A smart energy management and energy saving optimization system and method applied to a building

Abstract

The application relates to the technical field of building intelligent energy management, and discloses a system and method for intelligent energy management and energy-saving optimization applied to buildings. The method comprises the following steps: energy-carbon dual-dimension perception and hybrid energy storage twin modeling, adaptive granularity decision and variable resolution prediction, dynamic partition carbon constraint optimization deduction and integrated checking, physical-virtual mapping and edge layered autonomous execution, and energy-carbon dual-dimension monitoring and double-channel feedback evolution. Through energy-carbon dual-dimension modeling, adaptive granularity prediction and dynamic partition optimization, in combination with carbon emission tokens and safety checking, and with the aid of layered execution and double-feedback frequency division evolution, the building energy-saving and carbon-reducing effect and safety are improved.

Description

Technical Field

[0001] This invention relates to the field of building smart energy management technology, and in particular to a smart energy management and energy-saving optimization system and method for buildings. Background Technology

[0002] Building operation phases account for a significant proportion of total social carbon emissions, making smart energy management and energy-saving optimization key pathways to achieving low-carbon development goals in the building sector.

[0003] Currently, technologies such as digital twin-based building modeling, AI-based load forecasting, model predictive control (MPC)-based optimization scheduling, and carbon emission accounting based on carbon flow tracking generally suffer from the following structural defects: First, the prediction time domain, control time domain, and rolling cycle of MPC are usually fixed, which cannot adapt to the physical characteristics of building thermal inertia that change dynamically with the seasons and the state of the building envelope, nor can they respond to the high-frequency fluctuations of the dynamic carbon emission factor of the power grid. This results in ineffective computational redundancy during transitional seasons and a loss of pre-cooling and pre-heating foresight under extreme conditions. Second, the control zoning is based on fixed physical boundaries, which is incompatible with the dynamic migration of people and the thermal environment within the building, leading to oversupply in uninhabited areas and undersupply in concentrated areas. If dynamic zoning is introduced, there will be command conflicts between dynamic boundaries and fixed terminal equipment. Third, global centralized optimization is slow to respond to local sudden disturbances, while granting complete autonomy to the terminal may deviate from the global carbon emission optimization target, and there is a lack of effective coordination mechanism between global and local decisions.

[0004] Therefore, there is an urgent need for a building-based intelligent energy management and energy-saving optimization system and method that can deeply couple temporal adaptation, spatial adaptation, and decision-level adaptation into a closed-loop whole. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a smart energy management and energy-saving optimization system and method for buildings. Through energy-carbon dual-dimensional modeling, adaptive granular prediction, and dynamic zoning optimization, combined with carbon emission tokens and security verification, and leveraging hierarchical execution and dual-feedback frequency-division evolution, it enhances the energy-saving and carbon-reduction effects and safety of buildings.

[0006] The objective of this invention can be achieved through the following technical solutions: The first aspect of this disclosure provides a method for intelligent energy management and energy-saving optimization applied to buildings, comprising the following steps: S1. Energy-carbon dual-dimensional sensing and hybrid energy storage twin modeling: Energy-carbon data stream is formed by spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors; thermal inertia parameter identification is performed in parallel based on energy-carbon data stream to construct hybrid energy storage model, and carbon stream collaborative correction is performed to output thermal inertia time constant τ, hybrid energy storage state of charge SOC and energy-carbon dual-label data package. S2, Adaptive Granularity Decision and Variable Resolution Prediction: First, an AI prediction engine with a built-in variable resolution mechanism is constructed. Then, the dynamic granularity adaptation layer uses the carbon emission factor volatility, thermal inertia time constant and model historical prediction error calculated from the dynamic carbon emission factor sequence in the energy-carbon dual-label data package as input. The time domain parameters are decided according to the mapping logic, so that the rolling period in the time domain parameters is used as the output granularity and the prediction time domain is used as the sequence length to generate the load, temperature field and carbon emission factor prediction sequences of the corresponding granularity. S3, Dynamic Partition Carbon Constraint Optimization Deduction and Verification Integration: Based on environmental data, virtual partitions are dynamically clustered to generate virtual partitions and a physical-virtual two-layer mapping is established; with time-domain parameters and predicted sequences, energy-carbon dual-label data packets and hybrid energy storage state of charge as inputs, the virtual partition control strategy and carbon emission budget token are output through rolling optimization; the virtual energy storage boundary is deduced and verified synchronously with the rolling cycle step size; if it fails, it reverts to the previous safety strategy and triggers a re-solution, outputting safety control instructions and tokens; S4, Physical-Virtual Mapping and Edge Layered Autonomous Execution: The verified control strategy is weighted and decomposed into physical grid control quantities through a two-layer mapping model and distributed using an incremental version number synchronization mechanism; the regional agent makes autonomous decisions on local control within the carbon emission budget token quota, and applies for additional quotas from the global coordinator if the limit is exceeded. S5. Energy-Carbon Dual-Dimensional Monitoring and Dual-Channel Feedback Evolution: A correction uncertainty display layer is superimposed on the carbon flow map, and a token remaining budget warning is superimposed on the carbon intensity map; a dual-channel feedback is formed by global prediction and actual deviation, and local execution and token deviation, and the global model parameters and regional model parameters are updated at different frequencies. The strategy cold start is accelerated by combining global and regional dual-layer case library retrieval.

[0007] Further, S1 includes: S11. Collect multi-source energy consumption and environmental sensing data across the entire domain at differentiated frequencies, achieve spatiotemporal structure alignment through time synchronization and spatial grid mapping, inject dynamic carbon emission factors in real time and multiply and transform them to form a raw energy-carbon dual-dimensional data stream with spatiotemporal dual labels. S12. Using the indoor and outdoor temperature sequences and heating / cooling power sequences contained in the energy-carbon dual-dimensional data stream as input, the thermal inertia of the building envelope is equivalent to a virtual thermal energy storage model using the lumped parameter method, with the equivalent heat storage capacity... E v It is a state variable and constitutes a hybrid energy storage system alongside physical energy storage; the comprehensive thermal inertia time constant τ is identified online, and based on the equivalent heat storage... E v The equivalent state of charge of the virtual thermal energy storage is calculated using normalized values ​​within the allowable heat storage range. SOC v ; S13. Based on the component power consumption, dynamic carbon emission factor and total carbon emission data in the energy-carbon dual-dimensional data stream, and based on the tree-like carbon flow topology, the initial allocation of the current flow tracking and the weighted recursive least squares correction with forgetting factor, a corrected energy-carbon dual-label data packet is generated.

[0008] Furthermore, the equivalent heat storage E v The update formula is: ; in, t The thermal inertia time constant, or s For heat storage efficiency, For rolling period, P ch ( t The value is determined by the current indoor-outdoor temperature difference Δ. T ( t ) and total thermal conductivity of the building envelope K s Calculated charge / discharge power: ;when hour, Defined as the equivalent charging state of thermal virtual energy storage, when hour, Defined as the equivalent energy release state of thermal virtual energy storage; The equivalent state of charge of the thermal virtual energy storage SOC v The calculation formula is: ; in, E v,min and E v,max It is determined by the safe temperature range of the building envelope, the heat storage coefficient, and the allowable condensation / overheating boundary.

[0009] Furthermore, the weighted recursive least squares correction with a forgetting factor includes: Constructing a tree-oriented directed graph The root node is the master table, and the leaf nodes are the energy-consuming devices at each end. The power flow tracing algorithm is used to initially allocate the carbon emissions of the master table according to the energy consumption ratio of each branch, and the initial carbon emission allocation coefficient vector is obtained. Construct the weighted least squares objective function: ; in, For the first The accuracy weight of the metering device in each branch, where λ is the regularization coefficient; Solve using a recursive approach: ; Introducing the forgetting factor r The update formula for the error covariance matrix is ​​exponentially weighted: ; Using the corrected carbon emission allocation coefficient vector Recalculate carbon emissions for each branch It is then packaged with the corresponding energy consumption data, spatial grid ID, timestamp, dynamic carbon emission factor, and correction uncertainty to form an energy-carbon dual-label data package.

[0010] Further, S2 includes: S21. Construct a variable-resolution AI prediction engine: A multi-dimensional time series prediction model with a physical-data fusion deep learning architecture is adopted. The model is trained by constructing samples with the past 24 hours of data as input and the next 6 hours as the target. The intrinsic output resolution of the model is 5 minutes. A variable-resolution output interface is configured. After receiving the target granularity instruction, the 5-minute prediction sequence is resampled on the time axis and the uncertainty is recalibrated. The corresponding coarse-grained prediction value and statistically consistent confidence interval are output. S22, Real-time decision-making time-domain parameters of the dynamic granularity adaptation layer: based on thermal inertia time constant. t Load forecast error rolling variance σ² and dynamic carbon emission factor volatility n As input, the time-domain prediction is output through mapping logic. Control Time Domain With rolling cycle ; S23. Generation of variable resolution prediction sequence: The AI ​​prediction engine receives the rolling period. and prediction time domain Instructions, Generate the Future Within a time period, the particle size is The system includes a load forecast sequence, an indoor temperature field forecast sequence, a dynamic carbon emission factor forecast sequence, and an indoor-outdoor temperature difference forecast sequence. The indoor-outdoor temperature difference forecast sequence is fed back to the thermal virtual energy storage model for rolling updates of the equivalent heat storage. The state trajectory.

[0011] Furthermore, the mapping logic is as follows: when Hours and and At that time, take Hour, Hour, minute; when Hours or or At that time, take Hour, Hour, minute; Take the remaining intermediate cases Hour, Hour, minute; And satisfy Hour, Hours, and , and All Integer multiples of; Take a value from the set {10 minutes, 15 minutes, 20 minutes}; The dynamic granularity adaptation layer has built-in hysteresis switching logic. When a change in the indicator triggers a switching condition, the switching is only performed after the condition is met within two consecutive fixed basic judgment periods.

[0012] Further, S3 includes: S31. Spatial Adaptive Partitioning and Mapping Modeling: Extract the three normalized feature values ​​of personnel density, heat flux density and CO2 concentration from the environmental and personnel data stream to form a feature matrix. Use the DBSCAN variant algorithm with adaptive neighborhood radius for clustering to divide the entire building area into several virtual partitions. Establish a two-layer mapping model of physical grid-virtual partition. Use a sigmoid membership smooth transition function to allocate control quantities to the boundary grid. S32. Multi-objective Rolling Optimization and Carbon Emission Budget Token Coexistence: A rolling optimization problem is constructed using time-domain parameters. The optimization variables are the control objective sequences of each virtual partition and the hybrid energy storage charging and discharging power plan. The objective function is the weighted sum of six predicted values: energy cost, carbon emission cost, thermal comfort deviation, equipment life loss, renewable energy consumption, and grid interaction benefits. The solver outputs the rolling optimization coexistence output for each rolling cycle in the future control time domain. The sequence of virtual partition control strategies at any given time and the carbon emission budget tokens for each virtual partition calculated based on carbon flow distribution, personnel prediction and hybrid energy storage allocation; S33. Variable Step Size Derivation and Virtual Energy Storage Boundary Closed-Loop Verification: Digital Twin Simulation Engine Receives Rolling Cycle The instruction sets the simulation step size to match the rolling period, enabling real-time simulation of the control strategy; it employs an early termination strategy, adjusting the equivalent heat storage step size by step. The virtual energy storage equivalent temperature is calculated by back-calculating the heat storage coefficient and compared point by point with the safety temperature limit of the building envelope; if the limit is exceeded, the simulation is terminated, the previous safety strategy is rolled back and the solution is triggered again; if the limit is not exceeded throughout the process, the verification is passed and the current control strategy sequence is stored in the strategy cache area.

[0013] Further, S4 includes: S41. Physical-Virtual Mapping Command Decomposition and Issuance: For each physical grid, retrieve all its virtual partitions and corresponding fusion weights. The final control quantity is the weighted average of the control targets of each virtual partition. ; in, The set of virtual partitions to which the physical grid belongs. To normalize the fusion weights, For virtual partitions The target value for control; the mapping table update adopts a version number incremental synchronization mechanism, which only generates incremental entries for physical meshes where the boundary changes and sends them to the edge nodes; S42. Edge-layered autonomous execution and token-constrained coordination: Regional agents receive physical grid-level control commands as execution baselines and physical grid set-level carbon emission budget tokens, and independently make micro-control adjustments within the token limit; if the autonomous decision is expected to cause actual carbon emissions to exceed the token limit, an additional quota application is sent to the global coordinator; the global coordinator decides to approve or reject based on the remaining global carbon emission budget and the real-time carbon emission pressure of each region.

[0014] Further, S5 includes: S51. Energy-Carbon Dual-Dimensional Enhanced Visualization and Carbon Emission Budget Early Warning: Overlaying a correction coefficient display layer on the carbon flow Sankey diagram to correct for uncertainty. As a color / linewidth modulation factor, the remaining carbon emission budget progress bar is overlaid on the regional carbon intensity map. When the actual carbon emission in the region reaches the preset warning ratio of the token limit, the color will automatically change to warn and trigger the regional energy-saving mode. S52. Dual-channel deviation feedback and dual-layer knowledge base co-evolution: The global deviation channel feeds back the load prediction deviation, carbon emission factor prediction deviation, and optimization target calculation deviation to the AI ​​prediction engine and the global model prediction control solver; the local deviation channel feeds back the regional comfort compliance rate, the number of token over-limits, and the cumulative deviation between actual carbon emissions and token carbon emissions to the regional intelligent agent decision model; a global case library and a regional case library are established, and during combined retrieval, the top 3 cases with the highest feature matching degree are selected respectively and the initial strategy parameters are generated by weighted average based on similarity; the global model performs an online incremental update every 24 hours, and the regional model performs an online incremental update every 1 hour.

[0015] A second aspect of this disclosure provides a smart energy management and energy-saving optimization system for buildings, performing the method described above, including: The data acquisition and twin modeling module is used to form an energy-carbon data stream through spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors. It performs parallel thermal inertia parameter identification to construct a hybrid energy storage model and carbon flow collaborative correction, and outputs thermal inertia time constant τ, hybrid energy storage state of charge (SOC), and energy-carbon dual-label data packets. The adaptive granularity decision-making and variable resolution prediction module is used to generate load, temperature field and carbon emission factor prediction sequences of corresponding granularity by using the built-in variable resolution mechanism of the AI ​​prediction engine and the dynamic granularity adaptation layer, with carbon emission factor volatility, thermal inertia time constant and model historical prediction error as input decision time domain parameters. The dynamic partition carbon constraint optimization deduction and verification module is used to dynamically cluster and generate virtual partitions based on environmental data and establish a physical-virtual two-layer mapping. It uses time-domain parameters and predicted sequences, energy-carbon dual-label data packages and hybrid energy storage state of charge as inputs to continuously optimize the symbiotic output of virtual partition control strategies and carbon emission budget tokens. It deduces and verifies the virtual energy storage boundary synchronously with the rolling cycle step size. The physical-virtual mapping and edge-layered autonomous execution module is used to decompose the verified control strategy into physical grid control quantities through a two-layer mapping model and distribute them using a version number incremental synchronization mechanism, and enable regional agents to make autonomous decisions on local control within the carbon emission budget token limit. The dual-dimensional monitoring and dual-channel feedback evolution module for energy and carbon is used to overlay a correction uncertainty display layer on the carbon flow map and overlay a token remaining budget warning on the carbon intensity map. It uses global prediction and actual deviation, local execution and token deviation to form a dual-channel feedback frequency division to update model parameters, and accelerates the cold start of the strategy through a combination of global and regional case libraries.

[0016] The beneficial effects of this invention are: This invention utilizes dual-dimensional energy-carbon sensing and hybrid energy storage twin modeling to equate building thermal inertia to virtual thermal energy storage and unify it with physical energy storage using a standardized SOC interface. Simultaneously, it incorporates weighted recursive least squares with a forgetting factor for online closed-loop carbon flow correction, generating high-fidelity dual-label energy-carbon data packages and hybrid energy storage states in real time, including correction uncertainties. This provides a precise and reliable data foundation for optimization decisions. Building upon this, a dynamic granularity adaptation layer jointly determines time-domain parameters based on three indicators: thermal inertia time constant, prediction error variance, and carbon emission factor volatility. This drives a single-model variable-resolution AI prediction engine to achieve single-training, multi-granularity output, reducing computational load while ensuring statistical consistency of prediction confidence intervals across different time scales. This achieves dynamic matching between prediction granularity and physical response characteristics and external uncertainties. Furthermore, by generating virtual partitions based on three-dimensional feature adaptive clustering and establishing a physical-virtual dual-layer mapping, it drives control strategies and carbon emission budget orders in multi-objective rolling optimization. The system outputs data in symbiotic fashion and, relying on a digital twin simulation engine, simulates virtual energy storage boundaries with a rolling cycle and synchronous step size. It proactively intercepts the risk of exceeding the equivalent temperature limit with an early termination strategy. If the limit is not met, it automatically reverts to a safety strategy and triggers a re-solution with strong penalties, achieving a balance between global economic optimization, carbon emission quota control, and the safety of the building envelope. Through weighted mapping decomposition and incremental version number synchronization mechanisms, it ensures continuous and conflict-free issuance of instructions under dynamic partition boundaries. At the same time, it introduces an edge autonomous decision-making mechanism and an over-limit quota arbitration mechanism under carbon emission token quota constraints, which takes into account both global carbon emission optimization and achieves second-level local disturbance response. Finally, through dual-channel feedback of global and local deviations, it drives the global model (24 hours) and regional model (1 hour) to evolve incrementally online with differentiated cycles. It also generates initial strategies by combining and weighting the global and regional dual-layer case libraries, significantly shortening the cold start time and forming a complete closed loop of accurate perception, adaptive decision-making, safety simulation, layered execution, and continuous evolution. Attached Figure Description

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

[0018] Figure 1 This is a schematic diagram illustrating the steps of a smart energy management and energy-saving optimization method for buildings provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the steps for the integrated dynamic partitioning carbon constraint optimization deduction and verification provided in the embodiments of the present invention. Detailed Implementation

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

[0020] Example 1

[0021] This embodiment provides a smart energy management and energy-saving optimization method for buildings, such as... Figure 1 As shown, it includes the following steps: S1. Energy-carbon dual-dimensional sensing and hybrid energy storage twin modeling: Energy-carbon data stream is formed by spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors; thermal inertia parameter identification is performed in parallel based on the energy-carbon data stream to construct a hybrid energy storage model, and the thermal inertia time constant τ, hybrid energy storage state of charge (SOC) and energy-carbon dual-label data package are output through carbon stream collaborative correction.

[0022] This step forms the data and model foundation of the overall technical solution, constructing a building digital twin with "self-awareness" capabilities. This digital twin is not merely a static data acquisition and storage process; rather, through an embedded physical model and data correction algorithms, it processes the raw data stream in real time and outputs three standardized information products directly usable for optimization calculations to subsequent steps: thermal inertia time constant τ, energy-carbon dual-label data packets, and hybrid energy storage state of charge (SOC). Specifically, it includes the following steps: S11. Collect multi-source energy consumption and environmental sensing data across the entire domain at differentiated frequencies, achieve spatiotemporal structure alignment through high-precision time synchronization and spatial grid mapping, inject dynamic carbon emission factors in real time and multiply and transform them to form a raw energy-carbon dual-dimensional data stream with spatiotemporal dual labels. Specifically, operational data is collected across the entire building using building automation, energy management systems, and independent sensor networks at differentiated frequencies: smart meters, heating and cooling meters, and gas meters provide millisecond to second-level energy consumption pulses; temperature (including indoor air temperature and building envelope wall temperature), humidity, CO2, and human infrared sensors provide second to minute-level environmental and personnel data streams; air volume is acquired by VAV terminal air volume sensors, providing second-level air volume data streams; and weather stations or API interfaces provide minute-level outdoor temperature and humidity, and solar irradiance. All data is stamped with a unified UTC timestamp using a high-precision time synchronization protocol and aggregated and aligned in a windowed manner according to a preset baseline step size (e.g., 5 minutes) to form a time series matrix without missing or misaligned data. Each data point is mapped to a unique identifier in the building space grid based on its location, achieving spatial structuring. Dynamic carbon emission factor signals corresponding to different energy types, such as electricity, gas, and purchased cooling / heating, are acquired in real time from the electricity market, grid interface, and energy supply-side interface. These signals are then multiplied in real time with the corresponding time-aligned energy consumption data to complete the instantaneous conversion from energy consumption to carbon emissions, generating a raw energy-carbon two-dimensional data stream for each moment and each spatial grid. This data stream serves as the common data foundation for S12 thermal inertia identification and S13 carbon flow correction. Indoor and outdoor temperatures, building envelope inner surface temperatures, and heating / cooling power sequences are supplied to the S12 parameter identifier, while individual electricity consumption and total meter carbon emission data are supplied to the S13 carbon flow allocation and closed-loop correction.

[0023] S12. Using the spatiotemporally aligned data provided by S11, which includes indoor and outdoor temperature sequences and heating / cooling power sequences, as input, the thermal inertia of the building envelope is parameterized into a thermal virtual energy storage model using the lumped parameter method. The equivalent heat storage is used as the state variable and is combined with physical energy storage to form a hybrid energy storage model. The comprehensive thermal inertia time constant is identified online and output to S2 as the source input signal for time-domain adaptive decision-making.

[0024] Specifically, the lumped parameter method is used to equate the thermal inertia of the building envelope, such as exterior walls, roof, and interior partitions, to a virtual thermal energy storage model, defining the heat storage coefficient. C s Exothermic time constant t With heat storage efficiency or s Three parameters; where τ is updated online in real time by the parameter identifier. An equivalent heat storage capacity is added to the digital twin state space. E v As a continuous state variable, it is used to quantify the actual thermal energy stored in the building envelope at the current moment; during the initial operation phase, the charging and discharging power is calculated and dynamically updated using the current measured indoor and outdoor temperature difference provided by S11. E v Once S2 feeds back the indoor-outdoor temperature difference prediction sequence, it uses this prediction sequence to continuously update future time periods. E vThe state trajectory. Update step size Δ t The initial base step size (5 minutes) is set using S11, and the subsequent rolling period Δ is dynamically determined by S2. t A unified adjustment is made to ensure that the virtual energy storage status update aligns with the optimized time-domain grid. The specific update formula is as follows: ; This enables the virtual energy storage state to dynamically respond to future pre-cooling and pre-heating strategies; where the charging and discharging power... Based on the current indoor and outdoor temperature difference Total thermal conductivity of the building envelope (Calculated from building thermal parameters such as the area of ​​the building envelope, the thermal conductivity of the material, and the thickness input in S11) Calculation: .when When the outdoor temperature is higher than the indoor temperature, This indicates that external heat is transferred to the indoor side through the building envelope, causing the virtual thermal energy storage state Ev to be updated in the positive direction, and is defined as the equivalent charging state of the virtual thermal energy storage; when hour, This indicates that indoor heat is lost to the outside through the building envelope, and... Updated according to the negative term, defined as the equivalent energy release state of thermal virtual energy storage. In this calculation formula... The convention of positive and negative signs and In the update equation The items remain consistent.

[0025] Thermal virtual energy storage and physical energy storage, representing batteries, are uniformly incorporated into the hybrid energy storage system at the optimization and scheduling level. Both are represented using a standardized State of Charge (SOC) interface and included in the S3 optimization solver as schedulable resources on par with batteries. The SOC of physical energy storage is calculated based on its rated capacity and current remaining capacity; the equivalent SOC of thermal virtual energy storage is based on the current... E v The normalized value is calculated within the allowable heat storage range, specifically as SOC. v =( E v - E v,min ) / ( E v,max - E v,min ),in E v,min and E v,maxThe temperature range of the building envelope, the heat storage coefficient, and the allowable condensation / overheating boundaries are all determined together. Simultaneously, a built-in parameter identifier utilizes historical time-series data of the building envelope wall surface and indoor / outdoor temperatures output by S11. By fitting the dynamic response curve of temperature to excitations (such as changes in outdoor temperature or switching of indoor heating / cooling power), it identifies and outputs the thermal inertia time constant in real time online. t , which serves as the source input signal for S2 time-domain adaptive decision-making.

[0026] S13. Based on the sub-item power consumption, dynamic carbon emission factor and total carbon emission data output by S11, and based on the tree-like carbon flow topology, the initial allocation of the current flow tracking and the weighted recursive least squares correction with forgetting factor, a corrected energy-carbon dual-label data packet is generated.

[0027] It should be noted that, utilizing the spatiotemporally tagged energy consumption and carbon emission data synchronously generated in S11, a tree-like carbon flow allocation model from the main meter to each terminal branch is established based on the building's power distribution and heating / cooling pipe network topology. The main meter's carbon emission is initially allocated to each terminal branch using a power flow tracing algorithm. A weighted recursive least squares closed-loop correction layer is added, minimizing the residual between the main meter's carbon emission and the reconstructed values ​​of each sub-item's carbon emission after initial allocation. A forgetting factor is introduced to track metering drift and weights are set according to the meter's accuracy level. The corrected carbon emission allocation coefficient vector is solved online to correct each carbon flow branch. The corrected carbon emission data is bound to time, space, and corresponding energy consumption data, encapsulated into an energy-carbon dual-tagged data package, serving as input for S3 carbon emission cost calculation and a data source for S5 carbon flow visualization. S12 and S13 run in parallel, characterizing the same digital twin from the two dimensions of building thermal inertia and carbon footprint, respectively. The hybrid energy storage state output by both, along with the energy-carbon dual-tagged data package, jointly supports the optimization decision of S3. Specifically, the thermal inertia time constant output by S12... t In addition to being directly output to S2 as input for time-domain adaptive decision-making, it is also embedded in the digital twin simulation engine for virtual energy storage boundary deduction and verification in S33; in the energy-carbon dual-label data package output by S13, the corrected sub-item carbon emissions are also included. Used for calculating the carbon emission cost term in the S32 objective function, correcting for uncertainty. The S32 constraint boundary is used for robust uncertainty correction and S51 for visualization and early warning. The digital twin constructed by S1 also provides a simulation interface for the thermal virtual energy storage model to S33, which is used to extrapolate the thermal process and virtual energy storage state before the control strategy is executed.

[0028] Specifically, the corrected energy-carbon dual-label data packet is generated, including: Initial carbon flow topology construction and power flow tracing: using the original energy-carbon two-dimensional data flow (including the power consumption of each component) Carbon emissions in total Using the building's electrical distribution system single-line diagram and hot / cold pipe network topology as input, a tree-structured directed graph is constructed. The root node represents the master table, and the leaf nodes represent the energy-consuming devices at each end. For branches without independent carbon emission metering devices, the energy consumption of each item is first converted into the estimated carbon emission of each item based on the carbon emission factor of the corresponding energy type. C sub,i Then, the power flow tracing algorithm is used to initially allocate the total carbon emissions according to the energy consumption ratio of each branch: ; in branch road Initial carbon emission allocation values, This represents the metered value of the branch's power consumption; the final output is the initial carbon emission allocation coefficient vector. ,in .

[0029] Residual observation and dynamic objective function construction: using initial allocation coefficients Sub-itemized carbon emissions Total carbon emissions Using the input as input, calculate the residual between the sum of the individual carbon emissions at the current time and the total table: ; Construct the weighted least squares objective function: ; in For the first The accuracy weight of the branch metering device is set according to the instrument accuracy class (e.g., the weight of a 0.2 class instrument is set to...). The weight of level 1.0 instruments is set to This allows the readings of high-precision instruments to play a leading role in optimization; The regularization coefficients are used to prevent the allocation coefficients from deviating excessively from the initial physical allocation ratio; the allocation coefficient vector k also satisfies k i ≥ 0, and To ensure that carbon emissions are non-negative and total emissions are conserved after correction for each branch, a projection normalization method is used to map them back to the feasible region when the coefficients obtained by recursion do not meet this constraint. The final output is the residual at the current time step. With weighted objective function .

[0030] Solving the problem using a recursive form with a forgetting factor: The weighted least squares problem is rewritten in recursive form using the residual sequence, objective function, and historical allocation coefficient estimates as inputs, avoiding the need to store all historical data at each step. ; in, Let be the observation matrix, and its first... The row element is the first The energy consumption composition of each branch at the current moment: , This is a vector of carbon emission values ​​for each component. The gain matrix is ​​calculated recursively from the error covariance; a forgetting factor is introduced. The update formula for the error covariance matrix is ​​exponentially weighted, causing the contribution of historical data to decay exponentially over time: ;in Let be the error covariance matrix. For the observation matrix, It is an identity matrix. The final output is the carbon emission allocation coefficient vector after online recursive correction. .

[0031] Standardized encapsulation of corrected carbon-carbon dual-tag data packets: to correct allocation coefficients The original energy-carbon two-dimensional data stream is used as input, and the carbon emissions of each branch are recalculated using the correction allocation coefficient: ; Calculate the correction uncertainty (Based on the diagonal elements of the error covariance matrix); the corrected carbon emission data is packaged with the corresponding energy consumption data, spatial grid ID, timestamp, dynamic carbon emission factor, and uncertainty to form an energy-carbon dual-label data package.

[0032] S2, Adaptive Granularity Decision and Variable Resolution Prediction: First, an AI prediction engine with a built-in variable resolution mechanism is constructed. Then, the dynamic granularity adaptation layer uses the carbon emission factor volatility, thermal inertia time constant, and model historical prediction error calculated from the dynamic carbon emission factor sequence in the S1 energy-carbon dual-label data package as input. It makes decisions on time-domain parameters according to mapping logic, thereby generating load, temperature field, and carbon emission factor prediction sequences of corresponding granularity with the rolling period in the time-domain parameters as the output granularity and the prediction time domain as the sequence length.

[0033] This step is the decision-making hub of the time-domain adaptive loop. Its core function is to determine in real time "the time granularity at which the system looks to the future" based on the current physical characteristics of the building and the uncertainty of external signals, and to generate a multi-dimensional prediction sequence that matches that granularity. This step reverses the traditional order of "predict first, then adapt" to "determine the granularity first, then generate the prediction." Specifically, it includes the following steps: S21. Construction of the Variable Resolution AI Prediction Engine: This involves constructing the variable resolution AI prediction engine (hereinafter referred to as the AI ​​prediction engine) upon which subsequent adaptation layer decisions and prediction generation rely. This engine is a multi-dimensional time-series prediction model employing a physical-data fusion deep learning architecture, responsible for outputting predicted sequences of load, indoor temperature field, and carbon emission factors at corresponding granularities under any specified rolling period. Specifically, this includes: The model employs a shared temporal convolutional-graph neural network hybrid encoder paired with three task-specific prediction heads: a fully connected-deconvolutional structure for the load prediction head, a graph neural network decoder for the temperature field prediction head, and a lightweight regression network for the carbon emission factor prediction head. The encoder is designed to be structurally robust to time-axis scaling.

[0034] Training data is taken from the historical database of the S1 digital twin, which contains energy-carbon dual-label data packages, measured indoor temperature field data, outdoor meteorological data, and dynamic carbon emission factor records accumulated over the past 12 months of operation. This historical database is formed by S1 continuously archiving energy-carbon dual-label data packages and corresponding sensor data during operation, all aligned with physical grid spatial identifiers at a 5-minute granularity. A sliding window is constructed using the past 24 hours (288 steps) as input and the next 6 hours (72 steps) as the target, and the training / validation / test sets are split chronologically. Training employs a negative log-likelihood loss to jointly learn the predicted mean and uncertainty, with the AdamW optimizer combined with cosine annealing and early stopping mechanisms. The three task losses are inversely weighted according to the convergence speed of the validation set. The intrinsic output resolution of the trained model is 5 minutes, and the target granularity of the subsequent adaptation layers is... The interval is limited to an integer multiple of 5 minutes to ensure that the variable resolution output interface can achieve misaligned resampling through integer window aggregation. The loss function is in the form of: ; In the formula, N is the total sample size, and y i For the true value, To predict the mean, To predict variance, This represents the predicted standard deviation.

[0035] The model is then configured with a variable resolution output interface: this interface receives the target granularity instruction and performs time-axis resampling (according to the target granularity) on the 5-minute prediction sequence. The process involves sliding window aggregation of 5-minute sequences (e.g., obtaining coarse-grained predicted values ​​by taking the window mean or median) and uncertainty recalibration (transferring the uncertainty of the 5-minute predictions (confidence intervals from the model output or historical error statistics) to the aggregated coarse-grained values ​​using the square root rule of the sum of variances, and recalibrating their confidence intervals). The output includes corresponding coarse-grained predicted values ​​and statistically consistent confidence intervals, all without requiring retraining or parallel processing of multiple models. Specifically, in the variable resolution output interface, uncertainty recalibration directly uses the values ​​learned during the training phase. Values ​​are subjected to variance and propagation, for values ​​within the window. The independent prediction formula is: ; in, This is the prediction variance of each 5-minute prediction point within the window.

[0036] Understandably, traditional prediction engines operate at a single fixed resolution, requiring retraining or redeploying multiple models to switch granularities. The single-model variable-resolution architecture in this embodiment, through structurally robust design to time scales and a resampling-recalibration mechanism at the output, achieves one-time training and multi-granularity output, ensuring that the prediction confidence interval does not statistically jump when the S22 adaptation layer dynamically switches rolling cycles. In this embodiment, the AI ​​prediction engine is only responsible for outputting the corresponding prediction result at a given target granularity and is not responsible for the target granularity decision; the target granularity is independently decided by the S22 dynamic granularity adaptation layer.

[0037] S22. Real-time decision-making of time-domain parameters by the dynamic granularity adaptation layer: The dynamic granularity adaptation layer serves as the decision-making module in this step, using the thermal inertia time constant... The dynamic carbon emission factor sequence and the rolling variance of load forecast error calculated from the historical forecast output of the AI ​​forecast engine and the actual load value input from S1. As input, optimal three time-domain parameters are generated through a pre-defined mapping logic, serving as a unified time reference for prediction generation and subsequent optimization. This includes: Input signal acquisition and processing: Thermal inertia time constant The output is generated in real time by the online identifier of S1 and can be directly read through the digital twin state interface of S1 to characterize the speed at which the building envelope responds to energy changes.

[0038] Load forecast error rolling variance The AI ​​prediction engine predicts the rolling period Δt of the previous control cycle within the past evaluation window (e.g., the past 2 hours). old At the granularity level, the moving average is calculated by the square of the difference between the predicted values ​​of electricity, cooling, and heating loads output and the corresponding actual load values ​​input by S1, which is used to quantify the uncertainty level of the current prediction model.

[0039] Dynamic carbon emission factor volatility The coefficient of variation (COP) is calculated from the time series of the dynamic carbon emission factor input by S1, which is the ratio of the standard deviation to the mean in the most recent short time window (e.g., the past 30 minutes). This COP represents the intensity of the external power grid carbon emission signal.

[0040] Rolling variance of load forecasting error during system initial startup or cold start Take empirical initial values (Set by historical statistical values ​​of similar buildings), and switch to the measured sliding variance value after the AI ​​prediction engine has run for a full evaluation window.

[0041] Mapping Logic and Temporal Parameter Output: The adaptation layer maps the three metrics mentioned above to their corresponding temporal parameters using a pre-defined mapping logic, outputting the predicted temporal domain. Control Time Domain With rolling cycle One specific implementation of this mapping logic is a set of time-domain parameter decision rules, including: When thermal inertia is large, prediction error is small, and carbon emission factor is stable, a longer prediction time domain and a longer rolling period are adopted to reduce computational load. When thermal inertia is low, prediction error is large, or carbon emission factor fluctuates drastically, a short prediction time domain and a short rolling cycle are adopted to improve foresight and response speed.

[0042] Furthermore, when Hours (high thermal inertia) and (Small prediction error) and When the carbon emission factor is stable, take Hour, Hour, minute; when Hours (low thermal inertia) or (Large prediction error) or When (carbon emission factor fluctuates drastically), take Hour, Hour, minute; Take the remaining intermediate cases Hour, Hour, minute.

[0043] The specific output is strictly constrained to be: Hour, Hours, and , and All Integer multiples of; Values ​​are taken from the set {10 minutes, 15 minutes, 20 minutes}. This integer multiple constraint ensures that the predicted sequence, the optimization solution step size, and the digital twin inference step size are perfectly aligned in the grid, thus eliminating time mismatch in principle.

[0044] Hysteresis Switching and Notification Mechanism: To avoid frequent jumps in time-domain parameters caused by minor fluctuations in the index near the threshold, the adaptation layer incorporates hysteresis switching logic. This logic applies when the rate of change of the thermal inertia time constant... Exceeding a set threshold (e.g., 20%) or carbon emission factor volatility Not triggered immediately Instead of switching, it requires that the condition be met within two consecutive fixed base decision cycles (each decision cycle length equal to the current one). If all conditions are met within the specified range, a switchover will proceed. If, before the hysteresis determination is completed, manual intervention, fault protection, or a higher-priority scheduling instruction causes the currently valid switchover to fail, the switchover will proceed. If a change occurs, the decision counter is reset to the new one. Restart counting from the baseline.

[0045] After the S22 adaptation layer completes the time-domain parameter decision, it simultaneously executes three distributions: Send to the S21 variable resolution output interface; The data was sent to the S12 thermal virtual energy storage model; The entire process is deployed to the S32 global model predictive control solver and the S33 digital twin simulation engine.

[0046] S23. Generation of Variable Resolution Prediction Sequence and Feedback of S1 State Update: This sub-step, guided by the time-domain parameters determined in S2.2, involves the AI ​​prediction engine generating the final prediction sequence and feeding back the key sequence to S1 to form a closed loop. Specifically, this includes: The AI ​​prediction engine built by S21 receives the rolling cycle issued by the adaptation layer of S22. and prediction time domain Commands (Control Time Domain) By S22 and , The results are passed to the global model prediction control solver in S3 (but do not participate in the prediction sequence generation step of the AI ​​prediction engine). Its variable resolution output interface generates the future according to the resampling and recalibration mechanism described in S21. Within a time period, the particle size is The following predicted sequences: Load forecast sequence: forecast values ​​and confidence intervals of electricity, cooling and heating loads for each sub-item (air conditioning, lighting, power, etc.).

[0047] Indoor temperature field prediction sequence: indoor temperature prediction values ​​for each spatial grid.

[0048] Dynamic carbon emission factor prediction sequence: future carbon emission factor prediction curve.

[0049] Indoor and outdoor temperature difference prediction sequence: The indoor and outdoor temperature difference sequence of each grid is calculated from indoor temperature field prediction and outdoor meteorological prediction data.

[0050] The predicted indoor-outdoor temperature difference sequence is fed back to the thermal virtual energy storage model of S1 in real time. S1 calculates future... The charging and discharging power of the building envelope during the time period And continuously update the equivalent heat storage capacity. The state trajectory forms a two-way data coupling closed loop of "prediction → virtual energy storage state update → next optimization". The remaining prediction sequences are output to S3 as input for global model predictive control rolling optimization. At the same time, the first prediction step temperature difference value in the indoor-outdoor temperature difference prediction sequence generated by S23, or the current indoor-outdoor temperature difference value obtained by real-time state estimation by S1, is used to calculate the heat flux density characteristics in the spatial adaptive partitioning of S31.

[0051] S3, Dynamic Partition Carbon Constraint Optimization Deduction and Verification Integration: Based on S1 environmental data, virtual partitions are dynamically clustered to generate virtual partitions and a physical-virtual two-layer mapping is established; with S2 time-domain parameters and prediction sequences, S1 energy-carbon dual-label data packets and hybrid energy storage state of charge as inputs, the virtual partition control strategy and carbon emission budget token are output in a rolling optimization manner; the virtual energy storage boundary is deduced and verified with the synchronous step size of S2, and if it fails, it reverts to the previous safety strategy and triggers a re-solution, outputting safety control instructions and tokens.

[0052] This step is the core decision-making and security verification stage of the overall technical solution. It encapsulates dynamic spatial partitioning, multi-objective global optimization, co-generation of carbon emission budget tokens, and digital twin simulation verification considering the safety of the building envelope into an atomic transaction. Its inputs are the energy-carbon dual-tag data package provided in real-time by S1 and the hybrid energy storage state (including equivalent heat storage capacity of thermal virtual energy storage). (with state of charge) and people-environment sensor data streams, as well as time-domain parameters for S2 decision-making (prediction time domain) Control Time Domain Rolling cycle The system outputs securely validated, directly executable physical grid-level control commands, regional carbon emission budget tokens for coordinating autonomous edge decisions, and a complete record of the equivalent temperature trajectory for S5 visualization and bias calculation. Figure 2 As shown, the specific steps include: S31. Spatial Adaptive Zoning and Mapping Modeling: From the environmental and personnel data stream of S1, using spatial grid identifiers consistent with energy-carbon dual-label data packets, the personnel density of each grid is extracted in real time. The heat flux density is calculated by combining the real-time indoor-outdoor temperature difference of S1 or the temperature difference value of the first prediction step of S23 with the VAV terminal air supply volume collected in S11 (the calculation formula is q = c). p · ·ΔT, where Here, ΔT is the supply air mass flow rate at the VAV terminal, and c is the temperature difference between the supply and return air. pThe feature matrix is ​​composed of three normalized feature values: specific heat at constant pressure of air, CO2 concentration, and three other features. Clustering is performed at preset intervals (e.g., every 5 minutes) using a variant of the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm that adapts to local k-nearest neighbor distances based on neighborhood radius. This preset interval is fixed at 5 minutes and is synchronized with the rolling period of the S2 decision. The interval should satisfy Multiples of 5 minutes (by S2) The set of values ​​{10 minutes, 15 minutes, 20 minutes} naturally satisfies this constraint, which ensures that each S2 rolling optimization window contains an integer number of spatial partition updates, avoiding unexpected changes in partition status during the optimization window that could lead to mismatches between control commands and physical partitions. The entire building area is divided into several virtual partitions that are homogeneous in terms of three-dimensional features: personnel, heat flow, and air quality. The partition boundaries change dynamically as the features evolve. A two-layer mapping model of physical grid and virtual partition is established: the physical layer retains the original spatial grid corresponding to the fixed controllable equipment group, serving as the unique address anchor point for command issuance; the virtual layer consists of dynamically generated partitions and their unified control objectives; the mapping layer weights the virtual partition control objectives to each physical grid according to the proportion of personnel density and the contribution rate of heat flow (e.g., a weight ratio of 0.6:0.4), and uses a sigmoid-type membership smooth transition function to allocate control quantities to the boundary grids, eliminating abrupt changes in partition boundaries.

[0053] S32, Multi-objective Rolling Optimization and Coexistence of Carbon Budget Tokens: Harvesting the Future from S2 Within a time period, the particle size is The system generates load forecast sequences, indoor temperature field forecast sequences, and dynamic carbon emission factor forecast sequences; it also obtains the current-time energy-carbon dual-label data package and the unified state of charge of hybrid energy storage from S1. Among these, the corrected sub-item carbon emission data in the energy-carbon dual-label data package... It can be directly used to calculate the carbon emission cost term in the objective function to correct for uncertainty. Robust correction for uncertainty in constraint boundaries. Prediction time domain of S2 decision. Control Time Domain Rolling cycle A rolling optimization problem is constructed for the time-domain parameters, with the optimization variables being the control target sequences for each virtual zone and the hybrid energy storage charging and discharging power plan. The objective function is a weighted sum of six predicted values: energy cost, carbon emission cost, thermal comfort deviation, equipment lifespan loss, renewable energy consumption, and grid interaction benefits. Energy cost is calculated based on sub-load forecasts and real-time electricity prices; the carbon emission cost is specifically defined as... ,in In the dynamic carbon emission factor prediction sequence input to S2 Carbon emission weighting factor at any given time The branch in the S1 energy-carbon dual-tag data packet after correction by S13. exist Carbon emissions at any given time (if) For future moments, the corresponding predicted values ​​in the S2 prediction sequence are used instead; thermal comfort deviation is calculated based on the sum of squares of the deviations between the predicted indoor temperature field and the setpoint; equipment lifespan loss is calculated based on the number of equipment start-ups and shutdowns and the rate of output change; renewable energy consumption and grid interaction benefits are calculated based on the power interaction plan with the grid. Constraints include the upper limit of the setpoint difference between adjacent virtual zones, the upper and lower limits of the hybrid energy storage SOC, the equipment output limit, and the virtual energy storage SOC boundary defined by the safe temperature range of the S1 building envelope. The solver outputs two types of results in each rolling optimization: future... Within each The system consists of a sequence of virtual zone control strategies at any given time, and carbon emission budget tokens for each virtual zone calculated based on optimal carbon flow distribution, personnel forecasting, and hybrid energy storage allocation. Tokens and strategies are generated from the same source, in the same domain, and under the same constraints, ensuring that regional quotas naturally meet global optimal requirements.

[0054] S33, Variable Step Size Simulation and Virtual Energy Storage Boundary Closed-Loop Verification: The simulation engine within the digital twin receives synchronous signals from S2. The command sets the simulation step size to match the rolling cycle, and performs real-time simulations of the control strategy and charge / discharge plan output by S32. The virtual energy storage part adopts the same equivalent heat storage capacity as the thermal virtual energy storage model in S1. The equations are updated dynamically.

[0055] Meanwhile, the digital twin simulation engine has a built-in policy cache that continuously stores the complete control policy sequence that most recently passed the virtual energy storage boundary verification and its corresponding equivalent temperature trajectory for subsequent rollback operations. Upon initial system startup, the cache pre-sets a conservative safety policy based on the current steady-state operating state (e.g., maintaining the current setpoint and zero energy storage output) to ensure a usable rollback baseline in case of initial simulation failure. This conservative safety policy also serves as an initial feasible solution reference for the first S32 solution, accelerating solver convergence.

[0056] Employing an early termination strategy: the simulation step size is changed from... The virtual energy storage equivalent temperature is calculated by inversely using the heat storage coefficient, and then compared point-by-point with the safe temperature limit of the building envelope. If in If the equivalent temperature exceeds the limit at any simulation step within the time domain, the simulation is immediately terminated, the verification is deemed failed, and the current cycle instruction is rolled back to the last verified safety policy read from the policy buffer to maintain uninterrupted control. During this safe re-solution cycle, the temporary overwrite is set to the lower limit of the S2 allowable range (if the current...). =20 minutes, then the lower limit is 40 minutes, which is twice the normal value. To ensure integer multiple constraints; the rest The lower limit of the value is 0.5 hours), and an equivalent temperature over-limit penalty term or tightening of the SOC constraint boundary is added to the objective function, triggering a re-solution in S32. This temporary overlay result is synchronously sent back to the S2 adaptation layer for the next cycle's time-domain parameter decision. After the re-solution is completed, the newly generated strategy sequence is resubmitted to S33 for deduction and verification. If the verification fails, the re-solution continues, but the number of re-solutions does not exceed the preset limit. If the equivalent temperature of all simulation steps in the time domain does not exceed the limit, the verification is deemed successful, and the current control strategy sequence and its equivalent temperature trajectory are stored in the strategy buffer as a safety benchmark for the next rollback.

[0057] It should be noted that the final output data structure includes the following fields: control policy sequence (indexed by virtual partition ID, containing the control target value for each partition). This data structure includes: its corresponding timestamp, the corresponding carbon emission budget token sequence, the verification status identifier (pass / fail, with the version number incrementing for pass), and a complete record of the equivalent temperature trajectory (used for visualization and deviation calculation in S5). After being decomposed by the physical-virtual mapping in S41, this data structure is transformed into physical grid-level control instructions. The carbon emission budget token sequence is converted into the equivalent token carbon emission limit for each physical grid set by the two-layer mapping model in S31, which is used for regional autonomous decision-making constraints in S42.

[0058] S4, Physical-Virtual Mapping and Edge-Layered Autonomous Execution: The control strategy verified in S3 is weighted and decomposed into physical grid control quantities through a two-layer mapping model and distributed using an incremental version number synchronization mechanism; the regional agent makes autonomous decisions on local control within the S3 carbon emission token quota, and applies for additional quotas from the global coordinator if the limit is exceeded, achieving second-level response and coordination with the global carbon emission target.

[0059] This step is the implementation phase of the overall technical solution. It is responsible for transforming the security-verified virtual partition control strategy output by S3 into deterministic commands executable by the end devices, and for achieving rapid autonomous response to local disturbances at the edge within the constraint framework of carbon emission budget tokens. Its inputs are the virtual partition control strategy sequence verified by S3, the carbon emission budget tokens for each region, and the physical grid-virtual partition two-layer mapping model established by S3. The outputs are the end device control commands issued via the building automation protocol, and the local micro-control adjustments generated by edge autonomous decision-making. Specifically, it includes the following steps: S41. Physical-Virtual Mapping Command Decomposition and Issuance: The control objectives generated by S3 optimization, oriented towards dynamic virtual partitions, are transformed into conflict-free deterministic commands for fixed physical devices. A reliability mechanism ensures command continuity during dynamic partition boundary changes, including: Mapping and decomposition input: Obtain the current control cycle virtual partition control strategy after passing the virtual energy storage boundary verification from S3. It includes the control target value of each virtual partition (such as temperature setpoint, lighting brightness level, fresh air volume, etc.) and uses the physical grid-virtual partition two-layer mapping model established in S31.

[0060] Weighted mapping decomposition calculation: For each physical grid, retrieve all virtual partitions to which the grid belongs and their corresponding fusion weights in the two-layer mapping model. The fusion weights are determined by S31—for grids within a partition, calculated based on the ratio of personnel density to heat flux contribution (e.g., 0.6:0.4); for boundary grids, calculated using a sigmoid-type membership smoothing transition function. The final control quantity of this physical grid is the weighted average of the control objectives of each virtual partition to which it belongs: ; in, This refers to the set of virtual partitions to which the physical grid belongs. To normalize the fusion weights, For virtual partitions The control target value. The physical grid is the unique address anchor point for issuing control commands. Each grid, after mapping and decomposition, obtains a uniquely determined control quantity, thus structurally eliminating command conflicts.

[0061] Edge node caching and incremental version number synchronization mechanism: To address potential changes in dynamic partition boundaries every 5 minutes, edge nodes locally cache the latest physical grid-virtual partition mapping table, supporting millisecond-level queries. Mapping table updates employ an incremental version number synchronization mechanism: Each time the S31 updates the mapping model, it only generates incremental entries for physical grids with changed boundaries, along with a monotonically increasing version number, and sends them to the edge nodes, rather than pushing them to all nodes. Before executing commands in each control cycle, edge nodes verify that the local mapping table version number matches the cloud version. If a version inconsistency occurs due to communication delays, the edge node executes control using the previously valid locally cached version and immediately sends a synchronization request back to the cloud, ensuring uninterrupted control commands during partition switching.

[0062] Protocol encapsulation and distribution: The decomposed physical grid control quantities are grouped according to device type (VAV terminal, lighting circuit, socket circuit, etc.), encapsulated into BACnet / IP or Modbus TCP standard messages, and distributed to the corresponding terminal controllers for execution via the building automation backbone network.

[0063] S42. Edge-layered autonomous execution and token-constrained coordination: Based on the global optimal baseline command issued in S41, regional agents deployed on the edge gateway are given local autonomous decision-making power within the carbon emission budget token limit to solve the problem of excessive delay in response to sudden local disturbances by global model predictive control.

[0064] Deployment and Baseline Reception of Regional Agents: Regional agents are deployed at the same layer as the S41 command delivery module on the building edge gateway. Each regional agent is responsible for one or more physical grid sets. It receives physical grid-level control commands from S41 as the execution baseline, and simultaneously receives physical grid set-level carbon emission budget tokens converted by the S31 physical grid-virtual partition two-layer mapping model. The tokens originally output by S32, indexed by the virtual partition ID, are converted into equivalent token carbon emission limits executable by the regional agent in S4 according to the fusion weights. S4, through the physical grid-virtual partition two-layer mapping model established by S31, weights and aggregates the virtual partition tokens according to the fusion weights of the partitions to which the physical grids belong, converting them into equivalent token carbon emission limits for each physical grid set.

[0065] Autonomous decision-making within the token limit: The autonomous decision-making cycle of the regional agent is fixed at 5 minutes (this cycle is no greater than the rolling cycle of S2). And satisfy (Constraints are integer multiples thereof), this cycle is consistent with the S31 partition update cycle; based on real-time sensor data of this area (changes in personnel density, changes in local illumination, etc.), within the carbon emission budget token quota, it independently solves a small-scale local optimization (using linear programming or rule-based lookup table decision-making, with the optimization variables being the control quantity fine-tuning values ​​of each physical grid in this area, and the constraints being the token carbon emission upper limit and the equipment physical limit) or executes preset rules, autonomously making micro-control adjustments—such as VAV terminal airflow fine-tuning, adaptive lighting brightness, and socket standby disconnection, etc. This autonomous decision-making does not trigger the global model predictive control re-solution, and the response latency is on the order of seconds.

[0066] Excess Quota Request and Global Decision: If a regional agent's autonomous decision is expected to cause its actual carbon emissions to exceed the token quota, the adjustment is suspended, and an additional quota request is sent to the global coordinator deployed in the cloud (running concurrently with the S32's global model predictive control solver). The global coordinator decides to approve, partially approve, or reject the additional quota based on the current remaining global carbon emission budget and the real-time carbon emission pressure of each region. If rejected, the regional agent must either resolve within the token quota or maintain the baseline execution. This token excess approval mechanism ensures that local autonomous decisions do not compromise global carbon emission optimality.

[0067] Global bias penalty term in the reward function: To guide the regional agent to proactively align with the global carbon emission target through the learning mechanism, when the regional agent adopts a reinforcement learning strategy for local autonomous decision-making, its reward function, in addition to including the comfort achievement rate and local energy saving terms, also includes a penalty term for the square of the difference between the actual regional carbon emission and the token carbon emission. ; In the formula, Defined as a region The negative normalized value of the temperature setpoint tracking deviation of all physical grids within the autonomous decision-making cycle: ,in The dead zone half-width for the temperature setpoint; Defined as a region The normalized positive value of the difference between actual energy consumption and baseline energy consumption (reference energy consumption at the time of token allocation) during the autonomous decision-making cycle: This means that when actual energy consumption exceeds the baseline, the contribution of energy-saving items is zero rather than negative, thus avoiding the abnormal situation of giving negative incentives to non-energy-saving behaviors. , , This is the weighting coefficient. When the regional agent adopts a linear programming or rule lookup table strategy, this squared penalty term is used as a constraint penalty term in the local objective function or rule priority.

[0068] Token balance recovery and secondary allocation: The global coordinator recovers unused token balances from each region on a rolling cycle and merges them into the global token pool. The recovered balances are then weighted and secondary allocated to carbon-scarce regions based on the real-time carbon emission pressure of each region (the ratio of current carbon emission intensity to the token limit), eliminating quota accumulation and improving the utilization rate of the global carbon budget.

[0069] S5, Energy-Carbon Dual-Dimensional Monitoring and Dual-Channel Feedback Evolution: The S1 correction uncertainty display layer is superimposed on the carbon flow map, and the S3 token remaining budget warning is superimposed on the carbon intensity map; the global prediction and actual deviation, and the local execution and token deviation constitute dual-channel feedback, and the global model parameters of S2 and S3 and the regional model parameters of S4 are updated at different frequencies. The cold start of the strategy is accelerated by combining global and regional case library retrieval.

[0070] This step is the feedback loop and continuous evolution link of the overall technical solution, undertaking two functions: first, presenting the system's operating status and carbon emission budget execution in an enhanced visualization format, providing decision support and early warning for operators; second, constructing a dual-channel deviation feedback and dual-layer knowledge accumulation architecture to drive the frequency-based evolution of the global model and regional models, enabling the system strategy to continuously optimize with changes in building usage patterns and the external environment. Its inputs are the energy-carbon dual-label data package and its correction uncertainty output in real time by S1, the prediction sequence of the variable-resolution AI prediction engine by S2, the carbon emission budget tokens and global model predictive control solver parameters output by S3, and the actual execution results and token consumption data of each regional agent by S4. The outputs are the updated global model parameters (feedback to the AI ​​prediction engine in S2 and the global model predictive control solver in S3) and regional model parameters (feedback to the regional agent reward function and decision model in S4), as well as early warning signals on the visualization interface. Specifically, this includes: S51. Enhanced Energy-Carbon Dual-Dimensional Visualization and Carbon Emission Budget Early Warning: Building upon traditional energy and carbon monitoring, a data quality perception layer and a carbon emission budget execution monitoring layer are overlaid, enabling operators to intuitively assess the reliability of carbon flow accounting and the progress of carbon emission budget execution in various regions, and trigger automatic energy-saving linkages when necessary. Specifically, this includes: The correction coefficient display layer in the carbon flow Sankey diagram: Building upon the traditional carbon flow Sankey diagram (showing the distribution path and flow magnitude from the overall carbon emissions to each sub-item and region), a new visualization dimension—the correction coefficient display layer—is added, represented by color depth or line thickness. The input data for this display layer is the energy-carbon dual-label data package output from S13—specifically, the corrected sub-item carbon emissions. Used to determine the flow width of each branch in the Sankey diagram; spatial grid ID is used to determine the start and end nodes of the carbon flow path; and uncertainty is corrected. As a superimposed color / linewidth modulation factor, branches with higher uncertainty have darker Sankey diagram streamlines or thicker lines, visually indicating to operators that the metering data quality of that branch is lower and that the calibration status of the corresponding sub-metering devices needs to be monitored.

[0071] Remaining carbon emission budget overlay on regional carbon intensity maps: On a gridded carbon intensity distribution heatmap of building space, a remaining carbon emission budget progress bar based on the carbon emission budget token output by S32 is overlaid for each region. The progress bar reflects in real time the percentage of carbon emissions consumed in the current token cycle for that region. When the actual carbon emissions of a region reach the preset warning ratio of the token limit (e.g., 85%), the interface automatically changes color to issue a warning; when it reaches or exceeds 95%, within the upper limit of the difference between adjacent virtual partition settings and the allowable range of thermal comfort constraints set in the S32 optimization problem, the region's energy-saving mode is automatically triggered—such as moderately relaxing the temperature setpoint dead zone and reducing the lighting brightness of non-critical areas. This warning-linkage mechanism connects the monitoring end of S5 with the optimization constraint end of S3 and the token execution end of S4 into a complete data-driven control path.

[0072] Measurement, verification, and anomaly diagnosis are linked: The energy-saving effect measurement and verification framework is retained, and the actual effect of energy-saving measures is quantitatively verified based on the corrected energy-carbon dual-label data output by S1. When the carbon emission deviation exceeds the set threshold for multiple consecutive token cycles (e.g., three consecutive token cycles), the anomaly diagnosis process is automatically triggered, and the root cause analysis is performed by tracing back to the data quality of S1, the prediction accuracy of S2, or the execution deviation of S4.

[0073] S52. Dual-channel deviation feedback and dual-layer knowledge base co-evolution: Constructing a hierarchical feedback and hierarchical knowledge accumulation mechanism, enabling global and local strategies to continuously evolve according to their own time scales, and achieving rapid cold start under similar working conditions through combined retrieval. Specifically, this includes: Construction of dual-channel bias feedback: Global Deviation Channel: At the end of each token update cycle or day, global deviation metrics are calculated—including the deviation between the load prediction value of the S2 variable resolution AI prediction engine and the actual load value of S1, the deviation between the predicted carbon emission factor and the actual carbon emission factor, and the deviation between the calculated value and the actual value of the carbon emission cost term in the global model predictive control optimization objective function. These global deviations are fed back to the AI ​​prediction engine described in S21 of S2 (used to update the weight parameters of the shared encoder and the task-specific prediction head based on gradient backpropagation using the negative log-likelihood loss function) and the global model predictive control solver of S32 (used to correct the weight coefficients of each penalty term in its objective function and the safety margin of the constraint boundary online based on the deviation residuals).

[0074] Local Deviation Channel: Based on the actual execution data of each regional agent in S42, local deviation indicators are calculated—including the regional comfort compliance rate (the percentage of time the actual temperature is within the setpoint dead zone), the number of token overruns, and the cumulative deviation between the actual carbon emissions of the region and the token carbon emissions. These local deviations directly drive the weighting coefficients of the reward function of the regional agents in S42. , , The policy is updated every hour, using a policy gradient approach to continuously converge the regional decision-making policy towards the global carbon emission target. This slower timescale compared to the autonomous decision-making cycle (5 minutes) of S42 is intended to smooth out the impact of random perturbations on policy updates.

[0075] The establishment and collaboration of a two-tier knowledge base architecture: Global Case Library: Stores scenario-policy mapping pairs for "Date Type (Weekday / Holiday / Season) + Weather Pattern (Sunny / Cloudy / Rainy) + Carbon Emission Factor Curve Shape (Morning Peak / Midday Peak / Bimodal) → Optimal Global Strategy". Each case record shows the optimal combination of time-domain parameters used by S2. The optimal objective function weight vector and constraint boundary that have been verified by the S32 global model predictive control solver in this scenario.

[0076] Regional Case Library: Stores scenario-action mapping pairs of "personnel density pattern (sparse / normal / gathered / meeting) + thermal interference type (direct sunlight / equipment heat dissipation / door and window draft) → local adjustment action sequence". Each case records the optimal micro-control action sequence (such as VAV airflow adjustment rate, lighting dimming slope) of the S42 regional agent in that local scenario, verified by the reward function.

[0077] Combined Search and Initial Strategy Generation: When the system starts up or encounters a new operating condition, a similar day search is performed. The matching rules for the global search include date type, weather pattern, and carbon emission factor curve shape; in one implementation, weather patterns can be categorized by total cloud cover and precipitation, and carbon emission factor curve shape can be categorized by peak time period. The matching rules for the regional search are: personnel density is based on the average number of people per grid. Mapped as "sparse" ( "Normal" (person / m²) people / m²), "gathering" ( people / m²), "meeting" (person / m²). Thermal interference type by direct solar irradiance. Equipment heat dissipation power Air leakage through doors and windows The comprehensive threshold determination is performed. During combined retrieval, the top three cases with the highest feature matching degree from both the global case library and the regional case library are selected, and the initial strategy parameters are generated by weighted averaging based on similarity, rather than selecting only a single best-matching case. The combination of both to generate the initial strategy significantly shortens the system's cold start time.

[0078] Differentiated online incremental updates: Matching the fundamental differences in time scale between the global model and the regional models, differentiated update frequencies are implemented. Specifically, the AI ​​prediction engine in S2 performs online incremental training every 24 hours using global bias (updating network weights via gradient backpropagation based on the negative log-likelihood loss function); the objective function weight coefficients and constraint boundary parameters of the global model predictive control solver in S32 undergo parameter correction every 24 hours using global bias (weighted recursive updates based on bias residuals); and the regional agent decision model and reward function weights in S42 undergo online incremental updates every hour using local bias (policy gradient updates based on reinforcement learning). High-frequency updates enable regional agents to quickly adapt to changes in human behavior patterns, while low-frequency updates ensure the stability and generalization ability of the global model.

[0079] Example 2

[0080] This embodiment also provides a smart energy management and energy-saving optimization system for buildings, including: The data acquisition and twin modeling module is used to form an energy-carbon data stream through spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors. It performs parallel thermal inertia parameter identification to construct a hybrid energy storage model and carbon flow collaborative correction, and outputs thermal inertia time constant τ, hybrid energy storage state of charge (SOC), and energy-carbon dual-label data packets. The adaptive granularity decision-making and variable resolution prediction module is used to generate load, temperature field and carbon emission factor prediction sequences of corresponding granularity by using the built-in variable resolution mechanism of the AI ​​prediction engine and the dynamic granularity adaptation layer, with carbon emission factor volatility, thermal inertia time constant and model historical prediction error as input decision time domain parameters. The dynamic partition carbon constraint optimization deduction and verification module is used to dynamically cluster and generate virtual partitions based on environmental data and establish a physical-virtual two-layer mapping. It uses time-domain parameters and predicted sequences, energy-carbon dual-label data packages and hybrid energy storage state of charge as inputs to continuously optimize the symbiotic output of virtual partition control strategies and carbon emission budget tokens. It deduces and verifies the virtual energy storage boundary synchronously with the rolling cycle step size. The physical-virtual mapping and edge-layered autonomous execution module is used to decompose the verified control strategy into physical grid control quantities through a two-layer mapping model and distribute them using a version number incremental synchronization mechanism, and enable regional agents to make autonomous decisions on local control within the carbon emission budget token limit. The dual-dimensional monitoring and dual-channel feedback evolution module for energy and carbon is used to overlay a correction uncertainty display layer on the carbon flow map and overlay a token remaining budget warning on the carbon intensity map. It uses global prediction and actual deviation, local execution and token deviation to form a dual-channel feedback frequency division to update model parameters, and accelerates the cold start of the strategy through a combination of global and regional case libraries.

[0081] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A method for intelligent energy management and energy-saving optimization applied to buildings, characterized in that: Includes the following steps: Energy-carbon dual-dimensional sensing and hybrid energy storage twin modeling: forming an energy-carbon data stream through spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors; The thermal inertia parameter identification is performed in parallel based on the energy-carbon data stream to construct a hybrid energy storage model. At the same time, the thermal inertia time constant, hybrid energy storage state of charge, and energy-carbon dual-label data packets are output after carbon flow collaborative correction. Adaptive Granularity Decision Making and Variable Resolution Prediction: First, an AI prediction engine with a built-in variable resolution mechanism is constructed. Then, the dynamic granularity adaptation layer uses the carbon emission factor volatility, thermal inertia time constant, and historical prediction error of the model calculated from the dynamic carbon emission factor sequence in the energy-carbon dual-label data package as input. The time domain parameters are decided according to the mapping logic, so that the rolling period in the time domain parameters is used as the output granularity and the prediction time domain is used as the sequence length to generate the load, temperature field, and carbon emission factor prediction sequences of the corresponding granularity. Dynamic partitioning carbon constraint optimization simulation and verification integration: virtual partitions are generated dynamically based on environmental data and a physical-virtual two-layer mapping is established; Using time-domain parameters and predicted sequences, energy-carbon dual-label data packets and hybrid energy storage state of charge as inputs, the rolling optimization co-output virtual partition control strategy and carbon emission budget tokens are used. The virtual energy storage boundary is deduced and verified with the synchronous step size of the rolling cycle. If it fails, it will fall back to the previous security policy and trigger a re-solution, and output security control instructions and tokens. Physical-virtual mapping and edge-layered autonomous execution: The verified control strategy is weighted and decomposed into physical grid control quantities through a two-layer mapping model and distributed using an incremental version number synchronization mechanism; the regional agent makes autonomous decisions on local control within the carbon emission budget token quota, and applies for additional quotas from the global coordinator if the limit is exceeded. Energy-carbon dual-dimensional monitoring and dual-channel feedback evolution: a correction uncertainty display layer is superimposed on the carbon flow map, and a token remaining budget warning is superimposed on the carbon intensity map; a dual-channel feedback is formed by global prediction and actual deviation, and local execution and token deviation, and the global model parameters and regional model parameters are updated at different frequencies; the strategy cold start is accelerated by combining global and regional case library retrieval. The integrated dynamic partitioning carbon constraint optimization deduction and verification includes: Spatial adaptive partitioning and mapping modeling: The feature matrix is ​​formed by extracting three normalized feature values ​​of personnel density, heat flux density and CO2 concentration from the environmental and personnel data stream. The DBSCAN variant algorithm with adaptive neighborhood radius is used for clustering to divide the entire building area into several virtual partitions. A two-layer mapping model of physical grid-virtual partition is established. The control quantity is assigned to the boundary grid using a sigmoid membership smooth transition function. Multi-objective rolling optimization and carbon emission budget token coexistence: A rolling optimization problem is constructed using time-domain parameters. The optimization variables are the control objective sequence of each virtual zone and the hybrid energy storage charging and discharging power plan. The objective function is a weighted sum of six predicted values: energy cost, carbon emission cost, thermal comfort deviation, equipment life loss, renewable energy consumption, and grid interaction benefits. The solver outputs each rolling optimization coexistence output for each rolling cycle in the future control time domain. The sequence of virtual partition control strategies at any given time and the carbon emission budget tokens for each virtual partition calculated based on carbon flow distribution, personnel prediction and hybrid energy storage allocation; Variable step size simulation and virtual energy storage boundary closed-loop verification: Digital twin simulation engine receives rolling cycle The instruction sets the simulation step size to match the rolling period, enabling real-time simulation of the control strategy; it employs an early termination strategy, adjusting the equivalent heat storage step size by step. The virtual energy storage equivalent temperature is calculated by back-calculating the heat storage coefficient and compared point by point with the safety temperature limit of the building envelope; if the limit is exceeded, the simulation is terminated, the previous safety strategy is rolled back and the solution is triggered again; if the limit is not exceeded throughout the process, the verification is passed and the current control strategy sequence is stored in the strategy cache area.

2. The method according to claim 1, characterized in that: The energy-carbon dual-dimensional sensing and hybrid energy storage twin modeling includes: Multi-source energy consumption and environmental sensing data are collected across the entire domain at differentiated frequencies. Spatiotemporal structure alignment is achieved through time synchronization and spatial grid mapping. Dynamic carbon emission factors are injected in real time and multiplied and converted to form a raw energy-carbon dual-dimensional data stream with spatiotemporal dual labels. Using the indoor and outdoor temperature sequences and heating / cooling power sequences contained in the energy-carbon dual-dimensional data stream as input, the thermal inertia of the building envelope is equivalent to a virtual thermal energy storage model using the lumped parameter method, with the equivalent heat storage capacity as the input. E v It is a state variable and constitutes a hybrid energy storage system alongside physical energy storage; the comprehensive thermal inertia time constant τ is identified online, and based on the equivalent heat storage... E v The equivalent state of charge of the virtual thermal energy storage is calculated using normalized values ​​within the allowable heat storage range. SOC v ; Based on the component power consumption, dynamic carbon emission factor and total carbon emission data in the energy-carbon dual-dimensional data stream, a corrected energy-carbon dual-label data package is generated using tree-like carbon flow topology, initial allocation for power flow tracking and weighted recursive least squares correction with forgetting factor.

3. The method according to claim 2, characterized in that: The equivalent heat storage E v The update formula is: ； in, τ The thermal inertia time constant, η s For heat storage efficiency, For rolling period, P ch ( t The value is determined by the current indoor-outdoor temperature difference Δ. T ( t ) and total thermal conductivity of the building envelope K s Calculated charge / discharge power: ;when hour, Defined as the equivalent charging state of thermal virtual energy storage, when hour, Defined as the equivalent energy release state of thermal virtual energy storage; The equivalent state of charge of the thermal virtual energy storage SOC v The calculation formula is: ； in, E v,min and E v,max It is determined by the safe temperature range of the building envelope, the heat storage coefficient, and the allowable condensation / overheating boundary.

4. The method according to claim 2, characterized in that: The weighted recursive least squares correction with forgetting factor includes: Constructing a tree-oriented directed graph The root node is the master table, and the leaf nodes are the energy-consuming devices at each end. The power flow tracing algorithm is used to initially allocate the carbon emissions of the master table according to the energy consumption ratio of each branch, and the initial carbon emission allocation coefficient vector is obtained. Construct the weighted least squares objective function: ； in, For the first Accuracy weight of branch line metering devices. To estimate carbon emissions by item, The total carbon emissions are represented by λ, where λ is the regularization coefficient. These are the initial allocation coefficients; Solve using a recursive approach: ； In the formula, This is the carbon emission allocation coefficient vector after online recursive correction. For the observation matrix, This is a vector of carbon emission values ​​for each component. This is the gain matrix; Introducing the forgetting factor ρ The update formula for the error covariance matrix is ​​exponentially weighted: ； In the formula, It is the identity matrix; Using the corrected carbon emission allocation coefficient vector Recalculate carbon emissions for each branch It is then packaged with the corresponding energy consumption data, spatial grid ID, timestamp, dynamic carbon emission factor, and correction uncertainty to form an energy-carbon dual-label data package.

5. The method according to claim 1, characterized in that: The adaptive granularity decision-making and variable resolution prediction include: Constructing a variable-resolution AI prediction engine: A multi-dimensional time series prediction model with a physical-data fusion deep learning architecture is adopted. The model is trained by constructing samples with the past 24 hours of data as input and the next 6 hours as the target. The intrinsic output resolution of the model is 5 minutes. A variable-resolution output interface is configured. After receiving the target granularity instruction, the 5-minute prediction sequence is resampled on the time axis and the uncertainty is recalibrated. The corresponding coarse-grained prediction value and statistically consistent confidence interval are output. Real-time decision-making time-domain parameters for dynamic granularity adaptation layer: based on thermal inertia time constant τ Load forecast error rolling variance σ² and dynamic carbon emission factor volatility ν As input, the time-domain prediction is output through mapping logic. Control Time Domain With rolling cycle ; Generation of variable resolution prediction sequences: The AI ​​prediction engine receives the rolling period. and prediction time domain Instructions, generating the future Within a time period, the particle size is The system includes a load forecast sequence, an indoor temperature field forecast sequence, a dynamic carbon emission factor forecast sequence, and an indoor-outdoor temperature difference forecast sequence. The indoor-outdoor temperature difference forecast sequence is fed back to the thermal virtual energy storage model for rolling updates of the equivalent heat storage. The state trajectory.

6. The method according to claim 5, characterized in that: The mapping logic is as follows: when Hours and and At that time, take Hour, Hour, minute; when Hours or or At that time, take Hour, Hour, minute; Take the remaining intermediate cases Hour, Hour, minute; And satisfy Hour, Hours, and , and All Integer multiples of; Take a value from the set {10 minutes, 15 minutes, 20 minutes}; The dynamic granularity adaptation layer has built-in hysteresis switching logic. When a change in the indicator triggers a switching condition, the switching is only performed after the condition is met within two consecutive fixed basic judgment periods.

7. The method according to claim 1, characterized in that: The physical-virtual mapping and edge-layered autonomous execution include: Physical-Virtual Mapping Command Decomposition and Issuance: For each physical grid, retrieve all its virtual partitions and corresponding fusion weights. The final control quantity is the weighted average of the control targets of each virtual partition. ； in, The set of virtual partitions to which the physical grid belongs. To normalize the fusion weights, For virtual partitions The target value for control; the mapping table update adopts a version number incremental synchronization mechanism, which only generates incremental entries for physical meshes where the boundary changes and sends them to the edge nodes; Edge-layered autonomous execution and token-constrained coordination: Regional agents receive physical grid-level control commands as execution baselines and physical grid set-level carbon emission budget tokens, and independently make micro-control adjustments within the token quota; if the autonomous decision is expected to cause actual carbon emissions to exceed the token quota, an additional quota application is sent to the global coordinator; the global coordinator decides to approve or reject based on the remaining global carbon emission budget and the real-time carbon emission pressure of each region.

8. The method according to claim 1, characterized in that: The energy-carbon dual-dimensional monitoring and dual-channel feedback evolution includes: Energy-Carbon Dual-Dimensional Enhanced Visualization and Carbon Emission Budget Early Warning: Overlaying a correction coefficient display layer on the carbon flow Sankey diagram to correct for uncertainty. As a color / linewidth modulation factor, the remaining carbon emission budget progress bar is overlaid on the regional carbon intensity map. When the actual carbon emission in the region reaches the preset warning ratio of the token limit, the color will automatically change to warn and trigger the regional energy-saving mode. Dual-channel deviation feedback and dual-layer knowledge base co-evolution: The global deviation channel feeds back load prediction deviation, carbon emission factor prediction deviation, and optimization target calculation deviation to the AI ​​prediction engine and the global model predictive control solver; the local deviation channel feeds back the regional comfort compliance rate, token over-limit times, and the cumulative deviation between actual carbon emissions and token carbon emissions to the regional intelligent agent decision model; a global case library and a regional case library are established, and during combined retrieval, the top 3 cases with the highest feature matching degree are selected and the initial strategy parameters are generated by weighted average based on similarity; the global model performs an online incremental update every 24 hours, and the regional model performs an online incremental update every 1 hour.

9. A smart energy management and energy-saving optimization system for buildings, comprising the method described in any one of claims 1-8, characterized in that: include: The data acquisition and twin modeling module is used to form an energy-carbon data stream through spatiotemporal alignment of multi-source data and injection of dynamic carbon emission factors. It performs parallel thermal inertia parameter identification to construct a hybrid energy storage model and carbon flow collaborative correction, and outputs thermal inertia time constant τ, hybrid energy storage state of charge (SOC), and energy-carbon dual-label data packets. The adaptive granularity decision-making and variable resolution prediction module is used to generate load, temperature field and carbon emission factor prediction sequences of corresponding granularity by using the built-in variable resolution mechanism of the AI ​​prediction engine and the dynamic granularity adaptation layer, with carbon emission factor volatility, thermal inertia time constant and model historical prediction error as input decision time domain parameters. The dynamic partition carbon constraint optimization deduction and verification module is used to dynamically cluster and generate virtual partitions based on environmental data and establish a physical-virtual two-layer mapping. It uses time-domain parameters and predicted sequences, energy-carbon dual-label data packages and hybrid energy storage state of charge as inputs to continuously optimize the symbiotic output of virtual partition control strategies and carbon emission budget tokens. It deduces and verifies the virtual energy storage boundary synchronously with the rolling cycle step size. The physical-virtual mapping and edge-layered autonomous execution module is used to decompose the verified control strategy into physical grid control quantities through a two-layer mapping model and distribute them using a version number incremental synchronization mechanism, and enable regional agents to make autonomous decisions on local control within the carbon emission budget token limit. The dual-dimensional monitoring and dual-channel feedback evolution module for energy and carbon is used to overlay a correction uncertainty display layer on the carbon flow map and overlay a token remaining budget warning on the carbon intensity map. It uses global prediction and actual deviation, local execution and token deviation to form a dual-channel feedback frequency division to update model parameters, and accelerates the cold start of the strategy through a combination of global and regional case libraries.

Images