Control method and device for heating, ventilation and air conditioning system migrated across buildings

Abstract

The invention provides a control method and device for a heating ventilation air conditioning system capable of moving across buildings, and relates to the technical field of heating ventilation air conditioners. The method comprises the steps of collecting data of a reference building group and measured data of a target domain, constructing a building dynamic prediction model based on a Gaussian process dynamic model, performing meta-learning based on a reference building group data set, performing kernel learning on the building dynamic prediction model based on a target building group fine tuning data set and kernel initial hyper-parameters, and performing kernel learning on the building dynamic prediction model. And obtaining an optimized building dynamic prediction model, and establishing a multi-level security constraint optimization framework. According to the method, collaborative optimization of the multi-unit start-stop strategy and the water supply temperature set point of the HVAC system can be remarkably achieved, and the deployment efficiency of the algorithm is remarkably improved while safe operation of the system is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of heating, ventilation and air conditioning technology, and in particular to a control method and device for a heating, ventilation and air conditioning system that can be moved across buildings. Background Technology

[0002] Modern buildings are a major energy consumer globally, and HVAC systems are a core source of building energy consumption. Within HVAC systems, energy consumption of source-end equipment (such as air-source heat pumps) is particularly prominent, accounting for approximately 60%-70% of total HVAC system energy consumption. Therefore, optimizing the operation strategies of HVAC source-end equipment is a key technological entry point for achieving building energy conservation, emission reduction, and low-carbon transformation. However, achieving low-cost cross-building transfer of energy-saving control algorithms for HVAC systems still faces significant technical challenges. Currently, the most popular research methods are MPC and MFRL. MPC relies heavily on high-quality data and has poor cross-building transferability. MFRL suffers from low sample efficiency, often requiring weeks or even months of massive amounts of interactive data to converge. Furthermore, MFRL typically lacks the ability to quantify the uncertainty of model predictions, and is prone to attempting dangerous actions during the exploration process (such as causing frequent equipment start-ups and shutdowns or severely exceeding indoor temperature limits), leading to significant safety and stability risks. In addition, its strategy transferability across buildings is extremely poor. Neither MPC nor RL, the models or strategies developed for a specific building, can be effectively migrated to another new building. In summary, existing technologies cannot simultaneously solve the three intertwined key technical challenges of "cross-building portability," "data efficiency," and "security control." Therefore, there is an urgent need in this field for a novel technical solution to systematically overcome all of the aforementioned shortcomings. Summary of the Invention

[0003] To address the technical problem that existing optimized control models for HVAC systems struggle to simultaneously achieve the three objectives of "cross-building portability," "data efficiency," and "safety control," this invention provides a cross-building portability control method and apparatus for HVAC systems. The technical solution is as follows:

[0004] On the one hand, a method for controlling a HVAC system that can be moved across buildings is provided. This method is implemented by a control device for the HVAC system that can be moved across buildings, and the method includes: S1: Collect historical data of the reference building complex and measured data of the target domain, and preprocess the historical data of the reference building complex and the measured data of the target domain to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. S2: Construct a building dynamic prediction model based on the Gaussian process dynamic model, calculate the covariance matrix using the exact regression method and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. S3: Obtain initial hyperparameters of the kernel by performing meta-learning based on the training dataset of the reference building complex. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. S4: Based on the fine-tuned dataset of the target building complex and the initial kernel hyperparameters, perform kernel learning on the building dynamic prediction model to obtain the optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the fine-tuned dataset of the target building complex as the task-level loss and using gradient descent to optimize the kernel hyperparameters. S5: Multiple candidate future control trajectories are generated through Monte Carlo sampling, and a future state prediction sequence is obtained based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain multiple control sequences. S6: Establish a multi-level safety constraint optimization framework, perform hierarchical screening and verification on the future state prediction sequence, and obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

[0005] Preferably, step S1 involves collecting historical data of the reference building complex and measured data of the target domain, and preprocessing the historical data of the reference building complex and the measured data of the target domain to obtain a training dataset of the reference building complex and a fine-tuning dataset of the target building complex, including: S11: Select a reference building group and obtain data on the building HVAC system at different times in history to obtain historical data of the reference building group. The data of the building HVAC system includes environmental conditions, region-specific variable data and action vectors. The environmental conditions and region-specific variable data include indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, humidity, occupancy status, start-up and shutdown sequence of the unit at the previous moment, and regional air temperature. The action vector includes the supply water temperature setpoint. S12: Select the target domain's measured building complex and collect data on the building's HVAC system at different times to obtain historical data of the target domain's measured building complex; S13: Based on the historical data of the reference building complex, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the reference building complex training dataset. The output variable is the region state at the next moment. S14: Based on the measured historical data of the target building complex, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the target building complex fine-tuning dataset.

[0006] Preferably, the building dynamic prediction model based on the Gaussian process dynamic model in S2 is constructed by using the exact regression method to calculate the covariance matrix and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses a radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix, including: S21: Starting from the candidate function set, a Gaussian process prior is set, a kernel function is selected, and a dynamic model of the Gaussian process is obtained. The kernel function is a radial basis function kernel, which is used to measure input similarity. The hyperparameters of the kernel function consist of a scalar and a matrix, and contain a total of 101 real parameters. S22: Based on the Gaussian process dynamic model, a building dynamic prediction model is constructed. The building dynamic prediction model is used to take the characteristics of building HVAC as system input to obtain model output. The system input includes indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, supply water temperature set point, and the start-up and shutdown sequence of the unit at the previous moment. The model output is the energy consumption and indoor comfort of the building at the next moment. S23: Based on the training dataset of the reference building complex, the covariance matrix is ​​calculated using the exact regression method. The identity matrix and noise variance are introduced to characterize the uncertainty noise, and the Gaussian process posterior distribution is obtained. The Gaussian process posterior distribution is used to obtain the result that follows the Gaussian distribution when making predictions.

[0007] Preferably, in step S3, the initial hyperparameters of the kernel are obtained through meta-learning based on the training dataset of the reference building complex. This meta-learning includes initializing the kernel parameters of the building dynamic prediction model, including: S31: Construct meta-learning initialization task instructions based on the training dataset of the reference building complex; S32: Based on the meta-learning initialization task instruction, the kernel parameters of the building dynamic prediction model are initially optimized, the common thermal dynamic laws of multiple buildings are learned, and the kernel initial hyperparameters are obtained. S33: Initialize the building dynamic prediction model based on the kernel initial hyperparameters.

[0008] Preferably, in step S4, kernel learning is performed on the building dynamic prediction model based on the target building complex fine-tuning dataset and the initial kernel hyperparameters to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the target building complex fine-tuning dataset as the task-level loss and optimizing the kernel hyperparameters using gradient descent, including: S41: Based on the measured data of the target domain and the initial hyperparameters of the kernel, construct the task-level loss function and calculate the loss function. For each specific prediction task, use the mean square error calculated based on the measured data of the target domain as the task-level loss to obtain the loss value. S42: Based on the loss value, calculate the gradient of the kernel hyperparameter with respect to the loss to obtain the loss gradient; S43: Based on the initial kernel hyperparameters, loss gradient and learning rate, aggregate the model error, perform gradient descent, and after multiple iterations, obtain the optimized kernel hyperparameters that minimize the mean square error of the model prediction. S44: Based on the optimized kernel hyperparameters and the measured data of the target domain, the model is fitted to obtain the optimized building dynamic prediction model.

[0009] Preferably, in step S5, multiple candidate future control trajectories are generated through Monte Carlo sampling, and a future state prediction sequence is obtained based on the optimized building dynamic prediction model. These multiple candidate future control trajectories include various control sequences, including: S51: Based on GP, ​​the MBRL building dynamic environment is constructed, the state is initialized, and the initial control state is obtained; S52: Using the MPC multi-step prediction idea and Monte Carlo sampling method, parallel sampling is performed. Based on the initial control state, the trajectory dynamics are simulated through GP forward propagation to generate multiple control sequences and obtain K state trajectory sets. S53: Input the set of K state trajectories into the optimized building dynamic prediction model to obtain the future state prediction sequence.

[0010] Preferably, the establishment of a multi-level safety constraint optimization framework in step S6, which performs hierarchical screening and verification on the future state prediction sequence to obtain the optimal control sequence, includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism, including: S61: Calculate the uncertainty threshold based on the training dataset of the reference building complex to obtain the uncertainty threshold; S62: Based on the uncertainty threshold, implement a safety screening to eliminate trajectories whose prediction variance of the future state prediction sequence exceeds the threshold; S63: Based on the screening of the multi-objective comprehensive reward function, select the trajectory with the highest comprehensive reward. The multi-objective comprehensive reward function is used to balance comfort, energy consumption and uncertainty indicators. S64: Set minimum running time and minimum stopping time constraints to suppress unnecessary device switching. Before the control action is executed, the minimum running time and minimum stopping time constraints of the first control action of the trajectory with the highest comprehensive return are checked. If the control action violates the minimum time constraint, the start-stop state of the previous moment is forcibly maintained to obtain the optimal control sequence. The setting of minimum running time and minimum stopping time constraints is used to solve the sensitivity problem of frequent start-stop.

[0011] On the other hand, a control device for a HVAC system that can be moved across buildings is provided. This device is applied to a control method for an HVAC system that can be moved across buildings. The device includes: Data acquisition module: used to collect historical data of the reference building complex and measured data of the target domain, and to preprocess the historical data of the reference building complex and the measured data of the target domain to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. Gaussian Process Module: Used to construct a building dynamic prediction model based on the Gaussian process dynamic model. It uses the exact regression method to calculate the covariance matrix and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. Meta-learning module: used to perform meta-learning based on the reference building group training dataset to obtain the initial hyperparameters of the kernel. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. Kernel learning module: used to perform kernel learning on the building dynamic prediction model based on the target building group fine-tuning dataset and the initial kernel hyperparameters, to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the target building group fine-tuning dataset as the task-level loss and using gradient descent to optimize the kernel hyperparameters. Control trajectory module: used to generate multiple candidate future control trajectories through Monte Carlo sampling, and obtain future state prediction sequences based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain a variety of control sequences. Safety constraint module: used to establish a multi-level safety constraint optimization framework, perform hierarchical screening and verification on the future state prediction sequence, and obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

[0012] On the other hand, a control device for a HVAC system that can be moved across buildings is provided. The control device for a HVAC system that can be moved across buildings includes: a processor; and a memory storing computer-readable instructions. When the computer-readable instructions are executed by the processor, they implement the method described in any of the above-described methods for controlling a HVAC system that can be moved across buildings.

[0013] On the other hand, a computer-readable storage medium is provided, characterized in that the computer-readable storage medium stores program code, which can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention enables efficient cross-building migration capabilities of the model by learning the "common knowledge" of multi-building data under the condition of "lightweight energy-saving renovation"; achieves rapid adaptation under "few sample" data in new buildings (target domain); and guides safe and robust decision-making by explicitly quantifying the prediction uncertainty to avoid high-risk actions; ultimately realizing the coordinated optimization of the start-up and shutdown strategy of multiple HVAC units and the water supply temperature setpoint, significantly improving the deployment efficiency of the algorithm, the energy-saving effect of the building, and the indoor comfort while ensuring the safe operation of the system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a control method for a HVAC system that can be migrated across buildings, provided by an embodiment of the present invention. Figure 2 This is a flowchart of a method for establishing multi-level security constraints provided by an embodiment of the present invention; Figure 3 This is a block diagram of a HVAC system control device that can be moved across buildings, provided in an embodiment of the present invention. Figure 4 This is a structural schematic diagram of a HVAC system control device that can be moved across buildings, provided by an embodiment of the present invention.

[0017] List of key term definitions 1. MPC Model Predictive Control 2. HVAC (Heating, Ventilation and Air Conditioning) 3. Reinforcement Learning (RL) 4. MFRL (Model-Free Reinforcement Learning) 5. MBRL: Model-Based Reinforcement Learning 6. GP Gaussian Process 7. MPPI Model Predictive Path Integral Control Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] This invention provides a method for controlling a HVAC system that can be moved across buildings. This method can be implemented by a control device for the HVAC system that can be moved across buildings, which can be a terminal or a server. Figure 1 The flowchart shown illustrates a method for controlling a HVAC system that migrates across buildings. This method's processing flow may include the following steps:

[0024] Historical data of the reference building complex and measured data of the target domain are collected. The historical data of the reference building complex and the measured data of the target domain are preprocessed to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. Preferably, historical data of a reference building complex and measured data of the target domain are collected. Preprocessing of these two datasets yields a training dataset for the reference building complex and a fine-tuning dataset for the target building complex, including: Select a reference building group and obtain data on the building HVAC system at different times in history to obtain historical data of the reference building group. The data of the building HVAC system includes environmental conditions, region-specific variable data and action vectors. The environmental conditions and region-specific variable data include indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, humidity, occupancy status, start and stop sequence of the unit at the previous moment, and regional air temperature. The action vector includes the supply water temperature setpoint. By selecting a target area of ​​measured building groups and collecting data on the HVAC systems of the buildings at different times, historical data of the target area of ​​measured building groups are obtained. Based on historical data of the reference building complex, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the reference building complex training dataset. The output variables are the regional state at the next moment. Based on the historical data of the target building complex measured in the target domain, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is then combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the target building complex fine-tuning dataset.

[0025] In some embodiments, the system state includes environmental conditions (such as outdoor air temperature, humidity, and occupancy status) and region-specific variables (such as region air temperature), which are integrated into a state vector. The state vector is then linked to the action vector. (Including control parameters such as setpoints) are combined to form input variables. The output variable is the region state at the next time step. For simplicity and consistency, the input and output of each sample are denoted as follows: and Historical datasets are defined as .

[0026] A building dynamic prediction model is constructed based on the Gaussian process dynamic model. The covariance matrix is ​​calculated using the exact regression method to obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. Preferably, a building dynamic prediction model is constructed based on a Gaussian process dynamic model. The covariance matrix is ​​calculated using the exact regression method to obtain the posterior distribution of the Gaussian process. The Gaussian process dynamic model uses a radial basis function kernel as the prior distribution of the Gaussian process. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix, including: Starting from the candidate function set, a Gaussian process prior is set, and a kernel function is selected to obtain the dynamic model of the Gaussian process. The kernel function is a radial basis function kernel, which is used to measure input similarity. The hyperparameters of the kernel function consist of a scalar and a matrix, and contain a total of 101 real parameters. Based on the Gaussian process dynamic model, a building dynamic prediction model is constructed. The building dynamic prediction model is used to take the characteristics of building HVAC as system input to obtain model output. The system input includes indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, supply water temperature setpoint, and the start-up and shutdown sequence of the unit at the previous moment. The model output is the energy consumption and indoor comfort of the building at the next moment. Based on the training dataset of the reference building complex, the covariance matrix is ​​calculated using the exact regression method. The identity matrix and noise variance are introduced to characterize the uncertainty noise, and the Gaussian process posterior distribution is obtained. The Gaussian process posterior distribution is used to obtain results that follow a Gaussian distribution when making predictions.

[0027] In some embodiments, modeling using Gaussian processes involves two stages: first, starting from a set of candidate functions (i.e., Gaussian process priors), and then calculating the conditional probabilities (i.e., Gaussian process posteriors) based on the training data. These two stages will be described separately below.

[0028] First, we choose a function that can intuitively measure the similarity between two inputs, denoted as . This function is the kernel function of a Gaussian process. Its basic idea is that buildings that perform similar actions under similar conditions should also have similar next states. It's important to note that each variable in the input contributes differently to the similarity, depending on their role in the unknown dynamic function. The kernel function reflects this difference through a set of adjustable parameters (i.e., hyperparameters).

[0029] The MPC-RL method employs radial basis function kernels due to their powerful expressiveness in modeling complex systems such as building thermal dynamics. The hyperparameters of this kernel function are... ,in It is a scalar. It is The matrix contains a total of 101 real parameters. The choice of these parameters is crucial because they significantly affect the modeling accuracy and the precision of uncertainty estimation.

[0030] It should be noted that, when using the exact regression method, the covariance matrix is ​​first calculated. Then, the posterior GP is calculated using formulas (2) and (3): (1) (2) (3) in, As a feature, As another feature, For the input feature matrix, Represents the identity matrix. It is a scale parameter. It is a scale matrix. It is the covariance matrix. This represents the noise variance (default value is zero), used to characterize uncertain noise. Based on this posterior, for new input points... The output prediction follows a Gaussian distribution, as shown in Equation (4).

[0031] (4) in, It is a posterior probability distribution. It follows a normal distribution, and the predicted mean is... The variance is This means that given the data... and input In this case, the level of uncertainty associated with the model output. As an indicator of cognitive uncertainty, if The values ​​are higher in regions of sparse data in the input space, thus highlighting areas of greater uncertainty and providing guidance for decision-making under these conditions.

[0032] The initial hyperparameters of the kernel are obtained by meta-learning based on the training dataset of the reference building complex. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. Preferably, the initial hyperparameters of the kernel are obtained through meta-learning based on the training dataset of the reference building complex. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model, including: Based on the training dataset of the reference building complex, we construct meta-learning initialization task instructions; Based on the meta-learning initialization task instruction, the kernel parameters of the building dynamic prediction model are initially optimized, the common thermal dynamic laws of multiple buildings are learned, and the kernel initial hyperparameters are obtained. The building dynamic prediction model is initialized based on the kernel initial hyperparameters.

[0033] Based on the fine-tuned dataset of the target building complex and the initial kernel hyperparameters, kernel learning is performed on the building dynamic prediction model to obtain the optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the fine-tuned dataset of the target building complex as the task-level loss and using gradient descent to optimize the kernel hyperparameters. Preferably, based on the fine-tuned dataset of the target building complex and the initial kernel hyperparameters, kernel learning is performed on the building dynamic prediction model to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the fine-tuned dataset of the target building complex as the task-level loss and optimizing the kernel hyperparameters using gradient descent, including: Based on the measured data of the target domain and the initial hyperparameters of the kernel, a task-level loss function is constructed and the loss function is calculated. For each specific prediction task, the mean square error calculated based on the measured data of the target domain is used as the task-level loss to obtain the loss value. Based on the loss value, the gradient of the kernel hyperparameter with respect to the loss is calculated to obtain the loss gradient; Based on the initial kernel hyperparameters, loss gradient, and learning rate, the model error is aggregated, gradient descent is performed, and after multiple iterations, the optimized kernel hyperparameters that minimize the mean square error of the model prediction are obtained. Based on the optimized kernel hyperparameters and the measured data of the target domain, the model is fitted to obtain the optimized building dynamic prediction model.

[0034] In some embodiments, a customized kernel learning method is used to fine-tune the hyperparameters of the GP kernel. First, the task typically associated with Markov decision processes is redefined, focusing instead on modeling building data within a specific period. The loss function then becomes... In this equation, Representative task The input features (e.g., historical temperature readings, occupancy levels), and This represents the corresponding output target (e.g., future indoor temperature). The equation quantifies the optimized kernel hyperparameters after application. At that time, the difference between GP predictions and actual observations.

[0035] The essence of meta-learning methods lies in minimizing the aggregated model error of the task set. By implementing this strategy, the initial hyperparameter set of the kernel is systematically optimized, enabling it to more accurately adapt to the specific thermal dynamics of different buildings.

[0036] Multiple candidate future control trajectories are generated through Monte Carlo sampling, and a future state prediction sequence is obtained based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain a variety of control sequences. Preferably, multiple candidate future control trajectories are generated through Monte Carlo sampling, and a future state prediction sequence is obtained based on the optimized building dynamic prediction model. These multiple candidate future control trajectories include various control sequences, including: The MBRL building dynamic environment was constructed based on GP, ​​and the state was initialized to obtain the initial control state. Using the MPC multi-step prediction concept and Monte Carlo sampling method, parallel sampling is performed. Based on the initial control state, the trajectory dynamics are simulated through GP forward propagation to generate multiple control sequences and obtain K state trajectory sets. Input the set of K state trajectories into the optimized building dynamic prediction model to obtain the future state prediction sequence.

[0037] A multi-level safety constraint optimization framework is established to perform hierarchical screening and verification on the future state prediction sequence to obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

[0038] Preferably, such as Figure 2 As shown, a multi-level safety constraint optimization framework is established to perform hierarchical screening and verification on the predicted future state sequence to obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on constraint mechanisms, including: The uncertainty threshold is calculated based on the training dataset of the reference building complex. Based on the aforementioned uncertainty threshold, a safety screening is implemented to eliminate trajectories whose prediction variance of future state prediction sequences exceeds the threshold. Based on the screening of the multi-objective comprehensive reward function, the trajectory with the highest comprehensive reward is selected. The multi-objective comprehensive reward function is used to balance comfort, energy consumption and uncertainty indicators. Minimum running time and minimum stopping time constraints are set to suppress unnecessary device switching. Before the control action is executed, the minimum running time and minimum stopping time constraints of the first control action of the trajectory with the highest comprehensive return are checked. If the control action violates the minimum time constraint, the start and stop state of the previous moment is forcibly maintained to obtain the optimal control sequence. The setting of minimum running time and minimum stopping time constraints is used to solve the sensitivity problem of frequent start and stop.

[0039] In some embodiments, the multi-level safety constraint optimization framework is also called the Multi-Safety Constraint MPPI control module. This module, based on the multi-step prediction idea of ​​MPC and combined with Monte Carlo sampling and multi-trajectory reward evaluation methods, realizes the control of multi-source main units and set temperatures in building HVAC systems. Specifically, MPPI is a highly efficient online trajectory planning algorithm that generates [data / data] through Monte Carlo sampling. The invention provides a list of candidate future control trajectories. The core of this invention lies in the fact that these candidate trajectories must pass the screening and verification of the following three layers of security constraints before being executed.

[0040] First layer: Uncertainty screening constraint. The algorithm utilizes the prediction variance of the model output. As an uncertainty indicator, the algorithm first determines an optimal uncertainty labeling threshold offline based on historical data. During online control, the system will check... Each of the candidate trajectories in the future at every prediction step ( ,in The uncertainty value (for the prediction step size) is used. Based on this uncertainty value and the uncertainty threshold, candidate trajectories are selected from the M trajectories that satisfy the condition that the uncertainty value is less than the threshold at all prediction steps. If none of the trajectories meet the condition, a rule-based control strategy is automatically adopted as a safety net to prevent the system from going out of control.

[0041] The second layer: Multi-objective reward evaluation constraints. The "credible trajectories" selected in the first layer proceed to the second layer of evaluation. A multi-objective comprehensive reward function is constructed. This reward function includes not only traditional comfort and energy consumption rewards, but also an "uncertainty reward term."

[0042] The third layer: Unit start-up and shutdown safety switching constraints. This is the final and most critical physical safety check before control commands are executed. HVAC ASHP units (especially compressors) are extremely sensitive to frequent start-up and shutdown operations. The first control action corresponding to the optimal trajectory output from the second layer includes "water supply set temperature". "and "unit start-up and shutdown sequence" This layer will... A secondary verification is performed. This invention incorporates physical constraints of "minimum running time" and "minimum stopping time." If the start / stop commands calculated from the optimal trajectory... Violating these physical limitations (e.g., attempting to start a unit that has just stopped 5 minutes ago, while...) If the time limit is 1 hour, the start / stop command will be rejected, and the system will forcibly maintain the start / stop state from the previous moment. ).and The instructions were executed normally.

[0043] Through the complete safety closed loop described above, the technical solution of this invention ultimately achieves efficient, safe, and reliable operation under the collaborative optimization of multiple HVAC units.

[0044] The above is an introduction to the method embodiments. The following describes the solution described in this application through device embodiments.

[0045] Figure 3 This is a block diagram illustrating a control device for a HVAC system that migrates across buildings, according to an exemplary embodiment. The device is used in a method for controlling an HVAC system that migrates across buildings. (Refer to...) Figure 3 The device includes a data acquisition module, a Gaussian process module, a meta-learning module, a kernel learning module, a control trajectory module, and a safety constraint module.

[0046] Data acquisition module: used to collect historical data of the reference building complex and measured data of the target domain, and to preprocess the historical data of the reference building complex and the measured data of the target domain to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. Gaussian Process Module: Used to construct a building dynamic prediction model based on the Gaussian process dynamic model. It uses the exact regression method to calculate the covariance matrix and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. Meta-learning module: used to perform meta-learning based on the reference building group training dataset to obtain the initial hyperparameters of the kernel. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. Kernel learning module: used to perform kernel learning on the building dynamic prediction model based on the target building group fine-tuning dataset and the initial kernel hyperparameters, to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the target building group fine-tuning dataset as the task-level loss and using gradient descent to optimize the kernel hyperparameters. Control trajectory module: used to generate multiple candidate future control trajectories through Monte Carlo sampling, and obtain future state prediction sequences based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain a variety of control sequences. Safety constraint module: used to establish a multi-level safety constraint optimization framework, perform hierarchical screening and verification on the future state prediction sequence, and obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

[0047] A control device for a HVAC system that can be moved across buildings, the control device comprising: a processor; and a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the method described in any one of the above-described methods for controlling a HVAC system that can be moved across buildings.

[0048] A computer-readable storage medium, characterized in that the computer-readable storage medium stores program code, the program code being invoked by a processor to execute the method as described in any one of claims 1 to 7.

[0049] Figure 4 This is a structural schematic diagram of a HVAC system control device that can be moved across buildings, as provided in an embodiment of the present invention. Figure 4 As shown, the control equipment for HVAC systems that are relocated across buildings may include the above-mentioned Figure 3 The illustrated HVAC system control device for cross-building relocation. Optionally, the cross-building relocation HVAC system control device 410 may include a first processor 2001.

[0050] Optionally, the HVAC system control device 410 that migrates across buildings may also include a memory 2002 and a transceiver 2003.

[0051] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0052] The following is combined Figure 4 A detailed description of each component of the HVAC system control equipment 410 that is relocated across buildings is provided below: The first processor 2001 is the control center of the HVAC system control device 410 that is relocated across buildings. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0053] Optionally, the first processor 2001 can perform various functions of the HVAC system control device 410 that migrates across buildings by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0054] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0055] In a specific implementation, as one example, the HVAC system control device 410 that migrates across buildings may also include multiple processors, for example... Figure 4 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0056] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0057] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be connected via the interface circuit of the HVAC system control device 410 that is migrated across the building. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0058] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0059] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0060] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be controlled via the interface circuit of the HVAC system control device 410 that is migrated across buildings. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0061] It should be noted that, Figure 4 The structure of the HVAC system control device 410 shown in the diagram does not constitute a limitation on the router. The actual knowledge structure identification device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0062] Furthermore, the technical effects of the HVAC system control device 410 that can be moved across buildings can be referred to the technical effects of the HVAC system control method that can be moved across buildings as described in the above method embodiments, and will not be repeated here.

[0063] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0064] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0065] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0066] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0067] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0068] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0069] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0071] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0072] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0074] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A control method for a HVAC system that migrates across buildings, characterized in that, The method includes: S1: Collect historical data of the reference building complex and measured data of the target domain, and preprocess the historical data of the reference building complex and the measured data of the target domain to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. S2: Construct a building dynamic prediction model based on the Gaussian process dynamic model, calculate the covariance matrix using the exact regression method and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. S3: Obtain initial hyperparameters of the kernel by performing meta-learning based on the training dataset of the reference building complex. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. S4: Based on the fine-tuned dataset of the target building complex and the initial kernel hyperparameters, perform kernel learning on the building dynamic prediction model to obtain the optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the fine-tuned dataset of the target building complex as the task-level loss and using gradient descent to optimize the kernel hyperparameters. S5: Multiple candidate future control trajectories are generated through Monte Carlo sampling, and a future state prediction sequence is obtained based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain multiple control sequences. S6: Establish a multi-level safety constraint optimization framework, perform hierarchical screening and verification on the future state prediction sequence, and obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

2. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, S1 involves collecting historical data of the reference building complex and measured data of the target domain. Preprocessing these data yields a training dataset for the reference building complex and a fine-tuning dataset for the target building complex, including: S11: Select a reference building group and obtain data on the building HVAC system at different times in history to obtain historical data of the reference building group. The data of the building HVAC system includes environmental conditions, region-specific variable data and action vectors. The environmental conditions and region-specific variable data include indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, humidity, occupancy status, start-up and shutdown sequence of the unit at the previous moment, and regional air temperature. The action vector includes the supply water temperature setpoint. S12: Select the target domain's measured building complex and collect data on the building's HVAC system at different times to obtain historical data of the target domain's measured building complex; S13: Based on the historical data of the reference building complex, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the reference building complex training dataset. The output variable is the region state at the next moment. S14: Based on the measured historical data of the target building complex, the environmental conditions and region-specific variables at each moment are integrated into a state vector. The state vector is combined with the subsequent action vector to form multiple sets of input variables and multiple sets of output variables, thus obtaining the target building complex fine-tuning dataset.

3. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, The S2-based Gaussian process dynamic model constructs a building dynamic prediction model. It uses an exact regression method to calculate the covariance matrix and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses a radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix, including: S21: Starting from the candidate function set, a Gaussian process prior is set, a kernel function is selected, and a dynamic model of the Gaussian process is obtained. The kernel function is a radial basis function kernel, which is used to measure input similarity. The hyperparameters of the kernel function consist of a scalar and a matrix. S22: Based on the Gaussian process dynamic model, a building dynamic prediction model is constructed. The building dynamic prediction model is used to take the characteristics of building HVAC as system input to obtain model output. The system input includes indoor temperature and humidity, outdoor temperature and humidity, supply and return water temperature, supply water temperature set point, and the start-up and shutdown sequence of the unit at the previous moment. The model output is the energy consumption and indoor comfort of the building at the next moment. S23: Based on the training dataset of the reference building complex, the covariance matrix is ​​calculated using the exact regression method. The identity matrix and noise variance are introduced to characterize the uncertainty noise, and the Gaussian process posterior distribution is obtained. The Gaussian process posterior distribution is used to obtain the result that follows the Gaussian distribution when making predictions.

4. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, The initial hyperparameters of the kernel are obtained by meta-learning based on the training dataset of the reference building complex in S3. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model, including: S31: Construct meta-learning initialization task instructions based on the training dataset of the reference building complex; S32: Based on the meta-learning initialization task instruction, the kernel parameters of the building dynamic prediction model are initially optimized, the common thermal dynamic laws of multiple buildings are learned, and the kernel initial hyperparameters are obtained. S33: Initialize the building dynamic prediction model based on the kernel initial hyperparameters.

5. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, S4, based on the fine-tuned dataset of the target building complex and the initial kernel hyperparameters, performs kernel learning on the building dynamic prediction model to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the fine-tuned dataset of the target building complex as the task-level loss and employing gradient descent to optimize the kernel hyperparameters, including: S41: Based on the measured data of the target domain and the initial hyperparameters of the kernel, construct the task-level loss function and calculate the loss function. For each specific prediction task, use the mean square error calculated based on the measured data of the target domain as the task-level loss to obtain the loss value. S42: Based on the loss value, calculate the gradient of the kernel hyperparameter with respect to the loss to obtain the loss gradient; S43: Based on the initial kernel hyperparameters, loss gradient and learning rate, aggregate the model error, perform gradient descent, and after multiple iterations, obtain the optimized kernel hyperparameters that minimize the mean square error of the model prediction. S44: Based on the optimized kernel hyperparameters and the measured data of the target domain, the model is fitted to obtain the optimized building dynamic prediction model.

6. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, The S5 process generates multiple candidate future control trajectories through Monte Carlo sampling, and obtains a future state prediction sequence based on the optimized building dynamic prediction model. These multiple candidate future control trajectories include various control sequences, including: S51: Based on GP, ​​the MBRL building dynamic environment is constructed, the state is initialized, and the initial control state is obtained; S52: Using the MPC multi-step prediction idea and Monte Carlo sampling method, parallel sampling is performed. Based on the initial control state, the trajectory dynamics are simulated through GP forward propagation to generate multiple control sequences and obtain K state trajectory sets. S53: Input the set of K state trajectories into the optimized building dynamic prediction model to obtain the future state prediction sequence.

7. The control method for HVAC systems relocated across buildings according to claim 1, characterized in that, The S6 establishes a multi-level safety constraint optimization framework to perform hierarchical screening and verification of the future state prediction sequence to obtain the optimal control sequence. This multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on constraint mechanisms. S61: Calculate the uncertainty threshold based on the training dataset of the reference building complex to obtain the uncertainty threshold; S62: Based on the uncertainty threshold, implement a safety screening to eliminate trajectories whose prediction variance of the future state prediction sequence exceeds the threshold; S63: Based on the screening of the multi-objective comprehensive reward function, select the trajectory with the highest comprehensive reward. The multi-objective comprehensive reward function is used to balance comfort, energy consumption and uncertainty indicators. S64: Set minimum running time and minimum stopping time constraints to suppress unnecessary device switching. Before the control action is executed, the minimum running time and minimum stopping time constraints of the first control action of the trajectory with the highest comprehensive return are checked. If the control action violates the minimum time constraint, the start-stop state of the previous moment is forcibly maintained to obtain the optimal control sequence. The setting of minimum running time and minimum stopping time constraints is used to solve the sensitivity problem of frequent start-stop.

8. A control device for a HVAC system that can be moved across buildings, the control device being used to implement the control method for a HVAC system that can be moved across buildings as described in any one of claims 1-7, characterized in that, The device includes: Data acquisition module: used to collect historical data of the reference building complex and measured data of the target domain, and to preprocess the historical data of the reference building complex and the measured data of the target domain to obtain the training dataset of the reference building complex and the fine-tuning dataset of the target building complex. Gaussian Process Module: Used to construct a building dynamic prediction model based on the Gaussian process dynamic model. It uses the exact regression method to calculate the covariance matrix and obtain the Gaussian process posterior distribution. The Gaussian process dynamic model uses the radial basis function kernel as the Gaussian process prior distribution. The hyperparameters of the radial basis function kernel consist of a scalar and a matrix. Meta-learning module: used to perform meta-learning based on the reference building group training dataset to obtain the initial hyperparameters of the kernel. The meta-learning includes initializing the kernel parameters of the building dynamic prediction model. Kernel learning module: used to perform kernel learning on the building dynamic prediction model based on the target building group fine-tuning dataset and the initial kernel hyperparameters, to obtain an optimized building dynamic prediction model. The kernel learning includes using the mean squared error obtained from the target building group fine-tuning dataset as the task-level loss and using gradient descent to optimize the kernel hyperparameters. Control trajectory module: used to generate multiple candidate future control trajectories through Monte Carlo sampling, and obtain future state prediction sequences based on the optimized building dynamic prediction model. The multiple candidate future control trajectories contain a variety of control sequences. Safety constraint module: used to establish a multi-level safety constraint optimization framework, perform hierarchical screening and verification on the future state prediction sequence, and obtain the optimal control sequence. The multi-level safety constraint optimization framework includes safety screening based on a preset uncertainty threshold, screening based on a multi-objective comprehensive reward function, and screening based on a constraint mechanism.

9. A control device for a HVAC system that can be moved across buildings, characterized in that, The HVAC system control processor for cross-building migration; a memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images