Using a probabilistic surrogate model to control the vapor compression system

By optimizing the setpoint of the vapor compression system using a data-driven closed-loop control method and a probabilistic surrogate model, the problems of model complexity and dynamic instability in existing technologies are solved, achieving efficient energy optimization and improved system stability.

CN116964532BActive Publication Date: 2026-11-17MITSUBISHI ELECTRIC CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202180095115.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-03-07
Filing Date
2021-12-24
Publication Date
2026-11-17
Estimated Expiration
2041-12-24

AI Technical Summary

Technical Problem

Existing control methods for vapor compression systems suffer from model complexity and dynamic system instability, making it difficult to efficiently optimize energy efficiency and update the model, resulting in energy waste and performance degradation.

Method used

A data-driven closed-loop control method is adopted, and a setpoint optimizer for the steam compression system is constructed using a probabilistic surrogate model and Bayesian optimization techniques. The setpoint is adjusted and the model is updated in real time through a feedback controller. The setpoint combination is optimized by combining gradient descent and Gaussian process regression.

Benefits of technology

This achieves efficient energy optimization of the vapor compression system, reduces energy consumption, improves system flexibility and stability, and reduces the complexity of model building and updating.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116964532B_ABST
    Figure CN116964532B_ABST
Patent Text Reader

Abstract

A controller for controlling a vapor compression system is provided. The controller is configured to control the operation of the VCS with different combinations of setpoints for different actuators of the VCS, to estimate the operating cost of the VCS for each of the different combinations of setpoints, and to compute a probabilistic surrogate model using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, wherein the probabilistic surrogate model constrains at least the first two moments of the operating cost in a probabilistic mapping. The controller is further configured to select an optimal combination of setpoints based on a function for obtaining the first two moments of the operating cost, the optimal combination of setpoints having a maximum likelihood that serves as the global minimum at the surrogate model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure generally relates to vapor compression systems, and more particularly to apparatus and methods for controlling the operation of vapor compression systems. Background Technology

[0002] Vapor compression systems, such as heat pumps, refrigeration, and air conditioning systems, are widely used in industrial and residential applications. The introduction of variable-speed compressors, variable-position valves, and variable-speed fans for vapor compression cycles has greatly improved the operational flexibility of vapor compression systems. The efficiency of a vapor compression system can be improved by controlling its variable components. In other words, control inputs can be provided to the vapor compression system to increase efficiency. For example, the compressor speed can be adjusted to modulate the refrigerant flow rate. The speeds of the evaporator and condenser fans can be changed to alter the heat transfer between the air and the heat exchanger. Changes in the expansion valve opening can affect the pressure drop between the high and low pressure points in the vapor compression system, which in turn affects the refrigerant flow rate and the corresponding superheat temperature at the evaporator outlet.

[0003] The combination of control inputs that delivers a specific amount of heat to a vapor compression system is usually not unique, and different combinations of control inputs consume different amounts of energy. Therefore, it is desirable to operate the vapor compression system using a combination of control inputs that minimizes energy consumption and maximizes the energy efficiency of the vapor compression system.

[0004] Some methods for maximizing energy efficiency rely on mathematical models of the physics of a vapor compression system. This model-based approach attempts to describe the effects of control inputs on the thermodynamic behavior and energy consumption of the vapor compression system. In this model-based approach, existing analytical models are used to predict combinations of control inputs that satisfy heat load requirements and minimize energy consumption.

[0005] However, the use of such predictive mathematical models for the combination of control inputs has drawbacks. For example, the mathematical models rely on simplifying assumptions to generate mathematically tractable representations. These assumptions ignore important effects or fail to consider specific installation characteristics (such as room dimensions), causing the mathematical models to deviate from the actual behavior of the vapor compression system. Furthermore, mathematical models are difficult to derive and calibrate. For instance, the parameters describing the operation of the components of the vapor compression system (e.g., the compressor) are determined experimentally for the various types of compressors used, and the mathematical model of the vapor compression system can have multiple such parameters. Therefore, determining the values ​​of the parameters for each mathematical model is a lengthy process. Moreover, vapor compression systems can change over time. When the vapor compression system changes, for example due to slow refrigerant leakage or corrosion accumulation on the heat exchanger, a mathematical model that accurately describes the operation of the vapor compression system at one time may be inaccurate at a later time.

[0006] Some methods aim to optimize control inputs without employing mathematical models. For example, an Extremum Search Controller (ESC) is used to optimize control inputs in a model-free manner. However, an ESC can take several hours to converge to the optimal setpoint. Moreover, because disturbances acting on a vapor compression system have faster kinetics, the optimal setpoint can change before the ESC converges. This slow convergence characteristic of ESC represents a barrier to solutions for real-time optimization of the performance of vapor compression systems.

[0007] Furthermore, ESC constructs local gradients numerically and chases these local gradients to a local optimum. This process can be slow. For example, if the control gain of ESC is too small and the basis function to be optimized includes small gradients, convergence will be slow. Conversely, if the control gain of ESC is too large and the basis function is nonlinear, the iterative solution of ESC may exceed the optimal setpoint due to large gradients. Therefore, ESC is incompatible with nonconvex problems with highly nonlinear dynamics.

[0008] To overcome the slow convergence characteristic of ESC, a filter is used. The filter separates the phase effect of the transient portion of the response to disturbances. However, the design of the filter is complex and requires a large amount of information about the vapor compression system. Summary of the Invention

[0009] One objective of some implementations is to provide a data-driven approach for determining the optimal combination of setpoints that minimizes the power consumption of a vapor compression system without using an analytical model of the underlying system dynamics. This objective can be achieved by formulating a surrogate model. Furthermore, one objective of some implementations is to provide a Bayesian optimization hot-start method that involves a surrogate model to accelerate the convergence of the extreme value search controller (ESC).

[0010] Vapor compression systems (VCS), such as heat pumps, refrigeration, and air conditioning systems, are widely used in industrial and residential applications. The introduction of variable-speed compressors, variable-position valves, and variable-speed fans for the vapor compression cycle has improved the operational flexibility of VCS. Specifically, the combination of control inputs that delivers a specific amount of heat to the VCS is often not unique, and different combinations of control inputs can have different operating costs. For example, a VCS operating according to two different combinations of control inputs controlling the state of the VCS's actuators can deliver the same heat capacity but can consume different amounts of energy. Therefore, it is desirable to operate the VCS using a combination of control inputs that optimizes operating costs, for example, by minimizing energy consumption and / or maximizing the VCS's energy efficiency.

[0011] Control methods for the energy efficiency of VCSs can be divided into two categories: model-based control methods and data-driven control methods. Model-based methods attempt to describe the impact of control inputs on the thermodynamic behavior and energy consumption of the VCS. In this approach, existing analytical models are used to predict combinations of control inputs that meet heat load requirements and minimize energy consumption. However, analytical models for different VCS configurations are difficult to derive, calibrate, and update over time. For example, it is difficult to formulate analytical models for multi-zone vapor compression systems (MZ-VCS) that include at least a single compressor connected to multiple heat exchangers arranged in one or more indoor zones. In other words, the complexity of MZ-VCS can prevent the use of corresponding analytical models.

[0012] In the absence of an accurate model of the dynamic system (such as a VCS), some control methods utilize operational data generated in real time by the dynamic system to construct control strategies that stabilize the system's dynamics. Designing control strategies using operational data is known as data-driven control. Two types of data-driven control methods exist: indirect and direct methods. Indirect methods involve constructing a surrogate model of the dynamic system and using that model to design the controller. Direct methods involve directly constructing the control strategy based on operational data without intermediate model-building steps. A disadvantage of indirect methods is the potential need for a large amount of data during the model-building phase. Conversely, direct methods require less data, but even leading-edge direct control methods experience difficulties in handling state and input constraints, which are essential for maintaining safe operation in closed-loop control systems such as VCSs.

[0013] Therefore, some implementations aim to provide an indirect data-driven control method for closed-loop control of a VCS. This closed-loop control method may avoid the complexity of model-based control methods and the instability of direct data-driven control methods. However, to provide such indirect data-driven control, it is necessary to generate a proxy model that can efficiently manage the operation of the VCS.

[0014] Some implementations are based on the understanding that several obstacles exist in constructing such a surrogate model for controlling the VCS. For example, to construct the surrogate model, there exists a large and virtually infinite continuous space of values ​​for the control inputs of the VCS actuators that needs to be explored. To address this problem, some implementations replace the surrogate model of the VCS actuator control inputs with a surrogate model of the setpoints of the different actuators that the VCS control needs to achieve and / or track. In this way, the optimization of the VCS control is separated from the actual control, which allows for the sampling space of setpoints without jeopardizing the continuous control of the different actuators of the VCS.

[0015] The proxy model provides a mapping between various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the VCS actuators. As used herein, "setpoint" refers to the desired value of a variable of the VCS. The term setpoint applies to any specific value of a particular combination of thermodynamic and environmental parameters, as well as control signals. For example, the term "setpoint" can indicate one or more specific states of components of the VCS. Different actuators of the VCS may include one or more actuators controlling the position of the expansion valve of the VCS, the indoor fan of the VCS, the outdoor fan of the VCS, damping actuators, etc. According to an implementation, the operating cost corresponding to the combination of setpoints is the power consumption of the VCS operating according to the combination of setpoints. Furthermore, the proxy model 102 can be used to control the operation of the VCS.

[0016] Some implementations are based on the understanding that various search techniques, such as gradient descent, can be used in the surrogate model to select the optimal combination of setpoints, thereby reducing the operating cost of the VCS. However, the slow dynamic response of the VCS leads to another problem: the separation of the control from the actual VCS control. For example, the surrogate model needs to map the setpoint to the cost of operating the VCS in its steady state. However, it takes time (e.g., 15 minutes) for the VCS to reach a steady state corresponding to the setpoint. Moreover, when setpoint optimization and control are separated, additional feedback controllers are needed to iteratively change the actuator states to their setpoints, and additional time is required afterward to stabilize the VCS output.

[0017] Therefore, the separation of setpoint optimization from actual VCS control raises internal problems: (1) how to estimate the cost of VCS operation during transient time without waiting for the steady state of VCS operation, and (2) how to reduce the corresponding operating costs and the number of sample pairs of setpoint combinations required to construct the surrogate model.

[0018] Some implementations address these internal problems by constructing probabilistic surrogate models using Bayesian optimization of combinations of setpoints and their corresponding estimated operating costs. A probabilistic surrogate model provides a mapping between various combinations of setpoints for different actuators of the VCS and probability distributions regarding the corresponding operating costs. A surrogate model corresponds to a probabilistic surrogate model. The probabilistic nature of the surrogate model includes uncertainty, which can be used to (1) estimate the cost of VCS operation without waiting for the VCS to reach a steady state, since transient measurements include uncertainties that can be incorporated into the probabilistic surrogate model, and (2) guide the surrogate model construction process by taking this uncertainty into account, thereby reducing the number of samples required to construct the surrogate model.

[0019] According to one implementation, a probabilistic surrogate model can be constructed using a probabilistic machine learning algorithm (such as Gaussian process regression). This algorithm maps combinations of setpoints to probability distributions about the corresponding estimated operating costs, such that the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping. In one implementation, the at least two moments may include the mean and variance of the operating costs (also referred to as the confidence range). For example, for a given combination of setpoints, the surrogate model provides not only a power consumption prediction but also a confidence range centered on the predicted power consumption (e.g., if the power consumption prediction is 1000W, the 95% confidence range for the predicted power consumption would be between 960W and 1040W).

[0020] Furthermore, some implementations aim to select a combination of setpoints (also referred to as "data points") that must be queried next. As used herein, querying a combination of setpoints (or data points) refers to using the combination of setpoints to evaluate an objective (e.g., operating the vapor compression system based on the combination of setpoints). Some implementations use a function that obtains the first two moments of the operating cost to select a combination of setpoints for the next query.

[0021] The acquisition function uses a probability mapping provided by the surrogate model to select a combination of setpoints for the next query. Therefore, the acquisition function serves as a guide for determining the data points for the next sampling / query. In the implementation, the acquisition function is maximized to select the optimal combination of setpoints with the maximum likelihood that serves as the global minimum of the surrogate model, for querying and optimizing the operation of the VCS. This surrogate model construction and utilization are part of Bayesian optimization. In other words, Bayesian optimization comprises: a surrogate model using Gaussian process regression to provide the probability mapping, and an acquisition function utilizing the probability mapping provided by the surrogate model to orient the query toward the resulting setpoint.

[0022] The optimal combination of selected setpoints is applied to the feedback controller associated with the VCS. The feedback controller is configured to determine the control command corresponding to the optimal combination of selected setpoints. Furthermore, the feedback controller controls the operation of the VCS based on the control command to change the state of the VCS actuators according to the optimal combination of selected setpoints. The feedback controller is one or a combination of a PI controller, PID controller, MPC controller, or robust controller.

[0023] Furthermore, the vapor compression system produces the performance output (e.g., power consumption) of the optimal combination of selected setpoints. The surrogate model is then updated using the optimal combination of selected setpoints and the corresponding performance output, leveraging Bayesian optimization. This results in an updated surrogate model. The same acquisition function is then applied to update the surrogate model to select the optimal combination of new setpoints for controlling the VCS and to update the surrogate model. The surrogate model is updated in a similar manner in each iteration. This iterative process is performed online (i.e., during the real-time operation of the vapor compression system) until a termination condition is met. The termination condition can be the number of iterations. The number of iterations can be arbitrary or user-defined. The iterative upgrade of the surrogate model yields an accurate surrogate model. Since the surrogate model is updated in each iteration and used accordingly with the acquisition function to determine the optimal setpoint, the surrogate model combined with the acquisition function acts as a setpoint optimizer.

[0024] Some implementations are based on the understanding that for a non-convex function that includes multiple minima, if the ESC is initiated in a region that includes local minima, then the ESC can determine only the local minima without determining the global minima.

[0025] Some implementations are based on the understanding that, for non-convex functions, determining the global minimum involves a well-identified initial guess—that is, initiating ESC in the region where the function is locally convex and includes the global minimum. Furthermore, gradient estimation is used within this region to identify the global minimum. Without a good initial guess, gradient-based algorithms may converge slowly and may get trapped in local minima.

[0026] Some implementations are based on the idea that a surrogate model can be used to initiate ESC in the region most likely to contain the global minimum. Specifically, a Bayesian optimization (BO) framework using the surrogate model can be formulated to initiate ESC in the region most likely to contain the global minimum. The solution obtained by the BO framework can then be used as an initial guess for an Adam-accelerated extremum search control, which can be used to obtain the global minimum.

[0027] Some implementations are based on the understanding that when a function of operating costs in analytical form is available, the gradient of the operating costs used by the extreme value search controller can be easily computed to tune the optimal combination of selected setpoints. However, such an analytical function is typically unavailable for controlling the VCS. However, some implementations are based on the realization that a probabilistic surrogate model can be used instead of such an analytical function to compute the gradient of the operating costs based on a continuous estimate of the operating costs for the combination of tuned setpoints. Specifically, one implementation involves taking the gradient of the kernel matrix of the Gaussian process directly with respect to each setpoint. In this way, the probabilistic surrogate model is used for both warm-starting the ESC and controlling the ESC.

[0028] Proxy modeling-based optimization (such as ESC with a Bayesian optimization warm start) offers decisive advantages. For example, machine learning methods enable the construction of surrogate models of a vapor compression system's input-to-output performance, for instance, via Gaussian process regression, regardless of noise in the measurements. However, model-free methods (such as ESC) estimate gradients directly from noisy data, which can lead to undesirable effects such as oscillations or even divergence in the neighborhood of the optimal input. Proxy modeling also learns the input-output model while implementing termination criteria that can be stored, for example, in the cloud, to achieve a good initial guess of the input (warm start) to reach the peak in subsequent runs. Furthermore, the convergence rate is improved with surrogate modeling via smoothing regressors because smoothing optimization methods, such as first-order and second-order methods, can be directly applied to the surrogate model.

[0029] Therefore, one embodiment discloses a controller for controlling the operation of a vapor compression system (VCS). The controller includes: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the controller to: control the operation of the VCS with different combinations of setpoints for different actuators of the VCS; estimate the operating cost of the VCS for each of the different combinations of setpoints; compute a probabilistic surrogate model using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, the probabilistic surrogate model providing a probabilistic mapping between the various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the VCS actuators, wherein the probabilistic surrogate model defines at least the first two moments of the operating cost in the probabilistic mapping; select an optimal combination of setpoints based on a function for obtaining the first two moments of the operating cost, the optimal combination of setpoints having a maximum likelihood that is the global minimum at the surrogate model; and submit the selected optimal combination of setpoints to a feedback controller to change the state of the actuators according to the corresponding setpoints of the VCS actuators in the selected optimal combination.

[0030] Another embodiment discloses a method for controlling the operation of a vapor compression system (VCS). The method includes the following steps: controlling the operation of the VCS with different combinations of setpoints for different actuators of the VCS; estimating the operating cost of the VCS for each of the different combinations of setpoints; computing a probabilistic surrogate model using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, the probabilistic surrogate model providing a probabilistic mapping between the various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the VCS actuators, wherein the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping; selecting an optimal combination of setpoints based on a function for obtaining the first two moments of the operating costs, the optimal combination of setpoints having a maximum likelihood that is the global minimum at the surrogate model; and submitting the selected optimal combination of setpoints to a feedback controller to change the state of the actuators according to the corresponding setpoints of the VCS actuators in the selected optimal combination.

[0031] Another embodiment discloses a non-transitory computer-readable storage medium embodied on which a program is executed by a processor to perform a method for controlling the operation of a vapor compression system (VCS). The method includes the following steps: controlling the operation of the VCS with different combinations of setpoints for different actuators of the VCS; estimating the operating cost of the VCS for each of the different combinations of setpoints; calculating a probabilistic surrogate model using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, the probabilistic surrogate model providing a probabilistic mapping between the various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the VCS actuators, wherein the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping; selecting an optimal combination of setpoints based on a function for obtaining the first two moments of the operating costs, the optimal combination of setpoints having a maximum likelihood that is the global minimum at the surrogate model; and submitting the selected optimal combination of setpoints to a feedback controller to change the state of the actuators according to the corresponding setpoints of the VCS actuators in the selected optimal combination. Attached Figure Description

[0032] Figure 1A A schematic diagram is shown illustrating the principles employed in some implementations of control separation for a vapor compression system.

[0033] Figure 1B A schematic diagram is shown illustrating the principles employed by some implementations for controlling the operation of a vapor compression system.

[0034] Figure 1CExamples of mean predictions and confidence ranges provided by surrogate models according to some implementations are shown.

[0035] Figure 1D The curves of the acquisition function according to some implementations are shown.

[0036] Figure 2 A block diagram of a control device for controlling the operation of a vapor compression system according to some embodiments is shown.

[0037] Figure 3 A block diagram of a Gaussian process for determining a proxy model, according to some implementations, is shown.

[0038] Figure 4 A block diagram is shown for selecting data points for the next query, according to some implementation methods.

[0039] Figure 5 A schematic diagram of a time-varying extremum search controller (TV-ESC) using the Adam algorithm principle according to some implementations is shown.

[0040] Figure 6 An example is given of an adaptive step size for the mean and standard deviation of gradient data according to some implementations.

[0041] Figure 7A An exemplary convex function according to some implementations is shown.

[0042] Figure 7B An exemplary nonconvex function according to some implementations is shown.

[0043] Figure 8 A flowchart of a Bayesian-optimized warm-start ESC according to some implementations is shown.

[0044] Figure 9A A block diagram of a multi-zone vapor compression system (MZ-VCS) according to some embodiments is shown.

[0045] Figure 9B A block diagram of a multi-zone vapor compression system (MZ-VCS) according to some embodiments is shown.

[0046] Figure 10 Performance results are shown for a standard time-varying ESC, a time-varying ESC accelerated by Adam, and a Bayesian-optimized hot-start ESC, according to some implementations.

[0047] Figure 11 Performance results of Bayesian-optimized warm-start ESC and standard time-varying ESC for complex systems, according to some implementations, are illustrated.

[0048] Figure 12 Examples of using a controller to control heating, ventilation, and air conditioning (HVAC) systems according to some implementations are shown. Detailed Implementation

[0049] In the following description, numerous specific details are set forth for purposes of explanation in order to provide a thorough understanding of this disclosure. However, those skilled in the art will understand that this disclosure can be practiced without requiring these specific details. In other instances, apparatus and methods are shown only as block diagrams to avoid obscuring the disclosure.

[0050] As used in this specification and claims, when the terms “for example,” “e.g.,” and “such as,” and the verbs “comprising,” “having,” “including,” and other verb forms thereof are used in conjunction with a list of one or more components or other items, these terms are interpreted as open-ended, meaning that the list is not considered to exclude other additional components or items. The term “based on” means at least partially based on. Furthermore, it is to be understood that, unless specifically defined as limiting, the language and terminology used herein are for descriptive purposes and should not be construed as limiting. Any headings used in this specification are merely for convenience and have no legal or restrictive effect.

[0051] Figure 1A This diagram illustrates the principles employed in several implementations of a vapor compression system based on control separation. Vapor compression systems (VCS), such as heat pumps, refrigeration, and air conditioning systems, are widely used in industrial and residential applications. The introduction of variable-speed compressors, variable-position valves, and variable-speed fans for the vapor compression cycle enhances the operational flexibility of the VCS. Specifically, the combination of control inputs delivering a specific amount of heat to the VCS is typically not unique, and different combinations of control inputs can result in different operating costs. For example, a VCS operating based on two different combinations of control inputs controlling the state of the VCS's actuators can deliver the same heat capacity but consume different amounts of energy. Therefore, it is desirable to operate the VCS using optimized combinations of control inputs that minimize energy consumption and / or maximize the VCS's energy efficiency, thereby optimizing operating costs.

[0052] Control methods for the energy efficiency of VCSs can be divided into two categories: model-based control methods and data-driven control methods. Model-based methods attempt to describe the impact of control inputs on the thermodynamic behavior and energy consumption of the VCS. In this model-based approach, existing analytical models are used to predict combinations of control inputs that meet heat load requirements and minimize energy consumption. However, analytical models for different VCS configurations are difficult to derive, calibrate, and update over time. For example, it is difficult to formulate analytical models for multi-zone vapor compression systems (MZ-VCS) that include at least a single compressor connected to multiple heat exchangers arranged in one or more indoor zones. In other words, the complexity of MZ-VCS can prevent the use of corresponding analytical models.

[0053] In the absence of an accurate model of the dynamic system (such as a VCS), some control methods utilize operational data generated in real time by the dynamic system to construct control strategies that stabilize the dynamics of the system. Designing control strategies using operational data is known as data-driven control. Two types of data-driven control methods exist: indirect and direct methods. Indirect methods involve constructing a surrogate model of the dynamic system and using that model to design the controller. Direct methods involve directly constructing the control strategy based on operational data without intermediate model-building steps. A disadvantage of indirect methods is the potential need for a large amount of data during the model-building phase. Conversely, direct methods require less data, but even leading-edge direct control methods experience difficulties in handling state and input constraints, which are essential for maintaining safe operation in closed-loop control systems such as VCSs.

[0054] Therefore, one objective of some implementations is to provide an indirect data-driven control method for closed-loop control of a VCS. This closed-loop control method may not suffer from the complexity of model-based control methods or the instability of direct data-driven control methods. However, to provide such indirect data-driven control, it is necessary to generate a proxy model that can efficiently manage the operation of the VCS.

[0055] Some implementations are based on the understanding that several obstacles exist in constructing such a surrogate model for controlling the VCS. For example, to construct the surrogate model, there exists a large and virtually infinite continuous space of values ​​for the control inputs of the VCS actuators that needs to be explored. To address this problem, some implementations replace the surrogate model 110 of the VCS actuator control inputs with a surrogate model 102 of the setpoints of the different actuators that the VCS control needs to achieve and / or track. In this way, the optimization of the VCS control is separated from the actual control, which allows for the sampling space of the setpoints without jeopardizing the continuous control of the different actuators of the VCS.

[0056] The proxy model 102 provides a mapping between various combinations of setpoints 102a for different actuators of the VCS and the corresponding operating costs 102b of the VCS actuators. As used herein, "setpoint" refers to the desired value of a variable of the VCS. The term setpoint applies to any specific value of a particular combination of thermodynamic and environmental parameters and control signals. For example, the term "setpoint" can indicate one or more specific states of components of the VCS. Different actuators of the VCS may include one or more actuators controlling the position of the expansion valve of the VCS, the indoor fan of the VCS, the outdoor fan of the VCS, damping actuators, etc. According to an implementation, the operating cost corresponding to the combination of setpoints is the power consumption of the VCS operating according to the combination of setpoints. Furthermore, the proxy model 102 can be used to control the operation of the VCS.

[0057] Figure 1B A schematic diagram illustrates the principles employed by some embodiments of controlling the operation of a vapor compression system 110 (VCS) using a surrogate model 102. Some embodiments are based on the understanding that various search techniques, such as gradient descent, can be used with the surrogate model 102 to select the optimal combination of setpoints, thereby reducing the operating cost of the VCS 110. However, the slow dynamic response of the VCS 110 leads to another problem: the separation of the control from the actual VCS control. For example, the surrogate model 102 needs to map the setpoint to the cost of operating the VCS in its steady state. However, it takes time (e.g., 15 minutes) for the VCS 110 to reach the steady state corresponding to the setpoint. Furthermore, when setpoint optimization and control are separated, an additional feedback controller is needed to iteratively change the actuator states to their setpoints, and additional time is required afterward to stabilize the VCS output.

[0058] Therefore, the separation of setpoint optimization from actual VCS control creates internal problems: significant time is consumed while waiting for the VCS to reach steady state, and a large number of sample sets and corresponding operating costs are required to construct the surrogate model 102. To address this, some implementations aim to estimate the cost of VCS operation during transient time without waiting for the VCS to reach steady state, while also reducing the number of samples required to construct the surrogate model 102.

[0059] According to the implementation method, this objective can be achieved by constructing a probabilistic surrogate model using Bayesian optimization of combinations of setpoints and their corresponding estimated operating costs. The probabilistic surrogate model provides a mapping between various combinations of setpoints for different actuators of the VCS110 and probability distributions regarding the corresponding operating costs. Surrogate model 102 corresponds to the probabilistic surrogate model. The probabilistic properties of surrogate model 102 include uncertainty, which can be used to (1) estimate the cost of VCS operation without waiting for the steady state of VCS operation, since transient measurements include uncertainties that can be incorporated into the probabilistic surrogate model, and (2) guide the surrogate model construction process by taking into account this uncertainty to reduce the number of samples required to construct surrogate model 102.

[0060] In one implementation, after the VCS110 has operated according to a combination of setpoints for a predetermined time (e.g., 5 minutes) during which the VCS110 has not yet reached a steady state, the operating cost for that combination of setpoints is estimated. In some implementations, at least some of the operating costs are estimated during the transient period of the VCS110 operating according to the corresponding combination of setpoints.

[0061] According to one implementation, a probabilistic surrogate model can be constructed using a probabilistic machine learning algorithm (such as Gaussian process regression) that maps combinations of setpoints to a probability distribution with respect to the corresponding estimated operating costs, such that the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping. In one implementation, the at least two moments may include the mean and variance of the operating costs (also referred to as the confidence range). For example, for a given combination of setpoints, surrogate model 102 provides not only a power consumption prediction but also a confidence range centered on the predicted power consumption (e.g., the power consumption prediction is 1000W, and the 95% confidence range of the predicted power consumption would be between 960W and 1040W).

[0062] Figure 1C The mean prediction and confidence range provided by surrogate model 102 according to some implementations are illustrated. Point 120 represents the sample / observation result. Curve 116 represents the mean prediction, and shaded area 118 represents the confidence range.

[0063] Furthermore, some implementations aim to select a combination of setpoints (also referred to as "data points") that must be queried next. As used herein, querying a combination of setpoints (or data points) refers to using the combination of setpoints to evaluate a target (e.g., operating the vapor compression system 110 based on the combination of setpoints). Some implementations use a function 104 for obtaining the first two moments of operating costs to select a combination of setpoints for the next query.

[0064] Figure 1DCurve 124 of the acquisition function 104 according to some embodiments is shown. The acquisition function 104 uses a probability mapping provided by the surrogate model 102 to select a combination of setpoints (i.e., data points) for the next query. For example, if the surrogate model 102 determines that the power consumption at a first data point (which has not been queried before) is at least 900W with an 85% confidence level, and at a second data point (which has also not been queried before) the surrogate model 102 determines that the power consumption is at least 990W with a 95% confidence level, then even though the first data point has a low confidence level, the first data point needs to be queried before the second data point, because 900W is more likely to be the optimal power consumption for the vapor compression system 110. Therefore, the acquisition function 104 is used as a guide to determine the data point for the next sampling / query.

[0065] In the implementation, the acquisition function 104 is maximized to select the optimal combination of setpoints 106 with the maximum likelihood of the global minimum 122 of the surrogate model 102 for querying and optimizing the operation of VCS 110. This surrogate model construction and utilization are part of Bayesian optimization 114. In other words, Bayesian optimization 114 includes: using the surrogate model 102 for Gaussian process regression to provide probability mappings, and the acquisition function 104 for using the probability mappings provided by the surrogate model 102 to direct the query to the result setpoint.

[0066] The optimal combination of selected setpoints 106 is applied to the feedback controller 108 associated with VCS 110. The feedback controller 108 is configured to determine a control command corresponding to the optimal combination of selected setpoints 106. Furthermore, the feedback controller 108 controls the operation of VCS 110 based on the control command to change the state of the actuators of VCS 110 according to the optimal combination of selected setpoints 106. Specifically, in this embodiment, the feedback controller 108 measures the output (or process variable) of VCS 110. The measured value is applied to the feedback controller 108 as a feedback signal 130. The feedback controller 108 calculates an error signal as the difference between the optimal combination of selected setpoints 106 and the feedback signal 130. The feedback controller 108 uses control parameters such as proportional gain, integral term, and / or derivative term. These control parameters can be applied to the error signal to determine the input of VCS 110, thereby driving the process variable or VCS output according to the optimal combination of selected setpoints 106. The feedback controller 108 is one or a combination of a PI controller, a PID controller, an MPC controller, or a robust controller.

[0067] Furthermore, the vapor compression system 110 generates a performance output 112 (e.g., power consumption) of the optimal combination of selected setpoints 106. Additionally, the surrogate model 102 is updated using Bayesian optimization with the optimal combination of selected setpoints 106 and the corresponding performance output 112. Thus, an updated surrogate model is obtained. Furthermore, the same acquisition function 104 is applied to update the surrogate model to select a new optimal combination of setpoints for controlling the VCS 110 and to update the surrogate model 120. The surrogate model 102 is updated in a similar manner in each iteration. This iterative process is performed online (i.e., during the real-time operation of the vapor compression system 110) until a termination condition is met. The termination condition can be the number of iterations. The number of iterations can be arbitrary or user-defined.

[0068] The iterative upgrade of the surrogate model produces an accurate surrogate model 102. Since the surrogate model 102 is updated in each iteration and used accordingly in combination with the acquisition function 104 to determine the optimal setpoint, the surrogate model 102 combined with the acquisition function 104 acts as a setpoint optimizer.

[0069] Figure 2 A block diagram of a control device 200 for controlling the operation of a vapor compression system 110, according to some embodiments, is shown. The control device 200 may have multiple interfaces for connecting the control device 200 to other systems and devices. For example, a network interface controller (NIC) 214 is adapted to connect the control device 200 to a network 216 via a bus 212. Through the network 216, wirelessly or wiredly, the control device 200 may receive sensor measurements 218 indicating one or more of the following: the speed of an indoor fan, the speed of an outdoor fan, the position of an expansion valve, the speed of a compressor, the performance output 112 of the vapor compression system 110, etc. Alternatively, the sensor measurements 218 may be received via an input interface 202. The control device 200 may monitor the operation of the components of the vapor compression system 110 based on the sensor measurements.

[0070] The control device 200 includes a processor 204 configured to execute stored instructions and a memory 206 storing instructions executable by the processor 204. The processor 204 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 206 may include random access memory (RAM), read-only memory (ROM), flash memory, or any other suitable memory system. The processor 204 is connected to one or more input and output devices via a bus 212. Furthermore, the control device 200 includes a storage device 208 adapted to store different modules that store executable instructions for the processor 204. The storage device 208 may be implemented using a hard disk drive, an optical drive, a thumb drive, a drive array, or any combination thereof.

[0071] Storage device 208 is configured to store proxy model 102, which provides a probabilistic mapping between various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the VCS actuators. Storage device 208 is also configured to store acquisition function 104. Acquisition function 104 is used to select different combinations of operating parameters of the vapor compression system 110 for the next query / sampling. In some embodiments, acquisition function 104 is used to select the optimal combination of setpoints having the maximum likelihood that serves as the global minimum of the proxy model.

[0072] In some implementations, processor 204 is configured to control the operation of VCS 110 with different combinations of setpoints for different actuators of VCS 110, to estimate the operating cost of VCS 110 or various combinations of different setpoints; and to compute a probabilistic surrogate model using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, the probabilistic surrogate model providing a probabilistic mapping between various combinations of setpoints for different actuators of VCS 110 and the corresponding operating costs of the actuators of the VCS, wherein the probabilistic surrogate model defines at least the first two moments of the operating cost in the probabilistic mapping. Processor 204 is also configured to select an optimal combination of setpoints based on an acquisition function 104 of the first two moments of the operating cost, the optimal combination of setpoints having a maximum likelihood that is the global minimum at the surrogate model. Processor 204 is also configured to submit the selected optimal combination of setpoints to feedback controller 108 via output interface 220 to change the state of the actuator of VCS 110 according to the corresponding setpoint of the actuator in the selected optimal combination. The feedback controller 108 is one or a combination of a PI controller, a PID controller, an MPC controller, or a robust controller.

[0073] In some implementations, processor 204 is further configured to estimate the operating cost of the optimal combination of selected setpoints 106, and to update surrogate model 102 using Bayesian optimization with the optimal combination of selected setpoints and the corresponding estimated operating cost. In some implementations, processor 204 is also configured to recursively update surrogate model 102 with the optimal combination of the most recently selected setpoints based on an acquisition function applied to the updated surrogate model 102, until a termination condition is met.

[0074] Figure 3 A block diagram of a Gaussian process for determining a surrogate model 102, according to some embodiments, is shown. According to one embodiment, the Gaussian process is used to define a prior distribution on a function. It is assumed that the steady-state power cost function J to be optimized is generated based on such a prior distribution, characterized by zero mean and kernelized covariance function. Nucleated covariance function This is responsible for defining characteristics of the function J, such as smoothness and robustness to additional noise. Some implementations are based on the understanding that using the Matern 3 / 2 function is beneficial because it provides a good approximation of the steady-state power function without oversmoothing.

[0075] At box 302, data samples are obtained. For this purpose, the user can select the expansion valve position (e.g., 150 counts), the indoor unit fan speed (e.g., 300 rpm), and the outdoor unit fan speed (e.g., 750 rpm). Furthermore, the vapor compression system 110 operates for several minutes (e.g., 5 minutes) based on the selected values ​​of the vapor compression system 110 parameters to calculate steady-state power consumption. Given a set of parameters (150 counts, 300 rpm, and 750 rpm), this process constitutes obtaining data samples from the vapor compression system 110. For example, five such data samples are obtained (each evaluated for 5 minutes). Let N θ This indicates the number of data samples. A data sample is represented by the following formula:

[0076]

[0077] Where v is additive white noise with zero mean and unknown covariance in the measurement channel.

[0078] At box 304, for a predetermined kernel, given a set of hyperparameters 300l and σ0, matrix K... D (θ) and Calculated as

[0079]

[0080] as well as

[0081]

[0082] At box 306, based on the calculated matrix K D (θ) and Calculate the mean function μ(θ) and variance function σ. 2 The posterior distribution represented by (θ). Mean function μ(θ) and variance function σ. 2 (θ) is given by the following formula:

[0083]

[0084]

[0085] According to the implementation method, the mean function μ(θ) and the variance function σ 2 (θ) Define proxy model 102. From equations (1) and (2), it can be noted that the posterior distribution depends on the kernel and such as l, σ0, and σ n A set of hyperparameters of 300 were selected.

[0086] In this implementation, the set of hyperparameters 300 is determined by maximizing the log-marginal likelihood function.

[0087]

[0088] in, And ξ=J(θ)+v. The problem given by equation (3) can be solved using the quasi-Newton method. In an alternative implementation, the problem given by equation (3) can be solved using the adaptive gradient method.

[0089] Alternatively, in some implementations, the surrogate model 102 may be determined by neural process regression or other variants of Bayesian neural networks and / or machine learning.

[0090] Figure 4 A block diagram is shown for selecting data points (i.e., combinations of parameters) for the next query, according to some implementations. The retrieval function 104 uses the surrogate model 102 to calculate the data points that must be queried. The data points that must be queried are given by the following formula:

[0091]

[0092] Since equation (4) depends on the Gaussian process approximation function but not on the actual function J, the function is obtained. Maximizing this involves computation rather than expensive function evaluation. In the implementation, the function is obtained... This is the Expected Improvement (EI) acquisition function. The EI acquisition function is given by the following equation:

[0093]

[0094] Where φ is the density function of the zero-mean, one-variance normal distribution, Φ is the cumulative distribution function, and

[0095]

[0096] Among them, J best This is the first time that data samples have been used. The lowest steady-state power measured.

[0097] According to the implementation method, the EI acquisition function (5) is maximized to compute the data point to be queried. At box 400, a set of random samples is generated on the permissible search space (Θ) of the parameters.

[0098] In addition, at box 402, the EI acquisition function is calculated for each sample.

[0099] At box 404, the EI retrieval function for each random sample is maximized. At box 406, the maximum value of the sample is selected as the data point to be queried.

[0100] To this end, an acquisition function (e.g., for equation (5)) provides data points for querying. After querying with the selected data points, the corresponding steady-state power is obtained. Some implementations are based on the understanding that the data points determined from the acquisition function can be used to update / retrain the surrogate model 102 to increase the accuracy of the surrogate model 102. In particular, the set of hyperparameters is recalculated using the data samples and data points selected using the acquisition function.

[0101] Similarly, in the next iteration, the updated / retrained surrogate model 102 can be used to select another data point to be queried, using the acquisition function (e.g., for equation (5)). Furthermore, the selected data points can be used to update / retrain the updated / retrained surrogate model 102 again. This iteration of updating the surrogate model 102 is performed until the termination condition is met. After an appropriate number of iterations, the surrogate model 102 learns the base function J.

[0102] Alternatively, in some implementations, an improved probability retrieval function or confidence cap retrieval function can be used to select the next data point to query.

[0103] Step size adjustment based on gradient estimation

[0104] According to the implementation method, a closed-loop vapor compression system (such as vapor compression system 110) can be modeled using dynamics.

[0105] x t+1=f(x) t ,θ t ),

[0106] Where t represents the time index. Let θ represent the state of the vapor compression system, and θ∈Θ represent the setpoint vector, which can be designed to drive the vapor compression system to equilibrium, where the steady-state power output is measured.

[0107]

[0108] To reach its minimum value. For example, θ could be the position of an electronically actuated valve, the speed of a fan, or a combination thereof. It is assumed that the function J is convex over a set of parameters in the permissible search space Θ, and that optimality is allowed. That is, for every θ∈Θ, there exists θmax. ★ ∈Θ, such that And there exists χ > 0 such that here, This represents a known, permissible set of parameters; for example, the range of fan speeds. It is assumed that the nonlinearity f is twice continuously differentiable, but it is not required to be known. The measured steady-state power output. This represents the steady-state power output of the vapor compression system, and π ∞ (·) is the mapping from the setpoint to the equilibrium state; that is, f(π) ∞ (θ), θ) = 0. Some implementations aim to drive θ t →θ ★ (Optimal set point).

[0109] Some implementations are based on the understanding that adaptive gradient algorithms, such as Adam, stochastic gradient descent, Adagrad, and RMSprop, not only utilize gradients but also correspondence statistics that lead to robust estimates of the gradients. Furthermore, adaptive gradient algorithms include momentum-based mechanisms to quickly identify step sizes to improve convergence speed. Adaptive gradient algorithms are commonly used to train neural networks. However, some implementations are based on the idea that adaptive gradient algorithms can be used for extreme value search control (ESC) to automatically adapt the step size based on the smoothness of the function J in its local neighborhood. To this end, one implementation aims to apply the Adam algorithm to extreme value search control to formulate an Adam-accelerated extreme value search control.

[0110] Figure 5A schematic diagram of a time-varying extremum search controller (TV-ESC) using the Adam algorithm principle according to some embodiments is shown. Processor 204 executes the TV-ESC. The TV-ESC causes an input perturbation to the feedback controller 108, starting from an optimal combination of selected setpoints 106, to tune the optimal combination of selected setpoints based on the gradient of the operating cost caused by the perturbation. The TV-ESC includes a gradient-based controller 500 of the following form:

[0111] θ t+1 =θ t -k g g t +d t ,

[0112] Among them, g t It is J(θ) t Gradient estimation of ), scalar k g It controls the gain or step size, and d t This is the bounded jitter signal 512 required to ensure the continuous stimulation.

[0113] Some implementations are based on the understanding that when a function of operating cost in analytical form is available, the operating cost J(θ) used by the extreme value search controller can be easily calculated. t The gradient g of ) t The optimal combination of selected setpoints is tuned. However, such an analysis function is typically unavailable for controlling the VCS. However, some implementations are based on the following: a probabilistic surrogate model 102 can be used instead of such an analysis function to calculate the gradient g of the operating cost based on a continuous estimate of the operating cost for the tuned combination of setpoints. t Therefore, in this implementation, the processor 204 uses a delay 510 to generate the gradient estimate g in a data-driven manner. t Specifically, one implementation involves taking the gradient of the kernel matrix of the Gaussian process directly with respect to each set point. The increment 504 in function J is given as:

[0114] ΔJ t :=J(θ) t )-J(θ t-1 )

[0115] Including incremental setpoints:

[0116] Δθ t :=θ t -θ t-1 .

[0117] The gradient is estimated in a data-driven manner by collecting historical data of the setpoint and the corresponding steady-state power values, and solving the following linear equation.

[0118]

[0119] Specifically, using N d The gradient is estimated using the history of each data point. In this implementation, the processor 204 is configured to estimate the operating cost J(θ) caused by the perturbation by mapping the perturbed value of the optimal combination of selected setpoints to a value relating to the operating cost of the surrogate model 102 according to the acquisition function 104. t The gradient g of ) t .

[0120] Some implementations are based on the understanding that, since the optimization problem is not static, the most recent data points effectively contribute to the gradient. Therefore, a recursive filter 506 is implemented to estimate the gradient 508, and includes a forgetting factor α∈(0,1) to assign importance to the most recent data points rather than past data. The recursive filter 506 is described by the following equation:

[0121]

[0122]

[0123]

[0124] g t =g t-1 +K t e t

[0125] And with P -1 =α -1 I and g -1 Initialize to 0.

[0126] According to some implementation methods, matrix P is applied after a user-defined number of iterations. t Reset to P -1 The numerical stability of the gradient estimator is improved. Since the recursive filter 506 requires a burn-in period denoted by τ to generate a meaningful gradient estimate, the gradient estimate is incorporated into the implementation of the ESC control law or setpoint update law 514 after the burn-in period, as follows:

[0127]

[0128] Since evaluating steady-state power requires data obtained through expensive and time-consuming experiments, it is necessary to accelerate ESC (to obtain the lowest power without many experiments). Because the ESC method requires numerical estimates of gradients, these estimates are often noisy and, on average, only replicate the true gradient. Therefore, maintaining a constant step size k... g This can lead to a degradation in closed-loop performance because of the step size k. g Small values ​​can lead to slow convergence, while large values ​​can lead to suboptimality due to overshoot centered on the optimal value.

[0129] Some implementations are based on adjusting the step size based on historical gradient data. Let m t and v t This indicates that the last N gradients are used to estimate g. t-N+1 , ...g t The current mean and standard deviation are calculated based on (also known as gradient data). This is based on the current mean m. t and standard deviation v t The accelerated controller can be formulated using a formula. The accelerated controller has the following forms:

[0130]

[0131] m t+1 =β1m t +(1-β1)g t

[0132] v t+1 =β2v t +(1-β2)g t ⊙g t

[0133] in,

[0134]

[0135] Where, k t It is the effective step size, and ∈>0 ensures v t +∈>0 scalars.

[0136] Figure 6 Examples of ratio-based implementations are illustrated according to some embodiments. Adaptive step size. Figure 6 The diagram shows the level set 600 of the cost function to be optimized. The star diagram 602 illustrates the optimizer. For trajectory 604, the estimated gradients are consistently in the same direction with small variations. Therefore, the estimated mean m... t The iterations are similar, and the variance v t Smaller. As a result, the ratio... It is larger and allows the controller to take positive steps 606 in the descent direction.

[0137] For trajectory 604, the estimated gradient is noisy, therefore the variance v t Larger. Therefore, the ratio. The smaller size results in a smaller step size and a cautious movement of 610 in the descent direction.

[0138] Figure 7A An exemplary convex function 710 according to some implementations is shown. In the convex function 700, initiating ESC anywhere in the domain 702 ensures that a minimum value can be obtained. Figure 7B An exemplary non-convex function 704 according to some implementations is shown. In the non-convex function 704 that includes multiple minima 710a and 710b, the determination of the global minimum depends on which domain the ESC is initiated in. For example, if the ESC is initiated in domain 706, the ESC determines the global minimum, i.e., minimum 710a. However, if the ESC is initiated in domain 708, the ESC may determine only the minimum 710b (i.e., a local minimum) instead of the global minimum.

[0139] Some implementations are based on the understanding that if the function J is nonconvex, determining the global minimum involves a well-identified initial guess, i.e., initiating ESC in the region where the function J is locally convex and includes the global minimum. Furthermore, gradient estimation is used within this region to identify the global minimum. Without a good initial guess, gradient-based algorithms may converge slowly and may get trapped in local minima.

[0140] Some implementations are based on the idea that a surrogate model 102 can be used to initiate the ESC in the region most likely to contain the global minimum (e.g., region 706). In particular, some implementations are based on the idea that the probabilistic nature of the surrogate model 102 and its ability to estimate the transient operating costs of the controlled VCS may not adequately capture the kinetics of the VCS in selecting the optimal combination of setpoints. To address this, some implementations use the optimal combination of setpoints selected according to the probabilistic model to warm-start the ESC exploring the VCS kinetics in a data-driven manner. Advantageously, the ESC can optimize the control input in a model-free manner. However, the ESC may take several hours to converge to the optimal setpoint. This slow convergence characteristic of the ESC represents a barrier to solutions for real-time optimization of the performance of a vapor compression system, a barrier that can be overcome by warm-starting according to different implementations.

[0141] Figure 8A flowchart of an ESC with Bayesian optimized warm start according to some implementations is shown. The ESC with Bayesian optimized warm start includes a Bayesian optimized warm start mechanism to provide an initial guess about the ESC.

[0142] In step 800, an initial set of setpoints is sampled within the setpoint tolerance space. The initial set of setpoints, along with subsequent setpoints and their corresponding steady-state power outputs, forms a dataset. In step 802, a surrogate model is determined based on the dataset. In step 804, an acquisition function is used to select the next optimal setpoint. In some implementations, the obtained next optimal setpoint may correspond to a combination of parameters with the maximum likelihood that serves as the global minimum of the surrogate model. Figure 3 and Figure 4 Steps 802 and 804 are described in detail.

[0143] In step 806, the obtained next optimal setpoint is applied to ESC as an initial guess.

[0144] Multi-zone vapor compression system

[0145] Some implementations are based on evaluations of the advantages of multi-zone vapor compression systems (MZ-VCS) in controlling residential and commercial buildings. In one implementation, VCS110 may correspond to an MZ-VCS. An MZ-VCS includes at least a single compressor connected to multiple heat exchangers arranged in one or more indoor zones. See below. Figure 9A and Figure 9B Detailed description of MZ-VCS.

[0146] Figure 9A and Figure 9B A block diagram of a multi-zone vapor compression system (MZ-VCS) according to some embodiments is shown. The MZ-VCS includes a set of heat exchangers and compressors configured to control the environment in a set of zones. For each zone, there is at least one heat exchanger. For example, in… Figure 9A In one implementation, each zone 925 or 935 corresponds to a room in the building, enabling the MZ-VCS to simultaneously provide cooling or heating to multiple zones. Figure 9B In the alternative embodiment shown, multiple heat exchangers are placed in a room or area 937 of the building, enabling the MZ-VCS to provide cooling or heating to different parts of the room.

[0147] In this disclosure, a two-zone MZ-VCS is described for clarity, but it should be understood that any number of zones can be used, subject to physical limitations such as refrigerant line length and capacity, compressor pumping power, and building codes. If the zone is an indoor zone (such as a room or part of a room), then the heat exchanger is an indoor heat exchanger.

[0148] The compressor 910 receives low-pressure refrigerant in a vapor state and performs mechanical work to increase the pressure and temperature of the refrigerant. Depending on the configuration of the four-way valve 909, the high-temperature refrigerant can be directed to the outdoor heat exchanger (in which case the system moves heat to the external environment and provides useful cooling, and is referred to as operating in cooling mode), or to the indoor heat exchanger (in which case the system moves heat to one or more indoor areas and provides useful heating, and is referred to as operating in heating mode).

[0149] For clarity and to simplify the following description, the cooling mode is generally considered, i.e., the compressor is connected to the rest of the vapor compression system, as shown by the solid line of the four-way valve 909. However, it should be understood that a similar statement can be made for a system operating in heating mode, wherein the condenser is appropriately replaced by an evaporator and the condensing temperature is replaced by the evaporating temperature.

[0150] In cooling mode, the high-temperature, high-pressure refrigerant moves to the outdoor heat exchanger 915, and an associated optional fan 916 blows air across the outdoor heat exchanger 915. This transfers heat from the refrigerant to the air, causing the refrigerant to condense from vapor into liquid.

[0151] In an idealized description of a vapor compression cycle, the phase change process of the refrigerant—from saturated vapor to a liquid and then back to a saturated liquid—is isothermal; that is, the phase change occurs at a constant temperature, and therefore there is no significant temperature change. However, if heat is further removed from the saturated liquid, the temperature of the saturated liquid decreases by a suitable amount, and the refrigerant is referred to as "subcooled." The subcooling temperature is the temperature difference between the subcooled refrigerant and the temperature of the saturated liquid refrigerant calculated at the same pressure.

[0152] Liquid, high-temperature refrigerant exits the outdoor heat exchanger 915 and is diverted by manifold 917 for distribution among subsequently connected indoor zones 925, 935, or 937. Individual expansion valves 926 and 936 are connected to the inlet manifold. These expansion valves are limiting elements and significantly reduce the refrigerant pressure. Due to the rapid pressure drop and the lack of significant heat exchange within the expansion valves, the refrigerant temperature decreases significantly, a phenomenon described ideally as "adiabatic" in the vapor compression cycle. The final refrigerant exiting the valves is a low-pressure, low-temperature two-phase mixture of liquid and vapor.

[0153] Two-phase refrigerant enters indoor heat exchangers 920 and 930, where fans 921 and 931 blow air across them, respectively. Heat 922 and 932, representing the heat load from the indoor space, is transferred from the zone to the refrigerant, causing it to evaporate from a liquid-vapor two-phase mixture into a saturated vapor state.

[0154] The phase change process of the refrigerant—evaporating from saturated vapor into a two-phase mixture of liquid and vapor, and then back into saturated vapor—is isothermal in an ideal description of the vapor compression cycle; that is, it occurs at a constant temperature, and therefore is a process that takes place without significant temperature changes. However, if further heat is added to the saturated vapor, the temperature of the saturated vapor increases by a suitable amount, and the refrigerant is referred to as "superheated." The superheated temperature is the difference between the superheated refrigerant vapor and the temperature of the saturated vapor calculated at the same pressure.

[0155] The low-pressure refrigerant vapor leaving the indoor unit heat exchanger is reconnected to the common flow path at the outlet manifold 918. Finally, the low-pressure refrigerant vapor returns to the compressor 910, and the cycle repeats.

[0156] The main actuators in the MZ-VCS include compressor 910, outdoor heat exchanger fan 916, indoor heat exchanger fans 921 and 931, and expansion valves 926 and 936. In some systems, the compressor speed can be fixed to one or more predetermined settings or continuously varied. Similarly, the outdoor heat exchanger fan 916 can operate at a fixed speed or continuously vary its speed. In some configurations, the indoor heat exchanger fans 921 and 931 can be determined by the MZ-VCS controller, or their speed can be determined by the occupant if they wish to directly control the indoor airflow. The expansion valves are controlled (e.g., electronically) by control device 200 to continuously vary from a fully closed position to a fully open position (including all possible intermediate positions). Some MZ-VCS implementations replace the electronically controlled expansion valves with a series combination of a solenoid valve for on / off control and a separate variable-opening valve for precise flow rate control.

[0157] The high and low refrigerant pressures are determined by thermodynamic conditions, such as the combined combination of outdoor and indoor air temperatures, compressor speed, and valve opening. Each expansion valve in expansion valves 926 and 936 can be set to different openings, but the total high and low pressures are determined by the total pressure drop across these valves connected in parallel in the refrigerant circuit. It should be noted that there are no pressure-reducing elements between the indoor heat exchangers 920 and 930 and the outlet manifold 918; therefore, all heat exchangers operate at substantially the same pressure. Furthermore, due to the previously mentioned isothermal characteristics of the phase change, all indoor heat exchangers are confined to evaporation at the same temperature.

[0158] According to the implementation, the heating or cooling capacity of indoor heat exchangers 920 and 930 can be modulated by the duty cycle that cycles each heat exchanger between "on" and "off" operating modes. When the inlet valve controlling the refrigerant flow is closed, heat exchangers 920 and 930 are off, or alternatively, the pumping of refrigerant through the system compressor 910 stops, so that heat exchangers 920 and 930 do not perform cooling or heating. When the inlet valve is open and the compressor 910 is running, heat exchangers 920 and 930 are on, so that heat exchangers 920 and 930 in indoor zones 925 and 935 operate at their full heat capacity. The controller determines how to alternate between modes based on the difference between the indoor zone temperature and the desired indoor zone temperature.

[0159] However, the action of switching heat exchangers on and off (especially in MZ-VCS, where zone heat exchangers can switch on and off independently of each other, resulting in continuous variations in system outputs such as zone temperature and heat exchanger temperature) is known to be energy inefficient and reduces occupant comfort. Therefore, a control system and method are needed to smoothly control the heat capacity of heat exchangers (such as those in MZ-VCS).

[0160] Some implementations are based on the understanding that the probabilistic surrogate model 102 can be advantageously used to reduce the complexity of controlling the MZ-VCS. Specifically, in an MZ-VCS with a large number of setpoints, brute-force sampling of the setpoints can lead to excessive computational overhead and long tuning times. Sample efficiency is greatly reduced by using a Bayesian optimization algorithm, which can balance exploration and exploitation via the probabilistic surrogate model 102. According to some implementations, scalable probabilistic machine learning algorithms are needed for MZ-VCS with a large number of setpoints. In some implementations, scalable probabilistic machine learning algorithms can be based on Bayesian deep neural networks or variational autoencoders.

[0161] result

[0162] Figure 10 Performance results for a standard time-varying ESC, an Adam-accelerated time-varying ESC, and a Bayesian-optimized hot-start ESC, according to some implementations, are shown. To evaluate performance, a system including a vapor compression cycle with a feedback controller is considered. The feedback controller is configured to regulate the cooling capacity to a setpoint by controlling the compressor speed. Such a system facilitates the study of optimization methods for the vapor compression cycle without involving the slower and lower-pass thermodynamics that consume space. Furthermore, optimization methods (such as standard time-varying ESC, Adam-accelerated time-varying ESC, and Bayesian-optimized hot-start ESC) are applied to adjust system inputs (such as expansion valve position, indoor fan speed (IFS), and outdoor fan speed (OFS)) to optimize power consumption.

[0163] For the case study, a PI loop was selected to regulate the cooling capacity of the vapor compression cycle to 2kW, and optimization methods were applied to determine the optimal values ​​for the expansion valve position and fan speed. The PI gain was tuned offline (i.e., in advance). The PI gain tuning mechanism for the setpoint was unknown.

[0164] In power curve 1000, 1002 represents the standard time-varying ESC, 1004 represents the Adam-accelerated time-varying ESC, and 906 represents the Bayesian-optimized warm-start ESC (hybrid). From power curve 1000, it can be observed that the standard time-varying ESC 1002 and the Adam-accelerated DE time-varying ESC 1004 exhibit slower convergence compared to the Bayesian-optimized warm-start ESC 906. The Bayesian-optimized warm-start ESC 906 converges to a steady-state cycle power of 0.335 kW within 3 hours, while other algorithms require more than 10 hours to approach the minimum cycle power.

[0165] According to the implementation, the difference in steady-state cycle power is caused by the difference in the optimal LEV (linear expansion valve) position determined by the standard time-varying ESC 1002, the Adam-accelerated time-varying ESC 1004, and the Bayesian-optimized hot-start ESC 906. Since the standard time-varying ESC 1002 and the Adam-accelerated time-varying ESC 1004 depend on the initial LEV position, they converge to a steady-state LEV position of approximately 300 counts. On the other hand, the Bayesian-optimized hot-start ESC 906 determines that a steady-state value of 260 counts yields significantly better energy efficiency. Because the Bayesian-optimized hot-start ESC 906 includes an exploration phase, at the end of this phase, the BO hot-start produces a locally effective set of tuning parameters. However, local gradient information still remains available online, leading to further reductions in power consumption after the Bayesian-optimized hot-start. Furthermore, it can be observed from curves 1008 and 1010 that the capacity setpoint was obtained throughout the setpoint tuning process, but the compressor frequency associated with the LEV position gradually decreased to approximately 17 Hz instead of maintaining 40 Hz as expected.

[0166] Figure 11 Performance results of Bayesian-optimized warm-start ESC and standard time-varying ESC for complex systems, according to some implementations, are illustrated. The complex system includes... Figure 10 The description describes the same loop, but incorporates two feedback loops: one loop regulates the room temperature to the setpoint via compressor speed control, while the other loop regulates the evaporator superheat temperature (CITE) to the setpoint of 5°C via expansion valve position control. This complex system is connected to a 27m³ volume... 3 A cubic room model with a 20cm thick cladding exhibiting the thermal properties of wood was constructed, with external and internal heat transfer coefficients of 1.5 W / m². -2 and 3.5Wm -2 This complex system operates based on a commercial office building that is occupied during the day and unoccupied at night. As a result, the sensible and latent heat loads in the space are set to 3200W and 800W respectively between 8:00 AM and 6:00 PM, and 2000W and 100W respectively at other times. It is assumed that the ambient temperature varies sinusoidally between 24°C and 40°C over a 24-hour period, with a peak temperature at 2:00 PM.

[0167] Furthermore, consider two tunable inputs: indoor fan speed (IFS) and outdoor fan speed (OFS). Assume both tunable inputs are bounded, and that the boundaries are known; for example, IFS is bounded at [200, 500] rpm, and OFS at [500, 900] rpm. While standard time-varying ESCs and Bayesian-optimized hot-start ESCs assume that the heat load can be measured at the current moment, neither provides prior information about temperature or heat distribution.

[0168] The Bayesian-optimized hot-start ESC uses the first 18 hours to collect data for each load mode (i.e., when the thermal load is high and when it is low). For each load mode, an energy-efficiency surrogate model (e.g., surrogate model 102) and the corresponding optimal IFS and OFS pair are learned. After 18 hours, the Adam-accelerated time-varying ESC takes over and applies gradient-based updates to the IFS and OFS. Performance results show that the standard time-varying ESC exhibits limited flexibility. For example, the IFS oscillates within a narrow range of 290 rpm to 310 rpm, while the OFS is stagnant. In contrast, in the Bayesian-optimized hot-start ESC, the surrogate model allows for rapid reset of the ESC position, which in turn leads to improved energy efficiency. Moreover, it can be noted that the Bayesian-optimized hot-start ESC results in improvements over several days; for example, the power reduction after 48 hours (approximately 60 to 70 hours) is more significant than that after 24 to 48 hours.

[0169] Figure 12 An example is illustrated using control device 200 to control heating, ventilation, and air conditioning (HVAC) system 1210 according to some embodiments. "HVAC" system refers to any heating, ventilation, and air conditioning (HVAC) system that implements a vapor compression cycle. HVAC system 1210 spans a very wide range of systems, from systems that only supply outdoor air to the occupants of a building to systems that only control the temperature of a building, or systems that control both temperature and humidity.

[0170] HVAC system 1210 is configured to condition room 1200. Room 1200 is occupied by occupants 1202, 1204, 1206, and 1208. Arrow 1214 indicates the air supplied by HVAC system 1210 to condition room 1200. Based on a proxy model (e.g., proxy model 102), a controller (e.g., control device 200) determines the optimal combination of setpoints that minimizes the power consumption of HVAC system 1210.

[0171] Furthermore, the optimal combination of setpoints is input to the feedback controller 1212 associated with the HVAC system 1210. The feedback controller 1212 generates control commands based on the optimal combination of setpoints. The feedback controller 1212 also controls the HVAC system 1210 according to the control commands to minimize the power consumption of the HVAC system 1210.

[0172] Optimization based on surrogate modeling (such as Bayesian optimization with a hot start using surrogate model 102 for ESC) yields decisive advantages. For example, machine learning methods enable the construction of surrogate models of a vapor compression system's input-to-output performance, for example via Gaussian process regression, regardless of noise in the measurements. However, model-free methods (such as ESC) estimate gradients directly from noisy data, which can lead to undesirable effects such as oscillations or even divergence in the neighborhood of the optimal input. Surrogate modeling also learns the input-output model while implementing a termination criterion that can be stored, for example, in the cloud, to obtain a good initial guess of the input (hot start) to reach the peak in subsequent runs. Furthermore, the convergence rate is improved with surrogate modeling via smoothing regressors because smoothing optimization methods, such as first-order and second-order methods, can be directly applied to the surrogate model.

[0173] The above description provides only exemplary embodiments and is not intended to limit the scope, applicability, or configuration of this disclosure. Rather, the above description of the exemplary embodiments will provide those skilled in the art with a description of what can be accomplished to implement one or more of the exemplary embodiments. Various changes to the function and arrangement of the elements are contemplated without departing from the spirit and scope of the disclosed subject matter set forth in the appended claims.

[0174] Specific details are set forth in the foregoing description to provide a thorough understanding of the embodiments. However, those skilled in the art will understand that these embodiments can be practiced without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in block diagram form to avoid obscuring these embodiments with unnecessary detail. In other cases, well-known processes, structures, and techniques may be shown without unnecessary detail to avoid obscuring these embodiments. Moreover, like reference numerals and designations in the various figures denote the same elements.

[0175] Furthermore, each implementation can be described as a process, depicted as a flowchart, block diagram, data flow diagram, structural diagram, or block diagram. Although a flowchart can describe operations as sequential processes, many operations can be executed in parallel or simultaneously. Additionally, the order of operations can be rearranged. A process may terminate upon completion of its operations, but may have additional steps not discussed or included in the diagrams. Moreover, not all operations in any particular described process may occur in all implementations. A process can correspond to a method, function, procedure, subroutine, subroutine, etc. When a process corresponds to a function, the termination of that function may correspond to the function returning to the calling function or the main function.

[0176] Furthermore, the implementation of the disclosed subject matter can be carried out at least partially, manually, or automatically. Manual or automatic implementation can be performed or at least assisted by using machines, hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments that perform the necessary tasks can be stored in a machine-readable medium. The processor can perform the necessary tasks.

[0177] The various methods or processes outlined herein can be encoded as software that can be executed on one or more processors employing any of a variety of operating systems or platforms. Furthermore, such software can be written using any of a variety of suitable programming languages ​​and / or programming or scripting tools, and can also be compiled into executable machine language code or intermediate code that executes on a framework or virtual machine. Typically, in various implementations, the functionality of program modules can be combined or distributed as needed.

[0178] Embodiments of this disclosure can be specifically implemented as the methods for which examples have been provided. Actions performed as part of this method can be arranged in any suitable manner. Therefore, even though actions are shown as sequential in the exemplary embodiments, embodiments can be constructed that perform actions in a different order than illustrated, which may include performing some actions simultaneously. Although this disclosure has been described with reference to specific preferred embodiments, it is to be understood that various other changes and modifications can be made within the spirit and scope of this disclosure. Therefore, aspects of the appended claims cover all such changes and modifications that fall within the true spirit and scope of this disclosure.

Claims

1. A controller for controlling operation of a vapour compression system, VCS, the controller comprising: At least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the controller to: The operation of the VCS is controlled by different combinations of setpoints for different actuators of the VCS, and the operating cost of the VCS is estimated for each combination of the different sets of setpoints. Using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, a probabilistic surrogate model is computed. The probabilistic surrogate model provides a probabilistic mapping between various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the actuators of the VCS, wherein the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping. Based on the acquisition function of the first two moments of the operating cost, an optimal combination of setpoints is selected, wherein the optimal combination of setpoints has the maximum likelihood of being the global minimum at the surrogate model; and The optimal combination of selected setpoints is submitted to the feedback controller to change the state of the actuator according to the corresponding setpoint of the actuator in the selected optimal combination. The processor is further configured to execute an extremum search controller that causes an input perturbation to the feedback controller, starting from an optimal combination of selected setpoints, to tune the optimal combination of selected setpoints based on a gradient of the operating cost caused by the perturbation. The processor is further configured to estimate the gradient of the operating cost caused by the perturbation by mapping the perturbed value of the optimal combination of selected setpoints to the value of the operating cost with respect to the surrogate model according to the acquisition function.

2. The controller of claim 1, wherein, After the VCS has been operating according to a combination of setpoints for a predetermined period of time during which the VCS has not reached a steady state, the operating cost for the combination of setpoints is estimated.

3. The controller of claim 1, wherein, At least some of the operating costs are estimated during the transient period of the VCS operating according to the corresponding combination of the setpoints.

4. The controller of claim 1, wherein, The processor is also configured to estimate the operating cost for the optimal combination of selected setpoints, and to update the surrogate model using the Bayesian optimization with the optimal combination of selected setpoints and the corresponding estimated operating cost.

5. The controller of claim 4, wherein, The processor is also configured to recursively update the proxy model with the optimal combination of the latest selected setpoints based on the acquisition function applied to the updated proxy model, until a termination condition is met.

6. The controller of claim 1, wherein, The Bayesian optimization uses one or a combination of Gaussian process regression, neural process regression, and machine learning to determine the probabilistic surrogate model.

7. The controller of claim 1, wherein, The Bayesian optimization includes using the probability mapping provided by the surrogate model to direct the query to the result setpoint.

8. The controller of claim 1, wherein, The VCS is a multi-zone vapor compression system (MZ-VCS), which includes at least a single compressor connected to multiple heat exchangers arranged in one or more indoor zones.

9. The controller of claim 1, wherein, The operating cost corresponding to the combination of the setpoints is the power consumption of the VCS operating according to the combination of the setpoints.

10. The controller of claim 1, wherein, The feedback controller is one or a combination of a PI controller, a PID controller, an MPC controller, or a robust controller.

11. A method for controlling the operation of a vapor compression system (VCS), the method comprising the following steps: The operation of the VCS is controlled by different combinations of setpoints for different actuators of the VCS, and the operating cost of the VCS is estimated for each combination of the different sets of setpoints. Using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, a probabilistic surrogate model is computed. The probabilistic surrogate model provides a probabilistic mapping between various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the actuators of the VCS, wherein the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping. Based on the acquisition function of the first two moments of the operating cost, an optimal combination of setpoints is selected, wherein the optimal combination of setpoints has the maximum likelihood of being the global minimum at the surrogate model; and The optimal combination of selected setpoints is submitted to the feedback controller to change the state of the actuator according to the corresponding setpoint of the actuator in the selected optimal combination. The method further includes executing an extremum search controller, which causes an input perturbation to the feedback controller, starting from an optimal combination of selected setpoints, to tune the optimal combination of selected setpoints based on a gradient of the operating cost caused by the perturbation. The gradient of the operating cost caused by the perturbation is estimated by mapping the perturbation value of the optimal combination of selected setpoints to the value of the operating cost with respect to the surrogate model according to the acquisition function.

12. A non-transitory computer-readable storage medium having a program implemented thereon, the program being executable by a processor to perform a method for controlling the operation of a vapor compression system (VCS), the method comprising the steps of: The operation of the VCS is controlled by different combinations of setpoints for different actuators of the VCS, and the operating cost of the VCS is estimated for each combination of the different sets of setpoints. Using Bayesian optimization of the combinations of setpoints and their corresponding estimated operating costs, a probabilistic surrogate model is computed. The probabilistic surrogate model provides a probabilistic mapping between various combinations of setpoints for different actuators of the VCS and the corresponding operating costs of the actuators of the VCS, wherein the probabilistic surrogate model defines at least the first two moments of the operating costs in the probabilistic mapping. Based on the acquisition function of the first two moments of the operating cost, an optimal combination of setpoints is selected, wherein the optimal combination of setpoints has the maximum likelihood of being the global minimum at the surrogate model; and The optimal combination of selected setpoints is submitted to the feedback controller to change the state of the actuator according to the corresponding setpoint of the actuator in the selected optimal combination. An extreme value search controller is executed, which causes an input perturbation to the feedback controller, starting from an optimal combination of selected setpoints, to tune the optimal combination of selected setpoints based on the gradient of the operating cost caused by the perturbation. The gradient of the operating cost caused by the perturbation is estimated by mapping the perturbation value of the optimal combination of selected setpoints to the value of the operating cost with respect to the surrogate model according to the acquisition function.

Citation Information

Patent Citations

  • Extremum seeking controller and method for controlling a vapor compression system

    CN106133462A

  • Vapor compression system, and control system and method for controlling operation of vapor compression system

    JP2013148336A

  • Search device, search method and manufacturing system

    JP2019152986A