System and method for data driven control of air conditioning system

Abstract

A system for controlling operation of an air conditioning system including a heat exchanger is provided. The system includes a processor executing a neural network trained to simulate operation of the heat exchanger for a test control input, generating a simulated output based on historical data defining a state of the heat exchanger. The historical data includes a historical control input sequence provided to the heat exchanger and a historical output sequence of operation of the heat exchanger corresponding to the historical control input sequence. The processor determines a control command to the air conditioning system based on a predicted test output for a simulation of operation of the heat exchanger for the test control input, and transmits the determined control command to an actuator of the air conditioning system.

Description

Technical Field

[0001] The present disclosure relates generally to a control system, and more particularly, to a control system and method for data-driven control of an air conditioner including a heat exchanger. Background Art

[0002] Control theory in control systems engineering is the subfield of mathematics that deals with the control of continuously operating dynamic systems in engineering processes and machines. Control systems aim to perform control operations on dynamic systems using control actions to ensure control stability.

[0003] For example, optimization-based control and estimation techniques such as model predictive control (MPC) allow for a model-based design framework in which system dynamics and constraints can be directly considered. MPC is used in many applications to control dynamic systems of varying complexity. Examples of such systems include production lines, automotive engines, robots, CNC machining, satellites, and generators. However, in some cases, the analytical model of the controlled system may be unavailable, difficult to update in real time, or inaccurate. Examples of this situation are prevalent in the fields of robotics, heating, ventilation, and air conditioning (HVAC) systems, vehicles, smart grids, factory automation, transportation, self-regulating machines, and transportation networks.

[0004] For example, heat exchangers are widely used to transfer heat between air and working fluid in refrigeration, comfort cooling, and heating applications, and play a major role in the performance of heating, ventilation, air conditioning, and refrigeration (HVAC&R) equipment. However, the design and optimization of heat exchangers can be challenging due to a large number of design variables, complex manufacturing constraints, and multiple conflicting objectives.

[0005] Typically, the dynamics of a heat exchanger can be described by conservation laws (i.e., balance of mass, energy, and momentum). In physics-based modeling approaches, in order to predict the heat transfer and fluid flow dynamics occurring in a heat exchanger, it may be necessary to solve the resulting differential algebraic equations (DAEs). However, DAE systems can be difficult to solve, and challenges can arise from exponential reduction, consistent initialization, and inherent stiffness. In addition, for large-sized heat exchangers with complex circuits, the model complexity (number of equations) increases significantly. Since the temporal dynamics of the HVAC system can often be dominated by the temporal dynamics of the heat exchanger, simulation-based equipment design processes and other applications of predictive models are limited by the computational complexity of the heat exchanger model. Summary of the invention

[0006] Technical issues

[0007] In the absence of an accurate model of a dynamic system, some control methods utilize operational data generated by these systems in order to construct feedback control strategies for stabilizing the system dynamics or embedding quantifiable control-related performance. Control using operational data is called data-driven control. Generally, there can be two types of data-driven control methods, such as indirect methods and direct methods. In indirect methods, a model of the system is first built, and then the model is used to design the controller. However, in direct methods, the control strategy from the operational data is directly constructed without an intermediate model building step. The disadvantage of indirect methods is that a large amount of data may be required during the model building stage. In contrast, direct methods require less data. However, even the most advanced direct control methods have difficulties in handling state and input constraints, which are essential for maintaining safe operation in closed-loop control systems.

[0008] Therefore, an indirect data-driven approach is needed to build a model using a smaller amount of data to control one or more heat exchangers.

[0009] Technical Solution

[0010] The present disclosure provides a control system and method for controlling the operation of an air conditioning system including a heat exchanger. The proposed control system and method can be used to generate a control strategy for controlling the dynamic behavior of the air conditioning system including at least one heat exchanger using control actions in an optimal manner without delay or overshoot, and ensuring control stability of the air conditioning system.

[0011] It is an object of some embodiments to provide a control system and method for data-driven control of an air conditioning system including a heat exchanger. Additionally or alternatively, it is an object of some embodiments to provide a control system and method for indirect data-driven control using a neural network model derived from actual and predetermined windows of historical data of the operation of the heat exchanger.

[0012] Some embodiments are based on the recognition that the dynamics of a heat exchanger can be replaced by a digital twin that simulates the operation of the heat exchanger and outputs the variables that a model of the heat exchanger (such as a neural network model) needs to estimate. The simulation of the operation of the heat exchanger by the digital twin model can reduce historical data compared to conventional indirect data-driven models of the heat exchanger.

[0013] Some embodiments are based on the recognition that conventional analytical models for air conditioning systems including heat exchangers are based on physical conservation laws and require complex differential algebraic equations (DAEs) that are often difficult to solve. Furthermore, the increase in the number and size of heat exchangers in air conditioning systems leads to an increase in the complexity of DAEs.

[0014] Some embodiments are based on the recognition that neural networks (such as convolutional recurrent networks) accelerate the simulation of nonlinear dynamics of heat exchangers. To this end, a deep state-space modeling framework is constructed that combines the feature extraction capabilities of convolutional neural networks (CNNs) with the sequence prediction properties of gated recurrent units (GRUs). The modeling framework describes the fluid flow and heat transfer dynamics of heat exchangers in vapor compression cycles used to construct energy systems.

[0015] Unlike modeling based on physical conservation laws, the proposed modeling approach develops the relationship between input and output data without explicit knowledge of the underlying behavior and does not involve any DAE. In addition, the number of model equations generated by the proposed modeling approach is independent of the geometry, size, circuit mode, and operating conditions of the heat exchanger. Therefore, the resulting model is effective for real-time solution, making it effective for online control of any air-conditioning system including a heat exchanger.

[0016] To this end, some embodiments disclose a neural network having a combination of convolutional and recurrent networks. The neural network is trained to simulate the operation of a heat exchanger for a test control input to generate a simulated output based on historical data defining the state of the heat exchanger. The historical data includes a historical control input sequence for the heat exchanger and a historical output sequence of the operation of the heat exchanger corresponding to the historical control input sequence.

[0017] The neural network includes a first arm that processes boundary inputs appended to a sequence of historical boundary inputs using a first combination of convolutional and recurrent networks, the first combination of convolutional and recurrent networks being trained to extract features from the historical boundary inputs that indicate changes in the boundary inputs. The neural network also includes a second arm that processes a sequence of historical boundary inputs paired with a sequence of historical outputs using a second combination of convolutional and recurrent networks, the second combination of convolutional and recurrent networks being trained to extract output features that indicate dynamic coupling between inputs and outputs of operation of a heat exchanger. The neural network also includes a third arm that processes control features and output features to predict a test output of the operation of the heat exchanger corresponding to the boundary inputs.

[0018] Some embodiments disclose that each of the first and second arms of a neural network includes a group of one-dimensional convolutional layers (such as a CNN) and a gated recurrent unit (GRU). In this way, the CNN-GRU uses the concept of a one-dimensional (1-D) convolutional network to perform a first-stage analysis of time series data. This enables better extraction of temporal features from input / output. The GRU is an efficient recurrent neural unit that analyzes the extracted temporal features to learn system dynamics. In addition, dividing the data into an input time series (analyzed by the first arm) and the corresponding output (analyzed by the second arm) enables the neural network to learn the impact of input changes on predicted outputs, as well as the coupling relationship between inputs and outputs in a dynamic system, respectively.

[0019] Some embodiments are based on the recognition that the third arm can be implemented as a fully connected deep neural network that accepts a tensor including control features and output features. The purpose of the fully connected deep neural network layer is to project the enhanced GRU state to the output space.

[0020] Thus, the first arm analyzes the input time series and the second arm analyzes the historical system outputs. Thus, the neural network is made aware of the immediately preceding inputs and corresponding outputs, thereby enabling it to understand the current state of the heat exchanger and / or air conditioning system. With knowledge of the system dynamics, the current system state, and the next input, the neural network can accurately predict the output for the next time step.

[0021] Some embodiments disclose a control system having a controller that determines control commands to actuators of an air conditioning system based on the predicted output of a system model. The control commands can be transmitted to the air conditioning system to change the speed of a compressor, open or close a valve, or change the speed of a fan of a machine, etc.

[0022] Some embodiments are based on the recognition that the path of the first arm and the path of the second arm are assigned different neural weights. Different neural weights may be assigned to improve the accuracy of the neural network.

[0023] Some embodiments are based on the recognition that a neural network can be trained by a data trajectory generated by running a physics-based model of a heat exchanger under one or more boundary conditions associated with the heat exchanger. During training of the neural network, the data trajectory is divided into multiple batches. Each batch includes at least one of a historical control input sequence, a historical output sequence, a first input to the machine, and a true output of the machine. Predefined data windows in the batch are processed instead of the entire data trajectory, which reduces the back-propagation time in GRU training. In addition, avoiding the requirement to learn from the entire time series at once and instead using batch processing can achieve scalability, parallelism, and improve storage efficiency when faced with large data sets.

[0024] Some embodiments disclose that the mean square error of each batch in a plurality of batches can be calculated based on the predicted output of the heat exchanger neural network model and the actual output of the air conditioning system. In addition, the training loss associated with the mean square error of each batch calculated is determined and optimized.

[0025] Some embodiments disclose that at each time step of the simulation, a model of the air conditioning system is initialized, the model including a room model, a compressor model, an expansion model, and a neural network model of a heat exchanger. In addition, the boundary conditions of the air side of the air conditioning system and the settings of the compressor and the expansion device can be determined. Then, the model of the air conditioning system and the room model are executed based on these conditions. One or more states including a refrigerant state or an air state are generated that are associated with the room model and multiple component models. Processing each component model instead of a single model for the entire machine can reduce the data requirements for predicting output data. In addition, processing a data-driven model with a similar neural network architecture for each heat exchanger is convenient in controlling the dynamic behavior of an air conditioning system with multiple heat exchangers.

[0026] Some embodiments are based on the recognition that a plurality of test control inputs can be generated based on a Gaussian process. The plurality of test control inputs are then provided as inputs to a model of an air conditioning system to generate a plurality of predicted test outputs corresponding to the plurality of test control inputs. Control commands are determined based on the predicted test outputs of the simulation of the air conditioning system for the selected test control inputs based on an optimization of a cost function associated with the plurality of test control inputs and the plurality of test outputs.

[0027] Some embodiments are based on the recognition that using one or more test control inputs at a previous time instant, a plurality of test control inputs for a next time instant are generated based on a Gaussian process, the next time instant being after the previous time instant.

[0028] Some embodiments are based on the recognition that a test control input sequence corresponding to a finite number of moments can be generated. The test control input sequence is provided as an input to a model of the system to generate a predicted test output sequence corresponding to the test control input sequence. Then, the gradient of a cost function associated with the test control input sequence and the test output sequence is calculated. In addition, the best test control input from the generated test control input sequence is selected based on the gradient of the calculated cost function, and a control command is determined based on the predicted test output of the simulation of the operation of the air conditioning system for the selected test control input. Such a process can select the best test control input for the model of the air conditioning system.

[0029] In the following detailed description, the present disclosure will be further described by way of non-limiting examples of exemplary embodiments of the present disclosure and with reference to the various drawings mentioned, in which like reference numerals represent similar parts in the several views of the drawings. The drawings shown are not necessarily drawn to scale, but rather generally emphasize the principles of the presently disclosed embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] [ Figure 1 ]

[0031] Figure 1 A schematic diagram of an air conditioning system according to some embodiments of the present disclosure is shown.

[0032] [ Figure 2A ]

[0033] Figure 2A A block diagram of a control system for controlling an air conditioning system using a model according to some embodiments of the present disclosure is shown.

[0034] [ Figure 2B ]

[0035] Figure 2B A block diagram of a neural network for predicting a test output of operation of a heat exchanger corresponding to a test control input in a single time step prediction is shown, according to some embodiments of the present disclosure.

[0036] [ Figure 3A ]

[0037] Figure 3A Detailed architecture of a neural network trained to simulate the operation of a heat exchanger included in an air conditioning system at a time step according to some embodiments of the present disclosure is shown.

[0038] [ Figure 3B ]

[0039] Figure 3B A flowchart illustrating an exemplary process for training a neural network according to some embodiments of the present disclosure is shown.

[0040] [ Figure 3C ]

[0041] Figure 3C A flow chart for optimizing training loss for training a neural network according to some embodiments of the present disclosure is shown.

[0042] [ Figure 4 ]

[0043] Figure 4 An example diagram of a model of an air conditioning system according to some embodiments of the present disclosure is shown.

[0044] [ Figure 5 ]

[0045] Figure 5 A flow chart for generating one or more states associated with a room model and multiple component models of an air conditioning system according to some embodiments of the present disclosure is shown.

[0046] [ Figure 6 ]

[0047] Figure 6A block diagram illustrating different components of a control system for controlling an air conditioning system according to some embodiments of the present disclosure.

[0048] [ Fig. 7A ]

[0049] Fig. 7A A flow chart of an exemplary method for selecting an optimal test control input from a plurality of test control inputs based on a Gaussian process according to some embodiments of the present disclosure is shown.

[0050] [ Figure 7B ]

[0051] Figure 7B A flow chart depicting the generation of a plurality of test control inputs according to some embodiments of the present disclosure is shown.

[0052] [ Figure 8 ]

[0053] Figure 8 A flow chart of an exemplary method for selecting an optimal test control input from a sequence of test control inputs based on a model predictive control process according to some embodiments of the present disclosure is shown.

[0054] [ Fig. 9 ]

[0055] Fig. 9 A flow chart of a method for determining a control command for an air conditioning system according to some embodiments of the present disclosure is shown.

[0056] [ Fig.10 ]

[0057] Fig.10 A block diagram illustrating a use case for controlling an air conditioning system using a control system according to some embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0058] In the following description, for the purpose of explanation, many specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent to those skilled in the art that the present disclosure can be implemented without these specific details. In other cases, the devices and methods are shown only in block diagram form to avoid obscuring the present disclosure. It is contemplated that various changes may be made to the functions and arrangements of the elements without departing from the spirit and scope of the subject matter disclosed in the appended claims.

[0059] As used in this specification and claims, the terms "for example," "for example," and "such as," and the verbs "include," "have," "comprise," and their other verb forms, when used in conjunction with a list of one or more components or other items, are interpreted as open-ended, meaning that the list should not be viewed as excluding other additional components or items. The term "based on" means based at least in part on. In addition, it should be understood that the words and terms used herein are for descriptive purposes and should not be considered limiting. Any headings used in this specification are for convenience only and have no legal or limiting effect.

[0060] Specific details are given in the following description to provide a comprehensive understanding of the embodiments. However, it will be appreciated by those skilled in the art that the embodiments can be implemented without these specific details. For example, the systems, processes, and other elements in the disclosed subject matter can be displayed as components in block diagram form to avoid obscuring the embodiments with unnecessary details. In other cases, well-known processes, structures, and techniques can be displayed without unnecessary details to avoid obscuring the embodiments. In addition, the same reference numerals and names in the various figures indicate the same elements.

[0061] When describing embodiments of the present disclosure, the following definitions apply throughout the disclosure.

[0062] A "control system" or "controller" may refer to a device or group of devices used to manage, command, direct or regulate the behavior of other devices or systems. A control system may be implemented by software or hardware and may include one or more modules. A control system including a feedback loop may be implemented using a microprocessor. A control system may be an embedded system.

[0063] "Vapor compression system" refers to a system that uses a vapor compression cycle to move a refrigerant through the components of the system based on the principles of thermodynamics, fluid mechanics, and / or heat transfer. Vapor compression systems can be, but are not limited to, heat pumps, refrigeration, and air conditioning systems. Vapor compression systems are used in applications other than conditioning residential or commercial spaces. For example, vapor compression cycles can be used to cool computer chips in high-performance computing applications.

[0064] "Components of a vapor compression system" means any component of a vapor compression system having operations that can be controlled by a control system. These components include, but are not limited to: a compressor with a variable speed for compressing the refrigerant and pumping the refrigerant through the system; an expansion valve for providing an adjustable pressure drop between the high pressure and low pressure portions of the system; and an evaporative heat exchanger and a condensing heat exchanger, each of which includes a variable speed fan for regulating the air flow rate through the heat exchanger.

[0065] "Evaporator" refers to the heat exchanger in a vapor compression system in which the refrigerant passing through the heat exchanger evaporates over the length of the heat exchanger so that the specific enthalpy of the refrigerant at the heat exchanger outlet is higher than the specific enthalpy of the refrigerant at the heat exchanger inlet, and the refrigerant generally changes from a liquid to a gas. There may be one or more evaporators in a vapor compression system.

[0066] "Condenser" means a heat exchanger in a vapor compression system in which the refrigerant passing through the heat exchanger condenses over the length of the heat exchanger so that the specific enthalpy of the refrigerant at the heat exchanger outlet is lower than the specific enthalpy of the refrigerant at the heat exchanger inlet, and the refrigerant generally changes from a gas to a liquid. There may be one or more condensers in a vapor compression system.

[0067] A "refrigerant circuit" refers to the interconnection of refrigerant piping and components into a closed structure so that the refrigerant flows in a closed path between a series of components. A refrigerant circuit can be used to construct a closed thermodynamic cycle to efficiently transfer thermal energy from one location to another. For example, the refrigerant circuit of a vapor compression system includes a compressor, a condensing heat exchanger, an expansion valve, and an evaporating heat exchanger, as well as piping for transferring the refrigerant from each component to the next.

[0068] A “circuit” refers to the interconnections of wires that carry electrical signals between components such as processors, memory, or actuators.

[0069] A "set of control signals" refers to specific values ​​of inputs used to control the operation of components of a vapor compression system. The set of control signals includes, but is not limited to, values ​​of compressor speed, expansion valve position, evaporator fan speed, and condenser fan speed.

[0070] "Set point" refers to the target value of a system, such as a vapor compression system, that is intended to be achieved and maintained by operation. The term "set point" applies to any specific value of a specific set of control signals and thermodynamic and environmental parameters.

[0071] "Computer" means any device that can accept structured input, process the structured input according to specified rules, and produce the processed results as output. Examples of computers include general-purpose computers, supercomputers, mainframes, supermicrocomputers, microcomputers, workstations, minicomputers, servers, interactive televisions, hybrid combinations of computers and interactive televisions, and special-purpose hardware for emulating computers and / or software. A computer may have a single processor or multiple processors, which may operate in parallel and / or non-parallel. A computer also refers to two or more computers connected together by a network to transmit or receive information between the computers. Examples of such computers include distributed computer systems for processing information via computers connected by a network.

[0072] A "central processing unit (CPU)" or "processor" refers to a computer or the component of a computer that reads and executes software instructions.

[0073] "Memory" or "computer-readable medium" refers to any storage device for storing data accessible by a computer. Examples include magnetic hard disks; floppy disks; optical disks such as CD-ROMs or DVDs; magnetic tapes; memory chips; and carrier waves for carrying computer-readable electronic data, such as those used to send and receive electronic mail or access the Internet, and computer storage such as random access memory (RAM).

[0074] "Software" refers to the prescribed rules for operating a computer. Examples of software include software, code segments, instructions, computer programs, and programming logic. The software of an intelligent system may have self-learning capabilities.

[0075] A "module" or "unit" refers to a basic component in a computer that performs a task or part of a task. A "module" or "unit" can be implemented by software or hardware.

[0076] "Controller", "control system" and / or "regulator" refers to a device or a group of devices used to manage, command, direct or regulate the behavior of other devices or systems. The controller can be implemented by hardware, a processor configured to operate by software, and a combination thereof. The controller can be an embedded system.

[0077] Figure 1A schematic diagram 100 of an air conditioning system 102 according to an embodiment of the present disclosure is shown. The schematic diagram 100 includes a space 104 and an air conditioning system 102 configured to provide heating or cooling to the space 104. The air conditioning system 102 may include various components, such as a variable setting actuator for performing an operation such as a refrigerant cycle. For example, the air conditioning system 102 may include a variable speed compression device 106, an outdoor heat exchanger 108, a variable speed outdoor unit fan 110, an expansion device 112, an indoor heat exchanger 114, and an air circulation blower 116. Various components may be connected in a closed loop series refrigerant flow arrangement. The air conditioning system 102 may act as a heating and cooling device, and may provide thermal comfort to a user by adjusting the required capacity to match the load in the space 104. The air is conditioned by the indoor heat exchanger 114 and transmitted to the space 104 by the air circulation blower fan 116. After absorbing or discharging the heat of the space 104, the air may be circulated back to the indoor heat exchanger 114 according to the operating mode of the air conditioning system 102.

[0078] Typically, the temporal dynamics of the air conditioning system 102 depends primarily on the dynamics of a heat exchanger of the air conditioning system 102. The heat exchanger of the air conditioning system 102 may be an indoor heat exchanger 114 or an outdoor heat exchanger 108. The dynamics of the air conditioning system 102 may be described by conservation laws, i.e., the balance of mass, energy, and momentum. In a physics-based analytical model of a heat exchanger, complex differential algebraic equations (DAEs) need to be solved to predict the heat transfer and fluid flow dynamics occurring in the heat exchanger. Furthermore, for large-sized heat exchangers with complex circuits, the complexity of the physics-based analytical model may increase significantly.

[0079] An object of some embodiments is to disclose an apparatus for data-driven control of a system including a heat exchanger, such as an air conditioning system 102. Another object of some embodiments is to disclose a method for data-driven control of a system including a heat exchanger. The proposed apparatus of the present disclosure may include a control system for data-driven control of the dynamic behavior of the air conditioning system 102 including a heat exchanger. The control system and method proposed in the present disclosure eliminate the necessity of solving these complex DAEs. Typically, a single data-driven model of the entire air conditioning system 102 requires a large amount of data to predict the output. In order to collect a large amount of data, a physics-based model of the air conditioning system 102 needs to be executed with various inputs, which can be a cumbersome task and increases the possibility of errors. In addition, even if the system architecture changes slightly, the data-driven system model cannot be reused unless additional inputs are included to accommodate these changes, which will inevitably increase the model dimension. Instead, a modular-based solution approach is more preferred. Specifically, at the component level, a data-driven model is used to completely or partially replace the physics-based component model. At the system level, the conservation of mass, energy, and momentum is achieved by following the first principles. Therefore, the data-driven model for controlling the components of the air conditioning system 102 according to the present disclosure can reduce data requirements and eliminate the complexity of model training. When any component in the air conditioning system 102 is modified or added, only the data-driven model of the specific component needs to be regenerated, while other models can remain unchanged. In addition, the multiple model equations generated by the data-driven model do not depend on the geometry, size, circuit mode, and operating conditions of components such as heat exchangers. Therefore, the data-driven model is effectively solved in real time, making it conducive to online control of any machine including a heat exchanger.

[0080] It is understood that in the present disclosure, the case of an air conditioning system is considered to adopt data-driven control. However, such data-driven control can be equally used for any machine including at least one heat exchanger. Examples of machines include, but are not limited to, vehicle systems, smart grid systems, factory automation systems, transportation systems, self-regulating machine systems, and transportation networks.

[0081] Figure 2A A block diagram 200A of a control system 202 utilizing a model 204 to control an air conditioning system 102 is shown according to an embodiment of the present disclosure. The block diagram 200A may also include a system 206. The system 206 includes the air conditioning system 102 and the space 104.

[0082] In an embodiment, control system 202, model 204, and system 206 are connected via a network ( Figure 2AThe network may include a public network (such as the Internet, a telephone network, and a satellite communication network) and / or a local area network (LAN), a wide area network (WAN, such as Ethernet (registered trademark)). In addition, examples of the network may include a dedicated line network such as an Internet Protocol-Virtual Private Network (IP-VPN). In addition, the network may include a wireless communication network of Wi-Fi (registered trademark), Bluetooth (registered trademark), etc.

[0083] The control system 202 is configured to control the functions of the system 206. The control system 202 includes at least one processor such as a processor 202A configured to execute computer executable instructions. The control system 202 also includes a memory 202B that stores computer executable instructions that can be executed by the processor 202A. The processor 202A can be a single-core processor, such as a central processing unit (CPU), a multi-core processor, a computing cluster, or any number of other configurations. The memory 202B can include a non-transitory memory, such as a random access memory (RAM), a read-only memory (ROM), a flash memory, or any other suitable memory system. The processor 202A is configured to control the functions of all components of the control system 202. In addition, the control system 202 is configured to generate actuator commands to control the system 206. In addition, the control system 202 controls the model 204 to execute and train the model 204 through the processor 202A.

[0084] Model 204 includes a simulation model of air conditioning system 102 and space 104. Test control inputs are transmitted to model 204 by control system 202. Model 204 includes a heat exchanger model having a neural network 208, which is configured to simulate the operation of the heat exchanger corresponding to the test control inputs in each time step of the multi-step forecast, as described later with reference to Figure 2B The data-driven heat exchanger model enables the system model 204 of the air conditioning system 102 to be more computationally efficient by eliminating the need to solve complex DAEs. Figure 2B and Figure 3A Further details of the neural network heat exchanger model are provided.

[0085] The control system 202 may receive a reference signal 210 (also indicated as “r”) and measurement data 212 (also indicated as “v”) from the system 206 and generate an actuator command output 214 (also indicated as “u”) to control the operation of the system 206. The actuator command output 214 is optimized to minimize a cost function while satisfying the reference signal 210. In one embodiment, the cost function may be the power consumption of the system 206. In another embodiment, the cost function may be a tracking error between the reference signal 210 and the measurement data 212. In one embodiment, the reference signal 210 may be associated with, for example, the air temperature of the space 104 and the suction superheat of the air conditioning system 102. In another embodiment, the reference signal 210 may be the air temperature of the space 104 and the discharge temperature of a component of the air conditioning system 102. The actuator command output 214 may be associated with at least one of a compressor speed, a valve opening, or a fan speed of the air conditioning system 102.

[0086] The processor 202A of the control system 202 generates and transmits the test control input 216 to the model 204 of the system 206. Unlike the first principle model of the vapor compression system, which is constructed based on conservation laws and is computationally very expensive due to the resulting DAE, the model 204 includes a data-driven component model that is very fast and can perform many simulations in a short period of time. The processor 202A is also configured to execute the model 204 based on the historical operating data 218 from the system 206 and the test control input 216 from the control system 202 to obtain the corresponding test output 220. The control system 202 can evaluate the test control input 216 based on various indicators and determine the actuator command output 214 (u) of the system 206.

[0087] Figure 2B A block diagram 200B of a neural network 208 for predicting a test output 222 of operation of a heat exchanger corresponding to a test control input 224 in a single time step prediction is shown.

[0088] The neural network 208 includes a first arm 226 having a first combination of convolutional and recurrent networks. The first combination of convolutional and recurrent networks includes a one-dimensional (1-D) convolutional neural network (CNN) group 228 and a recurrent neural network (RNN) 230. The neural network 208 also includes a second arm 232 having a second combination of convolutional and recurrent networks. The second combination of convolutional and recurrent networks includes a 1-D CNN group 234 and an RNN 236. The 1-D CNN group 228 of the first arm 226 is provided with a test control input 224 and a historical control input sequence 238 to extract a control feature 240. The control feature 240 indicates a change in the test control input 224 relative to the historical control input sequence 238. The first arm 226 is configured to process the test control input 224 appended to the historical control input sequence 238 using the first combination of the 1-D CNN group 228 and the RNN 230. The 1-D CNN group 234 of the second arm 232 is provided with a historical output sequence 242 paired with a historical input sequence 238 to extract output features 244. The output features 224 indicate a dynamic coupling between a test control input 224 of the operation of the heat exchanger and a corresponding test output 222. The second arm 232 is configured to process the historical output sequence 242 paired with the historical input sequence 238 using a second combination of the 1-D CNN group 234 and the RNN 236. The neural network 208 also includes a third arm 246, which is configured to process the control features 240 and the output features 244 to predict the test output 222 of the operation of the heat exchanger corresponding to the test control input 224.

[0089] Figure 3A The architecture 300A of a neural network 208 trained to simulate the operation of a heat exchanger included in the air conditioning system 102 at time step t is shown in accordance with an embodiment of the present disclosure. For example, the neural network 208 may be trained to simulate the operation of a heat exchanger.

[0090] The dynamics of a heat exchanger can be described in abstract input-output form by the following equation:

[0091] Y 0:T =Φ(X 0:T ,x0) (1)

[0092] Where Φ is an unknown function corresponding to an input-output mapping that generates a trajectory of observable outputs given the initial state and inputs of the heat exchanger of the air conditioning system 102. 0:T represents the observable output of the heat exchanger collected in the time range [0,T]. Similarly, X 0:T represents a set of inputs to the heat exchanger, which are often referred to as boundary conditions. The initial internal state of the heat exchanger is described by x0.

[0093] A set of boundary conditions X includes the inlet conditions for each fluid flow, such as mass flow rate, pressure, and specific enthalpy or other thermodynamic variables. The initial state variables x0 represent the initial values ​​of the differential variables of the neural network 208, such as the pressure and specific enthalpy of the fluid in each control volume or the temperature of the tube wall. Finally, a set of output variables Y may include variables defined in each volume, such as the exiting air temperature or moisture content of each volume, or may include comprehensive variables for the entire heat exchanger, such as the total heat power transferred from one fluid to another.

[0094] The neural network 208 incorporates a deep state-space modeling framework that combines the feature extraction capabilities of CNN with the sequence prediction properties of RNN.

[0095] According to some embodiments, a gated recurrent unit (GRU) may be used here as RNN 230 and / or RNN 236. GRU is an efficient recurrent neural unit configured to analyze the extracted temporal features to learn the system dynamics. According to some embodiments, 1-D CNN group 228 and 1-D CNN group 234 are used here for the first stage analysis of time series data. Therefore, the architecture of neural network 208 may be referred to as a CNN-GRU deep state space model (SSM). CNN-GRU SSM uses a "look back" window of length N, which includes the historical control input sequence (X t-N:t-1 )238 and the historical control input sequence (X t-N:t-1 )238 corresponding historical output sequence (Y t-N:t-1 )242. These past inputs and the current input (X t ) can be used to generate the test output (Y t )222 prediction. In addition, the historical control input sequence (X t-N:t-1 )238 and historical output sequence (Y t-N:t-1 ) 242 forms part of the operating data 208 and tests the control input (X t ) 224 forms part of the test control input 216, such as Figure 2B shown.

[0096] In addition, the processor 202A ( Figure 2A ) is configured to generate a windowed data sequence through two paths: for the upper X path 302, X t-N:t , for the lower (X; Y) path 304, X t-N:t-1 With Y t-N:t-1For the upper X path 302 and the lower (X; Y) path 304, the windowed data sequence passes through the CNN-GRU SSM of the first arm 226 and the second arm 232, respectively. To this end, the windowed data flows through the 1-D CNN group 228 and the 1-D CNN group 234, respectively, for feature extraction. The upper X path 302 is composed of the test control input (X t ) 224, and attempts to learn the effect of predicting the output from the changes in the inputs observed so far. The lower (X; Y) path 304 extracts features relevant to learning how the inputs and outputs are connected in a dynamic system. In addition, both the upper X path 302 and the lower (X; Y) path 304 continue into their respective RNNs 230 and RNNs 236, thereby modeling how the internal hidden states are updated based on the features extracted from the 1-D CNN group 228 and the 1-D CNN group 234, respectively.

[0097] Given a vector φ of CNN-extracted features t 、The current internal state of GRU h t-1 , reset gate vector r t , update gate vector z t and candidate activation The internal state update of GRU is given by:

[0098]

[0099] where ⊙ is the Hadamard product,

[0100]

[0101] where σ represents the sigmoid activation function and the W matrix is ​​the neural weights. Specifically, and are the reset weights corresponding to the features and the current internal state, and represents the bias of the gate. Similarly, and are the weights and biases of the update gate. Finally, is the weight of the feature, is the weight of the Hadamard product of the reset gate output vector and the current internal state, and is the bias vector at the activation gate. The upper X path 302 is a first path assigned to a first neural weight, and the lower (X; Y) path 304 is a second path assigned to a second neural weight. The first neural weight may be different from the second neural weight, so the first path and the second path may be trained simultaneously.

[0102] At time t, the internal state of the upper X path 302 as the extracted control feature 240 is represented by "h", and the internal state of the lower (X; Y) path 304 as the extracted output feature 244 is represented by "d". Therefore, the internal state of the RNN 230 is "h" 240, and the internal state of the RNN 236 is "d" 244. The internal states h 240 and d 244 are cascaded into a single enhanced state vector and passed through the final set of the third arm 246. The third arm 246 includes a fully connected layer in which each input neuron is connected to each output neuron. The purpose of the fully connected layer is to project the enhanced RNN state into the output space. The enhanced state vector can be a tensor accepted by the third arm 246. The output of the third arm 246 is the predicted test output Y t 222.

[0103] Then, the predicted output Y t 222 is passed to the next time step 306 and used to generate (X; Y) data 308 for the next time step.

[0104] The architecture of the proposed neural network 208 is designed to promote some specific beneficial properties. For example, the addition of 1-DCNN layer groups promotes the extraction of temporal features and helps in the preprocessing of GRU units.

[0105] Figure 3B A flowchart 300B is shown of an exemplary process for training the neural network 208 according to an embodiment of the present disclosure. At block 310, training data is collected to train the neural network 208. To this end, data traces are obtained by running a physics-based model with boundary conditions associated with the heat exchanger as input. For example, the boundary conditions associated with a heat exchanger may be boundary conditions on the refrigerant side and the air side. Examples of refrigerant side inputs may be, but are not limited to, refrigerant mass flow rate and refrigerant enthalpy flow rate. Examples of air side inputs may be, but are not limited to, air mass flow rate, air enthalpy flow rate, and relative humidity. Data-driven models do not need to solve the full physical equations, but rather learn nonlinear relationships between inputs and outputs to capture their relationship. Therefore, training can be successfully performed using only a subset of the total state space that needs to be solved in the physical model. This reduces the need to solve the full set of complex DAEs that define the system, and models a subset that is useful for the needs of the modeling effort.

[0106] At block 312, based on the data trace, To train the neural network 208. The processor 202A can process the data trace A machine learning algorithm is applied to train the neural network of neural network 208. The machine learning algorithm can be supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. The details of the training process will be described in Figure 3C The neural network 208 is trained to learn the mapping function from the input data of the system to the output data. To this end, the neuron weights of the neural network 208 are updated to minimize the error of the neural network 208 while processing the data trajectory. One or more data sets that are associated.

[0107] Additionally, at block 314 , training of the neural network 208 is completed and the refrigerant state and the air state are received as outputs.

[0108] Figure 3C Flowchart 300C is shown for optimizing the training loss of the neural network 208 according to an embodiment of the present disclosure. Flowchart 300C specifically shows Figure 3B An extension of step 312 .

[0109] At block 316, processor 202A is configured to generate multiple batches of data traces to generate one or more data sets. To this end, the entire data trace The data is divided into consecutive (in terms of time) blocks of training, validation, and test sets. Therefore, for scalability of large data sets, multiple batches (such as mini-batches) of training data for each of training, validation, and testing are constructed, and within each of the training and validation sets, the data is rearranged to construct windowed data of the historical control input sequence 238 and the historical output sequence 242. Each of the historical control input sequence 238 and the historical output sequence 242 has a length N and includes X t-N:t-1 and Y t-N:t-1 The batches also include a test control input X to the air conditioning system 102 t 224 and the actual output of the air conditioning system 102

[0110] By considering windows of length N instead of the entire dataset, the time back-propagation expansion required to train the GRU is not overly complex, ultimately reducing the training time of the GRU unit. In addition to time efficiency, avoiding the requirement to learn from the entire time series at once and instead adopting batch processing allows for scalability, parallelism, and improved storage efficiency when faced with large datasets. In addition to this, by forcing the algorithm to process each windowed time series segment independently of any other window (including consecutive windows), the proposed deep SSM learns a dynamic one-step update independent of the start time, while the temporal dependency is embedded in the internal state vector of the GRU unit. This type of modeling is often superior to autoregressive methods that use neural SSMs for prediction in future predictions, as the recursion of the estimation can often lead to error accumulation and degradation of model performance.

[0111] At block 318, the processor 202A is configured to calculate a mean square error for each of the plurality of batches based on the predicted output of the air conditioning system 102 and the actual output of the air conditioning system 102. To this end, the processor 202A performs a time step prediction for each of the plurality of batches. The predicted output Y is then used t And the real output Y t * To calculate the batch-by-batch mean squared error:

[0112]

[0113] Among them, B k is the kth batch, usually |B k | is the same across all k batches and represents the batch size.

[0114] At block 320, a training loss associated with the mean squared error computed for each of the plurality of batches is determined and optimized. To this end, based on the sum ∑ k ε k To calculate the training loss. The training loss analyzes the fit of the deep learning model to the training data. The training loss can be optimized using stochastic gradient descent or at least one of the variants of stochastic gradient descent.

[0115] In addition, the sum of all k batches based on the validation set ∑ k ε k to calculate the validation loss. Validation loss analyzes the requirement for further training or tuning in a model. Calculate validation loss for use in early stopping.

[0116] Figure 4 An example diagram 400 of a model 204 according to an embodiment of the present disclosure is shown. The model 204 may include a room model 402A of a room 402 and multiple models of corresponding components of the air conditioning system 102. Examples of components of the air conditioning system 102 may include a compressor 404, a condenser 406, an expansion device 408, and an evaporator 410. Examples of multiple component models may include a compressor model 404A based on a performance map of the compressor 404, a neural network 208 of the condenser 406, an expansion device model 408A based on a performance map of the expansion device 408, and a neural network 208 of the evaporator 410. In addition, the room model 402A may be a lumped parameter room model. In a lumped parameter model, spatial variations of parameters may be ignored, and the system may be described by adjustable parameters. In addition, the room 402 includes a space 104 in which the air conditioning system 102 is configured to provide heating or cooling.

[0117] Performance map based models map the performance of a component based on experimental data corresponding to the component. One example of a performance map based compressor model is the ARI 10 coefficient compressor model, which calculates the mass flow rate and compressor power of the compressor based on saturated suction and discharge temperatures. There are other types of performance map based compressor models that determine the mass flow rate and power consumption based on suction and discharge pressures. For performance map based expansion valve models, typically these models calculate the mass flow rate through the valve based on the pressure difference across the valve, where the flow coefficient is a function of the valve opening.

[0118] In addition, component models such as the performance map-based compressor model 404A, the neural network 208, and the performance map-based expansion device model 408A are connected together to form a closed-loop refrigerant cycle. The junction 412 between the compressor model 404A and the neural network 208 of the condenser 406, the junction 414 between the neural network 208 of the condenser 406 and the performance map-based expansion device model 408A, the junction 416 between the performance map-based expansion device model 408 and the neural network 208 of the evaporator 408, and the junction 418 between the performance map-based compressor model 404A and the neural network 208 of the evaporator 410 can correspond to the refrigerant mass flow rate and refrigerant enthalpy flow rate

[0119] All component models implement the same standard interface, so that model 204 can handle any system configuration and can adopt component-based solutions. Therefore, the outlet boundary condition of the upstream model is the inlet boundary condition of the downstream model. Specifically, the compressor model 404A is the upstream model of the condenser model 208. The condenser model 208 is the upstream model of the expansion device model 408A. The expansion device model 408A is the upstream model of the evaporator model 208. The evaporator model 208 is the upstream model of the compressor model 404A. Similarly, the condenser model 208 is the downstream model of the compressor model 404A. The expansion device model 408A is the downstream model of the condenser model 208. The evaporator model 208 is the downstream model of the expansion device model 408A. The compressor model 404A is the downstream model of the evaporator model 208.

[0120] As air is recirculated in the space 104, the interface 420 between the data-driven evaporator model 408A and the room model 402A includes the air mass flow rate associated with the room model 310 and the evaporator model. Air enthalpy flow rate and relative humidity (φ a). Model 204 is based on multi-step prediction. To this end, model 204 predicts the test output for each time step based on the corresponding test control input for that time step. Processor 202A of control system 202 can control system solver 422 to solve all component models in a systematic manner and ensure that mass, energy, and momentum are conserved at each joint of model 204 for each step of the test control input at the system level.

[0121] Figure 5 A flowchart 500 for generating one or more states associated with a room model 402A and multiple component models of an air conditioning system 102 is shown in accordance with an embodiment of the present disclosure. At box 502, the processor 202A is configured to initialize the model 204 of the system 206 and determine one or more conditions associated with the air side and the refrigerant side of the heat exchanger. The parameters of the air side conditions of the heat exchanger are the air mass flow rate, the air temperature, and the relative humidity. The initial conditions of the refrigerant side of the heat exchanger are the refrigerant pressure distribution and the refrigerant enthalpy distribution. Based on the initial conditions and boundary conditions, the model 204 of the system 206 is fully defined.

[0122] At block 504 , the historical control input sequence and the historical output sequence of the operation of the heat exchanger are provided as inputs to the neural network 208 of the heat exchanger model.

[0123] In addition, at block 506, the processor 202A executes all component models, such as the performance map-based compressor model 404A, the neural network 208 of the condenser 406, the performance map-based expansion device model 408A, and the neural network 208 of the evaporator 410, and the room model 402A. Since the refrigerant flow states at interfaces 412, 414, 416, and 418 of the system model 204 and the air flow states at interface 420 are iterative variables, these models are mathematically closed forms and can be executed independently.

[0124] At block 508, processor 202A is configured to collect the outputs obtained at block 506 from each component model of air conditioning system 102. System solver 422 then formulates residual equations at each interface, such as interface 412, interface 414, interface 416, and interface 418, to ensure conservation of mass, energy, and momentum.

[0125] At block 510 , the processor 202A is configured to execute the system solver 422 , which is used to determine whether convergence criteria based on conservation of mass, energy, and momentum are met at each interface.

[0126] At block 512, the processor 202A generates one or more states associated with the room model and each of the plurality of component models. The one or more states include at least one of a refrigerant state or an air state associated with the air conditioning system 102. In addition, the one or more states associated with each model are obtained based on the determination of the convergence criteria at block 510 and used as initial conditions for calculating the next time step.

[0127] At box 514, the process of setting the input to the neural network 208 is repeated at each time step until the entire simulation of the heat exchanger is completed.

[0128] Figure 6 A block diagram 600 is shown of different components of the control system 202 for controlling the air conditioning system 102 according to an embodiment of the present disclosure. The components of the system 206 include the compressor 404, a condenser fan 602 of the condenser 406, an expansion valve 604, and an evaporator fan 606 of the evaporator 410. The circuit 608 includes a condenser fan control 610 for controlling one or more parameters of the condenser fan 604, a compressor control 612 for controlling one or more parameters of the compressor 404, an expansion valve control 614 for controlling one or more parameters of the expansion valve 604, and an evaporator fan control 616 for controlling one or more parameters of the evaporator fan 606. For example, control devices such as condenser fan control device 610, compressor control device 612, expansion valve control device 614, evaporator fan control device 616 can control, but are not limited to, the speed of condenser fan 602 and evaporator fan 606; threshold, ratio, inflection point, attack time, release time or supplemental gain of compressor 404; valve body, diaphragm, pin or needle, spring, sensing ball and capillary line of expansion valve 604. In addition, the control device of circuit 608 can perform a control process to control the components of system 206 based on the actuator command output 214 received from control system 202. In addition, the components of system 206 generate corresponding outputs in response to the control process performed by circuit 608 based on actuator command output 214. Each component of air conditioning system 102 of system 206 has one or more sensors attached thereto. The group of sensors is collectively referred to as sensor 618 herein. Sensor 618 records measurement data 212 of air conditioning system 102 and transmits it back to control system 202. Control system 202 controls simulation of model 204 to reduce the error between reference signal 210 and measurement data 212 .

[0129] Fig. 7AA flowchart 700A of an exemplary method for selecting an optimal test control input from a test control input 216 based on a Gaussian process is shown according to an embodiment of the present disclosure. In this embodiment, at a current time, the control system 202 generates an optimal control input in advance for the next time step. To this end, at box 702, the processor 202A is configured to generate a test control input 216 based on a Gaussian process. At box 704, the test control input 216 is transmitted to the model 204. At box 706, the processor 202A receives a plurality of test outputs 220 corresponding to the test control input 216 predicted by the model 204. To this end, a simulation of the model 204 is performed for each test control input 216 to predict the corresponding test output 220 for the next time step. The processor 202A calculates a cost function for each test control input 216 for the corresponding test output 220 and the reference signal 210. The above reference Figure 2B An example of a cost function is described.

[0130] At block 708, the processor 202A is further configured to select an optimal test control input from the test control inputs 216 based on an optimization of a cost function associated with the test control inputs 216 and the test outputs 220. Thus, the selected optimal test control input optimizes the cost function and is selected as the actuator command output 214 for the next time step. The processor 202A also determines a control command based on the selected optimal test control input. The control command can be transmitted to the air conditioning system 102 for use in changing the speed of the compressor 404, opening or closing the expansion valve 604, changing the speed of the condenser fan 602, and / or changing the speed of the evaporator fan 606, etc.

[0131] Figure 7B A flow chart 700B depicting the generation of a test control input 216 according to an embodiment of the present disclosure is shown. The test control input 216 is generated by a one-step optimization based control.

[0132] Initially, the control system 202 may consider a uniform distribution of test control inputs since there is no available information about the test control inputs from previous time steps. These uniform test control inputs are sent to the system model 204. The processor 202A is configured to perform simulations to evaluate each of these uniform test control inputs. The control input that minimizes the cost function is selected as the actuator command 214 for the first time step. The processor 202A also utilizes one or more test control inputs at previous time instants to generate multiple test control inputs 216 for the next time instant based on a Gaussian process. To this end, at box 710, the control system 202 collects test control inputs from previous time steps. At box 712, the processor 202A is configured to generate test control inputs 216 for the current time step based on the test control inputs of the previous time step. The processor 202A may be based on Fig. 7A The Gaussian process of is used to generate the test control input 216. To this end, the mean and variance in the Gaussian process are calculated. For example, the generated test control input 216 should have a mean as the control input of the previous time step, and the variance can be determined as follows.

[0133] Given all test control inputs at the previous time step, corresponding predicted outputs are available because they are the outputs of the execution of the model 204 based on the previous time step.

[0134] Compute the mean and variance of all predicted outputs at previous time steps and discard outputs that differ by more than two standard deviations from their mean, and keep the remaining outputs.

[0135] Calculate the variance σ of the test control input 216 corresponding to the predicted output maintained in step (2) t-1 .

[0136] σ t-1 The variance of the Gaussian distribution used to generate the test control input 216 at the current time step.

[0137] Figure 8 A flowchart 800 of an exemplary method for selecting an optimal test control input from a sequence of test control inputs based on a model predictive control process according to another embodiment of the present disclosure is shown. In this embodiment, the test control inputs 216 may be generated based on a nonlinear model predictive control. Unlike control based on one-step optimization, gradient-based control evaluates a cost function over a limited time horizon rather than at a certain moment.

[0138] At block 802, processor 202A generates an initial guess of a test control input sequence 216 corresponding to a finite number of time instants. Next, at block 804, processor 202A calculates the gradient of a cost function with respect to the test control input sequence 216 based on the predicted output 220 from model 204. To this end, the test control input sequence 216 is provided as an input to model 204, and model 204 predicts a test output sequence 220 corresponding to the test control input sequence 216. In addition, processor 202A uses numerical differentiation to calculate the gradient of the cost function with respect to the test control input sequence 216. Since model 204 is very computationally efficient, numerical differentiation does not incur significant overhead during the process. At block 806, the test control input sequence 216 is adjusted along the gradient of the cost function.

[0139] In addition, the processor 202A selects the best test control input from the generated test control input sequence 216 based on the gradient of the calculated cost function. To this end, at box 808, the processor 202A compares the gradient of the cost function with respect to the test control input sequence 216 with a predefined tolerance. Next, at box 810, if the gradient of the cost function associated with the test control input sequence 204 is less than the predefined tolerance, the first step of the test control input sequence 216 is selected as the best control input. If the gradient of the cost function is greater than the predefined tolerance, the process at box 806 is repeated until the entire process converges. The processor 202A executes the model 204 to predict the test output based on the test control input, and determines the control command for the air conditioning system 102 based on the predicted test output of the model 204.

[0140] Fig. 9 A flow chart 900 of a method for determining control commands for the air conditioning system 102 is shown.

[0141] At block 902, the processor 202A determines a control command for the air conditioning system 102 based on a predicted test output 222 of a simulation of the operation of the heat exchanger for a test control input 224. To this end, the model 204 predicts a test output 220 based on the predicted test output 222 of the neural network 208. Using the predicted test output 220 of the model 204, the reference signal 210, and the corresponding measurement data 212, the processor 202A of the control system 202 generates an actuator command output 214. The processor 202A determines a control command for the air conditioning system 102 based on the generated actuator command output 214. Referring to FIG. Fig. 7A Examples of control commands have been described.

[0142] At block 904 , the processor 202A transmits control commands to actuators, such as the condenser fan control 610 , the compressor control 612 , the expansion valve control 614 , the evaporator fan control 616 , etc., of the air conditioning system 102 .

[0143] Fig.10 A block diagram 1000 for controlling the air conditioning system 102 using the control system 202 is shown according to an embodiment of the present disclosure.

[0144] Air conditioning system 102 is deployed in space 104. Space 104 is occupied by occupants 1002, 1004, 1006, and 1008. Arrow 1010 represents air supplied to space 104 by air conditioning system 102. Control system 202 executes model 204 to receive test output 220.

[0145] Additionally, using the test output 220 of the system 206, the reference signal 210, and the measurement data 212, the control system 202 generates a control command 214 for the system 206. The control command also controls the components of the air conditioning system 102 in an optimal manner.

[0146] The above description provides only exemplary embodiments and is not intended to limit the scope, applicability or configuration of the present disclosure. Instead, the following description of the exemplary embodiments will provide a feasible description of implementing one or more exemplary embodiments for those skilled in the art. It is contemplated that various changes may be made to the functions and arrangements of the elements without departing from the spirit and scope of the subject matter disclosed in the appended claims.

[0147] Specific details are given in the following description to provide a comprehensive understanding of the embodiments. However, if one of ordinary skill in the art understands that the embodiments can be implemented without these specific details. For example, the systems, processes, and other elements in the disclosed subject matter can be displayed as components in block diagram form to avoid obscuring the embodiments with unnecessary details. In other cases, well-known processes, structures, and techniques can be displayed without unnecessary details to avoid obscuring the embodiments. In addition, the same reference numerals and names in the various figures represent the same elements.

[0148] In addition, various embodiments may be described as processes shown in flowcharts, flow diagrams, data flow diagrams, structure diagrams, or block diagrams. Although a flowchart may describe an operation as a sequential process, many operations may be performed in parallel or simultaneously. In addition, the order of the operations may be rearranged. A process may terminate when its operations are completed, but may have additional steps not discussed or included in the figure. In addition, not all operations in any particular described process may occur in all embodiments. A process may correspond to a method, function, program, subroutine, subprogram, etc. When a process corresponds to a function, the termination of the function may correspond to the function returning to the calling function or main function.

[0149] In addition, embodiments of the disclosed subject matter may be implemented at least in part manually or automatically. Manual or automatic implementation may be performed or at least assisted by the use of a machine, hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments for performing the necessary tasks may be stored in a machine-readable medium. A processor may perform the necessary tasks.

[0150] The various methods or processes outlined herein may be encoded as software that can be executed on one or more processors using any of a variety of operating systems or platforms. In addition, such software may be written using any of a variety of suitable programming languages ​​and / or programming or scripting tools, and may also be compiled into executable machine language code or intermediate code that is executed on a framework or virtual machine. Typically, in various embodiments, the functionality of the program modules may be combined or distributed as desired.

[0151] The above various embodiments are described as processes shown in flowcharts, flow diagrams, data flow diagrams, structure diagrams or block diagrams. Although flowcharts can display operations as sequential processes, many operations can be performed in parallel or simultaneously. In addition, the order of operations can be rearranged. A process can terminate when its operations are completed, but can have additional steps that are not discussed or included in the figure. In addition, not all operations in any particular described process can occur in all embodiments. A process can correspond to a method, function, program, subroutine, subprogram, etc. When a process corresponds to a function, the termination of the function can correspond to the function returning to the calling function or main function.

[0152] In addition, embodiments of the disclosed subject matter may be implemented at least in part manually or automatically. Manual or automatic implementation may be performed or at least assisted by the use of a machine, hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments for performing the necessary tasks may be stored in a machine-readable medium. A processor may perform the necessary tasks.

[0153] Benefiting from the teachings presented in the above description and the related drawings, those skilled in the art will appreciate many modifications and other embodiments of the present disclosure described herein. It should be understood that the present disclosure is not limited to the specific embodiments disclosed, and modifications and other embodiments are intended to be included within the scope of the appended claims. In addition, although the above description and the related drawings describe example embodiments in the context of certain example combinations of elements and / or functions, it should be understood that different combinations of elements and / or functions may be provided by alternative embodiments without departing from the scope of the appended claims. In this regard, for example, combinations different from the elements and / or functions explicitly described above are also envisioned, as may be set forth in certain appended claims. Although specific terms are used herein, they are used only in a general and descriptive sense, and not for limiting purposes.

Claims

1. A control system for controlling the operation of an air conditioning system including a heat exchanger, the control system comprising: at least one non-transitory memory configured to store computer-executable instructions; as well as at least one processor configured to execute the computer-executable instructions to: executing a neural network trained to simulate operation of the heat exchanger for test control inputs, generating outputs of the simulation based on historical data defining states of the heat exchanger, the historical data comprising a historical sequence of control inputs provided to the heat exchanger and a historical sequence of outputs of the operation of the heat exchanger corresponding to the historical sequence of control inputs, the neural network comprising: A first arm, wherein the first arm is configured to: processing the test control input appended to the sequence of historical control inputs using a first combination of convolutional and recurrent networks trained to extract control features indicative of changes in the test control input from the historical control inputs; A second arm, the second arm being configured as: processing the historical control input sequence paired with the historical output sequence using a second combination of convolutional and recurrent networks trained to extract output features indicative of dynamic coupling between inputs and corresponding outputs of the operation of the heat exchanger; and A third arm, the third arm being configured as: processing the control characteristic and the output characteristic to predict a test output of the operation of the heat exchanger corresponding to the test control input; determining a control command for the air conditioning system based on the predicted simulated test output of the operation of the heat exchanger for the test control input; and The determined control commands are transmitted to actuators of the air conditioning system.

2. The control system according to claim 1, wherein: Each of the first combination of the convolutional and recurrent networks of the first arm of the neural network and the second combination of the convolutional and recurrent networks of the second arm includes a one-dimensional convolutional layer group and a gated recurrent unit.

3. The control system according to claim 1, wherein: The third arm is a fully connected deep neural network that accepts a tensor comprising extracted control features and extracted output features.

4. The control system according to claim 1, wherein: The determined control command to the actuator of the air conditioning system is associated with at least one of a speed of a compressor, an opening or closing of a valve, or a speed of a fan of the air conditioning system.

5. The control system according to claim 1, wherein: A first path of the first arm that processes the test control input appended to the historical control input sequence is assigned a first neural weight, and a second path of the second arm that processes the historical control input sequence paired with the historical output sequence is assigned a second neural weight.

6. The control system according to claim 1, wherein: The processor is further configured to: generating one or more data sets associated with a data trace obtained based on executing a physics-based model of the heat exchanger having one or more boundary conditions associated with the heat exchanger; as well as A machine learning algorithm is applied to the generated data set to train the neural network.

7. The control system according to claim 6, wherein: The processor is further configured to: A plurality of batches of the data traces are generated to generate the one or more data sets, and wherein each batch of the plurality of batches includes at least one of the historical control input sequence, the historical output sequence, a first input of the air conditioning system, and an actual output of the air conditioning system.

8. The control system according to claim 7, wherein: The process of training the neural network includes the following operations: calculating a mean square error of each of the plurality of batches based on the predicted output of the air conditioning system and the actual output of the air conditioning system; determining a training loss associated with the calculated mean squared error for each of the plurality of batches; as well as Optimize the determined training loss.

9. The control system according to claim 1, wherein: The processor is further configured to: Initializing a model associated with the air conditioning system, the model comprising a room model and a plurality of component models of a plurality of components of the air conditioning system, wherein the plurality of component models comprises at least the neural network trained to simulate operation of the heat exchanger; determining one or more conditions associated with at least an air side of a space associated with the air conditioning system; executing the room model and the plurality of component models of the plurality of components of the air conditioning system based on the one or more conditions associated with at least an air side of the air conditioning system; and One or more states associated with the room model and each of the plurality of component models are generated, wherein the one or more states include at least one of a refrigerant state or an air state associated therewith.

10. The control system according to claim 1, wherein: The processor is further configured to: Generate multiple test control inputs based on Gaussian process; providing the plurality of test control inputs as inputs to a model of the air conditioning system; receiving a plurality of test outputs corresponding to the plurality of test control inputs predicted by the model of the air conditioning system; selecting the test control input from the generated plurality of test control inputs based on optimization of a cost function associated with the plurality of test control inputs and the plurality of test outputs; as well as The control command for the air conditioning system is determined based on the predicted test output of the simulation of the operation of the heat exchanger for the selected test control input.

11. The control system according to claim 10, wherein: The processor is further configured to utilize one or more test control inputs at a first time instant to generate a plurality of test control inputs for a second time instant based on the Gaussian process, wherein the first time instant is prior to the second time instant.

12. The control system according to claim 1, wherein: The processor is further configured to: generating a test control input sequence corresponding to a finite number of time instants; providing the test control input sequence as input to a model of the air conditioning system; receiving a test output sequence predicted by the model of the air conditioning system corresponding to the test control input sequence; calculating a gradient of a cost function associated with the test control input sequence and the test output sequence; selecting the test control input from the generated sequence of test control inputs based on the calculated gradient of the cost function; as well as The control command for the air conditioning system is determined based on the predicted test output of the simulation of the operation of the heat exchanger for the selected test control input.

13. A method for controlling the operation of an air conditioning system including a heat exchanger, the method comprising the steps of: executing a neural network trained to simulate operation of the heat exchanger for test control inputs, generating outputs of the simulation based on historical data defining states of the heat exchanger, the historical data comprising a historical sequence of control inputs provided to the heat exchanger and a historical sequence of outputs of operation of the heat exchanger corresponding to the historical sequence of control inputs, wherein the neural network comprises: a first arm configured to process the test control input appended to the sequence of historical control inputs using a first combination of convolutional and recurrent networks trained to extract control features indicative of changes in the test control input from the historical control inputs; a second arm configured to process the historical control input sequence paired with the historical output sequence using a second combination of convolutional and recurrent networks trained to extract output features indicative of dynamic coupling between inputs and corresponding outputs of operation of the heat exchanger; and a third arm configured to process the control characteristic and the output characteristic to predict a test output of the operation of the heat exchanger corresponding to the test control input; determining a control command for the air conditioning system based on the predicted simulated test output of the operation of the heat exchanger for the test control input; and The determined control commands are transmitted to actuators of the air conditioning system.

14. The method according to claim 13, wherein: Each of the first combination of the convolutional and recurrent network of the first arm and the second combination of the convolutional and recurrent network of the second arm of the neural network includes a one-dimensional convolutional layer group and a gated recurrent unit.

15. The method according to claim 13, wherein: The third arm is a fully connected deep neural network that accepts a tensor comprising extracted control features and extracted output features.

16. The method according to claim 13, further comprising the steps of: generating one or more data sets associated with a data trace obtained based on executing a physics-based model having one or more boundary conditions associated with the heat exchanger; as well as A machine learning algorithm is applied to the generated data set to train the neural network.

17. The method according to claim 16, further comprising the steps of: A plurality of batches of the data traces are generated to generate the one or more data sets, wherein each batch of the plurality of batches includes at least one of the historical control input sequence, the historical output sequence, a first input of the air conditioning system, and an actual output of the air conditioning system.

18. The method according to claim 13, further comprising the steps of: Initializing a model associated with the air conditioning system, the model comprising a room model and a plurality of component models of a plurality of components of the air conditioning system, wherein the plurality of component models comprises at least the neural network trained to simulate operation of the heat exchanger; determining one or more conditions associated with at least an air side of the air conditioning system; executing the room model and the plurality of component models of the plurality of components of the air conditioning system based on the one or more conditions associated with at least an air side of the air conditioning system; and One or more states associated with the room model and each of the plurality of component models are generated, wherein the one or more states include at least one of a refrigerant state or an air state associated therewith.

19. The method according to claim 13, wherein: The processor is also configured to: Generate multiple test control inputs based on Gaussian process; providing the plurality of test control inputs as inputs to a model of the air conditioning system; receiving a plurality of test outputs corresponding to the plurality of test control inputs predicted by the model of the air conditioning system; selecting the test control input from the generated plurality of test control inputs based on optimization of a cost function associated with the plurality of test control inputs and the plurality of test outputs; as well as The control command for the air conditioning system is determined based on the predicted test output of the simulation of the operation of the heat exchanger for the selected test control input.

20. A non-transitory computer-readable medium having computer-executable instructions stored thereon, wherein when the computer-executable instructions are executed by a computer, the computer performs operations, the operations comprising: executing a neural network trained to simulate operation of a heat exchanger for a test control input, generating an output of the simulation based on historical data defining states of the heat exchanger, the historical data comprising a historical sequence of control inputs provided to the heat exchanger and a historical sequence of outputs of the operation of the heat exchanger corresponding to the historical sequence of control inputs, the neural network comprising: a first arm configured to process the test control input appended to the sequence of historical control inputs using a first combination of convolutional and recurrent networks trained to extract control features indicative of changes in the test control input from the historical control inputs; a second arm configured to process the historical control input sequence paired with the historical output sequence using a second combination of convolutional and recurrent networks trained to extract output features indicative of dynamic coupling between inputs and corresponding outputs of operation of the heat exchanger; and a third arm configured to process the control characteristic and the output characteristic to predict a test output of operation of the heat exchanger corresponding to the test control input; determining a control command for an air conditioning system based on the predicted simulated test output of operation of the heat exchanger for the test control input; and The determined control commands are transmitted to actuators of the air conditioning system.

Images