Building thermal inertia cooperative scheduling method and device

Abstract

The application discloses a building thermal inertia cooperative scheduling method and device, introduces PINN (Physics-Informed Neural Networks) to ensure that a prediction model conforms to a building thermal balance physical mechanism, and utilizes MCTS (Monte Carlo Tree Search) to realize efficient optimization of an air conditioning unit operation sequence, and solves the technical problem that energy saving effect and power quality (power factor) are difficult to cooperate in virtual energy storage scheduling.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of smart energy management and electricity demand response, specifically relating to a building thermal inertial collaborative scheduling method and device based on physical information fusion. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy for new power systems, refined control of building load has become a key link in the coordinated interaction of power generation, grid, load, and storage. Heating, ventilation, and air conditioning (HVAC) systems, as the main energy consumers in large public buildings and commercial complexes (typically accounting for over 40%), are ideal flexible load resources due to their high power density and emphasis on energy conservation. Research based on building thermodynamics shows that building envelopes and indoor air have significant thermal inertia. This characteristic causes indoor temperature response to exhibit significant amplitude attenuation and phase lag relative to cold / heat input, thus giving buildings the ability to "store" and "release" heat energy within a specific comfort deadband. In engineering, this is equivalent to a "virtual energy storage system" (VES) that does not require additional electrochemical media.

[0003] Existing technologies for building virtual energy storage primarily focus on peak shaving and valley filling or real-time electricity price response at the active power level, i.e., shifting electricity consumption time-series through pre-cooling / pre-heating strategies. However, in modern HVAC systems with variable frequency drives (VFDs) as the core actuators, active and reactive power exhibit a strong nonlinear coupling relationship. When the system load rate is significantly reduced for energy saving or demand response, the nonlinear characteristics of the power electronic components inside the VFDs lead to a significant deterioration in the input power factor (PF) and a surge in harmonic content. Once the power factor at the grid connection point falls below the grid company's assessment threshold (e.g., 0.90), users will face hefty power regulation fees and penalties, and it may even trigger voltage dips in the distribution network, threatening equipment safety.

[0004] Current mainstream building energy management systems (BEMS) have significant technical limitations in addressing the above problems: 1. Fragmentation of control strategies: The common practice is to adopt an open-loop strategy of "energy-saving control first, reactive power compensation later", which relies solely on capacitor banks or SVG for passive compensation at the end. It lacks the proactive defense capability of coordinating active and reactive power planning from the load source, often resulting in negative economic benefits where "energy-saving benefits are offset by penalties".

[0005] 2. Physical deficiencies in prediction models: Existing load and temperature predictions are mostly based on purely data-driven "black box" models such as RNNs and LSTMs, lacking explicit constraints on thermodynamic mechanisms. Under extreme conditions or in regions with sparse training data, prediction results often violate the law of conservation of energy, rendering scheduling commands unexecutable.

[0006] 3. Limitations of optimization algorithms: When faced with mixed integer nonlinear programming (MINLP) problems that include continuous variables (frequency) and discrete variables (start and stop states), traditional gradient descent methods or genetic algorithms are prone to getting trapped in local optima and are difficult to quickly search for the globally optimal control sequence in the massive solution space while satisfying the strict power factor redline constraint.

[0007] In summary, developing a building thermal inertial scheduling method that can integrate physical mechanism constraints and synergistically consider thermal comfort, operating energy consumption, and power quality is a pressing problem that needs to be solved in the current technological field. Summary of the Invention

[0008] This invention aims to address the following technical bottlenecks in existing building energy management systems when dealing with variable frequency heating, ventilation, and air conditioning (HVAC) load regulation: 1. The fragmented nature of the electric-thermal coordinated control strategy: Existing technologies generally adopt an open-loop strategy of "energy-saving control first, reactive power compensation later", which lacks the proactive defense capability to plan active and reactive power coordination from the load source. This leads to the deterioration of the power factor at the grid connection point when deeply adjusting the load rate, resulting in the negative economic benefit of "energy-saving benefits being offset by penalties". 2. Physical deficiencies in prediction models: Existing load and temperature predictions are mostly based on purely data-driven "black box" models, lacking explicit constraints on thermodynamic mechanisms. Under extreme conditions or in regions with sparse data, the prediction results often violate the law of conservation of energy, leading to the inability to execute scheduling commands. 3. Limitations of optimization algorithms: When faced with mixed integer nonlinear programming (MINLP) problems that include continuous variables (frequency) and discrete variables (start-stop states), traditional algorithms struggle to quickly search for the globally optimal control sequence in a massive solution space while satisfying strict power factor constraints.

[0009] To address the problems existing in the prior art, this invention provides a method and apparatus for coordinated scheduling of building thermal inertia.

[0010] To achieve the above objectives, the present invention provides the following solution: A building thermal inertia cooperative scheduling method includes: S1. Real-time acquisition of indoor and outdoor environmental parameters of the target building, operating electrical parameters and status parameters of the HVAC system, time-of-use electricity price data, and real-time power factor of the building's grid connection point through the sensing network; S2. Construct the thermal inertia state space equation of the building, and equivalently map the heat storage and release capacity of the building envelope within the temperature comfort range to virtual energy storage resources, and define the virtual energy storage state of charge. Simultaneously, a nonlinear coupling characteristic model of active and reactive power of the HVAC system is constructed, and a correlation mapping matrix between load rate and instantaneous power factor is established. S3. Construct a neural network prediction model integrating physical prior constraints, embedding the partial differential equation of heat conduction of the building envelope as a regularization constraint term into the loss function; use the PINN model to predict the indoor temperature evolution trend and virtual energy storage SOCves in the future time domain, ensuring that the prediction output conforms to the thermodynamic law of conservation of energy. S4. Taking the minimization of the total cost of operating electricity and power regulation electricity, and the minimization of indoor thermal comfort deviation as the multi-objective function, under the premise of satisfying the upper and lower limits of indoor temperature, the lower limit of the power factor at the grid connection point is introduced as a hard constraint operator; at the same time, Monte Carlo tree search or improved model predictive control algorithm is used to solve the optimal active power scheduling sequence of the HVAC system in the future finite prediction time domain within each scheduling cycle. S5. The first control command of the optimal active power scheduling sequence is sent to the actuator, and the model parameters in S2 are dynamically corrected using a feedback compensation algorithm based on the real-time temperature and power factor deviation, thus completing the closed-loop control in the rolling time domain.

[0011] Preferably, in step S2, the construction of the building thermal inertia virtual energy storage model includes: Define virtual energy storage state of charge The normalized relative position of the current measured indoor temperature within the preset temperature comfort range is defined as: the process of increasing the cooling or heating power of the HVAC system to store energy in the building structure is defined as virtual energy storage charging; the process of reducing the power of the HVAC system to maintain room temperature using residual heat or cold from the building is defined as virtual energy storage discharging; a differential equation describing the thermal dynamic evolution process is established using the second-order equivalent thermal parameter ETP model, and the mapping relationship between virtual energy storage charging and discharging power and indoor temperature change rate is determined accordingly.

[0012] Preferably, in step S2, the construction of the active-reactive power coupling characteristic model of the HVAC system includes: Based on historical operating samples, the power factor dynamic curves of the HVAC system under different load rates are fitted; reactive power is constructed. With active power nonlinear functional relationship ; The power factor constraint in step S4 is defined as follows: , in, This is the preset power factor assessment threshold.

[0013] Preferably, the construction of the physical information neural network PINN in step S3 includes: The partial differential equations of building thermodynamics (PDEs) are embedded as residual terms in the loss function of a neural network; the physical loss term is defined. To predict the thermal equilibrium residuals of temperature with respect to time and space derivatives, the output of the neural network is constrained to conform to the law of conservation of energy by minimizing the comprehensive loss function containing physical residuals during training, thereby achieving physical robustness for predicting the state of charge of virtual energy storage under extreme weather conditions.

[0014] Preferably, step S5 further includes feedback correction logic: Real-time monitoring of the residual between actual indoor temperature and predicted temperature; When the residual exceeds a preset threshold, the input parameters of the prediction model at the next moment are corrected using the error integral term to reduce the steady-state error caused by model mismatch or random disturbances, thereby improving the system's temperature control stability and the ability to maintain the power factor at the grid connection point under strong disturbance conditions. This invention also provides a building thermal inertia cooperative scheduling device, comprising: The first processing module is used to acquire in real time the indoor and outdoor environmental parameters of the target building, the operating electrical parameters and status parameters of the HVAC system, time-of-use electricity price data, and the real-time power factor of the building's grid connection point through the sensing network. The second processing module is used to construct the thermal inertia state space equation of the building, which maps the heat storage and heat release capacity of the building envelope in the temperature comfort range to virtual energy storage resources and defines the virtual energy storage state of charge (SOCves); at the same time, it constructs the active-reactive power nonlinear coupling characteristic model of the HVAC system and establishes the correlation mapping matrix between load rate and instantaneous power factor. The third processing module is used to construct a neural network prediction model that integrates physical prior constraints. It embeds the partial differential equation of heat conduction of the building envelope as a regularization constraint term into the loss function. It uses the PINN model to predict the indoor temperature evolution trend and virtual energy storage SOCves in the future time domain, ensuring that the prediction output conforms to the thermodynamic law of conservation of energy. The fourth processing module is used to minimize the total cost of operating electricity and power regulation electricity, and minimize the deviation of indoor thermal comfort as multiple objective functions. Under the premise of satisfying the upper and lower limits of indoor temperature, the lower limit of the power factor at the grid connection point is introduced as a hard constraint operator. At the same time, Monte Carlo tree search or improved model predictive control algorithm is used to solve the optimal active power scheduling sequence of the HVAC system in the future finite prediction time domain within each scheduling cycle. The fifth processing module is used to send the first control command of the optimal active power scheduling sequence to the actuator, and dynamically correct the model parameters in S2 using a feedback compensation algorithm based on the real-time collected temperature and power factor deviations, thereby completing the closed-loop control in the rolling time domain.

[0015] Preferably, the fifth processing module is also used to execute feedback correction logic: Real-time monitoring of the residual between actual indoor temperature and predicted temperature; When the residual exceeds the preset threshold, the input parameters of the prediction model at the next moment are corrected using the error integral term to eliminate steady-state errors caused by model mismatch or random disturbances.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention delves into the inherent thermodynamic time-delay and decay characteristics (i.e., thermal inertia) of the building itself, using them as the core adjustment object for demand-side response. By establishing a refined representation model based on Physical Information Neural Network (PINN), this invention decouples and reconstructs the high specific heat capacity physical properties of the building envelope (such as walls and floors) from the volumetric thermal effect of the indoor air space. Thus, within the dynamic constraints of meeting human thermal comfort (PMV-PPD index), the instantaneous thermal demand of the building is transformed into a virtual energy storage (VES) resource with "high physical consistency, strong predictability, and deep scheduling potential".

[0017] By implicitly coupling the variable frequency output of the HVAC system with the building's thermal state equation, this invention achieves non-intrusive, flexible migration of building electrical load along the time axis. This "heat-for-electricity" scheduling mode essentially utilizes the energy storage capacity of the building's physical entities to replace expensive electrochemical energy storage hardware, achieving energy spatiotemporal offsetting while eliminating the attenuation losses and safety hazards of physical energy storage devices. Building upon this, and addressing industry pain points such as nonlinear surges in reactive power, power factor (PF) degradation, and resulting power regulation penalties caused by variable frequency drive air conditioning equipment during low load rates or frequency adjustment, this invention proposes an "electricity-heat" cross-domain collaborative constraint algorithm.

[0018] Unlike traditional passive end-of-pipe reactive power compensation schemes, this invention innovatively embeds the instantaneous power factor at the grid connection point as a hard physical constraint into the scheduling decision logic. Utilizing the Monte Carlo Tree Search (MCTS) algorithm, it performs heuristic optimization within a massive discrete control space, proactively constraining the operating frequency and start-up / shutdown sequence of the HVAC system from the source. This mechanism ensures that while pursuing energy-saving goals, the system can predict and avoid power factor drop ranges, enabling deep integration and synergy between high-precision temperature control, maximized energy saving, and power quality assurance within a unified rolling optimization framework (MPC).

[0019] Through the aforementioned multiphysics coupling and high-order heuristic control methods, this invention aims to achieve the following multidimensional technical objectives: Maximizing economic benefits: Without increasing investment in additional physical energy storage hardware, by precisely controlling the "charging and discharging" of thermal inertia, the operating electricity cost of buildings under time-of-use pricing can be significantly reduced, and the power factor regulation penalty caused by power factor default can be completely avoided, thus achieving Pareto optimality of overall operating cost.

[0020] System robustness: By utilizing the physical constraint characteristics of the PINN model, the system ensures that scheduling commands still conform to the basic laws of thermodynamics under extreme weather or complex disturbances, thus guaranteeing the long-term stable operation of the HVAC system.

[0021] Grid friendliness and compliance: Enhance the ability of buildings to interact with the grid as "flexible loads", reduce reactive power pressure on the distribution network by actively maintaining a high power factor, and enhance the economic compliance of building energy systems and the accuracy of load-side peak-shaving response. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0023] Figure 1 This is an overall flowchart of the building thermal inertia collaborative scheduling method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the system architecture and data interaction logic provided in the embodiments of the present invention; Figure 3 This is a flowchart of the multi-protocol digital twin interface data synchronization and alignment provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the building thermal model construction and order reduction process provided in the embodiments of the present invention; Figure 5This is a structural diagram of the Physical Information Neural Network (PINN) prediction model provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the relationship between load rate and power factor in a variable frequency HVAC system provided in an embodiment of the present invention; Figure 7 This is a three-dimensional surface diagram of the PQ nonlinear characteristic mapping of an HVAC system provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the scheduling decision generation process under constraints provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the logic for coordinated control of power quality constraints and building thermal inertia provided in an embodiment of the present invention; Figure 10 This is a heuristic scheduling optimization logic diagram based on MCTS provided in the embodiments of the present invention; Figure 11 This is a schematic diagram of the rolling time-domain scheduling optimization process provided in an embodiment of the present invention.

[0024] The following is an example: Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The overall execution flow of this method is as follows: Figure 1 As shown; the overall system architecture and data interaction logic provided in this embodiment of the invention are as follows: Figure 2 As shown.

[0027] Example 1: Building Thermal Inertia Coordinated Scheduling (General Process) This invention provides a building thermal inertia cooperative scheduling method, comprising: Step 1: Integration of Multi-Source Heterogeneous Data Acquisition and Synchronous Communication Layer The specific implementation method of the above steps is as follows: Step S11: Perform high-frequency power quality sensing and transient electrical characteristic extraction. In terms of hardware assembly in this process, the system first installs and deploys industrial-grade smart power meters with high sampling rates at the entrance of the building's main power distribution cabinet and on the core feeders of the variable frequency HVAC branch circuit. These meters integrate high-precision current transformers (CTs) and voltage transformers (PTs) sampling modules, and are equipped with a dedicated digital signal processor (DSP) and an ultra-low-power microcontroller (MCU).

[0028] Its specific sampling logic is as follows: the instrument uses its internal high-speed analog-to-digital converter (ADC) to sample the instantaneous voltage waveform at the input side of the frequency converter. and instantaneous current waveform Real-time pulse sampling is performed. The invention is characterized by setting the instrument's physical data sampling frequency to... to This ensures that harmonic noise generated by the high-frequency switching of the variable frequency compressor due to the insulated-gate bipolar transistor (IGBT) is captured. At the communication reporting level, a data reporting cycle is set. The physical significance of this high-frequency reporting mechanism lies in its ability to fully capture the variable frequency compressor's execution of scheduling commands (such as frequency changes from...). Step to The transient reactive power fluctuations on the order of seconds generated during the process.

[0029] The instrument's internal processing unit performs a Fast Fourier Transform (FFT) on the original sampled sequence based on instantaneous power theory to calculate the amplitude and phase information of the fundamental wave and each harmonic. This is achieved by calculating the voltage fundamental vector. With the fundamental current vector Phase difference angle between Real-time mapping of instantaneous active power Instantaneous reactive power The specific instantaneous power factor The formula for the calculation function is shown below: This data is transmitted in real time to the edge computing gateway via RS485 bus using the Modbus-RTU protocol or via industrial Ethernet using the Modbus-TCP protocol. This instantaneous... The value will serve as the baseline feedback value for the "power factor hard constraint operator" in subsequent scheduling optimization, used to quantitatively evaluate the performance of the inverter air conditioner under low load conditions (i.e., load factor). The risk of incurring power regulation fee penalties due to a decrease in displacement power factor.

[0030] Step S12: Perform a space thermal field topology sensing and initial optimization simulation based on computational fluid dynamics (CFD).

[0031] 1) Perform space thermal field topology sensing and initial optimization simulation based on computational fluid dynamics (CFD). This process aims to establish a "physical truth database" of the thermodynamic evolution of indoor spaces, overcoming the technical deficiency in traditional environmental monitoring where "single-point sensor readings cannot accurately characterize the average state of large spaces." The specific implementation process includes the following four refined tasks: Constructing a 3D digital twin geometric model: Using high-precision CAD tools, create a 1:1 3D geometric model of the controlled space (such as a large open-plan office, atrium, or chiller room) based on the building's as-built drawings. The model needs to clearly define the boundaries of the building envelope (exterior walls, windows, interior walls, and floors), the distribution of internal heat sources (power density of lighting fixtures, distribution of staff workstations, and heat dissipation of office equipment), and the physical interface coordinates of the HVAC system (including the grille size of the air supply vents, the location of the return air vents, and the installation coordinates of the inverter indoor units).

[0032] 2) Perform high-quality mesh generation and boundary layer configuration: The geometric space is discretized using a mesh generation algorithm. Prism-layer meshes with geometrically increasing thickness are configured for the inner surface of the building envelope (windward side of the wall). Based on the fluid dynamics y+ criterion, the height of the first mesh layer is set to ensure that the flow characteristics at the wall surface under typical supply air velocities can be fully analyzed, thereby accurately calculating the convective heat transfer coefficient h of the building envelope surface.

[0033] 3) Configure constraints for unsteady physics field calculations: Enable the energy equation and turbulence model in a CFD solver (such as ANSYS Fluent). Given the combined effects of forced convection from air conditioning and natural convection due to thermal buoyancy indoors, the Reynolds-averaged Navier-Stokes (RANS) equations are used to describe the fluid motion. The fluid motion must satisfy the laws of conservation of mass (continuity equation) and momentum.

[0034] To achieve flow field closure, the system adopts standard Turbulence model, by solving turbulent kinetic energy Transport equations and dissipation rates The transport equation characterizes the heat diffusion properties within space. Its core transport function is shown in the following equation: In the formula, The turbulent viscosity coefficient, air density, This represents the velocity vector. By performing unsteady-state simulations on multiple combinations of control variables (such as different frequency converters and different outdoor temperatures), a high-fidelity temperature field time-series snapshot dataset is output. The aforementioned building thermal dynamics prediction model incorporates the partial differential equation of heat conduction in the building envelope as a physical constraint in the model training process to ensure that the prediction conforms to the thermodynamic energy conservation relationship.

[0035] Step S13: Execute the initial optimization algorithm for the coordinate position of the sensor measurement point (based on the POD-DEIM architecture).

[0036] This process utilizes high-dimensional ground truth data generated by CFD to identify key observation nodes in space that contribute the most to thermodynamic characteristics through mathematical dimensionality reduction. The specific algorithm execution steps are as follows: 1) Construct a high-dimensional snapshot matrix: The CFD simulation generated Group full-field grid node temperature data (usually includes) to (each dimension) is arranged to form a snapshot matrix .

[0037] 2) Perform modal feature extraction: Using the Proper Orthogonal Decomposition (POD) technique to analyze the matrix Singular value decomposition (SVD) is performed. The system extracts the principal mode orthogonal basis of the spatial temperature distribution. By calculating the energy proportion of eigenvalues, the preceding values ​​are truncated. First mode (preferably satisfying the cumulative energy variance) This enables the mapping of redundant flow field information to a low-dimensional principal feature subspace.

[0038] 3) Perform observation node selection: The Discrete Empirical Interpolation Method (DEIM) or a greedy optimization operator based on QR decomposition is employed. Among all candidate grid nodes, recursive search enables the observation matrix to... The set of indices with the smallest condition number (ConditionNumber).

[0039] Its physical function is to minimize the condition number, ensuring the robustness of sensor readings to the reconstruction of the spatial temperature field. The system generates the physical installation coordinates of the physical sensor based on the index values ​​output by the algorithm. In actual engineering assembly, the sensor height is set at... (Seated breathing belt) to Within the characteristic plane of the (standing breathing zone), the horizontal coordinates must avoid the dead corners of the variable frequency air conditioner's air supply and the center of local high-power heat sources.

[0040] Step S14: Construct the normalized correlation mapping operator and the multi-source data correction unit.

[0041] This process is responsible for the actual physical installation. Local discrete readings of individual sensors Real-time conversion into the "indoor volume average operating temperature" required by the scheduling algorithm The specific logic is as follows: (1) Perform sensor reading preprocessing: In order to eliminate the errors of different sensor ranges and the numerical jumps caused by position differences, perform real-time readings of each measuring point. Min-Max Normalization is performed. Its normalization function is shown below: In the formula, and This represents the extreme value of temperature evolution at this measuring point within a typical historical meteorological cycle.

[0042] (2) Establishing a linear weighted mapping operator: Using the CFD spatial average temperature obtained in step S12 as the monitoring signal, a normalized sensor reading vector is constructed through multidimensional linear regression. To the target operating temperature The mapping operator. Its mapping equation is shown below: In the formula, the weighted coefficient vector The bias term represents the weighting coefficient of each physical monitoring location's contribution to the overall thermal state characterization of the building. This is used to compensate for steady-state measurement biases in the system. The mapping operator transforms the limitations of local perception into global physical consistency, providing globally representative and realistic feedback for subsequent PINN models.

[0043] Step S15: Perform heterogeneous protocol middleware integration and timestamp synchronization (PTP implementation) like Figure 3 As shown, this invention achieves data synchronization and alignment between frequency converters, environmental sensors, and power meters through a multi-protocol digital twin interface. The edge computing gateway, as the integrated core of the communication layer, is responsible for handling the alignment of data streams with different baud rates and protocol stacks.

[0044] 1) Multi-protocol middleware configuration: The gateway is internally configured with a Modbus bus driver, a BACnet / IP client, an OPC-UA server, and an MQTT message queue dispatcher. The system uses a unified data modeling description file (DeviceTemplate) to map inverter frequency, motor current, ambient temperature, and grid connection point power factor to a unified logical address space.

[0045] 2) Perform high-precision timestamp alignment: To ensure an accurate causal relationship between the instantaneous characteristics (second-level) of the power system and the inertial characteristics (minute-level) of the thermal system, the system adopts the IEEE 1588 Precision Time Protocol (PTP). The edge gateway acts as the grandmaster clock, synchronizing the clocks of all devices within the sensing network equipped with the PTP protocol stack, achieving sub-microsecond synchronization accuracy.

[0046] 3) Construct a consistent feature tensor: The gateway allocates a circular queue buffer in memory to correlate and tag the power flow from S11 with the heat flow from S14. For data with inconsistent frequencies (e.g., power meters update at 1Hz, temperature sensors update at 0.1Hz), the system uses a zero-order hold (ZOH) or linear interpolation operator for resampling. The final generated data contains... The feature tensor flow serves as the core input dataset for model construction and collaborative scheduling decision layer.

[0047] Through the aforementioned integration process, this invention achieves a transformation from "physical quantity acquisition" to "physical meaning association" at the communication sensing layer, providing reliable data support for the subsequent joint operation of the Physical Information Neural Network (PINN) and Monte Carlo Tree Search (MCTS). Compared to traditional methods, this hierarchical integration not only significantly improves the spatial representativeness of the data but also eliminates the risk of dynamic mismatch errors in the frequency conversion regulation process through electro-thermal spatiotemporal alignment.

[0048] Step 2: Initialization of Physical Information Augmentation (PINN) Modeling Unit Parameters and Multiphysics Coupling Calibration like Figure 4 As shown, this step is responsible for deeply coupling the static physical properties of the building entity with the dynamic heat transfer mechanism to construct a reduced-order state-space model with physical consistency. The specific workflow, geometric analytical algorithm, numerical calibration logic, and neural network construction details are described below: Step S21: Geometric-Material Semantic Analysis Based on BIM and CAD The system uses the geometry analysis module (preferably based on the Python-based ifcopenshell library interface) to directly read the original BIM files in IndustryFoundationClasses (IFC) format.

[0049] The parsing logic includes: Using the IfcRelSpaceBoundary entity as the association core, the physical contact relationships between the indoor controlled space (IfcSpace) and the surrounding enclosing structure entities (such as IfcWall, IfcSlab, IfcWindow) are recursively traversed.

[0050] In this process, the system is characterized by identifying the normal vector of each thermal boundary surface through geometric algorithms. The algorithm determines whether a component is an exterior wall (in contact with the atmosphere), an interior wall (in contact with adjacent controlled or uncontrolled rooms), or a floor level. It then uses B-Rep (boundary representation) to extract the effective heat transfer area of ​​each component entity. The area As the core geometric weight parameter for subsequent thermal resistance and heat capacity calculations, it solves the statistical errors in engineering quantities caused by complex geometric shapes in traditional manual modeling.

[0051] Step S22: Perform semantic retrieval and parameter extraction of material constitutive properties.

[0052] The system follows the physical hierarchy path of the IFC model, tracing back to the IfcMaterialLayerSetUsage entity, and extracts the material composition of each physical layer of the enclosure structure. For each layer's material... The system automatically retrieves its constitutive property tensor, including: material density. Specific heat capacity at constant pressure and thermal conductivity Simultaneously, the geometric thickness of each material layer is extracted. .

[0053] For each identified enclosure structure entity (such as an IfcWall), the system performs a deep search along the following standard IFC data path to accurately obtain the physical properties of its multi-layered material construction: Path: IfcWall → IfcRelAssociatesMaterial → IfcMaterialLayerSetUsage → IfcMaterialLayerSet → Traverse IfcMaterialLayer.

[0054] Parameter extraction: For each IfcMaterialLayer, the system retrieves a predefined property set (IfcPropertySet) from its associated IfcMaterial entity and precisely extracts the following key thermal property parameters: Thermal conductivity, denoted by λ, has the unit of W / (m·K).

[0055] Density, denoted by ρ, is measured in kg / m³.

[0056] Specific heat capacity (c) is measured in J / (kg·K).

[0057] Geometric parameter extraction: At the same time, the thickness of the material layer (LayerThickness) is directly obtained from IfcMaterialLayer, denoted as L, in meters.

[0058] Step S23: Perform physical parameter initialization conversion using thermal resistance and thermal capacity calculation operators.

[0059] The system transforms the analytically obtained discrete material properties into initial values ​​for a second-order resistive-capacitive (RC) equivalent thermal parameter model. Total equivalent thermal resistance per unit area. The calculation mapping process is shown in the following formula: In the formula, and The surface convection heat transfer coefficient is preset based on indoor and outdoor convection heat transfer conditions. Structural equivalent heat capacity. The calculation mapping process is shown in the following formula: In the formula, This refers to weighting coefficients assigned based on material layer location. These coefficients characterize the inconsistency in thermal response between the inner and outer surfaces of the building envelope due to its physical thickness. This step transforms static CAD / BIM information into lumped parameters with dynamic physical meaning. and This provides the initial state with physical priors for subsequent neural networks.

[0060] Step S24: Establish a high-fidelity unsteady-state heat transfer simulation model.

[0061] This process loads the geometric boundaries obtained in step S21 into the finite element simulation environment. This section uses ANSYS Thermal / Fluent as an example for illustration.

[0062] Its characteristic is that a transient Neumann boundary condition (Transient Neumann BC) evolving over time is applied to the outer surface of the enclosure structure to simulate outdoor weather fluctuations during actual operation. This boundary condition is defined by the following equation: In the formula, The imported temperature time series data is typical for a meteorological year. The item is used to account for the significant contribution of nighttime longwave radiation cooling to building thermal inertia.

[0063] Step S25: Perform thermal resistance With heat capacity High-precision numerical calibration.

[0064] The system utilizes the indoor temperature response sequence output from the aforementioned finite element simulation. As a reference for the "physical truth value", the Levenberg-Marquardt (LM) nonlinear least squares iterative algorithm is used to... and Starting from the search point, minimize the predicted value of the simplified second-order RC model. Sum of squared residuals between the actual values ​​and the simulation values The optimization function is shown in the following formula: This step involves dynamically correcting the static parameters of the BIM analysis using large-scale simulation data, ensuring that the model can reflect complex physical effects that are not explicitly labeled in the IFC model, such as material moisture and thermal bridging.

[0065] Step S26: Construct a linear time-invariant (LTI) reduced-order model (ROM).

[0066] Singular value decomposition (SVD) is used to truncate the high-dimensional simulation state matrix and extract the principal energy modes. The system then projects the original high-dimensional physical field equations onto a low-dimensional principal eigenspace, constructing a state-space model of the following form: In the formula, It contains only a few core modes that characterize the thermal properties of the entire field, thereby compressing the computation time required for simulation from several hours to the millisecond level, providing an efficient operator for MCTS pre-simulation in real-time scheduling.

[0067] Step S27: Perform topology construction and parameter configuration of the Physical Information Neural Network (PINN).

[0068] The PINN prediction engine constructed in this embodiment adopts a deep fully connected architecture, and the topology of the prediction engine is as follows: Figure 5 As shown, embedding a physical regularization operator ensures that temperature prediction conforms to the law of conservation of energy. The specific components include: 1) Input layer: Receives a 4-dimensional feature tensor, including a time index. Outdoor ambient temperature HVAC operating power and the virtual energy storage state of charge of the previous control cycle .

[0069] 2) Hidden Layers: Configured with a depth of 4 to 6 layers. The number of neurons in each layer is set between 50 and 128. In the feature extraction path, each layer is configured with a nonlinear activation operator. A key feature of this invention is the forced use of the hyperbolic tangent function tanh as the activation function. The physical function of choosing tanh is that its second derivative has analytical continuity, which directly determines whether the system can automatically calculate the second derivative term of the partial differential equation (PDE) residual. If piecewise linear activation functions such as ReLU are used, the gradient of the physical loss term will vanish at singularities, failing to effectively embed the energy conservation constraint.

[0070] 3) Output layer: Outputs the predicted indoor volume average operating temperature. Step S28: Construct a hybrid loss function and physical mechanism regularization unit. The system defines the total loss function of PINN. The weighted sum of the data-driven terms and the physical mechanism terms is shown in the following formula: in, Used to ensure the model's accuracy in fitting measured or simulated historical data. The core function lies in the physical residual term. The system constructs a regularization constraint for neural network weights by forcibly transforming the law of energy conservation into a regularization constraint for the neural network weights using the following formula: In the formula, and This is derived from the precise value calibrated in step S25. The physical meaning of this loss term is that it does not depend on the labeled data, but rather evaluates whether the slope of the predicted temperature conforms to the thermodynamic differential equation across the entire time domain coordinates. If the predicted curve violates energy balance (i.e., the left and right sides of the equation are not equal), a large loss penalty is incurred. Step S29: Execute the model training process based on automatic differentiation (AD) technology. The system utilizes the chain rule engine built into the deep learning framework (preferably PyTorch) to perform automatic differentiation during backpropagation.

[0071] Its technical implementation logic is as follows: The system directly calculates the prediction scalar from the neural network. Starting from, extract its relationship to the input variables exact derivative This method replaces the traditional finite difference method (FDM) and avoids numerical instability caused by improper grid step size selection.

[0072] This "prior knowledge embedding" training mode enables the PINN model to provide predictions consistent with thermodynamic logic even in extreme weather scenarios with extremely sparse or even missing samples, thanks to the self-sustaining effect of physical laws. The model was subsequently encapsulated as an "environmental state transition function" in the MCTS optimization layer, achieving highly robust simulation of indoor thermodynamic evolution under long-term, multi-action combinations.

[0073] Step 3: Integration of HVAC inverter drive electrical characteristics and PQ dynamic response mapping Power factor variation characteristics of variable frequency HVAC systems under different load rates, such as Figure 6 As shown; the nonlinear mapping relationship between its active power, reactive power and operating status is as follows: Figure 11 As shown.

[0074] This step is responsible for constructing an electrochemical-electromagnetic coupling characteristic image of the variable frequency HVAC system. Through digital twin technology and multidimensional tensor calibration, a precise analytical mapping relationship is established between the operating frequency and load rate of the HVAC equipment and the power quality indicators of the grid connection point.

[0075] Step S31: The system constructs a digital twin virtual image of the HVAC system through an edge computing gateway.

[0076] The variable frequency screw compressor image is obtained by retrieving the internal register address of the VFD (variable frequency drive) via RS485 bus to acquire the inverter bridge output frequency finv, DC bus voltage Udc, and effective current values ​​Irms for each phase. Its core logic lies in establishing a three-dimensional dynamic image of "speed-torque-power," ensuring that fluctuations in the compressor's mechanical load are reflected in real-time changes in electrical parameters.

[0077] Chilled water pump and cooling water pump mapping: The operating frequency fpump of the pump inverter and the pump head / flow characteristics are read via BACnet / IP protocol. The system uses the similarity criterion (Affinity Laws) to map the frequency to the pump shaft power Pshaft∝f3, and simultaneously collects the displacement power factor DPF on the pump side.

[0078] Cooling tower fan image: A linear mapping model integrating fan blade frequency and air-side heat exchange.

[0079] The gateway internally deploys a multi-protocol conversion engine to achieve transparent mapping from underlying proprietary protocols or Modbus-RTU protocols to upper-layer unified data standards (such as OPC-UA or MQTT). Its data synchronization feature is that the system is configured with a digital twin interface module with bidirectional communication capabilities. This module synchronously acquires the inverter's output frequency setpoint fset and actual operating frequency fact in fixed steps of no more than 500ms. Through this interface, scheduling layer instructions can be losslessly parsed into electrical settings for the execution layer, while simultaneously feeding back the mechanical load status to the electrical characteristic mapping unit in real time, ensuring that the time alignment error of the "electric-thermal" coupling calculation is controlled within milliseconds.

[0080] At the physical connection level, the gateway communicates with the controller of the variable frequency compressor via an RS485 bus or an industrial Ethernet interface. The gateway internally deploys a multi-protocol conversion engine to achieve transparent mapping from underlying proprietary protocols or Modbus-RTU protocols to upper-layer unified data standards (such as OPC-UA or MQTT).

[0081] Its data synchronization feature is that the system is equipped with a digital twin interface module with bidirectional communication capability, which synchronously acquires the inverter's output frequency setpoint in fixed step sizes. Actual operating frequency DC bus voltage And the effective value of the current on the inverter side.

[0082] Through this interface, the instructions of the scheduling layer can be parsed into the electrical settings of the execution layer without loss, while the mechanical load status of the HVAC system is fed back to the electrical characteristic imaging unit in real time, ensuring that the time alignment error of the "electric-thermal" coupling calculation is controlled within the millisecond level.

[0083] Step S32: Perform a full-condition electrical data acquisition test of the HVAC system.

[0084] To obtain the precise electromagnetic response of the frequency converter at different operating points, the system executes a controlled "frequency-power" scanning process. During the actual building commissioning phase, a series of step control commands are sent using the edge gateway to adjust the frequency setpoint of the variable frequency compressor. From the lowest operating frequency (e.g.) )by or The step size is gradually increased to the rated frequency (e.g.) ).

[0085] At each frequency step point, the system invokes the high-frequency power sensing module from step S11 to synchronously collect the three-phase active power under that steady-state condition. reactive power Total Harmonic Distortion Its key feature is that the system not only records steady-state values, but also records transient reactive power spikes during frequency switching via a high-speed cache. All collected raw electrical parameters... It is stored in the calibration database, where The normalized inverter load rate.

[0086] Step S33: Construct an active-reactive (PQ) nonlinear fitting model based on load rate.

[0087] As shown in Figure 7, the active power, reactive power, and operating status of a HVAC system exhibit a nonlinear mapping relationship, which describes the electrical characteristics of the system under different operating conditions. Therefore, a PQ dynamic response mapping model based on a deep neural network is constructed for this purpose. In a variable frequency HVAC system, its reactive power... The generation of reactive power mainly stems from two dimensions: the excitation requirements of the motor and the distortion reactive power generated in the rectification and filtering stages. This invention quantitatively characterizes this by establishing a nonlinear mapping operator. With active power The physical laws governing change. The system calls a multidimensional parameter fitting unit and constructs a multinomial regression algorithm. The analytic mapping function. Its specific mathematical expression is shown in the following formula: In the formula, the coefficients These are the fitting characteristic coefficients to be determined. The physical function of the fitting logic is explained as follows: 1) Quadratic terms This is used to characterize the capacitive reactive power contribution of the inverter bridge of a frequency converter due to the parasitic capacitance of the insulated gate bipolar transistor (IGBT) when handling different active power.

[0088] 2) Linear term Used to characterize the excitation loss and magnetic circuit saturation characteristics of compressor motors during variable load processes.

[0089] 3) √(s) Used to characterize low load rate The impact of the charging and discharging current of the large-capacity electrolytic capacitor on the DC-Link side on the total reactive power of the system input side.

[0090] The system determines the coefficient vector under specific physical assembly conditions through least-squares fitting. This minimizes the sum of squared residuals in the reactive power prediction.

[0091] Step S34: Construct a sensitivity correction unit for total harmonic distortion (THD) on power quality.

[0092] Considering the characteristic that the total power factor of a power system is affected by both displacement power factor (DisplacementPF) and distortion power factor (DistortionPF), the system utilizes the collected data... Data constructs correction operators. When the frequency converter operates in the low-frequency range (e.g., When switching noise occurs, it can cause a surge in harmonic content. The system can resolve this using the following formula: The effect is converted into the equivalent reactive power component: This step ensures that in the subsequent MCTS optimization process, the system's prediction of the power factor not only takes into account the fundamental phase difference, but also fully accounts for the waveform distortion loss caused by the operation of power electronic switches.

[0093] Step S35: Define and generate the power factor hard constraint operator.

[0094] This process transforms all the aforementioned electrical images into the underlying firewall for scheduling decisions. The system uses instantaneous active power and instantaneous total reactive power to define and establish the real-time power factor. The closed-form calculation expression is shown below: At the engineering implementation level, this invention sets this formula as a "hard constraint criterion" for coordinated scheduling. Its technical logic is as follows: the system presets a power factor assessment threshold for a grid connection point. (For example, the power supply bureau's power adjustment fee assessment standard is set at 0.92).

[0095] The specific way to invoke this constraint operator during system operation is as follows: When the MCTS optimization module previews a scheduling sequence, the system will automatically generate the predicted active power sequence for that sequence. Substitute into equations (7) and (9). If the calculated result is at any given time... instant If the result is not met, the dispatch action is determined to be an "electricity violation". This determination will trigger an exponential penalty term in the reward function, forcing the dispatch strategy to passively maintain the power factor within the safe range by adjusting the distribution of active power (such as pre-cooling and using thermal inertia to smooth power fluctuations) without adding any physical compensation hardware (such as SVG devices).

[0096] Step S36: Generate the multidimensional PQ property tensor index space.

[0097] To improve the efficiency of real-time decision-making, the system discretizes the fitted continuous function into a high-dimensional tensor index table (LookupTable). This table is indexed by frequency. Current indoor operating temperature Outdoor temperature Use the coordinate axes to store the corresponding reactive power loss and PF value.

[0098] During the MCTS simulation phase, the optimization operator does not need to re-execute complex nonlinear regression calculations, but instead performs fast table lookup addressing through this tensor space. This engineering optimization reduces the time overhead of single-step electrical verification to the microsecond level, enabling the system to support tens of thousands of path simulations with complex electrical constraints within a 15-minute rolling scheduling window. Through the above calibration and mapping process, this invention realizes the transformation of the HVAC system from a simple "thermal load" to an "electric-thermal coupled flexible load." This deep integration can not only predict the electricity cost reduction brought about by frequency conversion control, but also accurately warn of power quality risks under low load rates, providing a logical closed loop and data support for subsequent collaborative control of "using thermal inertia to offset power penalties."

[0099] For variable frequency heating, ventilation and air conditioning (HVAC) systems under specific operating conditions, especially when the inverter is at a low load rate ( When the displacement power factor (DPF) and distortion power factor (DistortionPF) are in the range of 0, a severe degradation phenomenon is commonly observed. This step performs depth trajectory modeling and sensitivity quantification analysis, and its specific implementation process is detailed below: Step S37: Construct the discrete state space of the power factor degradation trajectory.

[0100] The system first defines the output frequency of the frequency converter. DC bus ripple coefficient and grid voltage fluctuation A multidimensional feature vector space is constructed. Using historical full-condition data collected in step S32, the electrical transient envelope of the HVAC system during startup, shutdown, and load regulation is extracted. The system is characterized by identifying the power factor using the kernel density estimation (KDE) method. With load rate The critical region of rapid decline and collapse is known as the "deterioration singularity". The modeling process incorporates the nonlinear loss model of the power electronic switching transistor (IGBT) under high-frequency modulation, and at this point, the nonlinear jump trajectory of the total reactive power at the grid connection point in the low-power range is plotted.

[0101] Step S38: Perform frequency-power quality sensitivity analysis based on partial derivative operators.

[0102] To quantitatively assess the contribution weight of each branch unit (including but not limited to multiple variable frequency compressors, circulating water pumps, and cooling fans operating in parallel) to the overall power quality index at the grid connection point, a sensitivity matrix is ​​introduced into the system. For each scheduling cycle, the system calculates the real-time power factor at the grid connection point. For the Taiwan frequency converter operating frequency The partial derivatives are calculated and normalized to obtain the sensitivity coefficient. Its mathematical expression is shown in the following formula: This coefficient reflects the percentage fluctuation in power factor caused by a unit frequency change. If If the value is high, the unit is determined to be an "electrically sensitive source" under the current operating conditions.

[0103] Step S39: Generate a Pareto-optimal power factor safety boundary constraint base map.

[0104] Based on the aforementioned trajectory model and sensitivity distribution, the system constructs a two-dimensional discrete state space distribution map with "system operating energy consumption" as the horizontal axis and "power factor compensation margin" as the vertical axis. This base map divides the continuous operating frequency domain into three discrete states: "safe zone" (PF≥0.95), "warning zone" (0.92≤PF<0.95), and "exceeding limit forbidden zone" (PF<0.92).

[0105] Its technical effects are as follows: This constraint map serves as the underlying logic support for the subsequent Pareto optimization of the "energy saving-power quality" algorithm in the MCTS process. During path deduction by the optimization operator, if a step falls within the "boundary-crossing forbidden zone," the system will use a sensitivity coefficient... The system quickly locates and automatically corrects the frequency limiting command for the branch unit. This mechanism ensures that the system can offset the risk of power penalties solely through the thermal inertia of the building itself, without relying on external hardware such as static var compensators (SVG), thus achieving a technical closed loop from "passive management" to "active defense".

[0106] like Figure 3 As shown, this invention constructs a three-dimensional tensor image of "electricity-heat-environment" covering the entire operating condition range by nonlinear fitting of physical measured data.

[0107] Its innovative technological features are: Topological nonlinear visualization: 3D surfaces intuitively demonstrate the "collapse effect" of power factor (PF) under the coupling of low load ratio (LoadRatio < 30%) and extreme ambient temperature. This "topographical map" distribution pattern reveals electrical degradation singularities that traditional fixed-limit control cannot reach.

[0108] Pareto decision boundary: The dark ridges (PF=0.92) on the surface constitute the "hard constraint Pareto boundary" of the MCTS algorithm. In 3D space, this boundary dynamically shifts with temperature, reflecting the real-time dynamic alignment capability of this invention for power quality sensitivity.

[0109] Computational efficiency optimization: The surface is discretized into a high-dimensional tensor index table (LookupTable) before system operation. The MCTS optimization operator performs microsecond-level fast table lookup through this tensor space, transforming the solution of complex partial differential equations into simple spatial topology retrieval, thereby supporting tens of thousands of path pre-simulations with complex electrical constraints under limited computing power.

[0110] Step 4: Edge collaborative optimization scheduling decision based on MCTS This step is responsible for performing global optimal path optimization in the discrete control space within each scheduling cycle, based on the environmental state evolution trend predicted by the Physical Information Neural Network (PINN). This process aims to solve the nonlinear decision-making problem of variable frequency HVAC systems under complex electricity price signals and hard power factor constraints.

[0111] Step S41: The system deploys a heterogeneous computing platform with high-performance tensor operation capabilities (preferably using NVIDIA Jetson AGX Orin embedded development components) at the building edge control node.

[0112] The hardware environment includes a multi-core ARM CPU and a GPU unit equipped with TensorCores. The GPU unit is dedicated to performing parallel inference computations for the PINN prediction engine, while the CPU unit is responsible for maintaining the dynamic growth of the MCTS decision tree, node management, and action selection logic. Through this heterogeneous parallel architecture, the system can ensure the completion of over 10,000 path rehearsals lasting up to 6 hours within a 15-minute rolling scheduling window.

[0113] Step S42: Construct the state vector of the four-dimensional multivariable system .

[0114] The system abstracts the complex building energy efficiency environment into state points consisting of four core dimensions, and the state vector is defined as follows: The physical and engineering meanings of its components are as follows: 1) (Indoor operating temperature): The current average temperature of the space volume is calculated by the normalization operator in step S14.

[0115] 2) (Virtual energy storage state of charge): Equation (11) is used to characterize the percentage of “cold” or “heat” currently stored in the building’s thermal inertia.

[0116] In the formula, and This represents the preset comfort boundary. This item is used to assess the building's remaining potential to maintain comfort without turning on the air conditioning.

[0117] 3) (Power factor at grid connection point): The real-time feedback based on step S11 reflects the instantaneous contribution of the inverter's current load rate to the power grid quality.

[0118] 4) (Time-of-use electricity pricing): It represents the energy cost information released by the power grid at the current moment.

[0119] Step S43: Define the set of discrete control actions .

[0120] Because MCTS excels at handling discrete combinatorial optimization, the system discretizes the continuous active power regulation range of the variable frequency HVAC system into a series of action steps. .

[0121] Every movement This is directly mapped to a specific operating frequency setting of the variable frequency compressor. This discretization process avoids the problem that traditional gradient optimization algorithms are prone to getting trapped in local optima in non-convex, discontinuous electrical constraint spaces.

[0122] Step S44: Construct a multi-objective comprehensive evaluation reward function .

[0123] This invention introduces a collaborative control framework based on hard constraints for power quality, such as... Figure 8 As shown, the specific process for generating scheduling decisions is as follows: Figure 9 As shown. During the MCTS search process, the system needs a quantitative metric to evaluate the merits of a particular control path. Reward function. Designed as a weighted sum of three negative benefit terms, its goal is to achieve Maximize (i.e. minimize penalty).

[0124] 1) Calculation of energy consumption cost: ,in This represents the duration of a single-step search.

[0125] 2) Calculation of thermal comfort penalty: This study employs a squared loss, which aims to impose a strong recovery constraint on behavior that exceeds the comfort boundary at room temperature.

[0126] 3) Calculation of power factor exponent penalty term (core constraint): The system has a dedicated penalty operator for PF degradation at the grid connection point. The mapping relationship is shown in the following formula: when hour: ; when hour: .

[0127] In the formula, and The default penalty weighting coefficient and slope coefficient are defined. The exponential function design ensures that the penalty increases rapidly when the power factor drops slightly, and tends towards positive infinity when the power factor is severely exceeded (e.g., below 0.85). This operator functions as a "power safety firewall," forcing scheduling paths to avoid low-load operating ranges. The final total reward function is defined as follows: In the formula, The adjustment weights configured for users are used to find the optimal balance between "saving money", "comfort" and "power quality".

[0128] Step S45: Execute the MCTS four-step iterative process.

[0129] To achieve rapid solutions within a massive control space, the system employs the following approach: Figure 10 The Monte Carlo tree search logic shown employs heuristic optimization. The system searches for the optimal power sequence within the decision-making time domain over the next 6 hours through the following four iterative stages: 1) Selection Phase: The system starts from the root node at the current time and recursively selects child nodes using the Upper Confidence Interval (UCT) formula. The UCT value is calculated as follows: In the formula, This stage is for exploring constants. Its function is to balance "paths with known high returns (utilization)" with "paths that have not been fully explored (exploration)" to ensure search coverage.

[0130] 2) Expansion phase: When the algorithm traverses to a leaf node that has not yet been fully expanded, and the number of times that node has been visited reaches a preset threshold, the system will create a new valid child node for that node. Each child node represents a new air conditioning active power adjustment level action to be taken under the current thermal environment.

[0131] 3) Simulation / Rollout Phase: Starting with the newly expanded nodes, the system invokes the PINN prediction engine trained in step S27 as a virtual environment simulator. PINN extrapolates the indoor temperature evolution trajectory under this action sequence at a speed of up to seconds. At each step of the extrapolation process, the system synchronously invokes the PQ coupling model from step S32 to calculate the corresponding instantaneous power factor. .

[0132] If the calculation during the deduction process If the exponential penalty term in equation (12) is triggered, the cumulative reward of that path will be... It will be instantly lowered. This process achieves "physical constraint-guided path pre-selection", that is, filtering out non-compliant actions is completed during the policy generation process.

[0133] 4) Backpropagation phase: The system will accumulate the rewards obtained during the simulation phase. Backtrack along the selected path to the root node, synchronously updating the access count and average revenue value of all nodes along that path.

[0134] Step S46: Execute the active load shifting control strategy.

[0135] This step explains how the present invention achieves "heat-for-electricity" through a search tree. When MCTS detects that low-power operation will occur during future peak electricity price periods due to temperature drops, it triggers... When the limit is exceeded and a penalty is imposed, the algorithm will automatically look for alternative actions in the early path of the search tree (off-peak electricity price segment).

[0136] Specifically, this manifests as follows: the system increases the active power of the air conditioning system in advance during off-peak periods, causing the compressor to operate in a high-power, high-load range. At this time, by lowering the indoor operating temperature, inexpensive electrical energy is stored in the building envelope as cooling (increasing...). During the subsequent peak period, the system significantly reduces or even shuts down the air conditioning, relying on the "cooling" inertia of the walls to maintain indoor comfort and completely avoid the high-risk power quality zone under low inverter load.

[0137] This invention uses a collaborative modeling approach, combining the building’s thermal inertia regulation capability with the power quality constraints of the HVAC system, to regulate the load operating range during the scheduling process, thereby preventing the power factor at the grid connection point from entering an unfavorable range.

[0138] Through the aforementioned MCTS collaborative decision-making process, this invention implements a scheduling logic at the edge that combines physical realism and grid friendliness. It not only solves the multi-constraint coordination problem through the nonlinear search capability of MCTS, but also utilizes the rapid prediction characteristics of PINN to provide real-time feedback for the decision tree, thereby achieving the dual technical objectives of avoiding electricity penalties and optimizing energy efficiency for building thermal inertia resources at the source.

[0139] Step 5: Multivariate closed-loop execution, robustness verification, and system integration testing. This step, acting as the final execution and feedback hub of the entire system, is responsible for converting the high-order power commands output by the optimization decision layer into electrical parameters recognizable by the underlying hardware, and ensuring the system's operational stability in dynamic environments through a multi-level closed-loop correction mechanism. Figure 11 As shown, this embodiment employs a rolling time-domain scheduling execution method, executing only the currently optimal control instruction within each scheduling cycle, and dynamically updating the subsequent scheduling process based on the latest collected data to achieve closed-loop optimized control. Its specific workflow, instruction mapping logic, fault-tolerance strategy, and online identification algorithm are detailed below: Step S51: Execute the physical mapping of the optimal scheduling instruction.

[0140] The system is configured with an instruction conversion unit to receive the optimal active power target value for the next 15 minutes in the rolling time domain, obtained from step S46. Because the compressor and fan of a variable frequency air conditioning (HVAC) system are physically driven by a frequency converter. It is controlled, therefore the system must perform a mapping from the "energy domain" to the "electrical domain".

[0141] Its engineering implementation features are: the active power-frequency mapping function calibrated in system call step S32. For each controlled variable frequency unit Based on its current operating conditions (such as intake and exhaust pressure, ambient temperature), the corresponding frequency setpoint is calculated. This mapping operator is implemented through linear interpolation or multidimensional lookup table method to ensure that the active power command can be accurately converted into the compressor speed setting.

[0142] Step S52: Execute dynamic limiting and inverter protection constraints.

[0143] To prevent thermal shock to the inverter's power electronic components from sudden changes in dispatch commands and to suppress instantaneous current surges in the power grid caused by drastic frequency fluctuations, a dynamic limiting operator is embedded in the execution link. This operator ensures electrical safety by constraining the rate of change of frequency. The specific slope limiting logic is defined by the following formula: In the formula, The preset single-step frequency step threshold (preferably set to) to (per second). The function of this limiting strategy is to smooth the adjustment trajectory of active power, thereby reducing the risk of a sudden drop in the power factor at the grid connection point due to frequency abrupt changes at the source and ensuring robust execution at the electrical level.

[0144] Step S53: Perform fault injection and cascading reliability testing.

[0145] Before the system is officially put into operation, a full-link simulation verification must be performed. The specific implementation method is as follows: simulating and injecting typical engineering fault conditions in the edge control gateway, including: 1) Communication packet loss simulation: The sensor network connection is manually disconnected. At this time, the fault-tolerant unit must call on the physical self-sustaining capability of the Physical Information Neural Network (PINN) model. Using the heat conduction differential equation (PDE) embedded in the model, an "inertial deduction" based on energy conservation is performed during the period without measured feedback to replace the measured values ​​and maintain the continuity of the scheduling logic.

[0146] 2) Sensor drift deviation injection: Simulates indoor thermometer generation The system needs to verify whether the Extended Kalman Filter (EKF) module can identify changes in the statistical characteristics of observed noise through residual analysis and trigger alarms or automatic gain adjustment. This process, through this "stress test" mechanism, ensures the coordinated stability of the software algorithm and hardware actuators in a fully cascaded state.

[0147] Step S54: Construct an augmented state space model.

[0148] The system is configured with an identification unit based on Extended Kalman Filter (EKF). Unlike traditional state machines that only estimate temperature, this invention achieves simultaneous identification of parameters and states by incorporating some physical parameters into the state vector. The augmented state vector is defined as follows: See the following formula: In the formula, Indoor operating temperature, Temperature of the building envelope mass block. For the equivalent total thermal resistance, The equivalent heat capacity is given by the structure. The physical evolution logic of this model follows a nonlinear state transition equation. The evolution of the temperature term strictly follows the discretized form of the defined second-order differential equation; while the physical parameters... and The evolution is assumed to be a random walk process with random perturbations. This allows the filter to fine-tune these "constants" during the feedback process.

[0149] Step S55: Prediction phase of Extended Kalman Filter (EKF) In each discrete control cycle The system in the previous cycle Optimal state posterior estimation Starting from [a certain point], prior prediction is performed. Its characteristics are: State prediction: Estimating the state at the previous time step. and current HVAC system inputs Substitute into the system state transition function In the process, the prior state estimate for the current period is calculated. This function Described by a second-order RC equivalent thermal parameter model with regularization constraints by a Physical Information Neural Network (PINN), its discrete form can be expressed as: .

[0150] Covariance prediction: Simultaneously, update the prior estimation error covariance matrix. The calculation process is as follows: ,in State transition function exist Jacobian matrix at the location, This is the preset process noise covariance matrix. It is used to quantitatively characterize the uncertainty in system state prediction introduced by unmodeled dynamic disturbances (such as ventilation heat loss caused by people randomly opening doors and windows, and fluctuations in internal heat sources).

[0151] Step S56: Update and correction phase of extended Kalman filter (EKF).

[0152] The system acquires and normalizes the measured indoor operating temperature values ​​in real time from the data acquisition layer. As an observational input, prior predictions are corrected. The process is characterized by executing in the following order: Calculation of observation residuals: Calculate the observation residuals (or news). The calculation formula is as follows: in, This is the observation matrix. In this invention, since only indoor air temperature is considered... The temperature of the building envelope can be measured directly and continuously via a sensor network. and physical parameters , Since it cannot be directly observed, the observation matrix is... Designed as follows: The function of this matrix is ​​to extract information from the state vector. Extract the corresponding indoor temperature prediction value With the measured value Compare them.

[0153] Kalman gain calculation: based on the latest predicted covariance Calculate the Kalman gain matrix using the observation model. The calculation formula is as follows: In the formula, To observe the noise covariance matrix, which is a scalar in this embodiment, it is used to characterize the measurement noise level of the temperature sensor. Gain matrix. It determines the system's level of confidence in the current observation residuals and dynamically allocates its correction weights to each state variable (including observable and unobservable).

[0154] Posterior state update (correction): using the calculated Kalman gain and observation residuals For prior state estimation Perform optimal adjustments to obtain the current period. posterior state optimal estimation The updated formula is: This operation allows you to observe the residuals. The real-time information contained therein is backpropagated, thereby enabling the understanding of states that cannot be directly observed (such as...). ) and key physical parameters (such as , The estimated value is fine-tuned online and dynamically.

[0155] Posterior covariance update: Finally, update the posterior estimation error covariance matrix. This prepares for the forecasting steps in the next cycle.

[0156] Through the "prediction-update" recursive loop consisting of S55 and S56 described above, the system realizes the monitoring of building thermal parameters. , The mechanism enables the model to automatically track and compensate for model mismatch caused by time-varying building performance and environmental disturbances, thereby ensuring the accuracy, robustness, and adaptability of forward-looking scheduling decisions based on this model throughout the entire operational lifecycle.

[0157] Step S57: Real-time fine-tuning of physical parameters.

[0158] The engineering value of this step lies in establishing a parameter adaptive correction mechanism. When the filter continuously detects a systematic deviation between the measured indoor temperature and the model's predicted value—for example, if the measured temperature rise rate is consistently faster than the prediction—the Extended Kalman Filter (EKF) algorithm will automatically adjust the parameter estimates in the augmented state vector through the Kalman gain matrix Kk. The correction direction is manifested in reducing the estimated equivalent thermal resistance R and / or increasing the estimated equivalent heat capacity Cm. Mathematically, this dynamic process is equivalent to mapping and compensating for the time-varying physical properties of the building envelope (such as the increase in thermal conductivity due to material moisture) or the changes in thermal response caused by unknown disturbances (such as unconventional ventilation) into the lumped parameter model in real time.

[0159] The high-confidence parameters identified and updated online by EKF, including the corrected thermal resistance Rnew and heat capacity Cm,new, will immediately trigger a dual synchronous update of the system model: 1. Physical Information Neural Network (PINN) Model Update: Parameters Rnew and Cm,new are fed back to step S27 to update the corresponding physical constraint parameters in the PINN network. This essentially recalibrates the physical regularization term in the PINN loss function using online data flow, ensuring that its prediction basis is consistent with the current actual physical state of the building, thus improving the model's accuracy in subsequent predictions over the time domain.

[0160] 2. Optimize the decision model update: Simultaneously, the updated parameters are fed back to the Monte Carlo Tree Search (MCTS) path simulation environment in step S43. Based on this updated, more accurate building thermodynamics model, the MCTS algorithm will perform a new round of forward-looking rolling optimization simulations to generate an optimal scheduling strategy that matches the real-time thermal characteristics of the building.

[0161] Through the closed-loop execution process described above—"control command execution → multi-source state monitoring → EKF online parameter identification → synchronous update of prediction and decision-making models"—this invention constructs a self-sensing and self-correcting collaborative scheduling system. Without relying on additional expensive sensing devices, this system achieves high-precision, real-time tracking and dynamic modeling of the core state of building thermal inertia—a "virtual energy storage"—by deeply integrating real-time operational data with the inherent mechanisms of physical models.

[0162] Final Technical Results: This dynamic adaptive capability ensures that the core prediction model of the entire dispatching system maintains high reliability when facing gradual changes in building physical characteristics, uncertain disturbances in the external environment, and random fluctuations in internal loads. Based on this reliable model, MCTS optimization decisions can accurately predict the impact of different dispatching strategies on the power factor at the grid connection point, thus ensuring that the grid safety constraint (PF≥0.92) is not breached under any operating condition, achieving a closed-loop technology from static design modeling to dynamic adaptive optimization operation throughout the entire lifecycle. If the filter continuously detects that the measured indoor temperature rises faster than the predicted value, the EKF will automatically adjust the gain... Reduce thermal resistance in the state vector The value or increase in heat capacity The value of this parameter. This dynamic drift compensation essentially maps complex environmental disturbances (such as the increase in thermal conductivity caused by wall dampness) into online correction of thermal circuit parameters.

[0163] Corrected precise thermal resistance and heat capacity It will immediately feed back to step S27, update the loss function constraint terms of PINN, and feed back to step S43 as the new environmental parameters for the next round of MCTS path simulation.

[0164] Through the aforementioned multivariate closed-loop execution process of "command mapping - reliability verification - online adaptive correction," this invention achieves high-precision real-time tracking of building thermal inertia state by utilizing the residual correlation between data and physical models without adding physical measurement methods. This ensures that the system can maintain the grid connection point power factor even when facing unpredictable physical fluctuations. Within the safety red line, a technical closed loop from static modeling to dynamic adaptive control has been completed.

[0165] The working process of this invention: After the system is officially put into operation, this invention realizes the scheduling and control of virtual energy storage based on building thermal inertia through standardized operating procedures. During the operation preparation phase, the system initializes parameters based on the time-of-use pricing strategy issued by the power sector and the building's historical load data, and sets the indoor temperature comfort range and power factor constraint threshold.

[0166] After system startup, the predictive and rolling optimization control module performs load forecasting, electricity price analysis, and power factor trend assessment in each control cycle, and generates HVAC system operation scheduling instructions. The execution and feedback correction module adjusts the operating status of the air conditioning equipment according to the instructions, while simultaneously collecting indoor temperature and power factor data in real time. When a deviation is detected, the control module automatically corrects the strategy, forming a closed-loop operation.

[0167] 1. System Assembly and Hardware Deployment First, non-intrusive installation is carried out on the existing building automation system (BAS) and power distribution system. Smart meters with power factor measurement function are installed at the main distribution cabinet or key feeder circuits to collect active power, reactive power and power factor data in real time; in the control circuits of each main air conditioning unit, chiller unit or terminal equipment, operating power, load rate and start / stop status signals are connected through communication interfaces.

[0168] Secondly, temperature and humidity sensors should be deployed in typical hot zones of the building (such as office areas, commercial areas, or conference areas), preferably in areas with high human activity and away from direct airflow or sunlight, to ensure the representativeness of the collected data. These sensors are connected to the edge control gateway via wired bus or wireless means.

[0169] Secondly, an edge control gateway is installed in the low-voltage electrical room or a nearby control cabinet. This gateway integrates data acquisition, model calculation, and optimized control functions, and establishes a two-way communication link with the HVAC controller, smart meter, and sensors through standard industrial communication protocols. The gateway has the scheduling and control program of this invention pre-installed.

[0170] 2. Model initialization and parameter calibration After system assembly is completed, the model initialization phase begins. First, based on building structural data and historical operating data, initial estimates are made for parameters such as thermal resistance and heat capacity in the equivalent thermal parameter model. Then, under normal building operating conditions, a short-cycle system identification process is used to correct the model parameters by utilizing the relationship between indoor and outdoor temperature changes and HVAC operating power, ensuring the model accurately reflects the building's true thermal inertia characteristics.

[0171] Simultaneously, active and reactive power data of the HVAC system under different load rates are collected, and P-Q characteristic curves are fitted to form a power factor prediction model, which provides a basis for subsequent optimized scheduling.

[0172] 3. Scheduling, Operation, and Workflow After the system enters the formal operation phase, the scheduling control is executed cyclically according to a fixed time step. In each control cycle, the edge control gateway first collects the latest indoor and outdoor environmental data, electricity price information and power factor status, and then predicts the load and electricity price in the future forecast time domain.

[0173] Subsequently, the control system, aiming to minimize operating electricity costs and indoor thermal comfort deviations, solves for the optimal operating power sequence of the HVAC system while satisfying upper and lower limits of indoor temperature and lower limit of power factor. When the prediction results indicate that reducing HVAC active power during a certain period may lead to a decrease in power factor, the system will automatically adjust the scheduling strategy, such as pre-cooling during low-price periods or limiting the frequency reduction of some equipment, thereby preventing the power factor from exceeding the limit at the source.

[0174] After the solution is obtained, the system only issues the control command corresponding to the current cycle to the HVAC system for execution, while the remaining control commands are retained as a reference sequence. In the next control cycle, the system re-acquires data and continuously updates the model and optimization results to form a closed-loop control.

[0175] 4. Anomaly Handling and Feedback Correction Workflow During continuous system operation, the execution and feedback correction module compares the predicted results with the actual operating status in real time. The system automatically enters the correction process when any of the following situations occur: the indoor temperature deviation continues to exceed the set range; the deviation between the actual and predicted power factor values ​​exceeds the preset threshold; or there is a sudden change in the external environment or abnormal equipment operation.

[0176] The system eliminates the impact of deviations by re-identifying model parameters, adjusting prediction weights, or temporarily switching to a conservative operating strategy. After the anomaly is resolved, the system smoothly returns to the normal scheduling mode, avoiding the impact of sudden changes in operating status on buildings and equipment.

[0177] Example 2: Load Reduction Maximization Scheduling Operation for Commercial Complexes This example focuses on a commercial complex with a building area of ​​45,000 square meters. The response command was for emergency peak shaving during the summer evening rush hour (19:00-21:00). The system first profiled the equipment based on its features, and the classification results are shown in Table 1.

[0178] Table 1. Parameter Information of Controlled Equipment in Commercial Complexes 2. Scheduling strategy deduction The system executes the control strategy for objective 1 (maximizing load reduction): (1) SL load translation For EV charging stations, the system solves a nonlinear programming problem to redirect 80% of the charging demand to after 21:30 (off-peak hours), with the response power limited to 30kW. For electric boilers, the system utilizes the thermal inertia of the water tank to shift the heating period to 04:00-06:00 AM.

[0179] (2) IL load resection: The exterior landscape lighting and large screen will be forcibly shut down, while the basic guide lighting (15kW) will be retained.

[0180] (3) AL load regulation: The thermal environment of the atrium was predicted using the PINN model. Calculations showed that raising the air conditioning setpoint by 2°C at 19:00, utilizing the release of cold energy stored in the stone floor during the day, could keep the temperature rise in the activity area below 2.5m within 0.8°C. Based on this, the system reduced the unit load rate to 25% (112.5kW).

[0181] 3. Implementation Results Through the aforementioned scheduling, during the response period (19:00-21:00), the total load of the commercial complex was reduced from the baseline peak of 865 kW (210+120+85+450) to 277.5 kW. The load reduction rate reached 67.9%, effectively responding to the grid command without causing any substantial impact on commercial operations.

[0182] Example 3: Coordinated Dispatch of Economic Benefits and Power Quality for Office Buildings in the Industrial Park 1. Job Preparation and Constraint Definition This example uses an office building in a park equipped with an 1122 kW variable frequency centrifugal chiller unit as the object.

[0183] (1) Electricity price environment: RMB 1.2 / kWh during peak hours (12:00-14:00) and RMB 0.7 / kWh during off-peak hours.

[0184] (2) Core constraint: Power factor PF at grid connection point ≥ 0.90.

[0185] (3) Equipment parameters: see Table 2.

[0186] Table 2. Parameter Information of Core Equipment in the Park Office Building 1. Scheduling strategy deduction (Objective 2: Maximize economic benefits) 2. For the midday peak period of 12:00-14:00, the system compared two strategies: (1) Traditional strategy (focusing only on achievements): To save on electricity costs, the PID controller reduced the chiller frequency to 35Hz (approximately 30% load). Although the active power decreased, according to... Figure 7 The PQ surface curves, at which point the unit enters the "reactive power surge zone," causing the power factor (PF) to drop to 0.78. Consequences: This triggers a power regulation penalty of approximately 500 yuan / day, and causes voltage fluctuations on the grid side.

[0187] (2) The collaborative strategy of this invention (focusing on active power + reactive power): The MCTS algorithm identifies the aforementioned risks in the logic shown in Figure 9 and performs pruning. The system generates a "cold storage-energy release" strategy: A. Cold storage phase (10:00-12:00, flat electricity price): Control the chiller to run at full load (PF=0.96, highest efficiency), pre-cooling the office area to 23℃. B. Energy release phase (12:00-14:00, peak electricity price): The unit is shut down or maintained at 10% standby, utilizing the building's thermal inertia to maintain room temperature. Since there are no high-power inductive loads operating, PF remains above 0.92.

[0188] 3. Implementation Results: Implementation data shows: (1) Economic efficiency: By implementing peak shaving and valley filling, daily electricity costs were reduced by NT$2,400; due to full PF compliance, there were no penalties. Overall operating costs were reduced by 32.5%.

[0189] (2) Stability: During the load recovery phase at 14:00 (target 3), the power rise slope was controlled at 1.5 Hz / min using the rolling time domain control and dynamic limiting shown in Figure 11, and the secondary peak was successfully eliminated.

[0190] During the operation of the aforementioned office building, the system continuously monitored the temperature using the EKF module. One afternoon at 3:00 PM, due to a sudden increase in the number of people in the conference room, the measured temperature rise rate exceeded the PINN predicted value. EKF quickly identified the change in internal thermal gain parameters and corrected the model. The corrected MCTS algorithm automatically decided to activate the backup chiller in advance, successfully controlling the room temperature fluctuation within ±0.5℃, verifying the strong robustness of this invention in dynamic environments.

[0191] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

[0192] Example 4: Building Thermal Inertia Coordinated Scheduling Device The present invention also provides a building thermal inertia coordinated scheduling device, comprising: The first processing module is used to acquire in real time the indoor and outdoor environmental parameters of the target building, the operating electrical parameters and status parameters of the HVAC system, time-of-use electricity price data, and the real-time power factor of the building's grid connection point through the sensing network. The second processing module is used to construct the thermal inertia state space equation of the building, which maps the heat storage and heat release capacity of the building envelope in the temperature comfort range to virtual energy storage resources and defines the virtual energy storage state of charge (SOCves); at the same time, it constructs the active-reactive power nonlinear coupling characteristic model of the HVAC system and establishes the correlation mapping matrix between load rate and instantaneous power factor. The third processing module is used to construct a neural network prediction model that integrates physical prior constraints. It embeds the partial differential equation of heat conduction of the building envelope as a regularization constraint term into the loss function. It uses the PINN model to predict the indoor temperature evolution trend and virtual energy storage SOCves in the future time domain, ensuring that the prediction output conforms to the thermodynamic law of conservation of energy. The fourth processing module is used to minimize the total cost of operating electricity and power regulation electricity, and minimize the deviation of indoor thermal comfort as multiple objective functions. Under the premise of satisfying the upper and lower limits of indoor temperature, the lower limit of the power factor at the grid connection point is introduced as a hard constraint operator. At the same time, Monte Carlo tree search or improved model predictive control algorithm is used to solve the optimal active power scheduling sequence of the HVAC system in the future finite prediction time domain within each scheduling cycle. The fifth processing module is used to send the first control command of the optimal active power scheduling sequence to the actuator, and dynamically correct the model parameters in S2 using a feedback compensation algorithm based on the real-time collected temperature and power factor deviations, thereby completing the closed-loop control in the rolling time domain.

[0193] In one embodiment of the present invention, the fifth processing module is further configured to execute feedback correction logic: Real-time monitoring of the residual between actual indoor temperature and predicted temperature; When the residual exceeds the preset threshold, the input parameters of the prediction model at the next moment are corrected using the error integral term to eliminate steady-state errors caused by model mismatch or random disturbances.

[0194] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A building thermal inertia cooperative scheduling method, characterized in that, include: S1. Real-time acquisition of indoor and outdoor environmental parameters of the target building, operating electrical parameters and status parameters of the HVAC system, time-of-use electricity price data, and real-time power factor of the building's grid connection point through the sensing network; S2. Construct the thermal inertia state space equation of the building, and map the heat storage and heat release capacity of the building envelope in the temperature comfort range to virtual energy storage resources, and define the virtual energy storage state of charge (SOCves); at the same time, construct the active-reactive power nonlinear coupling characteristic model of the HVAC system, and establish the correlation mapping matrix between load rate and instantaneous power factor. S3. Construct a neural network prediction model integrating physical prior constraints, embedding the partial differential equation of heat conduction of the building envelope as a regularization constraint term into the loss function; use the PINN model to predict the indoor temperature evolution trend and virtual energy storage SOCves in the future time domain, ensuring that the prediction output conforms to the thermodynamic law of conservation of energy. S4. Taking the minimization of the total cost of operating electricity and power regulation electricity, and the minimization of indoor thermal comfort deviation as the multi-objective function, under the premise of satisfying the upper and lower limits of indoor temperature, the lower limit of the power factor at the grid connection point is introduced as a hard constraint operator to eliminate infeasible operating states; at the same time, Monte Carlo tree search or improved model predictive control algorithm is used to solve the optimal active power scheduling sequence of the HVAC system in the future finite prediction time domain within each scheduling cycle; S5. The first control command of the optimal active power scheduling sequence is sent to the actuator, and the model parameters in S2 are dynamically corrected using a feedback compensation algorithm based on the real-time temperature and power factor deviation, thus completing the closed-loop control in the rolling time domain.

2. The building thermal inertia cooperative scheduling method according to claim 1, characterized in that, In S2, the construction of the building thermal inertia virtual energy storage model includes: Define virtual energy storage state of charge The normalized relative position of the current measured indoor temperature within the preset temperature comfort range is defined as: the process of increasing the cooling or heating power of the HVAC system to store energy in the building structure is defined as virtual energy storage charging; the process of reducing the power of the HVAC system to maintain room temperature using residual heat or cold from the building is defined as virtual energy storage discharging; a differential equation describing the thermal dynamic evolution process is established using the second-order equivalent thermal parameter ETP model, and the mapping relationship between virtual energy storage charging and discharging power and indoor temperature change rate is determined accordingly.

3. The building thermal inertia cooperative scheduling method according to claim 1, characterized in that, In step S2, the construction of the active-reactive power coupling characteristic model of the HVAC system includes: Based on historical operating samples, the power factor dynamic curves of the HVAC system under different load rates are fitted; reactive power is constructed. With active power nonlinear functional relationship ; The power factor constraint in step S4 is defined as follows: , in, This is the preset power factor assessment threshold.

4. The building thermal inertia cooperative scheduling method according to claim 1, characterized in that, The construction of the physical information neural network PINN in step S3 includes: The partial differential equations of building thermodynamics (PDEs) are embedded as residual terms in the loss function of a neural network; the physical loss term is defined. To predict the thermal equilibrium residuals of temperature with respect to time and space derivatives, the output of the neural network is constrained to conform to the law of conservation of energy by minimizing the comprehensive loss function containing physical residuals during training, thereby achieving physical robustness for predicting the state of charge of virtual energy storage under extreme weather conditions.

5. The building thermal inertia cooperative scheduling method according to claim 1, characterized in that, Step S5 further includes feedback correction logic: Real-time monitoring of the residual between actual indoor temperature and predicted temperature; When the residual exceeds the preset threshold, the input parameters of the prediction model at the next moment are corrected using the error integral term to eliminate steady-state errors caused by model mismatch or random disturbances.

6. A building thermal inertia coordinated scheduling device, characterized in that, include: The first processing module is used to acquire in real time the indoor and outdoor environmental parameters of the target building, the operating electrical parameters and status parameters of the HVAC system, time-of-use electricity price data, and the real-time power factor of the building's grid connection point through the sensing network. The second processing module is used to construct the thermal inertia state space equation of the building, which maps the heat storage and heat release capacity of the building envelope in the temperature comfort range to virtual energy storage resources and defines the virtual energy storage state of charge (SOCves); at the same time, it constructs the active-reactive power nonlinear coupling characteristic model of the HVAC system and establishes the correlation mapping matrix between load rate and instantaneous power factor. The third processing module is used to construct a neural network prediction model that integrates physical prior constraints. It embeds the partial differential equation of heat conduction of the building envelope as a regularization constraint term into the loss function. It uses the PINN model to predict the indoor temperature evolution trend and virtual energy storage SOCves in the future time domain, ensuring that the prediction output conforms to the thermodynamic law of conservation of energy. The fourth processing module is used to minimize the total cost of operating electricity and power regulation electricity, and minimize the deviation of indoor thermal comfort as multiple objective functions. Under the premise of satisfying the upper and lower limits of indoor temperature, the lower limit of the power factor at the grid connection point is introduced as a hard constraint operator. At the same time, Monte Carlo tree search or improved model predictive control algorithm is used to solve the optimal active power scheduling sequence of the HVAC system in the future finite prediction time domain within each scheduling cycle. The fifth processing module is used to send the first control command of the optimal active power scheduling sequence to the actuator, and dynamically correct the model parameters in S2 using a feedback compensation algorithm based on the real-time collected temperature and power factor deviations, thereby completing the closed-loop control in the rolling time domain.

7. The building thermal inertia coordinated scheduling device according to claim 6, characterized in that, The fifth processing module is also used to execute feedback correction logic: Real-time monitoring of the residual between actual indoor temperature and predicted temperature; When the residual exceeds the preset threshold, the input parameters of the prediction model at the next moment are corrected using the error integral term to eliminate steady-state errors caused by model mismatch or random disturbances.

Images