Building energy efficiency optimization method and system based on digital twinning and multi-agent

Abstract

This application provides a method and system for building energy efficiency optimization based on digital twins and multi-agent systems, relating to the field of building energy efficiency management. The method includes: acquiring building information of the target building to construct a digital twin model and generating a spatialized point database; constructing a hierarchical metering system with multiple levels, binding the acquired real-time energy consumption data with corresponding spatial identifiers in the spatialized point database to generate spatialized energy consumption data; deploying corresponding agents for each control area, generating a regional state field, and establishing a data interaction channel with the digital twin basic model; using the digital twin model to simulate and verify candidate control strategies, obtaining simulation verification results; calculating the spatial correlation between spatial identifiers in the spatialized point database, and coordinating the simulation verification results based on the spatial correlation to generate a control strategy. This solves the technical problem of lacking refined sub-metering and making accurate energy consumption positioning difficult.

Description

Technical Field

[0001] This application relates to the field of building energy efficiency management, and in particular to a method and system for optimizing building energy efficiency based on digital twins and multi-agent systems. Background Technology

[0002] In the field of building energy efficiency management, energy monitoring and optimization technologies for public buildings such as office buildings, commercial complexes, hotels, hospitals, and schools are the core means to achieve building energy conservation and reduce energy costs. Existing technologies generally adopt a system-level centralized metering method for building energy consumption monitoring. This involves setting up metering equipment at the main incoming line of the building and at the main terminals of a few major energy-consuming systems. Only system-level energy consumption data such as total building energy consumption, total lighting system energy consumption, and total air conditioning system energy consumption are collected. Therefore, in actual use, only the total energy consumption value of each energy-consuming system can be output. It is impossible to break down the energy consumption data of different floors, different functional areas, different power distribution circuits, or even key terminal equipment within each system. When problems such as abnormally high energy consumption or energy waste occur, it can only be determined that the abnormality occurs in a certain overall system such as lighting or air conditioning, and it is impossible to further locate the specific energy consumption abnormal circuit, equipment, or area. For example, when the total energy consumption of the lighting system is high, it's impossible to distinguish whether the energy waste is due to office lights on a certain floor being constantly on, emergency lighting malfunctioning, or improper operation of landscape lighting. Similarly, when the energy consumption of the socket system is abnormal, it's impossible to identify whether the problem stems from excessive standby power consumption of office equipment in meeting rooms, leakage in socket wiring in a certain area, or prolonged operation of charging equipment in public areas. Due to the lack of granular, itemized metering data, maintenance personnel can only find the root cause through manual on-site inspections. This not only results in extremely low inspection efficiency but also reduces the accuracy of energy consumption diagnosis due to limitations in the scope and experience of manual inspections. Summary of the Invention

[0003] This application provides a building energy efficiency optimization method and system based on digital twins and multi-agent systems, which solves the technical problem that existing technologies lack refined sub-metering and make it difficult to accurately locate energy consumption.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a building energy efficiency optimization method based on digital twins and multi-agent systems includes: acquiring building information of the target building to construct a digital twin model, assigning a unique three-dimensional spatial code to each monitoring point, and generating a spatialized point database. Based on the building's power distribution system, a hierarchical metering system with multiple levels is constructed, and according to the physical location of the monitoring points, the acquired real-time energy consumption data is bound to the corresponding spatial identifiers in the spatialized point database to generate spatialized energy consumption data. Based on the digital twin model and the spatialized point database, corresponding agents are deployed for each control area. Each agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model. Each agent generates candidate control strategies based on the regional state field and uses the digital twin model to simulate and verify the candidate control strategies, obtaining simulation verification results. Each agent calculates the spatial correlation between itself based on the spatial identifiers in the spatialized point database, and coordinates the simulation verification results based on the spatial correlation to generate control strategies.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the digital twin model specifically includes: acquiring the building information of the target building to construct a digital twin model, the building information including the BIM model, thermal parameters of the building envelope, internal space division, and equipment and facility layout. Based on the BIM model, a unified building spatial coordinate system is established, each monitoring point is assigned a unique three-dimensional spatial code, and the three-dimensional spatial codes and corresponding associated attribute information of all monitoring points are summarized to generate a spatialized point database. Based on the thermal parameters of the building envelope and the spatialized point database, a physical model containing building thermal zones is constructed, thermal parameters are configured for each thermal zone, and the physical model is embedded within the digital twin model. Based on the three-dimensional spatial codes in the spatialized point database, historical spatialized energy consumption time-series data of all monitoring points within the same spatial area are extracted to form graph structure data with spatial topological relationships, a behavioral model is constructed, and the behavioral model is embedded within the digital twin model.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the sub-metering system specifically includes: performing topology analysis on the power distribution system of the target building to construct a two-level metering mapping model containing physical circuit levels and spatial function levels. The physical circuit level, based on the outgoing circuits of the distribution cabinet, divides the system into four system-level metering branches: lighting, air conditioning, sockets, and power. A dedicated sub-metering integrated module is deployed at the end of each branch to independently collect the raw energy consumption data of each circuit. The spatial function level, based on the three-dimensional spatial coding in the spatialized point database, categorizes all energy-consuming devices associated with monitoring points within the same spatial area, forming functional area-level metering units. The raw energy consumption data collected at the physical circuit level is decoupled by the recognition unit built into the dedicated sub-metering integrated module, decomposing the energy consumption components of different energy-consuming device types within the circuit. Based on the two-level metering mapping model, the decoupled energy consumption components are re-aggregated with the corresponding functional areas in the spatial function level to generate a sub-item energy consumption dataset decomposed according to both spatial function and energy-consuming device type.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves binding the acquired real-time energy consumption data to the corresponding spatial identifiers in the spatialized point database based on the physical location of the monitoring points. Specifically, this includes: each monitoring point collecting real-time energy consumption data at a preset frequency, and converting heterogeneous energy consumption data output from different metering devices into standard data packets via a multi-protocol conversion unit. The standard data packets are then encapsulated with their own preset or automatically acquired 3D spatial codes to generate standardized data messages with spatial tags. The standardized data messages are parsed, and using the 3D spatial code as a unique key, a mapping relationship between the time-series database and the spatialized point database is established. The real-time energy consumption values ​​are written into the data table corresponding to the spatial code, generating spatialized energy consumption time-series data containing precise spatial location information. Finally, the spatialized energy consumption time-series data containing precise spatial location information is synchronized to the digital twin model, driving the updating of the corresponding 3D primitives in the digital twin model.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the intelligent agent specifically includes: dividing the building physical space into multiple three-dimensional voxel-based control units based on the thermal zoning of the digital twin model and the three-dimensional spatial encoding in the spatialized point database, with each three-dimensional voxel-based control unit corresponding to a deployed edge intelligent agent. The edge intelligent agent includes an environmental perception fusion module, a graph neural network decision module, and a twin collaboration interface. The environmental perception fusion module acquires real-time spatialized energy consumption data and environmental parameters of all monitoring points within its managed three-dimensional voxel-based control unit, uses a three-dimensional spatial interpolation algorithm to perform spatial continuity processing on the discrete point data, generates a regional three-dimensional state field reflecting the multi-dimensional distribution of temperature, humidity, illuminance, and energy consumption fields within the unit, and synchronizes it to the digital twin model for visualization mapping. The graph neural network decision module takes the regional three-dimensional state field as input, extracts the spatial gradient features and temporal evolution features in the regional three-dimensional state field through graph convolution operations, generates a candidate control strategy set by combining a preset multi-objective optimization function, and encapsulates the candidate control strategy set and the spatial identifiers of its associated three-dimensional voxel-based control units through the twin collaboration interface. The twin-like collaborative interface calculates the spatial correlation degree with neighboring edge agents based on the 3D spatial encoding in the spatialized point database. When the spatial correlation degree exceeds a preset collaborative threshold, it sends a set of candidate control strategies to the corresponding neighboring edge agents and receives the set of candidate control strategies sent by the neighboring edge agents. The graph neural network decision module evaluates the coupling effect of the boundary influence of the received candidate control strategy set on the 3D state field of its own region. With the goal of minimizing global energy consumption and inter-regional comfort deviation, it coordinates the candidate control strategy set to generate regional autonomous control commands.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves each agent calculating its spatial correlation with others based on spatial identifiers in a spatialized point database, and then coordinating simulation verification results based on this spatial correlation. Specifically, this includes: acquiring the three-dimensional spatial codes of all monitoring points within the jurisdiction of each agent and neighboring agents, and extracting the three-dimensional coordinates of each point; calculating the spatial correlation between any two agents based on the three-dimensional coordinates using a spatial correlation calculation model; exchanging simulation verification results from neighboring agents whose spatial correlation exceeds a preset coordination threshold; and using the simulation verification results of neighboring agents as boundary constraints, with the goal of minimizing global energy consumption and comfort deviation, and generating the final control strategy through a distributed cooperative optimizer.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the expression for the spatial correlation calculation model is as follows: .in, Let be the spatial correlation degree between the i-th agent and the j-th agent. The preset spatial attenuation coefficient, The thermal impact distance is the corrected distance for the vertical thermal stratification effect of a building.

[0011] In conjunction with the first aspect mentioned above, one possible implementation also includes: acquiring real-time monitoring data from the dynamic calibration trigger, including model prediction bias, physical equipment change events, and seasonal environmental inflection points. When any monitored indicator exceeds a preset threshold, the calibration process is triggered. The digital twin model acquires the trigger signal, retrieves historical operational data from the previous time window, and retrains or calibrates the behavioral and physical models. The calibrated model parameters, along with corresponding calibration time, data period, trigger event, and other information, are encapsulated to generate a model version snapshot, which is then synchronized to the digital twin model for hot loading.

[0012] In conjunction with the first aspect mentioned above, one possible implementation involves retraining or calibrating the behavioral and physical models. Specifically, this includes: using spatialized energy consumption time-series data containing precise three-dimensional spatial coordinates to update the weights of the behavioral model via a spatiotemporal graph convolutional neural network, thereby improving the prediction accuracy of human activity patterns and equipment usage patterns in different spatial regions; and using spatialized energy consumption data and synchronously collected environmental parameters to inversely correct the thermal or electrical parameters of the physical model via Bayesian inference methods, thereby minimizing the overall error between the model output and the measured data.

[0013] Secondly, a building energy efficiency optimization system based on digital twins and multi-agent systems is provided, including: a digital twin model construction unit, used to acquire building information of the target building to construct a digital twin model, and assign a unique three-dimensional spatial code to each monitoring point to generate a spatialized point database; a sub-metering system and data mapping unit, used to construct a hierarchical metering system with multiple levels based on the building's power distribution system, and bind the acquired real-time energy consumption data with the corresponding spatial identifiers in the spatialized point database according to the physical location of the monitoring points to generate spatialized energy consumption data; a multi-agent cluster construction unit, used to deploy corresponding agents for each control area based on the digital twin model and the spatialized point database. Each agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model; and a simulation verification unit, used by each agent to generate candidate control strategies based on the regional state field, and call the digital twin model to simulate and verify the candidate control strategies to obtain simulation verification results. The digital twin intelligent agent collaborative decision-making unit is used by each intelligent agent to calculate the spatial correlation between each other based on the spatial identifiers in the spatialized point database, and to coordinate the simulation verification results based on the spatial correlation to generate control strategies.

[0014] This application provides a building energy efficiency optimization method and system based on digital twins and multi-agent systems. It can construct a multi-level hierarchical metering system covering the building's main entrance, system level, circuit level, and equipment level according to the building's power distribution system. This system enables independent metering of all energy-consuming systems, such as lighting, air conditioning, sockets, and power, from the main end to the end circuits and key equipment. This changes the traditional situation where only total system-level energy consumption data can be collected. It enables fine-grained decomposition and collection of energy consumption data, solving the problem of coarse metering granularity in traditional methods that cannot distinguish the energy consumption of specific energy-consuming units within the system. This achieves the technical effect of refined sub-item metering of building energy consumption at all levels. Simultaneously, each monitoring point in the target building is assigned a unique three-dimensional spatial code, and a spatialized point database is established. This precisely binds the real-time collected energy consumption data with the spatial identifiers corresponding to the physical locations of the monitoring points, generating spatialized energy consumption data. Each set of energy consumption data is associated with precise XYZ spatial coordinates, the floor, area, or equipment, achieving a one-to-one mapping between energy consumption data and physical space and energy-consuming equipment. This solves the problems of traditional energy consumption data being disconnected from spatial location and the inability to accurately locate anomalies. It achieves the technical effect of accurately linking energy consumption data with physical space and directly locating energy consumption anomalies to specific circuits, equipment, or areas. The above technical solutions are interconnected, first achieving refined energy consumption data collection, then completing the spatial binding of energy consumption data, forming a complete technical link from data collection to data association. This transforms energy consumption diagnosis from vague system-level judgments to precise circuit or equipment-level location, eliminating the need for manual on-site inspections and significantly improving the efficiency and accuracy of energy consumption diagnosis.

[0015] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0016] Figure 1 A system architecture diagram of the building energy efficiency optimization method based on digital twins and multi-agents provided in the embodiments of this application; Figure 2A flowchart illustrating the digital twin model construction process in the building energy efficiency optimization method based on digital twins and multi-agents provided in this application embodiment; Figure 3 A flowchart illustrating the process of constructing a sub-metering system in the building energy efficiency optimization method based on digital twins and multi-agents provided in this application embodiment; Figure 4 A flowchart illustrating the agent construction process in the building energy efficiency optimization method based on digital twins and multi-agents provided in the embodiments of this application; Figure 5 A schematic diagram of the structure of a building energy efficiency optimization system based on digital twins and multi-agents provided in an embodiment of this application. Detailed Implementation

[0017] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0018] like Figure 1 As shown in the embodiments of this application, the building energy efficiency optimization method based on digital twins and multi-agent systems includes: Step 101: Obtain the building information of the target building, construct a digital twin model, and assign a unique three-dimensional spatial code to each monitoring point to generate a spatialized point database.

[0019] The building information includes data on building structure, envelope parameters, internal space division, and equipment and facility layout. The 3D spatial code is a unique identifier assigned to each monitoring point, generated based on a unified building spatial coordinate system, containing the precise XYZ coordinates of that point and its spatial attributes such as the floor and area it belongs to. The spatialized point database stores the 3D spatial codes of all monitoring points and their associated attribute information, used to achieve accurate mapping between energy consumption data and physical space.

[0020] In some implementation methods, a unified architectural spatial coordinate system is first established based on the BIM model of the target building or on-site survey data, determining the origin and coordinate axis directions. Then, information such as the building structure, envelope, interior space division, and equipment layout is extracted to construct the geometric framework of the digital twin model. All monitoring points, including sensors and metering points, are then assigned precise XYZ coordinates in the digital twin model based on their actual physical location, and a unique three-dimensional spatial code is generated according to preset coding rules. This code includes information such as building ID, floor, area, equipment type, and serial number. Finally, the three-dimensional spatial codes of all monitoring points, along with their associated equipment type, system, metering level, and other attribute information, are summarized to construct a spatialized point database.

[0021] For example, in an office building project, the BIM model of the building is first imported, and a spatial coordinate system is established with the southwest corner of the building's boundary line as the origin. Then, a three-dimensional coordinate system (X=125.45, Y=78.32, Z=15.60) is assigned to the temperature sensor deployed in the west office area on the 5th floor within the model, and a unique three-dimensional spatial code "B3-05-ZONE-A-TEMP-12_X=125.45_Y=78.32_Z=15.60" is generated according to the coding rules. This code, along with its associated device type "temperature sensor" and system "HVAC," is stored in a spatialized point database to bind the sensor's real-time temperature data to that spatial location.

[0022] Step 102: Based on the building power distribution system, construct a hierarchical metering system with multiple levels, and bind the acquired real-time energy consumption data with the corresponding spatial identifiers in the spatialized point database according to the physical location of the monitoring points to generate spatialized energy consumption data.

[0023] The hierarchical metering system refers to an energy consumption metering architecture built upon the building's power distribution system topology, encompassing multiple levels such as building entrance, system level, circuit level, and equipment level. Spatial identifiers are unique three-dimensional spatial codes assigned to each monitoring point in the spatialized point database. Spatialized energy consumption data is a dataset generated by binding real-time energy consumption data with the spatial identifiers of monitoring points, containing both energy consumption values ​​and precise spatial location information.

[0024] In some implementation methods, a detailed survey of the building's power distribution system is conducted, and the system levels such as lighting, air conditioning, sockets, and power are divided according to the outgoing circuits of the distribution cabinet. Metering equipment of corresponding accuracy is deployed at each level to construct a four-level sub-item metering system from the main incoming line of the building to the terminal equipment circuit.

[0025] Then, real-time energy consumption data for each metering point is acquired at a preset time window and collection frequency. Based on the physical location of the monitoring point, the corresponding three-dimensional spatial code is matched as a spatial identifier in the pre-constructed spatialized point database. The data mapping engine then automatically binds the real-time energy consumption value with the successfully matched spatial identifier, generating spatialized energy consumption data that includes energy consumption data and spatial attributes such as XYZ coordinates, floor, and region.

[0026] The automatic binding process includes: collecting real-time energy consumption data at a preset frequency through IPme®-DT digital twin smart modules deployed at each monitoring point. At this time, the real-time energy consumption data is converted into a standard data packet by the multi-protocol conversion unit built into the IPme®-DT module.

[0027] The IPme®-DT module then encapsulates the real-time energy consumption data, converted into standard data packets, with its own 3D spatial code (either pre-set or automatically acquired), generating standardized data packets with spatial tags. These packets are then sent to the central processing unit (CPU) via the edge network. Upon receiving the standardized data packets, the CPU parses out the real-time energy consumption data and the corresponding 3D spatial code. Using the 3D spatial code as a unique key, it searches the time-series database and establishes a mapping relationship with the spatialized point database. The real-time energy consumption value is written into the data table corresponding to the spatial code, generating spatialized energy consumption time-series data containing precise spatial location information. Simultaneously, the CPU synchronizes the updated spatialized energy consumption data to the digital twin model, driving the 3D primitives corresponding to the spatial location in the digital twin model to refresh their status, achieving real-time mapping and visualization of energy consumption data in virtual space.

[0028] For example, in an energy efficiency optimization project for a commercial complex, the technical team completed BIM modeling and monitoring point deployment. For instance, in the lighting distribution box in Zone A on floor B3, a smart circuit breaker supporting the Modbus protocol was installed for each outgoing circuit. Based on the BIM model, the precise coordinates of the center point of the distribution box and the corresponding lighting area for each circuit were extracted. At this point, for the circuit controlling the lighting in Zone A, a three-dimensional spatial code "B3-05-ZONE-A-LIGHT-01, X=125.45, Y=78.32, Z=15.60" can be directly generated and written into the IPme®-DT-Basic module connected to the smart circuit breaker using a configuration tool.

[0029] Once powered on, the module polls the smart circuit breaker's real-time power data every minute using the Modbus-RTU protocol. Simultaneously, the module's embedded protocol conversion unit converts the read Modbus data into a unified JSON format.

[0030] The module then encapsulates the converted data with its internally pre-defined 3D spatial code into a data packet, which is sent to the central server via the building's Wi-Fi network. After parsing the packet, the central server identifies the code "B3-05-ZONE-A-LIGHT-01..." and the power value "2.5kW," and writes the power value at the current timestamp into the InfluxDB time-series database, using this code as a tag. Next, the server pushes this data to a Unity3D-based digital twin visualization platform, driving an increase in the brightness of lighting elements in area A of fire compartment 5 on floor B3, and displaying a real-time data tag of "2.5kW" above them. When the power of this circuit abnormally rises to "5.0kW" one afternoon, the area is immediately highlighted in red in the digital twin model, triggering an alarm. Maintenance personnel can directly locate the potential lighting malfunction or uninterrupted lights in area A of floor B3 on the 3D map without on-site investigation.

[0031] Step 103: Based on the digital twin model and spatialized point database, deploy corresponding intelligent agents for each control area; each intelligent agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model.

[0032] The control area is an independent management unit defined based on the building's spatial layout, energy system characteristics, and management needs, with each area corresponding to an independent intelligent agent. The intelligent agent is a distributed intelligent decision-making unit deployed in each control area, possessing sensing, decision-making, coordination, and communication capabilities. Monitoring point data includes real-time environmental parameters and equipment status information collected by various sensors deployed within the control area (such as temperature, humidity, and illuminance sensors). The regional state field is a multi-dimensional data field generated by the intelligent agent through fusion and interpolation of all monitoring point data and spatialized energy consumption data within its jurisdiction in three-dimensional space, reflecting the overall situation of the region's physical environment and energy consumption distribution.

[0033] In some implementation methods, the target building must first be divided into multiple logical control areas based on the building's spatial layout, HVAC zoning, lighting circuit division, and energy consumption characteristics, such as by floor, by fire compartment, or by functional area.

[0034] This allows for the deployment of a corresponding agent within each control area, with each agent configured with its assigned area and a list of all monitoring points within that area. Each agent then continuously acquires real-time monitoring data (such as temperature, humidity, and CO2 concentration) and corresponding spatialized energy consumption data (such as lighting and air conditioning socket energy consumption) from a real-time database via a data interface. Based on this discrete point data, a spatial interpolation algorithm is used to construct a three-dimensional state field reflecting the continuous distribution of temperature, humidity, and energy consumption fields within the three-dimensional spatial framework provided by the digital twin model.

[0035] At the same time, each intelligent agent establishes a two-way data interaction channel with the digital twin basic model. On the one hand, the generated regional state field is synchronized to the digital twin model in real time for visualization. On the other hand, the boundary state information of adjacent areas and the physical constraint parameters of the building as a whole can be obtained from the digital twin model.

[0036] For example, in an office building project, the western office area on the 5th floor is designated as an independent control zone, and an "HVAC-Agent-05" intelligent agent is deployed. This agent acquires real-time monitoring data from six temperature sensors, three CO2 sensors, and spatialized energy consumption data for lighting and socket circuits within this area. Based on this data and its three-dimensional coordinates (e.g., TEMP-12 point X=125.45, Y=78.32, Z=15.60), the agent uses the Kriging interpolation algorithm to construct a three-dimensional temperature field distribution map of the western office area on the 5th floor within the three-dimensional framework provided by the digital twin model.

[0037] Meanwhile, the intelligent agent establishes a connection with the digital twin base model through the WebSocket interface, pushes the generated temperature field data to the model for rendering in real time, and can obtain the boundary temperature data of the adjacent eastern office area from the model to optimize its own control strategy.

[0038] Step 104: Each agent generates candidate control strategies based on the regional state field, and uses a digital twin model to simulate and verify the candidate control strategies, obtaining simulation verification results.

[0039] Simulation verification refers to the process of executing candidate control strategies in a digital twin model and predicting their impact on building energy consumption, environmental comfort, and equipment safety through multiphysics coupling simulation. The simulation verification result is a dataset output after the simulation verification is completed, containing predicted energy consumption impact values, comfort impact indicators, and safety boundary verification results.

[0040] In some implementations, each multi-agent generates multiple candidate control strategies based on its current 3D state field and a pre-defined optimization objective function, using reinforcement learning or a rule engine. These strategies could include adjusting the air conditioner's set temperature, regulating lighting brightness, or optimizing the opening of the fresh air valve. Then, each agent inputs the parameters of these candidate control strategies into the digital twin model through established data interaction channels.

[0041] Once the digital twin model receives the strategy parameters, it calls its built-in thermal model, lighting model, airflow model, and electrical model to perform multiphysics coupling simulation in parallel in the virtual space. This simulation simulates the temperature field changes, illuminance distribution, energy consumption curves, and equipment operating status of various areas of the building after the strategy is executed.

[0042] During the simulation process, the digital twin model will iterate and calculate at a preset time step until it reaches a steady state or the preset simulation duration.

[0043] After the simulation, the digital twin model will output the simulation verification results for each candidate strategy, including the predicted energy consumption reduction, temperature fluctuation range, PMV-PPD comfort index, number of equipment start-ups and shutdowns, and safety boundary information such as whether current or temperature limits are triggered. Simultaneously, each agent will receive these simulation verification results as the basis for subsequent collaborative decision-making and strategy selection.

[0044] For example, a heating, ventilation, and air conditioning (HVAC) agent in the west office area on the 5th floor of an office building generates two candidate strategies based on the current temperature field (showing that the temperature near the window is relatively high): Strategy A is to lower the air conditioning set temperature in this area from 24℃ to 23℃; Strategy B, based on Strategy A, simultaneously lowers the west-facing blinds by 30% to block afternoon sunlight. The agent then sends the parameters of these two strategies to the digital twin model.

[0045] The digital twin model used thermal and lighting models for a 30-minute simulation. The simulation results for Strategy A were: the average regional temperature decreased to 23.5℃, energy consumption increased by 8%, and the temperature near the window was still 0.8℃ higher. The simulation results for Strategy B were: the average regional temperature decreased to 23.2℃, energy consumption increased by only 3%, and temperature field uniformity improved. After receiving these two simulation results, the HVAC intelligent system assessed that Strategy B was superior in terms of both energy consumption and overall comfort.

[0046] Step 105: Each agent calculates the spatial correlation between itself based on the spatial identifiers in the spatialized point database, and coordinates the simulation verification results according to the spatial correlation to generate a control strategy.

[0047] Specifically, this includes: reading a spatialized point database from each agent, obtaining the 3D spatial codes of all monitoring points within its own and neighboring agent's jurisdiction, and extracting the precise 3D coordinates of each point. Based on the extracted 3D coordinates, a spatial correlation calculation model is used. Calculate the spatial correlation between any two agents ,in, Let be the spatial correlation degree between the i-th agent and the j-th agent. The preset spatial attenuation coefficient, The heat-affected distance after correction for vertical thermal stratification of buildings, through The calculation yielded, where and The three-dimensional coordinates of the i-th and j-th agents are respectively. The preset vertical thermal coupling coefficient, This refers to the standard floor height of a building.

[0048] Simultaneously, each intelligent agent exchanges its generated simulation verification results with the edge computing gateways corresponding to neighboring intelligent agents whose spatial correlation exceeds a preset cooperation threshold via industrial Ethernet or wireless mesh networks, achieving state information sharing between adjacent areas in physical space. Each edge computing gateway then invokes its built-in collaborative optimization control logic to iteratively optimize candidate control parameters for actuators such as air conditioning terminals, lighting circuits, and electric dampers within its jurisdiction, aiming to minimize global energy consumption and inter-regional comfort deviations. Finally, each edge computing gateway packages the optimized control parameters to generate the final control command, which is then sent to the corresponding air conditioning terminal devices, intelligent lighting controllers, or electric regulating valves via its built-in relay output module or standard communication interface (such as RS485, BACnetMS / TP), driving their mechanical structures to perform corresponding actions and feeding back the execution status to the digital twin model.

[0049] Spatial correlation refers to the quantitative index used by two multi-agents to reflect their physical proximity and potential mutual influence intensity, calculated based on the spatial identifiers of all monitoring points within their jurisdiction. It is usually calculated using three-dimensional Euclidean distance and after considering the vertical thermal stratification effect.

[0050] In some implementations, the collaborative decision-making phase officially begins after the deployment of agents in each control area is completed, and each agent has generated its own candidate control strategies and their simulation verification results in the digital twin model. At this point, each agent (e.g., the agent responsible for the HVAC of the 5th floor west zone)... iIt will proactively access the spatialized point database, and based on its known jurisdiction area, query and retrieve other intelligent agents that may be spatially adjacent to or related to that area (such as the Agent in the eastern area of ​​the same layer). j Agents upstairs and downstairs k Identifiers for (etc.).

[0051] Subsequently, Agent i The system reads the 3D spatial codes of all monitoring points within the jurisdiction of itself and its neighboring agents from the database, and extracts representative coordinates to represent the spatial location of each agent. These coordinates are typically the coordinates of the geometric center of its jurisdiction or the average coordinates of all sensor points within that area. Once the coordinate data is obtained, the 3D Euclidean distance between them is calculated using a formula, and then combined with a preset vertical thermal coupling coefficient γ and a standard floor height. ,pass Calculate the corrected heat-affected distance Then, the heat-affected distance and the preset spatial attenuation coefficient are used. Substitute into the formula Ultimately, the spatial correlation between the two agents is obtained. .

[0052] Agent i Repeat this process to calculate its spatial correlation with all relevant neighboring agents. After all calculations are complete, the Agent... i It will proactively send its simulation verification results (such as the predicted energy consumption reduction and temperature impact range of its optimal candidate strategy "adjusting the set temperature to 23℃") to those spatially correlated entities through the previously established data exchange channel. Neighboring agents that exceed a preset collaboration threshold (e.g., 0.5).

[0053] At the same time, Agent i It also receives simulation verification results from these neighboring agents. At this point, the Agent... i Possessing its own policy information and constraint information from neighboring regions, the system inputs this information, particularly the potential boundary effects of neighboring agents' policies on its local area (e.g., thermal radiation from the eastern region causing a 0.3°C increase in the local boundary temperature), as new boundary constraints into its local distributed cooperative optimizer. The optimizer then uses the objective function of "minimizing the combined energy consumption and comfort deviation between its local and neighboring regions while satisfying all boundary constraints" to perform rapid iterative calculations, solving a local cooperative optimization problem.

[0054] After solving the problem, the optimizer outputs an adjustment plan for the initial candidate strategy. For example, it might keep the original set temperature at 23°C, but simultaneously increase the airflow velocity by 10% to enhance local airflow and counteract the effects of external heat. Finally, the Agent... i The final control strategy parameters, after collaborative optimization, are encapsulated and written into a local execution queue, awaiting system scheduling. They are then distributed to the corresponding air conditioning equipment for execution via edge hardware modules.

[0055] For example, in a high-rise office building, HVAC agents covering each standard floor are deployed. Taking the 25th and 26th floors as examples, the agent on the 25th floor... 25 And the 26-layer intelligent agent 26 Each candidate strategy was generated and simulation verification was completed.

[0056] During the collaboration phase, the Agent is obtained from the spatialized point database. 25 and Agent 26 The coordinates of points within the jurisdiction can be obtained. =0.15, γ=0.3, =4 meters, |Δz|=4 meters, substitute into the formula to calculate the three-dimensional Euclidean distance (assuming the horizontal projection distance is 0), and obtain D. 3d =4 meters. Then calculate the heat-affected distance. =4×(1+0.3×(4 / 4))=5.2 meters. Finally, the spatial correlation degree R is calculated as 1 / (1+0.15×5.2)≈0.56.

[0057] If the preset collaboration threshold is 0.5, then R > 0.5, triggering collaboration. 25 The simulation results of "reducing the supply air temperature by 1°C to cope with the afternoon load" (predicting a 2% increase in energy consumption, but potentially leading to a 0.5°C decrease in temperature in the ceiling area of ​​the 26th floor below) were sent to the Agent. 26 Agent 26 Upon receiving this boundary constraint, the distributed co-optimizer assessed that if the original strategy were followed, the temperature in its area would deviate from the setpoint. Therefore, it adjusted its strategy: the original plan of "maintaining the current setpoint" was slightly modified to "increasing the supply air temperature by 0.3°C" to offset the cold air infiltration from upstairs. Ultimately, Agent... 25 and Agent 26 Each generated a coordinated final control strategy, which ensured the cooling needs of the 25th floor while avoiding excessive cooling load on the 26th floor, thus achieving energy efficiency synergy in the vertical direction.

[0058] Another possible implementation includes acquiring real-time monitoring data from the dynamic calibration trigger, such as model prediction bias, physical equipment change events, and seasonal environmental inflection points. When any monitored indicator exceeds a preset threshold, the calibration process is triggered. The digital twin model acquires the trigger signal, retrieves historical operational data from the previous time window, and retrains or calibrates the behavioral and physical models. This includes using spatialized energy consumption time-series data containing precise three-dimensional spatial coordinates to update the weights of the behavioral model through a spatiotemporal graph convolutional neural network, thereby improving the prediction accuracy of human activity patterns and equipment usage patterns in different spatial areas. Using spatialized energy consumption data and synchronously collected environmental parameters, Bayesian inference methods are used to reverse-correct the thermal or electrical parameters of the physical model, minimizing the overall error between the model output and the measured data. Finally, the calibrated model parameters, along with corresponding calibration time, data period, trigger event information, etc., are encapsulated to generate a model version snapshot, which is then synchronized to the digital twin model for hot loading.

[0059] In some implementations, during the continuous operation of a building energy efficiency optimization system based on digital twins and multi-agent systems, a dynamic calibration trigger deployed on a central server polls and reads model prediction deviation indicators reported by each agent, equipment change order status from the Building Management System (BMS), and seasonal transition forecast data provided by a third-party meteorological service interface every minute. When the model prediction temperature of a certain 5A office building project deviates from the measured average value of six temperature sensors in the west-facing open office area from the midday period (12:00-14:00) for three consecutive days on July 15th, exceeding a preset threshold of 0.8℃, the dynamic calibration trigger immediately determines that the "model prediction deviation" indicator has been triggered and generates a calibration event message containing the trigger time, trigger type (deviation exceeding limit), a list of associated agents (the HVAC agent responsible for the west-facing area), and a list of associated monitoring points, and publishes it to the system message bus.

[0060] At this point, the twin model's evolutionary engine, which had subscribed to the calibration event, was activated. Based on the information in the event, it retrieved spatialized energy consumption time-series data, indoor temperature data, and concurrent outdoor meteorological parameters from the time-series database for all relevant monitoring points in the west-facing office area over the past 30 days. Meanwhile, the twin model's internal evolutionary engine also initiated two concurrent optimization threads: The first thread inputs the above data into a pre-trained spatiotemporal graph convolutional neural network. The graph structure of this network is pre-built based on the three-dimensional spatial encoding and spatial correlation of these monitoring points. The behavior model is retrained through the backpropagation algorithm to update its internal weights, so that it can more accurately predict the occupancy rate of people in the room in the next 30 minutes and the resulting changes in socket load.

[0061] The second thread uses the Bayesian inference method, taking the parameters of the current physical model (mainly the thermal model) as the prior distribution and the indoor temperature data collected at the same time as the observations. It uses the Markov chain Monte Carlo algorithm to perform iterative sampling and inversely solve for the posterior wall heat transfer coefficient and window shading coefficient that best match the model output with the observations.

[0062] Once both threads have completed their optimization calculations, the evolutionary engine packages the updated behavioral model weights and physical model parameters, the data period used for this calibration (30 days), details of the triggering event (temperature deviation exceeding limits), and the current timestamp into a unique model version snapshot (e.g., "Model_v2.3.1_20240715_HeatWave"), and writes it to the model version database. Subsequently, a hot-load command is sent to the digital twin core service, enabling it to seamlessly load and use the new model version for subsequent simulation verification without interrupting system operation.

[0063] Meanwhile, the version management and rollback module will record the snapshot's metadata (version number, creation time, rollback address, etc.) in the version history list and retain the 10 most recent historical versions. If the operations and maintenance personnel find that the model effect is abnormal in subsequent monitoring, they can simply click the "rollback" button on the management interface to instruct the system to quickly switch back to the previous stable version, ensuring the reliability of system decisions.

[0064] Based on the above technical solution, a digital twin model is constructed by building information and a unique three-dimensional spatial code is assigned to the monitoring point. A spatialized point database is built, which solves the problem of fuzzy spatial positioning and inability to accurately associate physical location in traditional building energy consumption monitoring. It achieves centimeter-level accurate positioning of energy consumption monitoring points and lays a structural foundation for subsequent spatial mapping of data.

[0065] Meanwhile, a multi-level hierarchical metering system is built based on the power distribution system. Real-time energy consumption data is bound with spatial identifiers to generate spatialized energy consumption data. This solves the problem of coarse granularity of traditional system-level metering and the inability to accurately dismantle and locate energy consumption anomalies. It enables energy consumption data to be deeply associated with physical space and equipment, realizing refined energy consumption metering and data traceability at the circuit or equipment level.

[0066] Secondly, by deploying intelligent agents in each control area and generating regional state fields and establishing data interaction channels, the problem of scattered energy consumption and environmental data and the inability to perceive the overall regional operating status is solved, thereby realizing three-dimensional and integrated perception of regional energy consumption and environmental status.

[0067] Furthermore, the design that utilizes intelligent agents to generate candidate strategies and verifies them through simulation using digital twin models avoids the problems of traditional control strategies lacking pre-planning and being prone to errors during execution, thereby improving the scientific rigor and reliability of strategy formulation.

[0068] Finally, by using intelligent agents to calculate spatial correlation and generate control strategies based on collaborative simulation results, the problem of independent decision-making and mutual interference between control strategies in different areas is solved. This achieves collaborative energy efficiency optimization of the entire building, enabling precise positioning, refined management and intelligent optimization of building energy consumption in a progressive manner, and significantly improving energy efficiency management efficiency and energy-saving effects.

[0069] In another possible implementation of the embodiments of this application, combined with Figures 1-2 As shown, the construction process of a digital twin model can be achieved through the following steps 201 to 204, which are explained in detail below: Step 201: Obtain the building information of the target building and construct a digital twin model. The building information includes the BIM model, thermal parameters of the building envelope, internal space division and equipment and facility layout.

[0070] Internal space segmentation refers to dividing the interior of a building into functional areas based on the BIM model, such as rooms, corridors, and open office areas, forming independent control or analysis units.

[0071] In some implementations, the BIM model of the target building is imported, and its geometric structure is extracted as the basic framework for a digital twin. This allows for the model's accuracy to be checked and corrected based on on-site survey data, ensuring it accurately reflects the building's current condition. Then, material and structural information for each component of the building envelope is extracted from the BIM model, and corresponding thermal parameters are configured. For example, by consulting a material library, a thermal conductivity of 0.45 W / (m³) can be assigned to the exterior walls. 2 The heat transfer coefficient (K) is used to construct the core parameters of the physical model. Based on elements such as room boundaries and functional zoning in the BIM model, the internal space can be segmented and semantically defined, dividing the building into independently identifiable control areas such as "5th Floor West Open Office Area" and "B3 Floor Core East Corridor." Finally, based on the as-built drawings and on-site positioning, the precise three-dimensional coordinates (XYZ) of all monitoring points (sensors, metering equipment, energy-consuming equipment) are marked in the BIM model, forming an equipment and facility layout layer. All the above information is summarized and stored in a spatialized point database, generating a unique three-dimensional spatial code for each monitoring point, thus completing the construction of a high-precision initial digital twin model integrating geometric, physical attributes, spatial semantics, and equipment positioning.

[0072] Step 202: Based on the BIM model, establish a unified building space coordinate system, assign a unique three-dimensional spatial code to each monitoring point, and summarize the three-dimensional spatial codes and corresponding associated attribute information of all monitoring points to generate a spatialized point database.

[0073] The unified building spatial coordinate system is a global three-dimensional measurement benchmark established with a specific corner point of the building (such as the southwest corner of the building's boundary line) as the origin (0,0,0), with the east direction as the X-axis, the north direction as the Y-axis, and the vertical upward direction as the Z-axis. This ensures that the coordinate dimensions of all points are consistent and can be interoperated. The three-dimensional spatial code is a unique identifier generated based on this coordinate system, containing the building ID, floor, area, equipment type, and precise XYZ values. Associated attribute information refers to non-spatial characteristic data related to the monitoring points, including technical parameters such as equipment type, system (e.g., HVAC / lighting), metering level (system level, loop level, equipment level), area type (office, corridor, core tube), orientation, and data acquisition frequency.

[0074] In some implementations, the BIM software sets the origin of the coordinate system at the intersection of the building's southwest corner and the ground level, based on the building's boundary line or benchmark points. The model's rotation angle is then calibrated so that the X-axis points due east and the Y-axis points due north, establishing a unified architectural spatial coordinate system. Subsequently, all deployed or planned monitoring points (sensors, metering devices, smart circuit breakers, etc.) in the BIM model are traversed, and the software's point-picking function automatically extracts the precise XYZ coordinate values ​​of each point. This allows for the generation of a 3D spatial code according to preset coding rules: the coding structure is "Building ID-Floor-Fire Compartment-Area-Equipment Type-Serial Number-Coordinate Value," where the Building ID uses a project abbreviation, the floor is coded with actual numbers, the area is named according to its spatial function (e.g., ZONE-A for office area), and the equipment type uses standardized abbreviations (TEMP for temperature sensors, LIGHT for lighting circuits).

[0075] The code is then structurally linked with the point coordinates, associated attribute information exported from the BIM model (floor number, area type, system, orientation), and technical parameters obtained from the design documents (equipment model, accuracy class, communication protocol). This allows all the above information for all points to be imported into a spatiotemporal database, creating a complete record for each point containing a 3D spatial code, XYZ coordinates, and all associated attributes, forming a spatialized point database that supports spatial indexing and rapid retrieval.

[0076] Step 203: Based on the thermal parameters of the building envelope and the spatialized point database, construct a physical model containing building thermal zones, configure thermal parameters for each thermal zone, and embed the physical model into the digital twin model.

[0077] Building thermal zoning divides a building into several thermally independent zones based on its spatial layout, orientation, building envelope type, and energy consumption characteristics. These zones can be south / north facing zones based on orientation, inner / outer zones based on function, or vertical zones based on floor levels. A physical model is a mathematical abstraction of the building's thermophysical processes, used to simulate the transfer, storage, and release of heat within the building and between the building and the external environment. Thermal parameters are core quantitative indicators describing the thermophysical characteristics of a thermal zoning zone, including wall heat transfer coefficient, window shading coefficient, thermal inertia index, solar heat gain coefficient, and air permeability, which determine the zone's thermal response characteristics under specific boundary conditions.

[0078] In some implementations, thermal zoning of a building can be carried out by combining the three-dimensional spatial coding and area type attributes in the spatialized point database with the spatial geometric information extracted from the BIM model: the space on the same floor is divided into four oriented areas: east, south, west and north, according to the orientation; and divided into a peripheral area (within 5 meters from the outer wall) and an internal area according to the distance from the outer envelope. At the same time, each independent floor constitutes a vertical thermal stratification.

[0079] After partitioning is completed, the thermal parameters of the building envelope corresponding to each partition can be obtained from the design documents or material library: for example, extract the exterior wall construction layers from the BIM model, obtain the thermal conductivity of each layer of material by looking up the table, and calculate the comprehensive heat transfer coefficient K value of the wall; obtain the solar heat gain coefficient SHGC and heat transfer coefficient U value of the glass from the window catalog; and estimate the air permeability ACH according to the air tightness level of the doors and windows.

[0080] The EnergyPlus simulation engine kernel can then be used to construct a set of heat balance equations for each thermal zone based on the aforementioned zone geometry and thermal parameters. This includes one-dimensional heat conduction equations for the walls, indoor air heat balance equations, and calculation models for solar radiation transmitted through windows, forming a physical model describing the thermal dynamics of each zone. Finally, this physical model is encapsulated in a component-based manner and embedded into the thermal simulation module of the digital twin model, establishing a mapping relationship between the physical model and three-dimensional geometric primitives. This ensures that each thermal zone in the twin model is associated with a corresponding set of thermal parameters and a set of heat balance equations.

[0081] It should be noted that the granularity of thermal zoning directly determines the balance between simulation accuracy and computational efficiency of the physical model. For large public buildings, it is recommended to adopt a multi-level thermal zoning strategy: first, coarsely divide the buildings by floor, then subdivide each floor into four orientation zones based on orientation, and finally divide each orientation zone into inner and outer zones based on distance from the outer wall, ultimately forming a three-level zoning system of floor-orientation-inner and outer.

[0082] Step 204: Based on the three-dimensional spatial coding in the spatialized point database, extract the historical spatialized energy consumption time series data of all monitoring points in the same spatial area to form a graph structure data with spatial topological relationships, construct a behavior model, and embed the behavior model into the digital twin model.

[0083] In this context, "same spatial area" refers to an independent control unit with functional consistency and spatial continuity, determined based on building space division and thermal zoning, such as the west-facing office area or the internal area of ​​the core tube on a certain floor. Historical spatialized energy consumption time series data is a sequence of energy consumption values ​​with precise three-dimensional spatial coding collected from each monitoring point over a historical time dimension, recording the energy consumption variation patterns of that point at different times.

[0084] In some implementations, the three-dimensional spatial codes of all monitoring points in the target area and the corresponding historical energy consumption time series data are read from a spatialized point database (e.g., the time window is preset to the past 3 months, and the data granularity adopts an hourly interval to balance the amount of information and the computational load).

[0085] The spatial Euclidean distance between each pair of points is calculated based on their 3D coordinates, and an adjacency matrix of a graph structure is constructed accordingly. When the 3D distance between two points is less than a preset threshold (such as the maximum span of a region), an edge is established between them, with the initial weight of the edge using a Gaussian kernel function. Calculation, where For three-dimensional distance, The decay rate is controlled by the scale parameter, so that the closer the points are in space, the stronger the correlation in the graph.

[0086] This allows the construction of a spatiotemporal graph with monitoring points as nodes, the aforementioned adjacency matrix as edges, and historical energy consumption time-series data as node features. This graph fully encodes the energy consumption patterns and spatial dependencies of each point within the region. A spatiotemporal graph convolutional neural network is used as the core architecture of the behavioral model. This network consists of three stacked spatiotemporal convolutional blocks. Each block contains a graph convolutional layer to capture spatial dependencies (achieved through Chebyshev polynomial approximation of the graph Laplacian operator) and a gated convolutional layer to capture temporal dependencies (performing one-dimensional convolution along the time dimension and using gated linear units to control information flow).

[0087] During model training, the goal is to predict energy consumption sequences for future time windows using historical time window data. Supervised learning is performed using the Adam optimizer and mean squared error loss function, with iterative training until the validation set loss converges. Finally, the trained behavioral model is encapsulated as an independent inference module and embedded as an API service within the digital twin model, alongside the geometric and physical models. This enables the digital twin model to predict future energy consumption behavior based on historical patterns.

[0088] Based on the above technical solution, by acquiring complete building information including BIM model, thermal parameters, spatial division, and equipment layout, the information of the building model can be unified, providing a real and multi-dimensional digital foundation for all subsequent operations. Secondly, by establishing a unified building spatial coordinate system through the BIM model, assigning a unique three-dimensional spatial code to each monitoring point and generating a spatialized point database, abstract energy consumption data can be rigidly bound to specific physical locations. This solves the spatial ambiguity problem of data location in traditional systems, achieving a leap from system-level statistics to point-level positioning, enabling the tracing of energy consumption anomalies to be precise down to specific loops or equipment. Furthermore, using the thermal parameters of the building envelope and the point database, a physical model containing thermal zoning is constructed. The thermophysical characteristics of the building are quantified and embedded in the twin model, solving the problem that traditional models cannot accurately simulate the heat transfer process. This gives the twin the ability to perform dynamic simulation and prediction based on physical laws. Finally, historical energy consumption data was extracted from the point database to construct a behavioral model with spatial topological relationships. This uncovered the human factors and energy consumption patterns during the building's use, solving the problem that traditional models can only reflect physical responses and cannot depict the impact of human activities. This allows the digital twin model to move from static geometry to dynamic data, and from physical mechanisms to behavioral patterns.

[0089] In another possible implementation of the embodiments of this application, combined with Figures 1-3 As shown, the construction process of the sub-item measurement system can be achieved through the following steps 301 to 305, which are explained in detail below: Step 301: Perform topology analysis on the power distribution system of the target building and construct a two-level metering mapping model that includes physical circuit level and spatial function level.

[0090] The physical circuit level refers to the metering level defined according to the electrical topology of the building's power distribution system and the physical connection relationships of distribution cabinets, outgoing circuits, and terminal equipment. It is used to independently collect raw energy consumption data for each circuit. The spatial function level refers to the metering level based on the building's spatial layout and functional zoning, using three-dimensional spatial coding to classify energy-consuming equipment according to its physical spatial area. This is used to form functional area-level metering units. The dual-level metering mapping model is a correlation model used to establish the correspondence between energy consumption data collected at the physical circuit level and specific functional areas in the spatial function level. It enables dual mapping and fusion of energy consumption data in both electrical and spatial dimensions.

[0091] In some implementations, the complete electrical topology from the main building's incoming power line to various distribution cabinets and terminal distribution circuits is obtained by analyzing the target building's power distribution system drawings and conducting on-site surveys. This allows for the construction of a physical circuit hierarchy. Dedicated integrated modules for sub-metering are deployed in each terminal circuit (e.g., lighting, air conditioning, socket circuits) to independently collect raw energy consumption data. Simultaneously, based on the three-dimensional spatial coding in a pre-built spatialized point database, all energy-consuming equipment monitoring points belonging to the same physical space area (e.g., the same office, meeting room) are categorized, forming a spatial functional hierarchy. Subsequently, the built-in identification unit of the dedicated integrated modules for sub-metering deployed in the circuits can be used to decompose the collected raw energy consumption data of the mixed circuits into energy consumption components for different equipment types (e.g., computers, lighting fixtures) through algorithms such as current characteristic analysis or harmonic decoupling. Finally, based on the pre-built two-level metering mapping model, the decoupled energy consumption components are re-aggregated with the corresponding functional areas in the spatial functional hierarchy to generate a set of detailed energy consumption datasets that are refined according to both spatial function (e.g., the 5th floor west zone) and energy-consuming equipment type (e.g., lighting, sockets).

[0092] Step 302: Based on the outgoing circuits of the distribution cabinet, the physical circuit hierarchy is divided into four system-level metering branches: lighting, air conditioning, sockets, and power. At the end of each branch, a dedicated integrated module for sub-metering is deployed to independently collect the raw energy consumption data of each circuit.

[0093] Among them, the outgoing circuit refers to the terminal power distribution line that leads out from the lower end of the circuit breaker of the building's power distribution cabinet (box) and is used to supply power to a specific area or a specific type of load (such as lighting and air conditioning). It is the final physical channel for power distribution. The dedicated integrated module for sub-item metering is a hardware unit that integrates multi-protocol conversion, data acquisition, edge computing, and communication functions. It can independently access and collect the raw energy consumption data (such as voltage, current, power, and electrical energy) of a single power distribution circuit and convert it into a unified standardized data packet.

[0094] In some implementation methods, a detailed electrical topology survey is conducted on all distribution cabinets in the target building to clarify the number of outgoing circuits, the types of loads they support, and the power supply range of each distribution cabinet. Based on the load type, each outgoing circuit can be classified into one of four system-level metering branches: lighting, air conditioning, sockets, or power. For example, circuits supplying power to fluorescent lights in the office area are classified as lighting, while circuits supplying power to fan coil units are classified as air conditioning.

[0095] Therefore, at the end of each predefined outgoing circuit, specifically at the connection point between the circuit breaker and the load cable within the distribution box, a dedicated integrated module for sub-metering, such as IPme®-DT-Meter, is physically connected in series. This module then independently collects the raw energy consumption data of the circuit at a preset high frequency (e.g., once per minute) using its built-in current transformer and voltage sampling circuit. This data includes real-time power, current, voltage, and cumulative energy consumption.

[0096] The collected raw data is uniformly converted into standard format data packets by the module's built-in multi-protocol conversion unit (supporting Modbus, BACnet, etc.). Finally, the module uses its built-in edge computing capabilities to perform preliminary cleaning and filtering on the raw data, and reports the processed data, along with its own device ID, to the central data server in real time via wired or wireless networks within the building, thus realizing a complete link from physical loop to data acquisition.

[0097] Step 303: Based on the three-dimensional spatial coding in the spatialized point database, the energy-consuming equipment associated with all monitoring points in the same spatial area is classified to form a functional area-level metering unit.

[0098] In this context, "same spatial area" refers to an independent control unit with functional consistency and spatial continuity, determined by building space division and thermal zoning, such as the west-facing office area or the interior area of ​​the core tube on a certain floor. A functional area-level metering unit refers to a virtual metering aggregate that independently reflects the overall energy consumption of an area by aggregating the energy consumption data of all energy-consuming equipment belonging to the same physical spatial area (such as an office, meeting room, or shop).

[0099] In some implementations, a pre-built spatialized point database is invoked. This database stores a unique 3D spatial code for each monitoring point (such as a lighting circuit, socket circuit, or air conditioning terminal device), along with its corresponding XYZ coordinates, floor level, area type, and other associated attribute information. Based on preset spatial division rules, such as physical partitions (walls), fire compartment boundaries, or functional area definitions (e.g., room divisions in the BIM model), the building space can be logically divided into multiple "same spatial areas," each assigned a unique functional area identifier.

[0100] After the system traverses all monitoring points in the database, it can extract the floor information and XYZ coordinates from the 3D spatial code of each point. Then, through spatial topology calculations, it determines whether the point falls within the geometric boundary of a predefined "same spatial area." Finally, all energy-consuming devices (and their corresponding energy consumption data streams) associated with monitoring points falling within the boundary of the same functional area are logically aggregated into a "functional area-level metering unit." At this point, the unit itself does not add physical hardware but acts as a virtual data aggregator, summarizing the spatialized energy consumption data reported by all points within it in real time. This generates a sub-aggregated dataset that accurately reflects the total lighting energy consumption, total socket energy consumption, and total air conditioning energy consumption of that specific functional area (e.g., the 5th floor west wing office).

[0101] Taking a technology company office building as an example, its 5th floor is divided into multiple functional areas, including "Conference Room 5001". Multiple monitoring points are deployed within this conference room area: one lighting circuit (3D spatial code pointing to the conference room ceiling), two socket circuits (3D spatial code pointing to the area below the conference table sockets), and one air conditioning terminal branch circuit (3D spatial code pointing to the interior of the conference room ceiling) serving the fresh air supply. Based on the spatial boundary of "Conference Room 5001" defined by the BIM model, the energy consumption data streams associated with these five monitoring points can be automatically logically aggregated using the 3D spatial codes of each point, forming a functional area-level metering unit named "Zone-Floor5-Room5001". When a video conference is held in the conference room, the total power displayed in real time by this metering unit is 0.8kW for lighting, 2.5kW for sockets (mainly for monitors and computers), and 1.2kW for air conditioning. Through this aggregation unit, administrative staff can directly see the energy consumption cost of a single meeting in the conference room, and property engineers can quickly determine whether the total power is abnormally high when the conference room is empty due to a device being left on.

[0102] Step 304: The raw energy consumption data collected at the physical circuit level is decoupled by the identification unit built into the sub-metering integrated module to decompose the energy consumption components of different types of energy-consuming equipment in the circuit.

[0103] Raw energy consumption data refers to unprocessed electrical parameters collected directly from the power distribution circuit by the dedicated integrated module for sub-metering. It typically manifests as an aggregated signal containing total power, total current, and other parameters that reflect the operating characteristics of all electrical devices connected to that circuit. The identification unit, an edge computing module embedded in the dedicated integrated module for sub-metering, possesses signal processing and pattern recognition capabilities, used to analyze and deconstruct the aggregated raw energy consumption data. Feature decoupling refers to the process of decomposing the aggregated energy consumption into individual device energy consumption components by analyzing the unique electrical characteristics of different energy-consuming devices in the total energy consumption signal (such as starting current waveform, steady-state harmonic composition, active and reactive power ratio, etc.). Each energy consumption component refers to an independent energy consumption value obtained after feature decoupling, belonging to a specific type of energy-consuming device (such as a computer, lighting fixture, or air conditioner indoor unit) or a single device.

[0104] In some implementations, once the dedicated integrated module for sub-metering collects the raw energy consumption data of a certain outgoing circuit, its built-in identification unit immediately initiates the feature decoupling process. At this time, the identification unit will first perform high-speed sampling (e.g., thousands of times per second) on the raw current and voltage signals, convert the time-domain signal into a frequency-domain signal through Fourier transform, and extract the harmonic spectrum characteristics of the total current of the circuit.

[0105] Then, it calls upon its pre-trained internal device load feature library and employs a non-intrusive load decomposition algorithm to perform pattern matching and combination optimization between the real-time extracted mixed signal features and the device fingerprints in the feature library, solving for the most likely device combinations and their respective power proportions at that moment. Finally, the energy consumption components of different energy-consuming device types (such as lighting power 500W and computer power 800W) calculated by decoupling are separated from the original aggregated data and encapsulated together with the original loop ID and timestamp to generate refined energy consumption data with device type labels, which is then used by the upper-level system for further aggregation in spatial and functional dimensions.

[0106] The equipment load feature library is constructed by individually testing various typical electrical devices (such as computers, LED lights, printers, and fan coil units) in a controlled environment, collecting and recording their unique electrical fingerprints, such as active power, reactive power, current waveform, harmonic content, and transient current spikes during startup.

[0107] Step 305: Based on the two-level metering mapping model, the decoupled energy consumption components are re-aggregated with the corresponding functional areas in the spatial function level to generate a sub-item energy consumption dataset that is decomposed according to the dual dimensions of spatial function and energy-consuming equipment type.

[0108] Load re-aggregation refers to the process of redistributing and summarizing the energy consumption components, which are obtained after feature decoupling and have equipment type labels, to the corresponding spatial functional areas according to the attribution relationships defined in the two-level metering mapping model. The sub-item energy consumption dataset refers to the collection of energy consumption data generated after load re-aggregation, which can be queried and displayed simultaneously according to both spatial function dimensions (e.g., an office) and energy-consuming equipment type dimensions (e.g., lighting, air conditioning, sockets). It is a fundamental data asset supporting refined energy consumption diagnosis and management.

[0109] In some implementations, a pre-built two-level metering mapping model is first invoked. The core of the model is a mapping association table, which clearly records the power supply attribution relationship between each physical circuit ID (such as "lighting circuit L-05") and one or more spatial functional area IDs (such as "5F-West-Office" and "5F-West-Corridor"), as well as the weight ratio of circuit energy consumption allocated to each area in this mixed power supply scenario (such as a preset ratio based on the area area or the number of lamps).

[0110] When the dedicated integrated module for sub-metering reports energy consumption component data that has been decoupled based on features (e.g., decoupling the "lighting" component of 2.5kW from the total energy consumption of loop L-05), the data aggregation engine begins to work. At this point, the engine takes this energy consumption component data as input and first parses out the physical loop ID from which it originates. Then, using this loop ID as an index, it searches for all corresponding spatial functional areas and assigns weights within the two-level metering mapping model. Finally, based on the found mapping relationships and weights, this "lighting" energy consumption component (2.5kW) is accumulated and added to the corresponding spatial functional area (e.g., "5F-West-Office" adds 2.0kW, "5F-West-Corridor" adds 0.5kW) under the lighting energy consumption category. By performing this aggregation operation on the energy consumption components of all loops and all equipment types, a sub-item energy consumption dataset organized by the dual dimensions of "spatial function + equipment type" is ultimately generated in the time-series database and continuously updated.

[0111] Based on the above technical solution, a two-level metering mapping model is constructed. First, at the physical circuit level, the four major systems—lighting, air conditioning, sockets, and power—are finely divided into branches. Then, a dedicated integrated module for sub-metering is deployed in each terminal circuit. This solves the metering blind spot problem of traditional power distribution systems, which can only obtain the total system-level energy consumption but cannot perceive the energy consumption of specific circuits. This achieves independent collection of raw energy consumption data across the entire link from the distribution cabinet to the terminal circuit. Simultaneously, at the spatial function level, three-dimensional spatial coding is used to group energy-consuming devices within the same area into functional area-level metering units. This solves the spatial ambiguity problem of energy consumption data being disconnected from physical location and the inability to quickly locate specific spatial areas after anomalies occur, enabling each independent area such as an office or meeting room to form a complete energy consumption profile. Then, through the identification unit built into the sub-metering module, the raw energy consumption data is decoupled by features, decomposing the total energy consumption in the mixed circuit into energy consumption components of different equipment types (such as lighting, computers, and air conditioning). This solves the problem of data aliasing in mixed power supply circuits, which cannot distinguish specific energy-consuming devices, allowing energy consumption diagnosis to delve deeper from the circuit level to the equipment type level. Finally, based on the two-level metering mapping model, the decoupled energy consumption components are re-aggregated with the corresponding functional areas to generate a sub-item energy consumption dataset that is broken down according to both spatial function and equipment type. This completely connects the data chain of electrical topology, physical space and energy-consuming equipment, allowing maintenance personnel to view the total energy consumption from the spatial dimension (such as the 5th floor west office) and also penetrate to the equipment type dimension (such as the lighting energy consumption and socket energy consumption in this area). This enables minute-level location of energy consumption anomalies and equipment-level traceability, providing a high-quality data foundation for precise energy efficiency optimization.

[0112] In another possible implementation of the embodiments of this application, combined with Figures 1-4 As shown, the process of constructing an intelligent agent can be achieved through the following steps 401 to 404, which are explained in detail below: Step 401: Based on the thermal zoning of the digital twin model and the three-dimensional spatial coding in the spatialized point database, the building physical space is divided into multiple three-dimensional voxelized control units. Each three-dimensional voxelized control unit is deployed with an edge agent. The edge agent includes an environmental perception fusion module, a graph neural network decision module, and a twin collaboration interface.

[0113] Among them, the three-dimensional voxelized control unit refers to the smallest spatial control unit formed by meshing the building's physical space in the X, Y, and Z dimensions based on the thermal zoning boundaries of the digital twin model and the three-dimensional spatial coding in the spatialized point database. Each unit is an independent cubic spatial voxel with a unique spatial identifier and clear spatial boundaries, used to achieve refined decomposition of regional state perception and control.

[0114] In some implementations, the thermal zones pre-divided in the digital twin model (such as south / north zones based on orientation, inner / outer zones based on function, or vertical zones based on floor level) and the 3D spatial codes of monitoring points stored in a spatialized point database are used. Each thermal zone can serve as the basic boundary of an independent voxel-based control unit. Within each thermal zone, it is further divided into multiple 3D voxel grids according to preset voxel sizes (e.g., 5m x 5m horizontally, 3-4m vertically based on floor height). Each voxel grid is defined as a 3D voxel-based control unit, and a corresponding unique voxel code is generated. Thus, an edge agent is deployed in physical space for each 3D voxel-based control unit using the IPme®-DT digital twin intelligent module. Each agent contains the following three core modules: Environmental perception fusion module: responsible for collecting real-time data from all monitoring points within the managed voxels and generating a three-dimensional state field of the region through interpolation algorithms; Graph Neural Network Decision Module: Based on state field data, it extracts spatial features through graph convolution operations to generate candidate control strategies; Twin Collaboration Interface: Responsible for data interaction and collaborative computing with neighboring intelligent agents and cloud-based digital twin platforms.

[0115] Step 402: The environmental perception fusion module acquires real-time spatialized energy consumption data and environmental parameters of all monitoring points within the three-dimensional voxelized control unit under its jurisdiction. It uses a three-dimensional spatial interpolation algorithm to perform spatial continuity processing on the discrete point data, generating a regional three-dimensional state field that reflects the multi-dimensional distribution of temperature field, humidity field, illuminance field and energy consumption field within the unit, and synchronizes it to the digital twin model for visualization mapping.

[0116] Among them, the three-dimensional spatial interpolation algorithm refers to an algorithmic model that uses mathematical methods to deduce the numerical values ​​of unknown spatial locations based on the precise three-dimensional coordinates of known discrete monitoring points and the collected values, and is used to convert discrete point data into a continuous spatial distribution field. The regional three-dimensional state field refers to a data field model generated by interpolation algorithms within a three-dimensional voxelized control unit, which can reflect the continuous distribution of multi-dimensional physical quantities such as temperature, humidity, illuminance, and energy consumption in three-dimensional space.

[0117] In some implementations, the environmental perception fusion module reads real-time spatialized energy consumption data and environmental parameters of all monitoring points within its managed 3D voxelized control unit via a data interface. Simultaneously, it obtains the precise 3D coordinates of each point from the spatialized point database. This allows it to directly call a pre-built 3D spatial interpolation algorithm (based on the Kriging interpolation method). By calculating the semivariance function between points, it determines the spatial correlation weights. Using discrete point data and their 3D coordinates as input, it performs gridded interpolation calculations on the continuous space within the voxel unit, calculating the values ​​of unknown locations point by point in the X, Y, and Z dimensions according to a preset interpolation resolution.

[0118] The four physical quantities—temperature, humidity, illuminance, and energy consumption—are then subjected to the interpolation process described above to generate four independent sets of continuous distribution field data. Finally, these four sets of distribution field data are spatially superimposed and fused to form a multi-dimensional regional three-dimensional state field. This field is then synchronized to a digital twin model via a twin collaboration interface. In the model, different numerical ranges are mapped to preset color gradients, enabling real-time refreshing and visualization of the three-dimensional primitives.

[0119] Step 403: The graph neural network decision module takes the three-dimensional state field of the region as input, extracts the spatial gradient features and temporal evolution features in the three-dimensional state field of the region through graph convolution operation, generates a set of candidate control strategies by combining the preset multi-objective optimization function, and encapsulates the spatial identifier of the candidate control strategy set and its associated three-dimensional voxelized control unit through the twin cooperative interface.

[0120] Spatial gradient characteristics refer to the rate of change distribution of physical quantities in the three-dimensional state field of a region along three-dimensional spatial directions, reflecting the intensity and directionality of spatial changes in physical quantities such as temperature and humidity. The preset multi-objective optimization function is a predefined mathematical optimization expression composed of multiple mutually constraining objectives such as minimizing global energy consumption, maximizing comfort, and ensuring equipment safety, used to guide the generation of candidate control strategies. The candidate control strategy set refers to the collection of multiple feasible combinations of control parameters generated by the graph neural network decision module based on the current state field and optimization objectives. Each strategy includes adjustment schemes for actuators such as air conditioning set temperature, lighting brightness, and fresh air valve opening.

[0121] In some implementations, the graph neural network decision module obtains regional three-dimensional state field data from the environmental perception fusion module. This data is organized in the form of a three-dimensional voxel grid, with each voxel containing values ​​for four dimensions: temperature, humidity, illuminance, and energy consumption. A graph structure can then be constructed based on the spatial adjacency relationships of the three-dimensional voxels, with each voxel serving as a node, the spatial adjacency relationships between voxels as edges, and the multidimensional physical quantity values ​​within each voxel as the initial feature vector of the node.

[0122] The pre-trained graph convolutional neural network model is then invoked, employing a stacked three-layer graph convolution structure. Each layer performs spectral domain graph convolution using a Chebyshev multinomial approximation graph Laplacian operator, aggregating feature information from neighboring voxels and extracting spatial gradient features. Simultaneously, a temporal convolutional network branch is connected in parallel to the model, performing one-dimensional temporal convolution on the state fields across multiple consecutive time windows to extract temporal evolution features. The extracted spatial gradient features and temporal evolution features are then fused to form a comprehensive state representation vector, which is input into a pre-defined multi-objective optimization function solver. This optimization function takes the comprehensive state representation vector as input and has three sub-objectives: minimizing energy consumption, minimizing comfort deviation, and minimizing the number of equipment start-ups and shutdowns. A weighted summation method is used to transform the multi-objective into a single objective, and a genetic algorithm iteratively searches within the feasible solution space to generate multiple Pareto-optimal candidate control strategies.

[0123] Finally, the generated candidate control strategy set and its associated spatial identifiers of the three-dimensional voxelized control unit are structurally encapsulated and prepared to be sent to a neighboring intelligent agent or cloud platform for collaborative optimization through a twin collaboration interface.

[0124] Step 404: The twin collaboration interface calculates the spatial correlation degree with the neighboring edge agents based on the three-dimensional spatial encoding in the spatialized point database. When the spatial correlation degree exceeds the preset collaboration threshold, it sends the candidate control strategy set to the corresponding neighboring edge agent and receives the candidate control strategy set sent by the neighboring edge agent.

[0125] In this context, "neighboring edge agents" refers to other edge agents that are adjacent or close to the 3D voxelized control unit managed by the current edge agent in terms of spatial topology. This includes agents that are adjacent horizontally on the same floor, vertically on different floors, and diagonally opposite each other. Spatial correlation is a quantitative index calculated based on the 3D spatial encoding of monitoring points within the jurisdiction of two edge agents. It characterizes their physical proximity and the intensity of mutual influence between them in physical fields such as thermal and airflow. The preset collaboration threshold is a pre-defined numerical limit for spatial correlation. When the calculated spatial correlation exceeds this threshold, it is determined that there is a significant mutual influence between the two agents, requiring the activation of a collaborative decision-making mechanism.

[0126] In some implementations, the twin collaboration interface reads the 3D spatial codes of all monitoring points within the 3D voxelized control unit managed by the current edge agent from a spatialized point database, and extracts the precise 3D coordinates of each point. This allows the edge computing gateway to broadcast query requests in the building's industrial Ethernet or wireless mesh network via its built-in neighbor discovery protocol, obtaining the device identifiers and jurisdictional area information of all online edge agents in the network.

[0127] For each discovered neighboring edge agent, the twin collaboration interface also reads the 3D spatial code and coordinates of the monitoring points within the agent's jurisdiction from its database. Then, it calls a pre-defined spatial correlation calculation model, using the coordinate sets of the current agent and neighboring agents as input, to calculate the average 3D Euclidean distance between the two point sets, and substitutes in a preset spatial attenuation coefficient. Through formula Calculate the spatial correlation degree ,in This represents the average distance.

[0128] The calculated spatial correlation is compared with a preset collaboration threshold. When R exceeds the threshold, collaboration is deemed necessary. The twin collaboration interface encapsulates the candidate control policy set generated by the current agent and its associated spatial identifier into a standard data message, which is then sent to the twin collaboration interface of the corresponding neighboring edge agent via MQTT or OPCUA protocol. Simultaneously, the interface opens a listening port to receive candidate control policy sets from other neighboring agents that have exceeded the collaboration threshold, and stores the received policy sets in local shared memory for subsequent collaborative optimization calls by the neural network decision module. The preset coordination threshold is an empirical value set during system initialization based on the building space layout, thermal characteristics of the building envelope, and type of air conditioning system. It is usually set in the range of 0.3-0.7.

[0129] Step 405: The graph neural network decision module evaluates the coupling effect of the boundary influence of the three-dimensional state field of its own region based on the received candidate control strategy set. With the goal of minimizing global energy consumption and inter-regional comfort deviation, the module coordinates the candidate control strategy set to generate regional autonomous control instructions.

[0130] The coupling effect assessment refers to the comprehensive process by which the graph neural network decision module analyzes the changes in boundary conditions and mutual influences of candidate control strategies of neighboring agents on the three-dimensional state field of the region. This includes the cross-influence analysis of multiple physical fields such as heat transfer, airflow infiltration, and complementary illumination. The regional autonomous control command refers to the set of specific operating parameters generated by the graph neural network decision module after collaborative optimization, used to directly control various actuators within the region. This includes executable commands such as equipment start / stop status, setpoint adjustment, and valve opening.

[0131] In some implementations, the graph neural network decision module reads the candidate control policy sets of neighboring agents received from the twin collaboration interface, as well as the candidate control policy set generated by the current agent itself, from local shared memory. Then, the digital twin sub-model for the current region can be loaded. This sub-model is a lightweight version synchronized from the cloud-based digital twin model to the local edge, containing the region's geometric boundaries, thermal parameters of the building envelope, and the current three-dimensional state field.

[0132] Each candidate strategy of the neighboring agents is used as a boundary condition and applied to the corresponding boundary surface of the digital twin sub-model of this region. For example, the cooling strategy of the agent upstairs is transformed into the cold radiation boundary of the ceiling of this region, and the shading strategy of the agent on the east side is transformed into the solar radiation attenuation boundary of the east wall of this region.

[0133] This allows for multiphysics coupling simulations of each strategy combination, using the finite difference method to solve the heat conduction equation, and iteratively calculating the changes in temperature, humidity, and energy consumption in the region over the next 15-30 minutes under the influence of neighboring strategies. This quantitatively assesses the degree of disturbance to the region's state field caused by each combination. A collaborative optimization objective function can then be constructed, with the two core objectives of minimizing global energy consumption and minimizing inter-regional comfort deviations. The multi-objective is transformed into a single objective using a weighted summation method, and a particle swarm optimization algorithm is used to iteratively search the candidate strategy combination space, minimizing the objective function value to find the optimal strategy combination solution. Finally, the corresponding control parameters for the region are extracted from the optimal solution, and instructions are encapsulated according to the control protocol format of the actuator to generate regional autonomous control instructions, which are written into the execution queue for issuance. Simultaneously, the selected collaborative strategies and their expected effects are fed back to the digital twin model for recording and visualization.

[0134] The global energy consumption consists of the energy consumption of this region plus the expected energy consumption parsed from the policies of neighboring agents, and the comfort deviation between regions is calculated from the average temperature difference and humidity difference between this region and neighboring regions.

[0135] Taking the 5th floor of an office building as an example, the edge agent Agent-5W in the west office area receives a set of candidate strategies from Agent-6W on the 6th floor above (Strategy A: Reduce temperature by 1℃, Strategy B: Keep it unchanged) and a set of strategies from Agent-4W on the 4th floor below (Strategy C: Increase fresh air volume, Strategy D: Keep it unchanged). At the same time, it generates three candidate strategies (Strategy E: Lower the set temperature by 0.5℃, Strategy F: Turn on the sunshade, Strategy G: Combine the two).

[0136] At this point, Agent-5W's graph neural network decision module applies these 12 combinations of 2×2×3 strategies to the boundaries of the digital twin sub-model for this region. Coupled simulations show that when the upstairs unit lowers the temperature by 1°C and the downstairs unit increases the fresh air volume, the ceiling temperature in this region decreases by 0.3°C, the feeling of cold air at the floor increases, leading to poorer temperature field uniformity and increased comfort deviation. The temperature field in this region is most stable when the upstairs and downstairs units remain unchanged, and the shading is activated.

[0137] Therefore, with the goals of minimizing global energy consumption (energy consumption in this area + expected energy consumption upstairs + expected energy consumption downstairs) and minimizing comfort deviations between areas, a particle swarm optimization algorithm was used to find the optimal combination: keep upstairs unchanged, keep downstairs unchanged, and activate shading in this area. Based on this, an autonomous control command for the area was generated: the motorized louvers on the west side were closed by 30%, the air conditioning temperature was set to 24℃, this command was written into the execution queue, sent to the actuator through the IPme®-DT module, and the collaborative results were fed back to the digital twin model for visualization.

[0138] Based on the above technical solutions, by constructing a three-dimensional voxelized control unit and deploying edge agents, the technical problems of fuzzy spatial positioning and inability to accurately perceive regional environmental conditions in traditional building energy management can be solved. This allows for centimeter-level fine-grained subdivision of building space and precise correspondence between distributed intelligent sensing hardware, providing a high-precision spatial data foundation for subsequent decision-making. Simultaneously, the environmental perception fusion module employs a three-dimensional spatial interpolation algorithm to transform discrete point data into a continuous multi-dimensional regional three-dimensional state field. This solves the problem that discrete sensors cannot fully reflect the continuous spatial distribution, achieving holographic visualization perception of multiple physical fields such as temperature, humidity, illuminance, and energy consumption, enabling maintenance personnel to intuitively grasp the real environmental conditions within the area. Furthermore, the graph neural network decision module extracts the spatial gradient features and temporal evolution features of the state field through graph convolution operations. This solves the problem that traditional control strategies rely on single thresholds and cannot perceive spatial change trends and future evolution patterns. It achieves intelligent generation of multi-objective strategies based on spatial distribution and temporal patterns, improving the scientific rigor and foresight of strategy formulation. The twin collaboration interface, based on three-dimensional spatial encoding, calculates the spatial correlation between agents, solving the problems of mutual interference and data silos caused by independent decision-making in adjacent areas. It achieves an efficient communication mechanism that initiates collaboration only in areas with significant mutual influence, greatly reducing unnecessary network communication load and computational overhead. Finally, the graph neural network decision module evaluates the coupling effect of the received neighboring policies, and performs collaborative optimization with the goals of minimizing global energy consumption and inter-regional comfort deviation. This solves the problem of global imbalance caused by local optima, realizes cross-regional joint optimization and autonomous control, and ultimately achieves the effects of improving overall building energy efficiency, improving temperature uniformity, and shortening anomaly response time, forming a complete technical closed loop from spatial partitioning, state perception, policy generation, efficient collaboration to global optimization.

[0139] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as a building energy efficiency optimization system based on digital twins and multi-agent systems, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in a hardware-driven or software-driven manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0140] When using integrated units, Figure 5 The above-described building energy efficiency optimization system based on digital twins and multi-agent systems is illustrated, comprising: a digital twin model construction unit, used to acquire building information of the target building to construct a digital twin model, and assign a unique three-dimensional spatial code to each monitoring point to generate a spatialized point database; a sub-metering system and data mapping unit, used to construct a hierarchical metering system with multiple levels based on the building's power distribution system, and bind the acquired real-time energy consumption data with the corresponding spatial identifiers in the spatialized point database according to the physical location of the monitoring points to generate spatialized energy consumption data; a multi-agent cluster construction unit, used to deploy corresponding agents for each control area based on the digital twin model and the spatialized point database. Each agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model; and a simulation verification unit, used by each agent to generate candidate control strategies based on the regional state field, and call the digital twin model to perform simulation verification of the candidate control strategies to obtain simulation verification results. The digital twin intelligent agent collaborative decision-making unit is used by each intelligent agent to calculate the spatial correlation between each other based on the spatial identifiers in the spatialized point database, and to coordinate the simulation verification results based on the spatial correlation to generate control strategies.

[0141] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A building energy efficiency optimization method based on digital twins and multi-agent systems, characterized in that, include: The building information of the target building is obtained to construct a digital twin model, and each monitoring point is assigned a unique three-dimensional spatial code to generate a spatialized point database. Based on the building power distribution system, a hierarchical metering system with multiple levels is constructed, and according to the physical location of the monitoring points, the acquired real-time energy consumption data is bound to the corresponding spatial identifier in the spatialized point database to generate spatialized energy consumption data. Based on the digital twin model and spatialized point database, a corresponding intelligent agent is deployed for each control area; each intelligent agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model. Each of the intelligent agents generates a candidate control strategy based on the regional state field, and calls the digital twin model to simulate and verify the candidate control strategy to obtain the simulation verification results; Each of the intelligent agents calculates the spatial correlation between themselves based on the spatial identifiers in the spatialized point database, and coordinates the simulation verification results according to the spatial correlation to generate a control strategy.

2. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, The construction process of the digital twin model specifically includes: Obtain the building information of the target building to construct a digital twin model. The building information includes BIM model, thermal parameters of the building envelope, internal space division and equipment and facility layout. Based on the BIM model, a unified building space coordinate system is established, each monitoring point is assigned a unique three-dimensional spatial code, and the three-dimensional spatial codes and corresponding associated attribute information of all monitoring points are summarized to generate a spatialized point database. Based on the thermal parameters of the building envelope and the spatialized point database, a physical model containing building thermal zones is constructed, thermal parameters are configured for each thermal zone, and the physical model is embedded in the digital twin model. Based on the three-dimensional spatial coding in the spatialized point database, historical spatialized energy consumption time series data of all monitoring points in the same spatial area are extracted to form graph structure data with spatial topological relationships, construct a behavior model, and embed the behavior model into the digital twin model.

3. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, The process of constructing the sub-item measurement system specifically includes: The power distribution system of the target building is analyzed for topology, and a two-level metering mapping model including physical circuit level and spatial function level is constructed. The physical circuit hierarchy is based on the outgoing circuits of the distribution cabinet, and is divided into four system-level metering branches: lighting, air conditioning, sockets, and power. A dedicated integrated module for sub-metering is deployed at the end of each branch to independently collect the raw energy consumption data of each circuit. The spatial functional hierarchy is based on the three-dimensional spatial coding in the spatialized point database, which classifies the energy-consuming equipment associated with all monitoring points in the same spatial area to form a functional area-level metering unit. The raw energy consumption data collected at the physical circuit level is decoupled by the identification unit built into the sub-metering integrated module to decompose the energy consumption components of different types of energy-consuming equipment in the circuit. Based on the dual-level metering mapping model, the decoupled energy consumption components are re-aggregated with the corresponding functional areas in the spatial function level to generate a sub-item energy consumption dataset that is decomposed according to both spatial function and energy-consuming equipment type.

4. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, The process of binding the acquired real-time energy consumption data with the corresponding spatial identifier in the spatialized point database based on the physical location of the monitoring points specifically includes: Each monitoring point collects real-time energy consumption data at a preset frequency, and the heterogeneous energy consumption data output by different metering devices is uniformly converted into standard data packets through a multi-protocol conversion unit. The standard data packet is encapsulated with its own three-dimensional spatial code, either pre-set or automatically acquired, to generate a standardized data packet with a spatial tag. The standardized data message is parsed, and the three-dimensional spatial code is used as a unique key value. The mapping relationship between the time series database and the spatialized point database is searched and established. The real-time energy consumption value is written into the data table corresponding to the spatial code to generate spatialized energy consumption time series data containing accurate spatial location information. The spatialized energy consumption time-series data containing precise spatial location information is synchronized to the digital twin model, driving the three-dimensional primitives corresponding to the spatial location in the digital twin model to be updated.

5. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, The construction process of the intelligent agent specifically includes: Based on the thermal zoning of the digital twin model and the three-dimensional spatial coding in the spatialized point database, the building physical space is divided into multiple three-dimensional voxelized control units, and each three-dimensional voxelized control unit is deployed with an edge agent; the edge agent includes an environmental perception fusion module, a graph neural network decision module, and a twin collaboration interface; The environmental perception fusion module acquires real-time spatialized energy consumption data and environmental parameters of all monitoring points within the three-dimensional voxelized control unit under its jurisdiction, and uses a three-dimensional spatial interpolation algorithm to perform spatial continuity processing on the discrete point data to generate a regional three-dimensional state field that reflects the multi-dimensional distribution of temperature field, humidity field, illuminance field and energy consumption field within the unit, and synchronizes it to the digital twin model for visualization mapping. The graph neural network decision module takes the three-dimensional state field of the region as input, extracts the spatial gradient features and temporal evolution features in the three-dimensional state field of the region through graph convolution operation, generates a candidate control strategy set by combining a preset multi-objective optimization function, and encapsulates the candidate control strategy set and the spatial identifier of the associated three-dimensional voxelized control unit through the twin cooperative interface. The twin collaboration interface calculates the spatial correlation degree with the neighboring edge agent based on the three-dimensional spatial encoding in the spatialized point database. When the spatial correlation degree exceeds the preset collaboration threshold, it sends the candidate control strategy set to the corresponding neighboring edge agent and receives the candidate control strategy set sent by the neighboring edge agent. The graph neural network decision module evaluates the coupling effect of the boundary influence of the three-dimensional state field of its own region based on the received candidate control strategy set. With the goal of minimizing global energy consumption and inter-regional comfort deviation, it coordinates the candidate control strategy set to generate regional autonomous control instructions.

6. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, The process by which each intelligent agent calculates the spatial correlation degree between itself based on the spatial identifiers in the spatialized point database, and coordinates the simulation verification results according to the spatial correlation degree, specifically includes: Obtain the three-dimensional spatial codes of all monitoring points within the jurisdiction of the intelligent agent and neighboring intelligent agents, and extract the three-dimensional coordinates of each point; Based on the three-dimensional coordinates, the spatial correlation between any two agents is calculated using a spatial correlation calculation model. The simulation verification results of neighboring agents whose spatial correlation exceeds a preset cooperation threshold are exchanged. With the goal of minimizing global energy consumption and comfort deviation, the simulation verification results of neighboring intelligent agents are used as boundary constraints, and the final control strategy is generated through a distributed cooperative optimizer.

7. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 6, characterized in that, The expression for the spatial correlation calculation model is: ； in, Let be the spatial correlation degree between the i-th agent and the j-th agent. The preset spatial attenuation coefficient, The thermal impact distance is the corrected distance for the vertical thermal stratification effect of a building.

8. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 1, characterized in that, Also includes: The system acquires real-time monitoring data on model prediction bias, physical equipment change events, and seasonal environmental inflection points from the dynamic calibration trigger. When any of these monitoring indicators exceeds a preset threshold, the calibration process is triggered. The twin model acquires the trigger signal, retrieves historical running data from the previous time window, and retrains or calibrates the behavior model and physical model. The model parameters after calibration are encapsulated along with the corresponding calibration time, data period, trigger events, and other information to generate a model version snapshot, which is then synchronized to the digital twin model for hot loading.

9. The building energy efficiency optimization method based on digital twins and multi-agent systems according to claim 8, characterized in that, The process of retraining or calibrating the behavioral and physical models specifically includes: Using spatialized energy consumption time-series data containing precise three-dimensional spatial coordinates, the weights of the behavior model are updated through a spatiotemporal graph convolutional neural network to improve the prediction accuracy of human activity patterns and equipment usage patterns in different spatial regions. Using the spatialized energy consumption data and synchronously collected environmental parameters, the thermal or electrical parameters of the physical model are corrected in reverse using Bayesian inference methods to minimize the overall error between the model output and the measured data.

10. A building energy efficiency optimization system based on digital twins and multi-agent systems, characterized in that, The building energy efficiency optimization method based on digital twins and multi-agent systems, applied to any one of claims 1-9, specifically includes: The digital twin model building unit is used to acquire the building information of the target building to build a digital twin model, and assign a unique three-dimensional spatial code to each monitoring point to generate a spatialized point database. The sub-metering system and data mapping unit are used to construct a hierarchical metering system with multiple levels based on the building power distribution system, and bind the acquired real-time energy consumption data with the corresponding spatial identifier in the spatialized point database according to the physical location of the monitoring points to generate spatialized energy consumption data. A multi-agent cluster construction unit is used to deploy corresponding agents for each control area based on the digital twin model and the spatialized point database; each agent acquires monitoring point data and corresponding spatialized energy consumption data within its jurisdiction, generates a regional state field, and establishes a data interaction channel with the digital twin basic model. The simulation verification unit is used for each of the intelligent agents to generate candidate control strategies based on the regional state field, and to call the digital twin model to perform simulation verification on the candidate control strategies to obtain simulation verification results. The digital twin intelligent agent collaborative decision-making unit is used for each of the intelligent agents to calculate the spatial correlation between each other based on the spatial identifiers in the spatialized point database, and to coordinate the simulation verification results according to the spatial correlation to generate a control strategy.

Images