Building energy-saving regulation method and system based on digital twinning and storage medium

Abstract

The application relates to the field of building energy saving and intelligent control technology, and discloses a building energy saving regulation and control method and system based on digital twinning and a storage medium. The method comprises the following steps: acquiring multi-source real-time data to construct a multi-dimensional data set, inputting a building thermal digital twinning model to determine an indoor thermal environment state; when the state exceeds a thermal comfort interval, predicting a thermal balance trend, and generating a regulation and control scheme set through an intelligent optimization algorithm with the thermal balance trend as a target; inputting the scheme set into the twinning model for simulation and deduction to obtain a thermal environment response curve group; comparing the thermal environment response curve group with an energy saving and thermal comfort constraint interval and performing multi-objective optimization to generate a decision basis index set containing a regulation and control strategy and a constraint range; and finally selecting an optimal regulation and control scheme to be converted into a bottom layer instruction and executed. The application solves the problems of traditional regulation and control methods, such as response lag, high energy consumption and difficulty in balancing comfort and energy saving, and significantly improves the dynamic balance ability of building thermal performance and energy utilization efficiency.

Description

Technical Field

[0001] This application relates to the field of building energy conservation and intelligent control technology, and in particular to a building energy conservation regulation method, system and storage medium based on digital twins. Background Technology

[0002] With the escalating global energy crisis and the advancement of "dual carbon" goals, the proportion of building energy consumption in total social energy consumption is receiving increasing attention. The field of building environmental control faces a persistent contradiction: balancing the reduction of HVAC system energy consumption with ensuring the thermal comfort of indoor occupants is difficult. How to effectively reduce building energy consumption while maintaining comfort has become a key research focus in current building energy conservation technologies.

[0003] In recent years, building energy-saving technologies have gradually developed towards the integration of digital twins and artificial intelligence. Digital twin technology, by constructing a real-time mapping of a building in virtual space, enables dynamic monitoring and optimized control of the building's operating status. However, existing digital twin applications still have the following problems: Existing models mostly focus on data visualization or single-dimensional load forecasting, lacking in-depth feature fusion and mechanism coupling analysis of multi-source heterogeneous data. This data includes local thermal disturbances indoors, dynamic temperature differences in the building envelope, and air humidity. Due to the lack of such analysis, the system struggles to fully understand the building's thermal state, making precise control difficult. Existing control algorithms often only pursue theoretically optimal solutions without incorporating these into forward physics fields for simulation and without verifying temporal boundary constraints. As a result, the generated energy-saving strategies tend to deviate from the thermal comfort baseline in practical applications, failing to stably guarantee the comfort of the building's thermal environment. Existing methods lack a fine-grained transformation mechanism from virtual decision-making to physical execution strategy constraints. Specifically, it is unable to generate low-level instruction sequences containing crucial information such as device priorities (e.g., the coordinated operation of air conditioning and window ventilation) and load adjustment ranges (e.g., the control range of large-inertia floor radiation). This results in a common problem of "heavy on prediction, light on execution" in the system, making it difficult to achieve effective linkage between virtual and physical spaces, thus significantly reducing the energy-saving control effect.

[0004] Therefore, how to construct a high-fidelity digital twin model of a building that can dynamically update thermal parameters, and on this basis, form a dynamic closed loop from virtual decision-making to physical execution through multi-objective optimization and forward simulation, is a key technical problem that needs to be solved in the field of building energy conservation. Summary of the Invention

[0005] This application provides a digital twin-based method, system, and storage medium for building energy conservation control. It aims to address issues such as slow response, low energy efficiency, and insufficient comfort in existing building energy conservation control systems by integrating digital twin technology, intelligent optimization algorithms, and multi-source data. This invention achieves a precise balance between building energy efficiency and comfort by constructing a real-time virtual mapping of the building's thermal state, providing a new approach to building energy conservation and comfort control.

[0006] In a first aspect, this application provides a building energy-saving control method based on digital twins, the method comprising:

[0007] S1. Obtain multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status.

[0008] S2. Input the multi-dimensional data set into the preset building thermal digital twin model, extract the multi-source coupling features, and then determine the indoor thermal environment state vector and the thermal performance parameters of the building envelope.

[0009] S3. When the indoor thermal environment state vector exceeds the preset thermal comfort range, predict the thermal balance trend sequence within the preset time range based on historical data and real-time meteorological information.

[0010] S4. Using the heat balance trend sequence as the predictive control target, an intelligent optimization algorithm is used to search for the control variables and generate a set of control schemes.

[0011] S5. Input the set of control schemes into the building thermal digital twin model for simulation and obtain the corresponding set of indoor thermal environment response curves;

[0012] S6. Compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in different time periods, and determine the decision basis index set corresponding to each control variable based on the comparison results.

[0013] S7. Select the optimal control scheme from the set of decision-making indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

[0014] Secondly, this application provides a building energy-saving control system based on digital twins, the system comprising:

[0015] The data acquisition module is used to acquire multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status.

[0016] The twin modeling module is used to input multi-dimensional data sets into a preset building thermal digital twin model, extract multi-source coupling features, and then determine the indoor thermal environment state vector and building envelope thermal performance parameters.

[0017] The trend prediction module is used to predict the thermal balance trend sequence within a preset time range based on historical data and real-time meteorological information when the indoor thermal environment state vector exceeds the preset thermal comfort range.

[0018] The intelligent optimization module is used to search for control variables and generate a set of control schemes by using the heat balance trend sequence as the predictive control target and employing intelligent optimization algorithms.

[0019] The simulation and deduction module is used to input the set of control schemes into the building thermal digital twin model for simulation and deduction, and obtain the corresponding set of indoor thermal environment response curves.

[0020] The constraint assessment module is used to compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in a time-by-time manner, and determine the decision basis index set corresponding to each control variable based on the comparison results.

[0021] The decision execution module is used to select the optimal control scheme from the set of decision-making basis indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

[0022] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned digital twin-based building energy-saving control method.

[0023] The beneficial effects of this application are at least as follows:

[0024] By using a digital twin model to perceive the building's thermal state in real time, and combining future thermal balance trend prediction with genetic algorithm optimization, multi-objective optimization is performed within the dual boundaries of a preset energy-saving constraint range and a preset thermal comfort constraint range. By coordinating various control variables such as air conditioning supply air temperature, fresh air volume and return air volume ratio, heat recovery device efficiency level, shading component opening angle, floor radiant cooling and heating load, and window ventilation opening, a globally optimal balance between energy efficiency and comfort is achieved, avoiding control deviations caused by over-optimization of a single objective. Based on a digital twin model synchronized with the actual building state, forward physics simulation is used to quickly respond to and adaptively adjust to dynamic disturbances such as sudden changes in outdoor weather, changes in human activity, and fluctuations in equipment heat dissipation. Compared to traditional static control methods, the control system of this invention can dynamically update control strategies based on real-time data, effectively reducing energy waste and indoor environmental fluctuations caused by control lag. Through a set of decision-based indicators, the optimization results of the virtual space are transformed into a sequence of underlying control instructions containing information such as equipment execution priority and load adjustment magnitude, thus establishing an execution link between the digital twin model and the building's physical equipment. This mechanism ensures that the optimized control strategy can be accurately implemented, forming a complete closed loop of "perception-analysis-prediction-optimization-simulation-decision-execution", avoiding the gap problem of "emphasizing prediction and neglecting execution" in traditional methods, and realizing dynamic closed-loop control of building thermal performance from digital twin space to physical space. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of the digital twin-based building energy-saving control method of this application;

[0027] Figure 2 This is a schematic diagram comparing and verifying the technical effects of the embodiments of this application;

[0028] Figure 3 This is a schematic diagram of the structure of the digital twin-based building energy-saving control system of this application. Detailed Implementation

[0029] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of the building energy-saving control method based on digital twins provided by the present invention. The flowchart specifically includes the following steps:

[0031] S1. Obtain multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status.

[0032] In one specific embodiment, the process of performing step S1 may specifically include the following steps:

[0033] Real-time sequence data of outdoor meteorological parameters are collected through a sensor network deployed on the exterior of the building. The meteorological parameters include at least outdoor temperature and humidity, wind speed and solar radiation data.

[0034] Temperature data is acquired layer by layer by temperature sensors installed in each layer of the building envelope, and temperature gradient distribution data of the building envelope is generated based on the temperature difference between each layer.

[0035] By deploying temperature and humidity sensors in different indoor areas, temperature and humidity data of each area are collected, and indoor temperature and humidity distribution data are generated by combining the area of ​​each area.

[0036] By deploying personnel detection devices in different indoor areas, the location and number of people in each area are collected, thereby generating indoor personnel density distribution data;

[0037] The power monitoring device records the time-series fluctuation data of the heat dissipation power of indoor equipment;

[0038] Real-time sequence data, temperature gradient distribution data, indoor temperature and humidity distribution data, indoor personnel density distribution data, and time-series fluctuation data are standardized and structured to generate a multi-dimensional dataset.

[0039] Specifically, the sensor network deployed outside the building includes an automatic weather station fixed to the building roof and a total radiation meter installed on the building facade. The automatic weather station has built-in temperature and humidity sensors to collect outdoor dry-bulb temperature and relative humidity; an anemometer to collect ambient wind speed; and a total radiation meter to collect solar radiation intensity on horizontal and vertical planes. The sensors continuously collect data according to a set sampling period (e.g., once per minute), forming a time-series of outdoor temperature values, outdoor relative humidity values, wind speed values, and solar radiation illuminance values.

[0040] Sensor arrays, such as thin-film platinum resistance temperature sensors, are pre-embedded in each structural layer of the building envelope, including exterior walls and roofs. For example, for a 300mm thick exterior wall, five sensors are arranged at 60mm intervals along the thickness direction, located on the outer surface of the exterior wall, the outer side of the insulation layer, the center of the insulation layer, the inner side of the insulation layer, and the inner surface of the wall. A data acquisition unit reads the temperature at each measuring point at a set frequency. After acquiring the temperature values ​​of each layer, the temperature sensor readings of two adjacent material layers are subtracted, and this temperature difference is divided by the physical thickness of the material between the two layers to calculate the temperature gradient at that point. This generates temperature gradient distribution data of the building envelope, which characterizes the real-time heat transfer rate and heat flow direction through the envelope. This allows the digital twin model to perceive the transient heat transfer characteristics of the building envelope, solving the technical problem of simplifying the wall as a single lumped parameter in traditional static thermal calculations, which fails to quantify the actual dynamic heat storage and heat conduction delay process inside the wall.

[0041] The building's interior space is divided into several independent HVAC control zones, such as open-plan offices, meeting rooms, and corridors in an office building. Temperature and humidity sensor nodes are placed at the breathing zone height of each control zone to collect temperature and relative humidity data at a set frequency. For example, two sensors are placed diagonally in each zone. After acquiring the temperature and humidity data from each node, the planar area data of the corresponding control zone is retrieved from the building information model. A spatial interpolation algorithm is used to convert the discrete temperature and humidity node data into a continuous planar distribution field. For example, an inverse distance weighted interpolation algorithm is used, where the closer a known sensor node is to the unknown estimation point, the greater its weight. This calculates the estimated temperature and humidity values ​​at each coordinate point inside the building, generating a two-dimensional temperature and humidity distribution data matrix covering the entire building interior, i.e., the indoor temperature and humidity distribution data.

[0042] The personnel detection device includes an infrared array sensor installed on the indoor ceiling and a personnel counting terminal. The infrared array sensor scans the heat source distribution contours of each sub-area at a set frequency. By mapping the captured heat source distribution contours to coordinates, it extracts the real-time two-dimensional planar coordinate information of each individual person within the building plan and counts the total number of people in each independent area. Dividing the total number of people in each area by the effective usable area of ​​the corresponding area generates indoor personnel density distribution data that includes spatial location mapping relationships.

[0043] The power monitoring device consists of smart monitoring meters installed in the main power supply lines and branch circuit distribution boxes on each floor. It collects active power values ​​from lighting and office socket circuits at a set frequency. The collected active power values ​​are multiplied by preset heat conversion coefficients for various types of equipment to convert them into an hourly fluctuation sequence of heat dissipation power of indoor electrical equipment. Recording this time-series fluctuation data of equipment heat dissipation power solves the technical problem of the inability to detect dynamic fluctuations in internal heat gain caused by the random operation of non-HVAC electrical equipment within the building.

[0044] Standardization processing includes: unifying timestamps across all data sources; aligning data from different sampling frequencies to the same time base using linear interpolation; removing outliers exceeding sensor ranges or physical limits; and filling in missing data using nearest-neighbor imputation. Since the multi-source data collected exhibits significant differences in dimensions and numerical distribution ranges, a maximum-minimum standardization method is used to linearly transform the numerical boundaries of various data types, proportionally mapping sequence values ​​with different physical units to a dimensionless range of 0 to 1. Structured integration refers to aligning the standardized data sequences along a predefined data structure in the time dimension, resampling and aligning timestamps with a set minimum time step. The aligned multi-class data are then tensor-concatenated according to the time series dimension, spatial coordinate dimension, and physical attribute feature dimension to construct a multi-dimensional data set indexed by time, containing spatial location labels, and covering external meteorological conditions, building envelope response, indoor thermal and humidity conditions, and internal heat source disturbances. This multi-dimensional data set is stored in system memory as a three-dimensional data tensor for subsequent input into the building thermal digital twin model for multi-source coupling feature extraction. For example, the multidimensional data set is constructed as a three-dimensional tensor X(i, j, k), where i represents the discrete time step index, j represents the spatial grid number or region number, and k represents the physical attribute feature channel number. Under different spatial discretization strategies, the spatial dimension can also be further subdivided into two-dimensional planar coordinates or three-dimensional voxel coordinates, and their equivalent implementations all fall within the scope of the multidimensional data set described in this application.

[0045] S2. Input the multidimensional data set into the preset building thermal digital twin model, extract the multi-source coupling features, and then determine the indoor thermal environment state vector and the thermal performance parameters of the building envelope.

[0046] In one specific embodiment, a pre-trained neural network model is set in the building thermal digital twin model, and the process of executing step S2 may specifically include the following steps:

[0047] The multidimensional dataset is divided into subsets corresponding to different thermal influencing factors;

[0048] Features are extracted from each data subset through the feature extraction layer of the neural network model, and the coupling relationship between multiple source factors is calculated through the coupling analysis layer to generate a comprehensive feature vector.

[0049] The current indoor thermal environment state vector is determined based on the comprehensive feature vector;

[0050] Sub-features related to the thermal performance of the building envelope are extracted from the comprehensive feature vector, and the thermal resistance degradation rate of the building envelope is calculated by combining historical thermal performance data. The thermal resistance degradation rate of the building envelope is used as the thermal performance parameter of the building envelope.

[0051] Specifically, a neural network model is pre-built and trained. A large-sample multidimensional dataset from the building's historical operating cycle is collected as the training input set, and synchronously measured indoor thermal environment calibration values ​​and physical thermal resistance values ​​calculated based on the energy balance equation are used as training labels. The multidimensional dataset specifically covers outdoor meteorological disturbance data (such as temperature, humidity, and solar radiation), indoor dynamic heat source data (such as occupancy density and equipment heat dissipation), historical operating parameters of HVAC equipment (such as supply air temperature and air volume ratio), and corresponding measured distribution data of the indoor three-dimensional thermal and humidity environment. The neural network model includes an input layer, a feature extraction layer, a coupling analysis layer, and a fully connected output layer. The feature extraction layer consists of a one-dimensional convolutional neural network for processing temporal data, a two-dimensional convolutional neural network for processing spatially distributed data, and a fully connected layer for processing sparse data. The coupling analysis layer employs a multi-head attention mechanism architecture to calculate the coupling relationships between multiple source factors. For example, during training, mean squared error is used as the loss function, the Adam optimizer is configured to iteratively update the network weights, the initial learning rate is set to 0.001, the batch size is 64, and the training is iterated for 500 rounds until the value of the loss function decreases and stabilizes within the set convergence threshold range. The network parameters are then solidified to obtain the pre-trained neural network model.

[0052] In the application phase, the multidimensional dataset is divided into subsets corresponding to different thermal influencing factors according to their physical attributes. Based on the physical attributes and heat transfer mechanisms of the data, the multidimensional dataset is further divided into outdoor meteorological data subsets, internal thermal disturbance data subsets, and building envelope thermal data subsets. The outdoor meteorological data subset includes at least outdoor temperature, outdoor humidity, wind speed, and solar radiation sequences; the internal thermal disturbance data subset includes indoor occupant density distribution data and time-series fluctuation data of indoor equipment heat dissipation power; and the building envelope thermal data subset includes temperature gradient distribution data for each layer of the building envelope. Dividing the multidimensional dataset into subsets corresponding to different thermal influencing factors solves the technical problem that mixed inputs of multi-source heterogeneous data make it difficult for neural network models to effectively learn the independent features of each factor, ensuring the independence and integrity of the features of various thermal influencing factors during the forward propagation process of the network.

[0053] The feature extraction layer employs multiple parallel feature extraction sub-networks. For example, for one-dimensional time-series data such as outdoor meteorological data subsets, a one-dimensional convolutional neural network is used as the feature extraction sub-network. The one-dimensional convolutional layer extracts the change patterns within a local time window, with a kernel size of 3, a stride of 1, and 32 output channels. After convolution, a preliminary feature vector containing local time-series features is obtained. For two-dimensional distributed data with spatial location labels, such as the thermal data subset of building envelope, a two-dimensional convolutional neural network is used as the feature extraction sub-network. The two-dimensional convolutional layer extracts spatial distribution features, with a kernel size of 3×3, a stride of 1, and 64 output channels. After convolution, a preliminary feature vector containing spatial local correlations is obtained. For sparsely distributed data such as the internal thermal disturbance data subset, a fully connected layer is used for feature dimensionality reduction and extraction. Each feature extraction sub-network outputs a preliminary feature vector for the corresponding data subset, representing external hidden layer features characterizing the temporal evolution of external climate, heat transfer hidden layer features characterizing the heat conduction delay characteristics of the walls, and internal hidden layer features characterizing the spatiotemporal aggregation characteristics of internal heat sources.

[0054] The initial feature vectors output from each feature extraction sub-network are input into the coupling analysis layer. The coupling analysis layer employs a multi-head attention mechanism to calculate the coupling coefficients between different data subsets. The attention mechanism is calculated as follows: each initial feature vector is mapped to a query matrix, a key matrix, and a value matrix, respectively. The attention score is calculated by the dot product of the query matrix and the key matrix, and after normalization using the softmax function, the coupling weight coefficients between each subset are obtained. These coupling weight coefficients quantify the degree of mutual influence between different thermal factors, such as the influence weight of solar radiation intensity on the temperature gradient of the building envelope, or the influence weight of personnel density distribution on indoor temperature and humidity distribution. The coupling weight coefficients are weighted and summed with the corresponding value matrices to obtain a fused feature vector that integrates the coupling relationships of multiple factors. Then, the fused feature vector is residually connected to the input of the coupling analysis layer (i.e., the aggregated representation of each initial feature vector), and then input into a fully connected layer for feature recombination, outputting a comprehensive feature vector. For example, the initial feature vectors can be concatenated along the feature dimensions to obtain an aggregated feature vector. If the dimensions are inconsistent, the dimensions can be aligned via a linear mapping layer, and then the aggregated feature vector and the fused feature vector can be added element-wise to complete the residual connection. For example, the comprehensive feature vector is set to 256 dimensions, with each dimension representing an abstract characteristic of the current building thermal state in the coupling space. Generating a comprehensive feature vector can solve the problem of state assessment bias caused by treating each influencing factor as an independent variable and ignoring their dynamic coupling relationship.

[0055] The comprehensive feature vector is input into the first fully connected output branch of the neural network model, which is specifically used for thermal environment state estimation. This branch consists of two fully connected layers: the first layer has 128 neurons, and the activation function is ReLU; the second layer has the same number of neurons as the dimension of the indoor thermal environment state vector, and the activation function is a linear function. The indoor thermal environment state vector contains at least the following three types of parameters: the first type is macroscopic statistical parameters, including indoor average temperature and indoor average humidity; the second type is spatial distribution parameters, including indoor temperature distribution feature matrix and indoor air humidity spatial distribution feature matrix, used to characterize the differences in thermal and humidity states in different areas; the third type is thermal comfort evaluation parameters, including the predicted average thermal perception index (PMV). For example, the PMV value is calculated according to the Fanger thermal comfort equation. The required input parameters, air temperature, relative humidity, and average radiant temperature, are obtained by the neural network model through regression of the comprehensive feature vector. Since we are currently in the initial evaluation stage before the issuance of control instructions, the air velocity parameter adopts the indoor benchmark wind speed value preset by the building design code, for example, 0.15 m / s; the human metabolic rate and clothing thermal resistance are determined based on personnel density distribution data and preset empirical values ​​for different time periods. Through the above mapping relationship, the high-dimensional comprehensive feature vector is compressed into a low-dimensional state vector that can be directly used for thermal comfort evaluation.

[0056] The neural network model includes a second fully connected output branch specifically for evaluating the performance of the building envelope. This branch extracts a subset of features related to the heat transfer characteristics of the building envelope from the comprehensive feature vector using an attention mechanism. The building envelope performance characterization vector is composed of the physical and thermal parameters of the building envelope, including parameters such as thermal resistance, thermal conductivity, material thickness, and degradation rate of each layer. After normalization, these parameters form a fixed-dimensional feature vector. This vector serves as a query term in the attention mechanism, measuring the correlation between the current comprehensive features and the standard physical performance of the building envelope. For example, by calculating the vector similarity between the comprehensive feature vector (as the Key) and the building envelope performance characterization vector (as the Query), corresponding attention weight coefficients are generated. These attention weight coefficients are then applied element-wise to each dimension of the comprehensive feature vector (as the Value) for weighted filtering (i.e., independent weighted filtering), thereby aggregating feature information strongly correlated with the heat transfer characteristics of the building envelope to form a feature subset. This feature subset is the sub-feature vector used to evaluate the performance of the building envelope. The similarity calculation method can employ cosine similarity or dot product. Before calculating vector similarity, a linear mapping layer projects the building envelope performance characterization vector and the comprehensive feature vector to the same feature dimension space. This sub-feature vector is then input into a linear regression model, which outputs an estimated value of the actual equivalent thermal resistance of the building envelope at the current moment. The linear regression model aims to establish a mapping relationship between hidden features and physical thermal parameters. During model training, samples of the same-dimensional sub-feature vectors extracted from the building's historical operating cycle are used as input to the linear regression model. These same-dimensional sub-feature vectors are high-dimensional abstract representations composed of feature dimensions strongly correlated with the heat transfer characteristics of the building envelope, selected from the comprehensive feature vector of historical operating data through the attention mechanism. Their specific values ​​are calculated by forward propagation of a neural network, with each dimension corresponding to the abstract feature component in the comprehensive feature vector that has the highest correlation with the thermal performance of the building envelope. The training label corresponding to this sub-feature vector is the historical true equivalent thermal resistance value obtained by back-calculating the measured heat transfer temperature difference between the inner and outer surfaces of the wall and the measured heat flux density penetrating the wall during the corresponding historical period, based on Fourier's law of thermal conductivity. The calculation formula is: True equivalent thermal resistance = Measured heat transfer temperature difference / Measured heat flux density. The weight coefficients of the linear regression model are fitted and optimized using the least squares method or gradient descent method until the mean square error meets the preset requirements.

[0057] During the model application phase, the sub-feature vectors containing temperature gradient and coupling correlation information extracted in real time are input into the model. Through linear mapping using the fitted and fixed weight coefficients, a unit of m can be output in real time. 2• Estimated actual equivalent thermal resistance per kWh. Historical thermal performance data for the corresponding building physical area is retrieved from the pre-existing database to obtain the baseline design thermal resistance value at the time of initial delivery or last major overhaul. The baseline design thermal resistance value is subtracted from the current estimated actual equivalent thermal resistance value of the building envelope, and then divided by the baseline design thermal resistance value to calculate a dimensionless percentage value. This percentage value is defined as the thermal resistance degradation rate of the building envelope. The thermal resistance degradation rate of the building envelope is loaded as a thermal performance parameter into the parameter library of the building thermal digital twin model to update the virtual material properties within the twin model. For example, the building thermal digital twin model is pre-constructed based on the building information model (BIM) of the target building, containing the building's geometric topology information (such as room dimensions and building envelope construction), initial physical material properties (such as thermal conductivity, density, and specific heat capacity), and the piping topology and equipment parameters of the HVAC system. This model uses a parametric modeling approach, storing the above attributes as readable and writable parameter variables in the model parameter library. Based on this static data, the digital twin model further integrates a state-space model constructed based on the resistive-capacitive (RC) equivalent heat network theory, or a dynamic heat transfer physical white-box model. Its specific derivation mechanism is as follows: the physical building is divided into multiple independent thermal zones, and a heat balance differential equation containing thermal resistance and heat capacity terms is established for heat transfer nodes such as walls, doors, and windows. When the thermal resistance degradation rate is written as an update parameter into the model parameter library, the system can correct the heat transfer coefficient boundary conditions in the above differential equations in real time. This allows the virtual model to strictly approximate the actual aging state of the physical entity in subsequent dynamic mathematical derivations, providing a high-fidelity physical simulation environment for subsequent heat balance prediction.

[0058] The calculation of the thermal resistance degradation rate of the building envelope solves the technical problem that the actual heat transfer coefficient of the physical building deviates from the initial design conditions due to the aging of insulation materials, dampness or cracking of the exterior walls during long-term operation, which in turn causes serious distortion of the cold and heat load calculation in the subsequent simulation and deduction stage of the digital twin model. It realizes the dynamic adaptive evolution of the physical parameters of the virtual model with the life cycle of the physical building.

[0059] S3. When the indoor thermal environment state vector exceeds the preset thermal comfort range, predict the thermal balance trend sequence within the preset time range based on historical data and real-time meteorological information.

[0060] In one specific embodiment, the process of performing step S3 may specifically include the following steps:

[0061] The indoor thermal environment state vector is compared with the preset thermal comfort range. When the indoor thermal environment state vector exceeds the preset thermal comfort range, historical data for continuous time periods corresponding to the current operating state are extracted from the historical thermal state multidimensional data set.

[0062] A pre-defined time series prediction model is used to process historical data over continuous periods to extract the temporal evolution characteristics of building thermal status.

[0063] Using real-time outdoor meteorological parameters and the dynamic rate of change of indoor and outdoor temperature difference as input conditions, and combining them with time series evolution characteristics, a time series prediction model is used to extrapolate the trend and generate a heat balance trend sequence within a preset time range in the future.

[0064] Specifically, the preset thermal comfort range can be set according to the recommended range of the expected average thermal perception index (PMV) specified in national standard GB / T 18049 or international standard ISO 7730. For example, the indoor temperature and humidity range corresponding to a PMV value between -0.5 and +0.5 can be taken as the preset thermal comfort range. The current PMV value in the indoor thermal environment state vector is extracted and compared with this range. When the current PMV value is greater than 0.5 or less than -0.5, it is determined that the current indoor thermal environment state vector exceeds the preset thermal comfort range, triggering subsequent prediction and control procedures.

[0065] The historical thermal performance data set is stored in the building history database, containing daily synchronous records of outdoor meteorological parameters, building envelope temperature gradients, indoor temperature and humidity distributions, occupant density distributions, and equipment heat dissipation power over a preset time period (e.g., one year). Extracting continuous historical data corresponding to the current operating state involves retrieving continuous time segments with similar operating conditions from the historical database, using the current moment as a baseline. Matching criteria for similar operating conditions include: the difference in external solar radiation intensity being within a set threshold range, similar indoor occupant density, and consistent building equipment start-up and shutdown schedules. For example, if the current time is a summer weekday afternoon, data from the past 30 days that are also weekday afternoons with similar outdoor temperatures and a solar radiation intensity difference not exceeding 50 W / m² will be extracted. 2 A continuous 72-hour historical data sequence is used as the input basis for the time series forecasting model.

[0066] The time series prediction model employs a Long Short-Term Memory (LSTM) network architecture combined with a multi-head attention mechanism. Its hierarchical modules sequentially include: an input layer, an LSM network layer with two hidden layers, a multi-head attention mechanism layer, and a fully connected output layer. For example, the input layer receives a normalized historical data sequence with a 24-hour time step, meaning each input sample contains hourly time-series data such as outdoor temperature, outdoor humidity, solar radiation, indoor temperature, and population density sampled over the past 24 hours. The first LSM network hidden layer has 128 memory units, returning a complete sequence state containing all time steps; the second LSM network hidden layer has 64 memory units, similarly returning a complete hidden state sequence containing all time steps to preserve local features at each historical moment. The multi-head attention mechanism layer uses eight attention heads to perform parallel self-attention computation and weighted fusion of features from different time steps in the hidden state sequence. The output layer is a fully connected layer, with the output dimension matching the prediction task requirements. During the training phase, large sample data from the historical operating period is collected as the training set. For example, historical operational data of the building over the past three complete cooling or heating seasons are extracted. Historical outdoor meteorological parameters, historical indoor occupancy density, historical indoor temperature and humidity, and the dynamic rate of change of temperature difference between the inner surface of the building envelope and the indoor air are resampled according to the time step and used as training input data. The actual changes in indoor heat load and indoor temperature drift after the corresponding time window are used as training label output data. The dataset is divided into training and validation sets in an 8:2 ratio. An adaptive moment estimation optimizer, Adam, is configured with an initial learning rate of 0.001, a batch size of 32, and mean squared error as the loss function. After multiple iterations of training, the weight parameters of the forget gate, input gate, and output gate in the long short-term memory network layer are updated until the loss function value of the validation set no longer decreases significantly. The model parameters are then saved, completing the training and solidification of the time series prediction model.

[0067] The extracted historical data from consecutive time periods is transformed into a time-series tensor and input into the input layer of the time-series prediction model. The data propagates forward along time steps within the Long Short-Term Memory (LSTM) network layer. During information transmission, the cell states within the LSM layer retain the heat transfer delay patterns caused by the thermal inertia of the building envelope and the cumulative effect of indoor occupant heating through a gating mechanism. After nonlinear mapping through two LSM network hidden layers, a hidden state sequence matrix containing the temporal evolution process is output. This hidden state sequence matrix is ​​input into a multi-head attention mechanism layer. By dynamically weighting the feature matrices at different time steps, historical abrupt changes highly correlated with the current prediction task are strengthened. The output is a high-dimensional numerical tensor after attention weighting. This high-dimensional numerical tensor is the temporal fusion feature vector, which represents the temporal evolution characteristics of the current building thermal state in mathematical space. Extracting this temporal evolution feature can solve the problem that relying solely on physical quantities at a single moment cannot characterize the dynamic delay characteristics of the building thermodynamic system.

[0068] The system acquires real-time sequences of outdoor meteorological parameters collected at the current moment, including weather forecast data for the next few hours (e.g., 24 hours) (such as outdoor temperature, humidity, wind speed, and solar radiation forecasts). It calculates the dynamic rate of change of indoor and outdoor temperature difference: using the difference between the current indoor and outdoor temperatures as the baseline temperature difference, and combining this with the temperature difference trend over the past few hours, it calculates the rate of change of temperature difference over time. It performs feature mapping between the outdoor meteorological forecast sequence and the dynamic rate of change of indoor and outdoor temperature difference to obtain an external condition feature representation. This external condition feature representation is then dimensionally aligned and fused with the temporal evolution features to generate a fused complete input condition. This complete input condition is fed into the fully connected output layer of the time series prediction model. The fully connected output layer decodes it into continuous numerical values ​​with physical dimensions through matrix mapping, thereby generating a heat balance trend sequence within a preset future time range, completing trend extrapolation. For example, it outputs the predicted building cooling and heating load demand values ​​(i.e., the total cooling / heating supply required to maintain a stable indoor thermal and humidity environment, including sensible heat load and latent heat load) and the predicted indoor temperature natural drift value every 15 minutes for the next 3 hours, forming a heat balance trend sequence. The thermal equilibrium trend sequence is mathematically defined as a time series vector Y containing prediction results from multiple time steps. trend = {(Q t ΔT t )| t = 1, 2, ..., N}, where N is the total number of prediction time steps (e.g., 12 15-minute steps). In this sequence: Q t ΔT represents the total heating and cooling load requirement needed to maintain the target indoor thermal and humidity environment at the t-th prediction time step. It explicitly includes the sensible heat load for controlling indoor temperature and the latent heat load for controlling indoor humidity. tThis represents the predicted natural drift of indoor temperature, i.e., the predicted change in indoor temperature caused by the combined effects of outdoor meteorological disturbances and indoor heat sources, under the assumption of zero output from the HVAC system equipment (i.e., no cooling / heating intervention). It is primarily used to help describe the future evolution of the thermal environment. The heat balance trend sequence is used as the predictive control target for the intelligent optimization algorithm in subsequent steps, solving the technical problem of severe response lag in traditional HVAC feedback control when facing sudden changes in outdoor weather. This achieves a shift from passive feedback regulation to feedforward predictive control based on dynamic prediction of future thermodynamic boundaries.

[0069] S4. Using the heat balance trend sequence as the predictive control target, an intelligent optimization algorithm is used to search for the control variables and generate a set of control schemes.

[0070] In one specific embodiment, the process of performing step S4 may specifically include the following steps:

[0071] The time-series distribution characteristics of solar radiation heat gain are extracted from the heat balance trend sequence, and indoor temperature and humidity distribution data are obtained from the multidimensional dataset to determine the gradient distribution characteristics of indoor air humidity.

[0072] Genetic algorithm is used as intelligent optimization algorithm to encode the control variables by chromosome. The control variables include at least the air conditioning supply air temperature, radiant cooling and heating load, external window ventilation opening, fresh air volume and return air volume ratio, shading component opening angle and heat recovery device efficiency adjustment level.

[0073] Combining the time-sharing distribution characteristics and the indoor air humidity gradient distribution characteristics, and taking the response prediction control target and minimizing thermal deviation as the fitness evaluation basis, the control variables are iteratively optimized and searched.

[0074] Extract various combinations of control variable parameters that satisfy the iteration termination condition, and generate a set of control schemes.

[0075] Specifically, the heat balance trend sequence includes the predicted building heating and cooling loads and the predicted natural drift of indoor temperature for each time period within a preset future time range. When analyzing and extracting the time-series distribution characteristics of solar radiation heat gain from the heat balance trend sequence, the predicted building heating and cooling loads in the heat balance trend sequence are used as a benchmark. While keeping other prediction input conditions unchanged, a first prediction scenario including the outdoor solar radiation forecast and a second prediction scenario with the outdoor solar radiation input set to zero or the benchmark value are constructed, and the corresponding building heating and cooling load prediction results are obtained respectively. The difference between the two prediction results at the corresponding time step is determined as the component of building heat gain change caused solely by outdoor solar radiation in the heat balance trend sequence. Then, the component of building heat gain change is sliced ​​according to the time axis, and combined with the orientation division of the building's external area and the radiation-receiving area of ​​the building envelope or windows for each orientation, the solar radiation heat gain of the building's external area for each time period is calculated. The proportion of solar radiation heat gain of the building's external area for each orientation to the total solar radiation heat gain for the corresponding time period is used to form the time-series distribution characteristic vector of solar radiation heat gain. Simultaneously, indoor temperature and humidity distribution data at the current moment are obtained from the multidimensional dataset constructed in S1. This data includes temperature and relative humidity values ​​for each area indoors. Based on the air thermodynamic equation of state, the dry-bulb temperature and relative humidity at each monitoring node are converted into absolute moisture content values, expressed in g / kg dry air. The spatial derivative of the absolute moisture content values ​​of adjacent nodes is calculated to obtain the indoor air moisture content gradient distribution characteristics. In specific implementation, the actual physical coordinates of each monitoring node in the building's plane coordinate system are obtained. A spatial finite difference algorithm is used to calculate the absolute moisture content difference between adjacent nodes, and this difference is divided by the straight-line physical distance between the two nodes to obtain the rate of change of moisture content in that spatial direction. The rates of change in each spatial direction are collected to calculate the indoor air moisture content gradient distribution characteristics.

[0076] Genetic algorithms include basic elements such as encoding, initial population generation, fitness evaluation, selection, crossover, mutation, and iteration termination. Controlled variables include at least the air conditioning supply air temperature setpoint, radiant cooling / heating load setpoint, window ventilation opening percentage, fresh air to return air ratio, shading component opening angle, and heat recovery device efficiency adjustment level. Based on the physical properties of the controlled variables, a chromosome model is constructed using a hybrid gene encoding strategy. For example, the air conditioning supply air temperature, radiant cooling / heating load amplitude, window ventilation percentage, fresh air to return air ratio, and shading component opening angle are set as continuous variables and mapped to their corresponding physical control upper and lower limit intervals using real-number encoding. For example, the value range of the air conditioning supply air temperature gene is 16℃ to 28℃; the value range of the radiant cooling and heating load gene is 0 to 100% of the rated load; the value range of the window ventilation opening gene is 0 to 100% fully open; the value range of the fresh air to return air ratio gene is 0 to 1, where 0 represents 100% fresh air mode and 1 represents full return air mode; and the value range of the shading component opening angle gene is 0° to 90°. The efficiency adjustment levels of the heat recovery device are set as discrete variables, using integer encoding to map to integer values ​​representing different operating levels. For example, levels 1 to 5 correspond to stop, low speed, medium speed, high speed, and full speed operation, respectively. The above real-valued gene sequences and integer gene sequences are concatenated in a fixed order to form a complete chromosome. Each chromosome represents a potential multi-machine collaborative control scheme for HVAC equipment. The initial population is randomly generated; for example, the population size is set to 100, meaning 100 chromosomes are randomly generated within the search space as the first generation population. The hybrid gene coding strategy solves the problem that traditional gradient descent optimization algorithms cannot simultaneously handle the non-convex optimization technique of mixed variables of continuous adjustment variables and discrete equipment gear positions in building control systems.

[0077] The design of the fitness function is a core evaluation criterion for guiding the search direction of the genetic algorithm. Each chromosome in the population is decoded into a corresponding combination of regulatory variable parameters, and this combination of regulatory variable parameters is input into a pre-defined algebraic equation for the energy efficiency of lightweight devices. The algebraic equation for the energy efficiency of lightweight devices is a steady-state calculation formula based on the laws of thermodynamics. For example, based on the air conditioning supply air temperature and the ratio of fresh air volume to return air volume decoded by the chromosome, combined with the specific heat capacity of air at constant pressure and the latent heat of vaporization, the total sensible cooling (heating) and latent cooling (heating) volume that the control combination can theoretically output to the room is calculated. The corresponding algebraic equation expression is: Qair = cp×m×(Tr - Ts) + r×m×(dr - ds), where Qair is the cooling and heating output of the air conditioning terminal, cp is the specific heat capacity of air at constant pressure, m is the supply air mass flow rate calculated based on the ratio of fresh air to return air, Tr and Ts are the dry-bulb temperatures of the indoor return air and the supply air under the corresponding control command, respectively, r is the latent heat of vaporization of water, and dr and ds are the air humidity under the corresponding conditions. At the same time, based on the radiant cooling and heating load parameters decoded by the chromosome, the theoretical heat exchange volume of the radiant terminal is calculated. The corresponding algebraic equation expression is: Qrad = K×A×ΔT, where Qrad is the heat exchange at the radiant terminal, K is the comprehensive heat transfer coefficient at the radiant terminal, A is the effective radiant area, and ΔT is the equivalent heat exchange temperature difference between the radiant surface and the indoor environment. Then, the total output cooling and heating capacity of the equipment obtained from the above theoretical calculation (i.e., Qout = Qair + Qrad) is subtracted from the predicted value Qt of the target cooling and heating load demand in the heat balance trend sequence generated by S3, and the absolute value is defined as the basic thermal deviation. The smaller the basic thermal deviation, the more closely the control scheme represented by the chromosome matches the load demand during the prediction period in terms of total cooling and heating capacity. Based on this, a dynamic penalty weight coefficient is introduced into the fitness function to transform the time-sharing distribution characteristics of solar radiation heat gain and the gradient distribution characteristics of indoor air humidity into physical constraints on specific control variables. For example, when the time-sharing distribution characteristics of solar radiation heat gain indicate that the heat gain of a certain orientation exceeds the set exposure threshold (e.g., 600W / m²) during a specific period... 2When the opening angle of the shading component corresponding to the decoded direction of a chromosome is too large (e.g., greater than 45°), a large radiation penalty coefficient is applied to the chromosome in the fitness function to amplify its thermal deviation value and increase its probability of being eliminated, thereby forcing the evolution direction to retain the scheme with a smaller shading angle to cut off direct heat disturbance. When the indoor air humidity gradient distribution characteristics indicate that the absolute humidity of a certain area exceeds the standard, if the fresh air ratio in the fresh air and return air ratio of the corresponding chromosome is too low or the air conditioning supply air temperature is too high (unable to meet the dew point dehumidification requirements), a humidity penalty coefficient is applied to the chromosome in the fitness function to cause the search process to converge towards increasing the fresh air ratio or decreasing the supply air temperature. The fitness function calculation formula is constructed as follows: the fitness value is equal to the reciprocal of the sum of the product of the basic thermal deviation, the solar radiation penalty coefficient, and the humidity gradient penalty coefficient, and a preset small normal number. The introduction of this preset small positive number aims to prevent the numerical explosion problem caused by the thermal deviation approaching zero, which results in a zero denominator. This small positive number has the same dimensions as the fundamental thermal deviation. A higher fitness value indicates that the candidate scheme not only responds to the predictive control objective in terms of total quantity but also perfectly satisfies the physical constraints for eliminating local environmental degradation.

[0078] After obtaining the fitness values ​​of each chromosome, the regulatory variables are iteratively optimized. The selection operation employs a tournament selection method, randomly selecting several individuals from the current population for comparison, and choosing the individual with the highest fitness value to enter the mating pool of the next generation. The crossover operation aims to explore new control combination spaces. For continuous variable gene segments such as air conditioning supply temperature, a simulated binary crossover algorithm is used, with a crossover probability set to 0.8; for discrete variable gene segments such as heat recovery device efficiency adjustment levels, a single-point crossover algorithm is used. The mutation operation aims to maintain population diversity and prevent the algorithm from getting stuck in local non-energy-saving extreme points. For continuous variable gene segments, multinomial mutation is used, with a mutation probability set to 0.1; for discrete variable gene segments, random integer mutation is used. After the above selection, crossover, and mutation operations, a new generation of regulatory strategy population is generated, and fitness evaluation and evolution are performed iteratively. For example, the maximum number of iterations is set to 200 generations, and a convergence monitor is set in the control flow. The monitor extracts the optimal fitness value from multiple generations of the population. When the change in the optimal fitness value over 20 consecutive generations is less than a preset convergence tolerance threshold (e.g., 0.01), the algorithm is determined to have converged to the global suboptimal or optimal solution interval, and the iterative search is terminated early. Introducing a dynamic penalty weight coefficient and performing iterative search solves the technical problem that blindly seeking optimization can easily lead to local non-energy-saving extreme points due to differences in response inertia and energy efficiency ratios of different devices. This forces the optimization direction to move closer to the physical goals of eliminating direct solar radiation disturbances and quickly balancing the humidity gradient.

[0079] After the genetic algorithm converges, the top N (e.g., 10) chromosomes with the highest fitness values ​​are selected from the last generation population, and their reverse decoding is performed to obtain the corresponding physical numerical sequences of regulatory variables. These parameter combinations represent multiple candidate schemes that can effectively balance energy saving and comfort while meeting solar radiation response and humidity constraints under the current thermal balance trend sequence prediction control objective. Each scheme in this set is sorted from high to low fitness value to construct a set of multiple alternative control schemes. Each scheme in this set contains a complete sequence of air conditioning supply air temperature setpoints, radiant cooling and heating load setpoints, window ventilation opening values, fresh air and return air ratio values, shading component opening angle values, and heat recovery device efficiency levels. The time resolution is set to 15 minutes for example. Generating a set of control schemes containing multiple alternative schemes solves the technical problem that a single absolutely optimal solution may fail in practical applications due to model errors, equipment execution deviations, or sudden disturbances.

[0080] S5. Input the set of control schemes into the building thermal digital twin model for simulation and obtain the corresponding set of indoor thermal environment response curves.

[0081] In one specific embodiment, the process of performing step S5 may specifically include the following steps:

[0082] The parameters of each set of control variables in the control scheme set are loaded into the thermal mirror simulation module built into the building thermal digital twin model;

[0083] By combining the thermal balance trend sequence, forward physical field simulation calculations are performed through the thermal mirror simulation module to deduce the dynamic response characteristics of the indoor thermal environment over time and generate basic time-series response curves.

[0084] During the simulation, based on the indoor local thermal disturbance distribution characteristics mapped by the digital twin model, the corresponding local thermal disturbance compensation air volume value is calculated, and the temperature distribution curve of the indoor space level is generated based on the air flow simulation.

[0085] By integrating the basic time-series response curve with the temperature distribution curve in a spatiotemporal dimension, a set of indoor thermal environment response curves with spatiotemporal response characteristics is generated.

[0086] Specifically, the building thermal digital twin model is a pre-constructed virtual three-dimensional spatial entity containing building geometry, physical material properties, and HVAC network topology. Thermal performance parameters such as the thermal resistance degradation rate of the building envelope extracted in S2 have been synchronously updated into this model to correct the heat transfer properties of the virtual materials. The loading operation refers to sequentially taking each set of parameters from the scheme set and mapping the values ​​in that set to the control interface of the corresponding virtual physical device in the thermal mirror simulation module, using them as boundary conditions for this simulation.

[0087] The thermal balance trend sequence includes the predicted building heating and cooling loads and the predicted natural drift of indoor temperature for each time period within a preset future time range. During simulation, the future outdoor weather forecast data in the thermal balance trend sequence is used as the external environment driving boundary, and the currently applied control variable parameters are used as the equipment control boundary. The thermal mirror simulation module uses the state-space method to solve the transient thermal balance differential equations of the building envelope and indoor air nodes at a set discrete time step (e.g., 5 minutes). During the solution process, the model calculates the amount of heating and cooling energy injected or extracted by the HVAC equipment into the indoor space and the amount of dehumidification at each time step based on the applied control variable parameters. Combining this with the heat storage delay characteristics of the building envelope, the model outputs a sequence of temperature and relative humidity changes over time at representative indoor points (e.g., the center point of each room) within a preset future time range (e.g., 24 hours). This sequence is the basic time-series response curve, reflecting the overall trend of the indoor thermal environment evolution over time under the candidate control scheme. Forward physics simulation calculations solve the technical problem that traditional simplified models cannot accurately simulate the building's thermal inertia and the dynamic process of heat exchange between equipment and the indoor environment.

[0088] The digital twin model internally stores indoor local thermal disturbance distribution characteristics that are synchronously updated with the real building. These characteristics include at least the location and intensity of heat sources mapped from personnel density distribution data, the location and intensity of heat sources mapped from equipment heat dissipation power distribution data, and the distribution of local radiative heat spots formed by solar radiation projected through exterior windows. At each time step of the simulation, the local thermal disturbance distribution characteristics at the current moment are read, and areas where the local heat source intensity exceeds a set threshold are identified. Specifically, based on the excess sensible heat power of the local area, combined with the air specific heat capacity, air density, and a preset supply air temperature difference, the air volume increment required to eliminate the local sensible heat accumulation is calculated using the heat balance equation, and this increment is used as the local thermal disturbance compensation air volume value. For example, to address the increase in carbon dioxide and moisture dissipation caused by areas with high personnel density, an additional 20m³ of air volume is set for each additional person. 3 / h of fresh air compensation; for solar radiation hotspot regions, the compensation air volume is calculated based on the hotspot area and instantaneous radiation intensity. This compensation air volume value is loaded into the subsequent flow field proxy model in the form of supply air boundary conditions to correct the initial velocity vector distribution of the spatial flow field.

[0089] Because traditional computational fluid dynamics (CFD) calculations in digital twin models are too time-consuming, this embodiment uses a flow field proxy model based on a three-dimensional convolutional neural network to replace the traditional iterative solver. The hierarchical structure of this flow field proxy model sequentially includes: a three-dimensional input layer, a downsampling encoder composed of multiple three-dimensional convolutional layers with a stride of 2, a residual connection bottleneck layer, an upsampling decoder composed of multiple three-dimensional transposed convolutional layers, and an output layer. Preferably, the flow field proxy model adopts a multi-task output structure, using the indoor three-dimensional temperature distribution matrix as the main output result, and simultaneously outputting the velocity vector field at each voxel node in the room as an auxiliary output result. During the model training phase, a large number of steady-state flow field calculation conditions accumulated from historical digital twin simulations are used as the training set; a three-dimensional voxelized mesh matrix containing the geometric boundaries of the building space, the coordinates and heating intensity of local thermal disturbance areas, and the air outlet velocity and temperature including the local thermal disturbance compensation airflow value are used as the model's training input data; the real temperature distribution matrix at each voxel node in the indoor three-dimensional space under the corresponding conditions is used as the main training label, and the velocity vector field under the corresponding conditions is used as the auxiliary training label. For example, a joint loss function combining mean squared error and physical conservation penalty term is adopted, the Adam optimizer is configured, the initial learning rate is set to 0.001, the batch size is 4, and iterative training is performed until the validation set loss converges, and the network parameters are solidified. In the application phase, the virtual building geometric boundary conditions, local heat source boundary conditions, and the calculated local thermal disturbance compensation air volume values ​​at the current simulation moment are spatially voxelized and encoded, and input into the pre-trained flow field proxy model. After forward feature extraction and physical mapping by a three-dimensional convolutional network, the flow field proxy model outputs the predicted three-dimensional temperature grid matrix of each spatial coordinate point in the room, and can simultaneously output the corresponding velocity vector field. Among them, the predicted three-dimensional temperature grid matrix serves as the temperature distribution curve at the indoor spatial level, which is used for subsequent thermal environment response analysis; the velocity vector field is mainly used to assist in characterizing the indoor air flow state and to enhance the physical consistency of the model output results. Performing air flow simulation based on the flow field proxy model can effectively solve the contradiction between the huge computing power required for high-fidelity three-dimensional spatial thermal field simulation and the real-time requirements of the control system.

[0090] Furthermore, to ensure the physical rationality and feasibility of the output of the 3D CNN flow field proxy model, the following settings are made for the indoor 3D spatial discretization method and the model's training loss function: First, in terms of spatial discretization, the target indoor space is divided into a 3D voxel grid with a fixed resolution. For example, the voxel spatial resolution can be set to 0.1m~0.5m, or the entire space can be discretized into standard tensor dimensions such as 32×32×32, to adapt to the receptive field of the 3D convolutional kernel. Second, during the model training phase, in addition to the temperature field reconstruction error, a physical conservation penalty term L is introduced into the total loss function. phyFor low-speed indoor airflow, approximating it as an incompressible fluid, the physical conservation penalty term includes at least a divergence penalty term based on the law of mass conservation, and its mathematical expression is: Where U = (u, v, w) is the velocity vector field of each voxel in three-dimensional space predicted by the model. The divergence of the velocity vector field is represented by , The square of the L2 norm; These are preset dynamic weighting coefficients. By introducing a physical conservation penalty term, a mass conservation constraint is imposed on the three-dimensional flow field output by the model, thus enabling the model to predict indoor temperature distribution while also considering the consistency of the physical boundaries of airflow, avoiding the irregular prediction results that violate fluid dynamics laws produced by a purely data-driven model.

[0091] Using the time step as an index, the indoor spatial temperature distribution curves (i.e., the three-dimensional temperature grid matrix) at a specific moment output by the flow field proxy model are superimposed and embedded into the corresponding time-series nodes of the basic time-series response curves. Tensor stitching technology is employed to construct a four-dimensional data tensor containing both a time evolution dimension and a three-dimensional spatial coordinate dimension. This four-dimensional data tensor constitutes the indoor thermal environment response curve set. This curve set comprehensively records the overall trend of indoor macroscopic temperature and humidity evolution over time within a preset time range under the influence of the candidate control scheme, as well as the differentiation of local temperature gradients at various spatial locations. Integrating the spatiotemporal dimensions solves the technical problem that response curves in a single time dimension cannot expose the risks of overcooling or overheating in spatial dead zones, providing high-dimensional physical fidelity data support for subsequent multi-objective comparative analysis and the generation of decision-making indicator sets.

[0092] S6. Compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in different time periods. Based on the comparison results, determine the decision basis index set corresponding to each control variable.

[0093] In one specific embodiment, the process of performing step S6 may specifically include the following steps:

[0094] The indoor thermal environment response curves were matched and compared with the preset energy-saving constraint range and the preset thermal comfort constraint range for each time period.

[0095] Based on the comparative analysis results, the deviation direction and deviation magnitude of the indoor thermal environment response curve set relative to the preset energy-saving constraint range and the preset thermal comfort constraint range in each time period are determined, and the energy-saving deviation sequence and comfort deviation sequence are generated.

[0096] Multi-objective optimization is performed based on energy-saving deviation series and comfort deviation series to determine the control strategies of each control variable under different time periods. The control strategies include at least the adjustment direction, adjustment magnitude and adjustment priority.

[0097] Based on the control strategy, the specific execution strategy constraint range for each control variable is generated. The specific execution strategy constraint range includes at least the adjustment range of cooling and heating load for floor radiant heating and the time-sharing control priority sequence for the opening of external windows.

[0098] The regulatory strategy and the specific implementation strategy are integrated to form a set of indicators for decision-making.

[0099] Specifically, the indoor thermal environment response curve set includes a sequence of predicted indoor temperature values, a sequence of predicted indoor relative humidity values, and a temperature distribution matrix at the indoor space level for each time period within a preset future time range. The preset energy-saving constraint range is set according to building energy consumption quota standards, representing a sequence of maximum allowable energy consumption thresholds for building HVAC equipment at each time step. For example, an energy-saving constraint upper limit of no more than 0.1 kW per unit building area for air conditioning is taken. The preset thermal comfort constraint range is set according to the recommended range of the expected average thermal perception index (PMV) specified in the national standard GB / T 18049. For example, an indoor temperature of 24°C to 26°C and a relative humidity of 40% to 60% corresponding to a PMV value between -0.5 and 0.5 are taken as the thermal comfort constraint range. Time-by-time comparison refers to using a set discrete time step as the search window to synchronously align and compare the predicted temperature and humidity values ​​corresponding to each time step in the indoor thermal environment response curve group with the preset thermal comfort constraint interval. Simultaneously, based on the operating status parameters of each HVAC device under the corresponding control scheme, the predicted energy consumption values ​​for each future time step are calculated, and these predicted energy consumption values ​​are synchronously aligned and compared with the preset energy-saving constraint interval. The predicted energy consumption values ​​are obtained by converting the equipment operating parameters under the corresponding control scheme. These operating parameters include at least the air conditioning supply air temperature setpoint, the ratio of fresh air volume to return air volume, the radiant cooling and heating load setpoint, the efficiency adjustment level of the heat recovery device, and the ventilation opening of external windows. Based on the power model, energy efficiency model, or preset energy consumption conversion relationship of each device under the corresponding operating parameters, a sequence of predicted energy consumption values ​​for each future time step is calculated. The matching and comparison analysis specifically includes: determining whether the predicted temperature value falls within the thermal comfort constraint range, determining whether the predicted humidity value falls within the thermal comfort constraint range, determining whether the energy consumption index exceeds the upper limit of the energy saving constraint, and identifying whether there are local areas in the temperature distribution matrix that exceed the thermal comfort constraint range.

[0100] Deviation direction refers to whether the physical quantity in the response curve group is higher than the upper limit or lower than the lower limit of the constraint interval. Deviation magnitude refers to the absolute value of the value exceeding the constraint boundary. For example, if the predicted indoor temperature for a certain period is 27℃ and the upper limit of the thermal comfort constraint interval is 26℃, then the deviation direction is "too high" and the deviation magnitude is 1℃; if the predicted indoor relative humidity for a certain period is 65% and the upper limit of the thermal comfort constraint interval is 60%, then the deviation direction is "too high humidity" and the deviation magnitude is 5%. Organizing the deviation direction and deviation magnitude of all periods in chronological order forms an energy-saving deviation sequence and a comfort deviation sequence. The energy-saving deviation sequence records the degree to which energy consumption indicators exceed the constraint boundary for each period within a future preset time range, while the comfort deviation sequence records the degree to which temperature and humidity indicators exceed the thermal comfort boundary.

[0101] A fast non-dominated sorting genetic algorithm with an elite retention strategy is adopted as the multi-objective optimization framework. Minimizing the sum of the quantitative indices of the energy-saving deviation sequence and the comfort deviation sequence are used as the two interplay objective functions. The algorithm's input parameters are the basic control variables in the control scheme set generated by S4, including the air conditioning supply air temperature setpoint, radiant cooling and heating load setpoint, window ventilation opening value, fresh air volume to return air volume ratio, shading component opening angle value, and heat recovery device efficiency adjustment level value. During the algorithm's optimization iteration process, the impact of the basic control variables on the above two objective functions under different mutation states is calculated. Based on the Pareto front principle, the non-dominated solution set that minimizes comfort deviation to the greatest extent without increasing system energy consumption deviation is selected. The non-dominated solution set is then analyzed into control strategies corresponding to specific physical devices. For example, for a specific time period, when the comfort deviation sequence indicates local overheating and the energy-saving deviation sequence indicates that the system energy consumption is approaching its upper limit, the optimization algorithm outputs the following control strategy: the adjustment direction for radiant cooling is to increase, with an adjustment range set at 10% of the baseline value; the adjustment direction for air conditioning supply temperature is to increase, with an adjustment range of 1°C; in terms of adjustment priority, the lowering of energy-efficient shading components and external window ventilation are given the highest priority, while increasing the mechanical cooling load is given the lowest priority. Multi-objective optimization based on the energy-saving deviation sequence and the comfort deviation sequence solves the technical problem that single-objective optimization cannot simultaneously address the dual constraints of energy saving and comfort, enabling the generated control strategy to respond to both control objectives simultaneously.

[0102] The specific execution strategy constraints are based on the adjustment direction and magnitude determined by the control strategy, combined with the physical limitations of the equipment, to define the feasible boundaries. For example, considering the hydraulic balance constraints of the pipe network, the adjustment range for the cooling load of floor radiant heating is limited to 15W / m². 2 Up to 30W / m 2Within the upper and lower limit range, and based on wind speed and rainfall sequence from outdoor meteorological data, a time-sharing control priority sequence for window ventilation is generated. This prioritizes the execution of the control command to open windows to a 45° angle during specific periods when outdoor enthalpy is lower than indoor enthalpy during the transition season, and forcibly postpones the execution priority of mechanical refrigeration valves. This generates specific execution strategy constraints, resolving the technical problem that abstract control strategies cannot directly guide physical equipment execution, and transforming the algorithm's abstract control intent into a securely identifiable physical boundary at the underlying level.

[0103] After multi-objective optimization and boundary constraint processing, the data is extracted and structured according to the dimensions of time, equipment spatial topology nodes, and control command type. For example, the decision-making basis index set can be organized into a table, with each row corresponding to a time period, each column corresponding to a control variable, and each cell containing the control direction indicator, numerical adjustment range, and priority level for that time period. This structured data set completely records all intermediate results from deviation analysis to control decision-making, carrying the priority action sequence to be followed by each subsystem of the building in future time periods, the upper and lower limits of the adjustable range, and the expected environmental target parameters. Integrating the control strategy with the constraints of the specific execution strategy to form the decision-making basis index set solves the technical problem of inconsistent data formats and potential timing conflicts when massive heterogeneous control commands are issued to the underlying programmable logic controller.

[0104] S7. Select the optimal control scheme from the set of decision-making indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

[0105] In one specific embodiment, the process of performing step S7 may specifically include the following steps:

[0106] Based on the characteristics of various indicators in the decision-making basis indicator set, candidate schemes that simultaneously meet the preset energy-saving constraint range and the preset thermal comfort constraint range are selected from the set of control schemes, and the optimal control scheme is selected from the candidate schemes according to the preset evaluation weight.

[0107] The optimal control scheme is analyzed, and its specific control parameters under the corresponding time period are extracted. The control parameters include the air conditioning supply air temperature setpoint sequence, the ratio of fresh air volume to return air volume, the efficiency adjustment level of the heat recovery device, the opening angle sequence of the shading component, the adjustment range of the floor radiation load, the ventilation opening of the external window, and the local thermal disturbance compensation air volume value.

[0108] Based on the preset building equipment underlying communication protocol, various specific control parameters are converted into a sequence of control instructions that can be recognized and executed by the building equipment control layer;

[0109] The control command sequence is sent to the corresponding building physical equipment to drive the operation, so as to realize the dynamic closed-loop control of building thermal performance from digital twin space to physical space.

[0110] Specifically, the decision-making basis indicator set includes the adjustment direction, adjustment magnitude, adjustment priority, and specific execution strategy constraints for each control variable at each time period within a future preset time range. The control scheme set includes multiple candidate schemes optimized by a genetic algorithm. Each scheme includes a complete sequence of air conditioning supply air temperature setpoints, a sequence of fresh air volume and return air volume ratios, a sequence of heat recovery device efficiency adjustment levels, a sequence of shading component opening angles, a sequence of floor radiant load adjustment magnitudes, a sequence of window ventilation opening degrees, and a sequence of local thermal disturbance compensation airflow values ​​(generated during the simulation phase). The decision-making basis indicator set generated by S6 is retrieved, and the energy-saving deviation sequence and comfort deviation sequence of each scheme within the future preset time range are read. Schemes whose equipment energy consumption exceeds the preset energy-saving constraint upper limit or whose temperature and humidity exceed the preset thermal comfort constraint range at any time step are eliminated, and the remaining schemes are selected as candidate schemes.

[0111] For example, a linear weighted method is used to construct a comprehensive evaluation function. The total predicted energy consumption of the candidate schemes is used as the energy-saving evaluation item, and the absolute difference between the predicted average thermal perception index of the candidate schemes and the ideal value of 0 is used as the comfort evaluation item. Evaluation weights are assigned based on the actual operational needs of the building. For example, during weekday daytime hours when personnel are mainly working, the comfort evaluation item is assigned a weight of 0.7 and the energy-saving evaluation item a weight of 0.3; during non-office hours when there are few or no personnel, the energy-saving evaluation item is assigned a weight of 0.8 and the comfort evaluation item a weight of 0.2. Each indicator is normalized, and the normalized item is multiplied by its corresponding weight and summed to calculate the comprehensive evaluation score for each candidate scheme. The candidate scheme with the highest comprehensive evaluation score is selected as the optimal control scheme.

[0112] The optimal control scheme is stored in a structured data format, for example, JSON or a three-dimensional matrix format. The parsing process specifically involves: degrading the highly structured data in the optimal control scheme to restore the parameter setpoints of each physical device at different time steps. The extracted control parameters include at least: a sequence of air conditioning supply air temperature setpoints, with a time resolution of 15 minutes for example; a sequence of fresh air volume to return air volume ratios, dimensionless, ranging from 0 to 1; a sequence of heat recovery device efficiency adjustment levels, taking integer values ​​from 1 to 5, corresponding to the rotor speed of the total heat exchanger or the opening degree of the bypass valve; a sequence of shading component opening angles, ranging from 0° to 90°, used to control the tilt angle of electric louvers to block direct radiation; a sequence of floor radiation load adjustment ranges, which are constrained within preset physical boundaries by the hydraulic balance of the pipe network; a sequence of external window ventilation openings, in percentage form, ranging from 0% to 100%, with priority determined according to outdoor meteorological conditions; and a sequence of local thermal disturbance compensation airflow values, calculated for high-density areas of people or equipment with concentrated heat generation.

[0113] The underlying communication protocols for building equipment include at least Modbus RTU, Modbus TCP / IP, BACnetMS / TP, or BACnet / IP protocols. The conversion process specifically involves retrieving the communication protocol library that matches the physical building's HVAC programmable logic controller (PLC), and encoding it according to the data frame format specified in the protocol. For example, for an air conditioning unit using the Modbus RTU protocol, the air conditioning supply temperature setpoint of 26.0°C is converted into a corresponding register write instruction. The conversion steps include: determining the Modbus slave address of the device to be 01; determining the PLC addressing address corresponding to the temperature setpoint to be 40001, whose corresponding protocol data address is 0x0000 (i.e., 40001 - 40001 = 0); multiplying 26.0℃ by 10 to convert it to the integer 260, corresponding to hexadecimal 0x0104; assembling the instructions according to the Modbus RTU frame format: address 01, function code 06, protocol data address high byte 00, protocol data address low byte 00, data high byte 01, data low byte 04, CRC checksum low byte, CRC checksum high byte. For devices using the BACnet protocol, according to the BACnet protocol message specification, the parsed air conditioning supply temperature setpoint is converted into a write attribute service request message representing an analog output object; discrete state parameters such as the efficiency adjustment level of the heat recovery device and the ventilation opening of the external window are converted into message instructions representing polymorphic value objects.

[0114] During the instruction conversion process, multiple messages generated within the same time step are queued in sequence, based on the control priority sequence provided by the decision-making indicator set in S6. For example, messages related to the operation of sunshade components and exterior windows without mechanical energy consumption are prioritized for distribution, while messages related to compressor refrigeration and water pump frequency conversion are placed at the end of the queue, forming a control instruction sequence that includes timestamps and device physical address identifiers.

[0115] The process of issuing control command sequences is completed through network communication. For example, through the control gateway of the building automation system, the converted protocol messages are sent to the communication ports of the corresponding field programmable logic controllers (PLCs) or direct digital controllers according to the timestamp nodes in the control command sequence. For air conditioning units, the commands are issued to the serial interface of the air conditioning controller; for shading systems, the commands are issued to the fieldbus interface of the shading control module; for floor radiant systems, the commands are issued to the electric heating actuator control module of the zoned water collector. After receiving the messages, the underlying controller parses them into corresponding analog voltage or digital relay opening / closing signals, driving the air conditioning water valve actuators, air valve actuators, electric shading motors, and other equipment to perform corresponding mechanical actions. After the physical equipment performs its actions, its state parameters and the changed indoor sensor measured temperature and humidity data are transmitted back to the multi-dimensional data set in step S1 via the original data bus, serving as the initial state data for the next round of digital twin model updates and control strategy optimization. By sending the control command sequence to the building physical equipment for driving operation and transmitting measured data back, the technical problems of virtual decision-making not being able to automatically act on physical equipment and open-loop control not being able to dynamically correct were solved. The complete control loop from data acquisition, virtual space synchronous deduction and optimization to physical space control and driving was completed, realizing dynamic closed-loop control of building thermal performance from digital twin space to physical space.

[0116] To verify the beneficial effects of the digital twin-based building energy-saving control method provided by this invention, a week-long comparative experiment was conducted in a typical office building. During the experiment, the building environment was controlled using both the traditional PID feedback control method (comparative) and the method of this application (example), and the total building energy consumption and indoor thermal comfort index (PMV) were recorded simultaneously. Figure 2 This is a comparison chart of the dual-objective synergistic optimization effects of the method of this invention and the traditional method on a typical day during the experiment. The horizontal axis represents time (hours), the left vertical axis represents the building's cumulative energy consumption (kWh), and the right vertical axis represents the deviation from indoor thermal comfort (|ΔPMV|, i.e., the absolute deviation of the measured PMV value from the ideal value of 0). Figure 2As shown in the figure, the dotted and dashed curves represent the performance of the traditional PID control method: its hourly energy consumption curve rises significantly during the midday peak load period, and the thermal comfort deviation fluctuates wildly, exceeding the absolute deviation threshold of 0.5 multiple times. The dotted and solid curves represent the performance of the method in this application: because the system uses a digital twin model for forward extrapolation and "peak shaving and valley filling" pre-cooling regulation, its energy consumption curve remains flat throughout the day. While significantly reducing the daytime peak load, the total energy consumption is reduced by 23.5% compared to the traditional method, and the thermal comfort deviation remains stable within 0.3, with an average deviation reduction of 42%. Experimental results show that this invention, through forward extrapolation and multi-objective optimization using a digital twin model, successfully achieves dual optimization of energy saving and comfort, significantly improving the dynamic balance capability of building thermal performance.

[0117] The above describes the building energy-saving control method based on digital twins in the embodiments of this application. The following describes the building energy-saving control system based on digital twins in the embodiments of this application. Please refer to [link / reference]. Figure 3 The present application provides a schematic diagram of the structure of a digital twin-based building energy-saving control system, which includes:

[0118] The data acquisition module 10 is used to acquire multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status.

[0119] The twin modeling module 20 is used to input a multi-dimensional data set into a preset building thermal digital twin model, extract multi-source coupling features, and then determine the indoor thermal environment state vector and the thermal performance parameters of the building envelope.

[0120] The trend prediction module 30 is used to predict the thermal balance trend sequence within a preset time range based on historical data and real-time meteorological information when the indoor thermal environment state vector exceeds the preset thermal comfort range.

[0121] The intelligent optimization module 40 is used to search for the control variables and generate a set of control schemes by using the heat balance trend sequence as the predictive control target and employing an intelligent optimization algorithm.

[0122] The simulation module 50 is used to input the set of control schemes into the building thermal digital twin model for simulation and to obtain the corresponding set of indoor thermal environment response curves.

[0123] The constraint assessment module 60 is used to compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in a time-by-time manner, and determine the decision basis index set corresponding to each control variable based on the comparison results.

[0124] The decision execution module 70 is used to select the optimal control scheme from the set of decision basis indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

[0125] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the digital twin-based building energy-saving control method.

[0126] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A building energy-saving control method based on digital twins, characterized in that, The method includes: S1. Obtain multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status. S2. Input the multidimensional data set into the preset building thermal digital twin model, extract multi-source coupling features, and then determine the indoor thermal environment state vector and the thermal performance parameters of the building envelope. S3. When the indoor thermal environment state vector exceeds the preset thermal comfort range, predict the thermal balance trend sequence within the preset time range based on historical data and real-time meteorological information. S4. Using the heat balance trend sequence as the predictive control target, an intelligent optimization algorithm is used to search for the control variables and generate a set of control schemes. S5. Input the set of control schemes into the building thermal digital twin model for simulation and obtain the corresponding set of indoor thermal environment response curves; S6. Compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in a time-by-time manner, and determine the decision basis index set corresponding to each control variable based on the comparison results. S7. Select the optimal control scheme from the set of decision-making indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

2. The method according to claim 1, characterized in that, S1 includes: Real-time sequence data of outdoor meteorological parameters are collected through a sensor network deployed on the exterior of the building. The meteorological parameters include at least outdoor temperature and humidity, wind speed and solar radiation data. Temperature data is acquired layer by layer by temperature sensors installed in each layer of the building envelope, and temperature gradient distribution data of the building envelope is generated based on the temperature difference between each layer. By deploying temperature and humidity sensors in different indoor areas, temperature and humidity data of each area are collected, and indoor temperature and humidity distribution data are generated by combining the area of ​​each area. By deploying personnel detection devices in different indoor areas, the location and number of people in each area are collected, thereby generating indoor personnel density distribution data; The power monitoring device records the time-series fluctuation data of the heat dissipation power of indoor equipment; The real-time sequence data, the temperature gradient distribution data, the indoor temperature and humidity distribution data, the indoor personnel density distribution data, and the time-series fluctuation data are standardized and structured to generate a multi-dimensional data set.

3. The method according to claim 1, characterized in that, A pre-trained neural network model is set in the building thermal digital twin model, S2 including: The multidimensional dataset is divided into data subsets corresponding to different thermal influencing factors; The feature extraction layer of the neural network model extracts features from each data subset, and the coupling analysis layer calculates the coupling relationship between multiple source factors to generate a comprehensive feature vector. The current indoor thermal environment state vector is determined based on the comprehensive feature vector. Sub-features related to the thermal performance of the building envelope are extracted from the comprehensive feature vector, and the thermal resistance degradation rate of the building envelope is calculated by combining historical thermal performance data. The thermal resistance degradation rate of the building envelope is used as the thermal performance parameter of the building envelope.

4. The method according to claim 1, characterized in that, S3 include: The indoor thermal environment state vector is compared with the preset thermal comfort range. When the indoor thermal environment state vector exceeds the preset thermal comfort range, historical data for continuous time periods corresponding to the current operating state are extracted from the historical thermal state multidimensional data set. The historical data of the continuous time period are processed using a preset time series prediction model to extract the temporal evolution characteristics of the building's thermal state; Using real-time outdoor meteorological parameters and the dynamic rate of change of indoor and outdoor temperature difference as input conditions, and combining the time series evolution characteristics, the time series prediction model is used to extrapolate the trend and generate the heat balance trend sequence within a preset time range in the future.

5. The method according to claim 1, characterized in that, S4 includes: The time-division distribution characteristics of solar radiation heat gain are extracted from the heat balance trend sequence, and indoor temperature and humidity distribution data are obtained from the multidimensional dataset to determine the gradient distribution characteristics of indoor air humidity. A genetic algorithm is used as the intelligent optimization algorithm to encode the control variables using chromosomes. The control variables include at least the air conditioning supply air temperature, radiant cooling and heating load, window ventilation opening, fresh air volume to return air volume ratio, shading component opening angle, and heat recovery device efficiency adjustment level. Combining the time-sharing distribution characteristics and the indoor air humidity gradient distribution characteristics, and taking the response to the predictive control objective and minimization of thermal deviation as the fitness evaluation basis, the regulation variables are iteratively optimized and searched. Extract various combinations of control variable parameters that satisfy the iteration termination condition to generate the control scheme set.

6. The method according to claim 1, characterized in that, S5 include: The parameters of each set of control variables in the control scheme set are loaded into the thermal mirror simulation module built into the building thermal digital twin model; By combining the heat balance trend sequence, forward physics field simulation calculations are performed through the thermal mirror simulation module to deduce the dynamic response characteristics of the indoor thermal environment over time and generate basic time-series response curves. During the simulation, based on the indoor local thermal disturbance distribution characteristics mapped by the digital twin model, the corresponding local thermal disturbance compensation air volume value is calculated, and the temperature distribution curve of the indoor space level is generated based on the air flow simulation. The basic time-series response curve and the temperature distribution curve are integrated in a spatiotemporal dimension to generate the indoor thermal environment response curve set with spatiotemporal response characteristics.

7. The method according to claim 1, characterized in that, S6 include: The indoor thermal environment response curve set was matched and compared with the preset energy-saving constraint range and the preset thermal comfort constraint range in a time-by-time analysis. Based on the comparative analysis results, the deviation direction and deviation magnitude of the indoor thermal environment response curve set relative to the preset energy-saving constraint range and the preset thermal comfort constraint range in each time period are determined, and the energy-saving deviation sequence and comfort deviation sequence are generated. Based on the energy-saving deviation sequence and the comfort deviation sequence, multi-objective optimization is performed to determine the control strategies of each control variable under different time periods. The control strategies include at least the adjustment direction, adjustment magnitude and adjustment priority. Based on the control strategy, a specific execution strategy constraint range is generated for each control variable. The specific execution strategy constraint range includes at least the adjustment range of cooling and heating load for floor radiant heating and the time-sharing control priority sequence for the opening degree of external windows. The control strategy and the specific implementation strategy are integrated to form the decision-making basis indicator set.

8. The method according to claim 1, characterized in that, S7 includes: Based on the characteristics of each indicator in the decision-making criteria indicator set, candidate schemes that simultaneously satisfy the preset energy-saving constraint range and the preset thermal comfort constraint range are selected from the control scheme set, and the optimal control scheme is selected from the candidate schemes according to the preset evaluation weight. The optimal control scheme is analyzed, and its specific control parameters under the corresponding time period are extracted. The control parameters include the air conditioning supply air temperature setpoint sequence, the ratio of fresh air volume to return air volume, the efficiency adjustment level of the heat recovery device, the opening angle sequence of the shading component, the adjustment range of the floor radiation load, the ventilation opening of the external window, and the local thermal disturbance compensation air volume value. According to the preset building equipment underlying communication protocol, the specific control parameters are converted into a sequence of control instructions that can be recognized and executed by the building equipment control layer; The control command sequence is sent to the corresponding building physics equipment for driving operation, so as to realize dynamic closed-loop control of building thermal performance from digital twin space to physical space.

9. A building energy-saving control system based on digital twins, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is used to acquire multi-source real-time data on the current building operation status and construct a multi-dimensional data set of the current building thermal status. The twin modeling module is used to input the multidimensional data set into a preset building thermal digital twin model, extract multi-source coupling features, and then determine the indoor thermal environment state vector and the thermal performance parameters of the building envelope. The trend prediction module is used to predict the thermal balance trend sequence within a preset time range based on historical data and real-time meteorological information when the indoor thermal environment state vector exceeds the preset thermal comfort range. The intelligent optimization module is used to search for the control variables using the heat balance trend sequence as the predictive control target and generate a set of control schemes by employing an intelligent optimization algorithm. The simulation module is used to input the set of control schemes into the building thermal digital twin model for simulation and to obtain the corresponding set of indoor thermal environment response curves. The constraint assessment module is used to compare and analyze the indoor thermal environment response curve set with the preset energy-saving constraint range and the preset thermal comfort constraint range in a time-by-time manner, and determine the decision basis index set corresponding to each control variable based on the comparison results. The decision execution module is used to select the optimal control scheme from the set of decision basis indicators and convert it into a sequence of control instructions that can be executed by the building equipment control layer to drive the operation of building equipment and realize closed-loop control of building energy conservation.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the building energy-saving control method based on digital twins as described in any one of claims 1 to 8.

Images