of each spatial unit and energy consumption item, and the output includes the target localization results including the location of anomalies, the degree of anomaly, the contribution ratio, and the type of difference.

7. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 6, characterized in that, Construct a combination of energy-saving renovation scenarios that include building envelope modification, equipment system performance optimization, and operation strategy optimization; Based on the calibrated graph neural network bidirectional mapping model, the carbon efficiency improvement curves under different modification intensities are simulated and calculated to quantify the marginal diminishing effect of each measure. The Pareto optimal solution set is obtained with the dual objectives of improving retrofit cost and carbon efficiency, and the optimal retrofit combination and corresponding BIM parameter adjustment list under different budget constraints are output.

8. The green building energy-saving carbon efficiency evaluation method based on BIM and energy consumption data fusion according to claim 7, characterized in that, The carbon efficiency evaluation results, anomaly markers, and remediation scheme parameters are back-mounted to the spatial index grid and component entities corresponding to the BIM model in the form of extended semantic attributes, so as to realize the spatial display and interactive query of carbon efficiency distribution. After each data iteration, the deviation between the BIM model parameters and the actual working conditions is calculated. When the deviation exceeds the threshold, the full calibration of the embedded parameters in the BIM model is triggered. The expected carbon efficiency parameters after the transformation are fed back to the dedicated semantic ontology for carbon efficiency evaluation, and the benchmark parameters of the dedicated semantic ontology for carbon efficiency evaluation are updated.

9. A green building energy-saving carbon efficiency evaluation system based on the fusion of BIM and energy consumption data, used to implement the method described in any one of claims 1-8, characterized in that, Includes the following modules: The semantic extension and spatial subdivision module is used to extend the standard IFC data model with a dedicated semantic ontology for carbon efficiency evaluation, subdivide the BIM space to generate a spatial index grid and bind BIM component attributes. The data registration and decoupling module is used to complete the time-series preprocessing and spatial attribution labeling of energy consumption data, decouple the energy consumption feature components, and establish the mapping relationship between the energy consumption data and the spatial index grid. The bidirectional mapping network module is used to construct a graph neural network bidirectional mapping model, which outputs the predicted value of the energy consumption distribution of the corresponding spatial unit in the forward direction and calibrates the embedded parameters of the BIM model in the reverse direction. The benchmark generation module is used to generate a carbon efficiency benchmark using an incremental learning framework, set a working condition drift detection mechanism to trigger the reconstruction of the carbon efficiency benchmark and calculate the benchmark confidence weight. The anomaly tracing and attribution module is used to locate carbon efficiency anomaly points by level, track spatial transmission paths, quantify the carbon efficiency contribution ratio, and distinguish the types of carbon efficiency differences. The renovation scheme optimization module is used to construct energy-saving renovation scenario combinations, calculate carbon efficiency improvement curves, solve for the optimal solution set of the two objectives, and generate a BIM parameter adjustment list. The semantic mounting closed-loop module is used to reverse mount the evaluation results and transformation parameters to the BIM model, set a model fidelity decay early warning mechanism, and update the benchmark parameters of the carbon efficiency evaluation-specific semantic ontology.

10. A green building energy-saving carbon efficiency evaluation system based on the fusion of BIM and energy consumption data as described in claim 9, characterized in that, The forward branch of the bidirectional mapping network module uses the geometric properties and thermal parameters of the spatial index grid as graph node features and the topological adjacency relationship as graph edge features to output the time-period energy consumption distribution prediction value of each spatial unit. The reverse branch sets up a gradient constraint calibration mechanism and parameter physical boundary constraints, and iteratively calibrates the embedded parameters of the BIM model in reverse along the spatial topology.

Images