TEMPERATURE SETPOINT CONTROL

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 62/554,785, filed on September 6, 2017. The entire disclosure of the above application is incorporated herein by reference.

FIELD

[0002] The present disclosure relates to systems and methods for managing data center cooling devices to optimize energy consumption, and more particularly to a system and method which utilizes machine learning and neural network models to control a supply air temperature setpoint for a cooling unit in a manner that optimizes energy usage.

BACKGROUND

[0003] This section provides background information related to the present disclosure which is not necessarily prior art.

[0004] Energy usage is a major concern for data centers. As data centers becomes larger and larger, often with thousands, or even tens of thousands, of components located in one or more rooms of a data center, the need to minimize energy consumption of CRAC (computer room air conditioner) units, while still providing adequate cooling, is becoming even more important.

[0005] One important consideration in managing the energy consumption of data center cooling equipment such as CRAC units is in monitoring and adjusting the supply air temperature ("SAT") to optimize room conditions. Traditionally this has involved a user manually, periodically adjusting a SAT input by the user, in an effort to best meet physical temperature conditions in the room while attempting to minimize energy usage. As will be appreciated, this approach relies on the attention, knowledge and experience of a user in manually selecting the optimal SAT setpoint, as well as other parameters for CRAC units being used within the room. An automated system that is capable of "learning", based on past system behavior, how to best select the SAT setpoint, would enable energy usage to be optimized. Importantly, such an automated, intelligent system would also eliminate reliance, or reduce reliance, on the experience and knowledge of a user in selecting an optimal SAT setpoint to best meet the physical room conditions at any given time.

SUMMARY

[0006] This section provides a general summary of the disclosure, and is not a comprehensive disclosure of its full scope or all of its features.

[0007] In one aspect the present disclosure relates to a system for controlling a supply air temperature adjustment for a cooling unit to optimize operation of the cooling unit with respect to at least one of room air temperature and humidity requirements. The system may comprise a controller for implementing a machine learning module configured to select which portion or portions of acquired data pertaining to operation of the cooling unit will be utilized. The controller may also implement a neural network model which uses information supplied by the machine learning module and learns an operational behavior of the cooling unit, and wherein the machine learning module performs supervised learning and regression for the neural network model, and wherein the neural network model uses information supplied by the machine learning module for generating an output. The controller may also implement an optimization module which receives the output from the neural network model and which implements a global optimization routine, using unit power consumption of the cooling unit as the objective function, to produce a supply air temperature set point for use by the cooling unit which optimizes an operating parameter of the cooling unit.

[0008] In another aspect the present disclosure relates to a system for controlling a supply air temperature adjustment for a cooling unit to optimize operation of the cooling unit with respect to room air temperature and humidity requirements. The system may comprise a controller for implementing a machine learning module, a neural network model, and an optimization module. The machine learning module is configured to select which portion or portions of acquired data pertaining to operation of the cooling unit are utilized. The neural network model uses information supplied by the machine learning module and learns an operational behavior of the cooling unit, and wherein the machine learning module also performs supervised learning and regression for the neural network model. The neural network model uses information supplied by the machine learning module for generating an output. The neural network model has a Unit Power neural network module which receives inputs from at least one other neural network model including:

a remote air temperature (RET) neural network model representing a rack inlet temperature of the cooling unit, the RET neural network model providing an output to the unit power neural network model; or

a return air temperature (RAT) neural network model for representing a temperature of air being returned to a given one of the cooling units; or

a supply air temperature (SAT) neural network model for representing a temperature of air being generated and output by the cooling unit.

[0009] The optimization module receives the output from the neural network model and implements a global optimization routine, using unit power consumption of the cooling unit as the objective function, to produce a supply air temperature set point for use by the cooling unit which optimizes an operating parameter of the cooling unit.

[0010] In still another aspect the present disclosure relates to a system for controlling a supply air temperature adjustment for a data center cooling unit to optimize operation of the cooling unit with respect to room air temperature and humidity requirements. The system may comprise a controller for implementing a machine learning module configured to select which portion or portions of acquired data pertaining to operation of the cooling unit are utilized. The controller may also implement a neural network model which uses information supplied by the machine learning module and learns an operational behavior of the cooling unit, and wherein the machine learning module performs supervised learning and regression for the neural network model. The neural network model may include a Unit Power neural network model representing cooling unit power consumption; a remote air temperature (RET) neural network model representing a rack inlet temperature of the cooling unit, the RET neural network model providing an output to the unit power neural network model; a return air temperature (RAT) neural network model for representing a temperature of air being returned to a given one of the cooling units; a supply air temperature (SAT) neural network model for representing a temperature of air being generated and output by the cooling unit; a fan percentage neural network model which represents a percentage of maximum fan speed that a fan of the selected cooling unit is running at; and a cooling capacity (CC) neural network model for representing an overall cooling capacity, in percentage units, of the cooling unit. The Unit Power neural network model may use information supplied by all of the RET, RAT, SAT, fan percentage and CC neural network models in providing an output. The controller may also implement an optimization module which receives the output from the neural network model and which implements a global optimization routine, using unit power consumption of the cooling unit as the objective function, and which produces a supply air temperature set point for use by the cooling unit which optimizes an operating parameter of the cooling unit.

[0011] Further areas of applicability will become apparent from the description provided herein. The description and specific examples in this summary are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure.

DRAWINGS

[0012] The drawings described herein are for illustrative purposes only of selected embodiments and not all possible implementations, and are not intended to limit the scope of the present disclosure.

[0013] Figure 1 is a high level flow diagram illustrating various operations performed by one embodiment of the present disclosure;

[0014] Figure 2 is chart listing the various neural network models used by the present disclosure as well identifying the outputs used by various models as inputs to other models;

[0015] Figure 3 is a list of various criteria used by the machine learning module of the present disclosure in carrying out a machine learning function;

[0016] Figure 4 is a diagram illustrating a neural network model implemented by the present disclosure and the relationship between neural network model inputs, the neurons of a hidden layer of network model, and the Unit Power output;

[0017] Figure 5 is a chart setting forth criteria that are implemented by the optimizer module of the present disclosure;

[0018] Figures 6A and 6B illustrate examples of the performance of the neural network models of the system of Figure 1 along with a graph showing how closely the Unit Power control achieved by the present disclosure follows a calculated optimal Unit Power for a cooling unit.

[0019] Corresponding reference numerals indicate corresponding parts throughout the several views of the drawings. DETAILED DESCRIPTION

[0020] Example embodiments will now be described more fully with reference to the accompanying drawings.

[0021] Referring to Figure 1 there is shown one example of a system 10 for implementing a method in accordance with the present disclosure. The system 10 may include a controller 12 having a processor 14, a memory 16 (e.g., non-volatile RAM or ROM) and an input/output subsystem 18. The controller 12 communicates with data center components 20, which typically will include computer room air conditioner ("CRAC") units that require a supply air temperature ("SAT") setpoint adjustment. By "setpoint" it is meant a user selected supply air temperature that one wishes the system 10 to maintain. One such widely used cooling system is made by Liebert (a Vertiv company) under the name "DSE Free Cooling System", which is a pumped refrigerant economization system, and the teachings of the present disclosure are fully applicable to this system as well as virtually any other type of room air conditioning system that requires a periodic SAT setpoint adjustment to meet changing room air temperature conditions. The system 10 enables optimal energy efficiency to be obtained in operating a CRAC unit by automatically determining, in part through a plurality of neural network models, an optimal SAT setpoint adjustment to be made that best meets current room temperature and humidity requirements.

[0022] With further reference to Figure 1 , the controller 12 acquires data, as indicated at box 22, which may involve applying one or more filters to filter the data to produce a data set which will be used by subsequent processing operations. At operation 24 the collected (and potentially filtered) data set is used via one or more machine learning modules to choose what data is kept for use by neural network models, to train neural network models using selected training algorithms, for example via the Levenberg-Marquardt training algorithm, as well as to decide when to retrain the neural network models and/or to start from scratch with new data. At operation 26, the data may be used in one or more neural network models, as will be described further in the following paragraphs. At operation 28 one or more optimization routines may run. This may involve choosing when to run optimization as well as running an interior-point global optimization algorithm using the collected data, as will also be described more fully in the following paragraphs. The results of operation 26 may be used to supply inputs to the global optimization algorithm. The optimization operation 28 produces a SAT setpoint 30 that may be used by the controller 12 in setting the SAT setpoint applied to one or more of the data center cooling components.

[0023] Figure 2 lists in chart form the different neural networks that the system 10 uses as well as the inputs to each neural network. Neural network ("NN") models are used for "Unit Power" (NN model 26a, also noted as "UP"), which represents cooling unit power consumption; "Remote Air Temp" (NN model 26b, as designated as "RET"), which is the rack inlet temperature of a given data center cooling unit under consideration; "Return Temp", also referred to as "Return Air Temp" or "RAT" (model NN 26c), which is the temperature of air being returned to a given one of the cooling units 20; "Supply Temp", which is also referred to as "Supply Air Temp ("SAT") (NN model 26d), which is the temperature of air being generated and output by the cooling unit; "Fan %" (NN model 26e), which is the percentage of maximum fan speed that a fan of the selected cooling unit is running at; and "Cooling Capacity" (model 26f, as referred to as "CC"), which is the overall cooling capacity (in percentage units, such as percentage compressor capacity or percentage valve opening, etc.) of a given one of the cooling units. It will be noted that the Unit Power neural network 26a may be viewed as the primary or main neural network model, as this model receives inputs from all of the Remote Temp NN model 26b, the Return Temp NN model 26c, the Supply Temp NN model 26d, the Fan% NN model 26e and the Cooling Capacity NN model 26f.

[0024] The chart of Figure 2 also illustrates which ones of the NN models 26b- 26f are used as inputs to other ones of the NN models. Some of the variables in the columns of the table shown in Figure 2 correspond to outputs of the NN models, and "X" in the row indicates that that variable is used as an input to that NN model. For example, it can be seen that outputs from the RAT NN model 26c and the SAT NN model 26d are inputs to the Cooling Capacity NN model 26f, along with the SAT setpoint and the Return Dew Point Evaporator Heat Rejection ("Qin"). The term "Qin" represents the heat rejected by the cooling unit. More specifically, it is the mass flow rate multiplied by specific heat, which is in turn multiplied by the temperature delta. The neural network models for Remote Temp (RET) 26b, Return Air Temp (RAT) 26c, Supply Air Temp (SAT) 26d, Fan% 26e and Cooling Capacity (CC) 26f are non-linear constraints to the Unit Power (UP) NN model 26a.

[0025] Referring to Figure 3, specifics of the machine learning module 24 are shown. The machine learning module 24 may perform supervised learning and regression for the neural network models 25a-26f. Essentially, the machine learning module 24 uses computational methods to "learn" information directly from collected data, as opposed to using empirical equations/models. Performance improves as the number of learning samples increase. Machine learning also differs from simply using empirical equations/models in that it acts as a "predictive" model which is based on both input and output data. The neural network models 25a-26f can be deployed on each unique unit (e.g., each CRAC unit) and learn its individual effect on system behavior. Unique system effects on the individual units are captured and reflected in the neural network model. System degradation over time is also captured in the data used to train the neural network models. The machine learning module 24 makes use of an algorithm in which weights applied to each neuron of a hidden layer may automatically be adjusted to help "train" the NN models 25a-26f, and the data collected may be used to further help train the NN models. In one specific implementation the machine learning module 24 makes use of one hidden layer having 10 neurons, a train ratio of 0.70, a validation ratio of 0.15, and a test ratio of 0.15. This means that 70% of the data is used to train the model while 30% is held in reserve for testing and verification of the model. A MATLAB® program fitnet function in its Neural Network Toolbox is utilized to execute the machine learning algorithm. The machine learning algorithm utilizes regression with neural network (NN) machine learning to model DSE unit power consumption (Unit Power) as a function of the variables: the supply air temperature setpoint (SAT SP), the Supply Air Temperature (SAT), the Return Air Temperature (RAT), Remote Air Temperature (RET), the unit fan speed (Fan%), the unit cooling capacity (CC), the Return Dew Point Evaporator Heat Rejection (Qin); and the parameters: Remote Air Temperature setpoint and ambient temperature. Ambient temperature is the temperature within a room that the cooling unit is located it. Figure 4 shows a diagram illustrating the relationships between the various neural network model inputs, the hidden layers represented by neurons 1 -9, and the output, which is the determined Unit Power (UP) for the cooling unit.

[0026] The global optimization module 28 implements a global optimization routine which is run with the DSE unit power consumption model as the objective function. Non-linear constraints used by the global optimization module 28 may include the Remote Air Temp (RET) NN model 26b, the Return Air Temp (RAT) NN model 26c, the Supply Air Temp (SAT) NN model 26d, the Fan% NN model 26e, the Cooling Cap (CC) NN model 26f, and the Return Dewpoint Evaporator Heat Rejection (Qin) calculation. The SAT SP is modified to minimize unit power consumption and RET, SAT, Remote T, Fan%, and Cooling Cap (CC) non-linear constraints must follow the rules of their respective NN models. The cooling unit energy consumption is reduced by optimizing the proper Supply Air Temperature Setpoint (SAT SP) to the physical room conditions in which the cooling unit is located, and the Remote Air Temperature (RET) requirement is maintained. The need for the user to manually, periodically input the Supply Air Temperature Setpoint (SAT SP) is eliminated. The bounds for the Supply Air Temperature Setpoint (SAT SP) may vary, but in one implementation the SAT may be set anywhere from 55° F to 75° F. One linear constraint is that the SAT must equal the SAT SP. Optimization points are realized where steady state conditions are achieved, and one hour time duration exists between points unless remote air temperature exceeds a danger threshold at which Supply Air Temp Set Point (SAT SP) can change before the one hour time duration. A MATLAB® program global search optimization algorithm may be used to perform the optimization operations. This algorithm uses an interior-point algorithm.

[0027] During testing, the system 10 has been found to reduce energy consumption of a cooling unit by up to 20% or more. Importantly, the system 10 accomplishes automatic control of the SAT SP, and thus does not require periodic user inputs for the SAT SP. The SAT SP is controlled by the system 10 to optimize (i.e., minimize) energy consumption of the cooling units being managed by the system 10 without violating any important operating or environmental criteria. The graphs of Figures 6A and 6B illustrate how a system with the control methodology described herein would work with varying IT loads and initial SAT setpoints as snapshots in time. The global search optimization algorithm of the global optimization module 28 outputs an optimized SAT setpoint to reduce unit power while meeting all constraints or to meet the remote air temperature requirement.

[0028] The foregoing description of the embodiments has been provided for purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure. Individual elements or features of a particular embodiment are generally not limited to that particular embodiment, but, where applicable, are interchangeable and can be used in a selected embodiment, even if not specifically shown or described. The same may also be varied in many ways. Such variations are not to be regarded as a departure from the disclosure, and all such modifications are intended to be included within the scope of the disclosure. Example embodiments are provided so that this disclosure will be thorough, and will fully convey the scope to those who are skilled in the art. Numerous specific details are set forth such as examples of specific components, devices, and methods, to provide a thorough understanding of embodiments of the present disclosure. It will be apparent to those skilled in the art that specific details need not be employed, that example embodiments may be embodied in many different forms and that neither should be construed to limit the scope of the disclosure. In some example embodiments, well- known processes, well-known device structures, and well-known technologies are not described in detail.

[0029] The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting. As used herein, the singular forms "a," "an," and "the" may be intended to include the plural forms as well, unless the context clearly indicates otherwise. The terms "comprises," "comprising," "including," and "having," are inclusive and therefore specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. The method steps, processes, and operations described herein are not to be construed as necessarily requiring their performance in the particular order discussed or illustrated, unless specifically identified as an order of performance. It is also to be understood that additional or alternative steps may be employed.