System with automatic chiller anti-surge and method of operating chiller to avoid future surge events

Abstract

A method of operating a chiller to avoid future surge events includes: applying chiller operating data associated with the chiller as input to one or more machine learning models; and generating a threshold for a controllable chiller variable based on an output of the one or more machine learning models to prevent a future chiller surge event from occurring, which further includes affecting operation of the chiller based on the threshold to prevent the future chiller surge event from occurring. The method can automatically control the cooler to avoid future cooler surge events.

Description

[0001] This application is a divisional application of an invention patent application with an application date of February 10, 2021, an international application number of PCT / US2021 / 017327, a national application number of 202180027586.9, and an invention name of "System with automatic cooler anti-surge and method of operating a cooler to avoid future surge events".

[0002] Cross-references to related patent applications

[0003] This application claims the benefit of and priority to U.S. Patent Application No. 16 / 805,365, filed on February 28, 2020, the entire disclosure of which is incorporated herein by reference. Background Art

[0004] The present disclosure generally relates to control systems for building heating, ventilation or air conditioning (HVAC) equipment, and more specifically to systems and methods for monitoring and controlling chillers in building HVAC systems. The HVAC system may include chillers (e.g., reciprocating chillers, centrifugal chillers, screw-driven chillers, absorption chillers, etc.) that can be used to cool a building. In order for the chiller to perform optimally, it may be desirable to operate the chiller at an operating point close to a surge state without entering a surge state. The chiller may enter a surge state when the compressor of the chiller cannot overcome the pressure difference between the inlet and outlet of the compressor, causing the chiller to operate in a surge state and causing the refrigerant to move backward through the compressor. The backward movement of the refrigerant may cause the chiller to vibrate and make a loud noise. If the chiller enters a surge state too frequently, the bearings of the compressor and the chiller may be damaged, thereby shortening their life. Summary of the invention

[0005] In an embodiment of the present disclosure, a method for operating a chiller to avoid future surge events is described. The method may include: applying chiller operating data associated with the chiller as input to one or more machine learning models; and generating thresholds of controllable chiller variables based on outputs of the one or more machine learning models to prevent future chiller surge events from occurring, further including affecting the operation of the chiller based on the thresholds to prevent future chiller surge events from occurring.

[0006] In some embodiments, the method further includes generating one or more machine learning models by: creating a training data set by detecting past surge events in historical cooler operating data, and marking a portion of the historical cooler operating data as surge cooler operating data associated with the past surge events; and training the one or more machine learning models with the training data set.

[0007] In some embodiments, the surge cooler operating data includes a time series including values ​​of one or more variables of historical cooler operating data associated with times before and during past surge events.

[0008] In some embodiments, affecting the operation of the cooler includes determining values ​​of one or more controllable cooler variables predicted to prevent future surge events based on outputs of the one or more machine learning models, the one or more controllable cooler variables including at least one of: variable speed drive frequency; pre-rotation vane position; or variable geometry diffuser position.

[0009] In some embodiments, the controllable cooler variable is a first controllable cooler variable, and affecting the operation of the cooler further includes adjusting the first controllable cooler variable to affect the cooling capacity of the cooler until the cooling capacity of the cooler reaches a desired cooling capacity; or the first controllable cooler variable reaches a threshold; and in response to the first controllable cooler variable reaching the threshold before the cooling capacity of the cooler reaches the desired cooling capacity, adjusting the second controllable cooler variable until the cooling capacity of the cooler reaches the desired cooling capacity.

[0010] In some embodiments, the chiller operating data includes at least one of refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, current variable speed drive frequency, current pre-rotation vane position, or current variable geometry diffuser position.

[0011] In some embodiments, an input to one or more machine learning models includes a value of a controllable chiller variable, and an output of the one or more machine learning models includes a probability that a future chiller surge event will occur if the value of the controllable chiller variable is used to operate the chiller, and wherein generating a threshold for the controllable chiller variable includes comparing the probability that a future chiller surge event will occur to a probability threshold; and in response to determining that the probability that a future chiller surge event will occur does not violate the probability threshold, setting the threshold for the controllable chiller variable to the value of the controllable chiller variable.

[0012] In some embodiments, inputs to the one or more machine learning models include a probability that a future chiller surge event will occur, and wherein outputs of the one or more machine learning models include values ​​for thresholds of controllable chiller variables.

[0013] In some embodiments, the one or more machine learning models include at least one of a neural network, a random forest, or a support vector machine.

[0014] In some embodiments, the output of one or more machine learning models includes probabilities of multiple values ​​of a controllable cooler variable, and wherein generating a threshold for the controllable cooler variable includes selecting one of the multiple values ​​of the controllable cooler variable based on the probabilities of the multiple values ​​of the controllable cooler variable.

[0015] In yet another embodiment of the present disclosure, a system is described. The system may include one or more memory devices having instructions stored thereon, which, when executed by one or more processors, cause the one or more processors to provide chiller operating data associated with a chiller as input to one or more machine learning models; and use the one or more machine learning models to provide thresholds for controllable chiller variables to prevent future chiller surge events from occurring.

[0016] In some embodiments, the instructions cause the one or more processors to generate the one or more machine learning models by creating a training data set by detecting past surge events in historical cooler operation data, and marking a portion of the historical cooler operation data as surge cooler operation data associated with the past surge events. The instructions may further cause the one or more processors to generate the one or more machine learning models by training the one or more machine learning models with the training data set.

[0017] In some embodiments, an input to one or more machine learning models includes a value of a controllable chiller variable, and an output of the one or more machine learning models includes a probability that a future chiller surge event will occur if the value of the controllable chiller variable is used to operate the chiller, and wherein the instructions cause one or more processors to generate a threshold value for the controllable chiller variable by comparing the probability that a future chiller surge event will occur to a probability threshold; and in response to determining that the probability that a future chiller surge event will occur does not violate the probability threshold, set the threshold value for the controllable chiller variable to the value of the controllable chiller variable.

[0018] In some embodiments, the instructions cause the one or more processors to affect operation of the cooler to avoid applying a control signal having a value exceeding a threshold.

[0019] In some embodiments, the controllable cooler variable is a first controllable cooler variable, and wherein the instructions cause one or more processors to affect the operation of the cooler by: adjusting the first controllable cooler variable to affect the cooling capacity of the cooler until the cooling capacity of the cooler reaches a desired cooling capacity; or the first controllable cooler variable reaches a threshold; and in response to the first controllable cooler variable reaching the threshold before the cooling capacity of the cooler reaches the desired cooling capacity, adjusting the second controllable cooler variable until the cooling capacity of the cooler reaches the desired cooling capacity.

[0020] In yet another embodiment, a method for operating a chiller is described. The method may include generating a first threshold value of a first controllable chiller variable corresponding to a probability of a future chiller surge event occurring based on outputs of one or more machine learning models; determining whether the chiller can handle the cooling capacity based on the first threshold value of the first controllable chiller variable; generating a second threshold value of a second controllable chiller variable corresponding to a probability of a future chiller surge event occurring based on a second output of the one or more machine learning models, based on determining that the chiller will not be able to handle the cooling capacity based on the first threshold value of the first controllable chiller variable; and affecting the operation of the chiller based on the first threshold value and the second threshold value.

[0021] In some embodiments, the first controllable cooler variable is a pre-rotation vane position or a variable geometry diffuser position and the first threshold is a first upper boundary corresponding to a probability of a future cooler surge event occurring, and wherein the second controllable cooler variable is a variable speed drive frequency and the second threshold is a second upper boundary corresponding to a probability of a future cooler surge event occurring.

[0022] In some embodiments, the first controllable cooler variable is the variable speed drive frequency and the first threshold is a first lower boundary corresponding to a probability of a future cooler surge event occurring, and wherein the second controllable cooler variable is the pre-rotation vane position or the variable geometry diffuser position and the second threshold is a second lower boundary corresponding to a probability of a future cooler surge event occurring.

[0023] In some embodiments, the probability is a first probability, and wherein generating a first threshold value of a first controllable cooler variable corresponding to the probability of a future cooler surge event occurring includes obtaining a first output from one or more machine learning models indicating a second probability that surge will occur; updating a value of the first controllable cooler variable of cooler operating data based on a first difference between the second probability and the first probability exceeding a third threshold value to obtain updated cooler operating data; providing the updated cooler operating data to the one or more machine learning models to obtain a third probability; and generating the first threshold value of the first controllable cooler variable based on the updated value based on a second difference between the third probability and the first probability being less than the third threshold value.

[0024] In some embodiments, the controllable cooler variable is a first controllable cooler variable, and affecting the operation of the cooler further includes adjusting the first controllable cooler variable to affect the cooling capacity of the cooler until either: the cooling capacity of the cooler reaches a desired cooling capacity; or the first controllable cooler variable reaches a threshold; and in response to the first controllable cooler variable reaching the threshold before the cooling capacity of the cooler reaches the desired cooling capacity, adjusting the second controllable cooler variable until the cooling capacity of the cooler reaches the desired cooling capacity.

[0025] Those skilled in the art will appreciate that the summary of the invention is illustrative only and is not intended to be limiting in any way. Other aspects, inventive features and advantages of the apparatus and / or process described herein, which are defined solely by the claims, will become apparent in the detailed description set forth herein and in conjunction with the accompanying drawings. Figure 1 Consider it. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Various objects, aspects, features and advantages of the present disclosure will become more apparent and better understood by reference to the detailed description in conjunction with the accompanying drawings, in which the same reference numerals always identify corresponding elements. In the accompanying drawings, the same reference numerals generally indicate the same, functionally similar and / or structurally similar elements.

[0027] Figure 1 is a perspective view of an exemplary cooler according to some embodiments.

[0028] Figure 2 is a block diagram of a model-based chiller control system according to some embodiments.

[0029] Figure 3 According to some embodiments, Figure 1 A block diagram of an exemplary chiller control system of a chiller controller associated with a chiller.

[0030] Figure 4 According to some embodiments Figure 3 Detailed block diagram of the chiller controller environment.

[0031] Figure 5 is a diagram of an exemplary neural network for generating surge probability in accordance with some embodiments.

[0032] Figure 6 is a flow chart of an exemplary process for generating and using boundaries for controllable chiller variables to operate a chiller, according to some embodiments.

[0033] Figure 7 is a flow chart of an exemplary process for generating and using boundaries for controllable chiller variables to operate a chiller, according to some embodiments.

[0034] Figure 8 is a flow chart of an exemplary process for generating boundaries for controllable chiller variables, according to some embodiments.

[0035] Fig. 9 is a flow chart of an exemplary process for generating boundaries for controllable chiller variables, according to some embodiments.

[0036] Fig.10 According to some embodiments Figure 3 Detailed block diagram of the chiller controller environment.

[0037] Fig.11 is a diagram of an exemplary neural network for generating surge probabilities using time series data, according to some embodiments.

[0038] Fig.12 is a flow chart of an exemplary process for generating and using boundaries for controllable chiller variables to operate a chiller, according to some embodiments.

[0039] Fig.13 According to some embodiments Figure 3 Detailed block diagram of the chiller controller environment.

[0040] Fig.14 is a diagram of an exemplary neural network for generating VSD frequency boundaries in accordance with some embodiments.

[0041] Fig.15 is a diagram of an exemplary neural network for generating PRV / VGD position boundaries in accordance with some embodiments.

[0042] Fig.16 is a flow chart of an exemplary process for generating and setting boundaries for controllable chiller variables, according to some embodiments.

[0043] Fig.17 According to some embodiments Figure 3 Detailed block diagram of the chiller controller environment.

[0044] Fig.18 is a diagram of an exemplary neural network for generating PRV / VGD position boundaries in accordance with some embodiments.

[0045] Fig.19 is a diagram of an exemplary neural network for generating VSD frequency boundaries in accordance with some embodiments.

[0046] Fig. 20 is a flow chart of an exemplary process for generating and setting boundaries for controllable chiller variables, according to some embodiments.

[0047] Fig.21 is a flow chart of an exemplary process for generating controllable chiller variable boundaries based on outputs of a machine learning model, according to some embodiments. DETAILED DESCRIPTION

[0048] Referring to the accompanying drawings in general, systems and methods for monitoring cooler variables to predict controllable cooler variable boundaries or thresholds and controlling cooler inputs to remain within such boundaries or thresholds are shown according to various embodiments. When the compressor of the cooler cannot overcome the pressure difference between the inlet and outlet of the compressor, the cooler may experience surge. The pressure difference can cause the refrigerant to move backward into the discharge line and return through the compressor, as well as cause the cooler to vibrate and emit a loud noise. When the cooler experiences surge, the amount of current required to drive the cooler may also swing more than during normal operation. If the cooler enters a surge state too frequently, the compressor and bearings of the cooler may be damaged, thereby shortening their lifespans, and shortening the lifespan of the cooler itself in some cases. By actively monitoring the cooler variables and determining the boundaries of the cooler operation, the controller can apply control signals to avoid such surges, which can increase the lifespan of the cooler and reduce its energy consumption by avoiding large current swings.

[0049] In some embodiments, the cooler operation data is used to generate (e.g., train) a machine learning model, and the machine learning model is used to generate thresholds for various controllable cooler variables to avoid exceeding the thresholds, thereby avoiding potential surge events. The cooler operation data used to train the machine learning model may include various operating parameters, measured variables, or other data that characterizes the performance and / or state of the cooler. For example, the cooler operation data may include discharge pressure, suction pressure, stall voltage, bearing position, input current, or any other type of data associated with the operation of the cooler (e.g., measured variables, calculated variables, control variables, etc.). Such cooler operation data may be associated with the time when the cooler is close to (e.g., within a threshold) a surge state. The cooler operation data may include historical data collected from a cloud database and / or a local database and / or live data collected from the cooler in real time. In some embodiments, the machine learning model is continuously trained via live data collected from sensors, which detect various values ​​of the cooler's variables (e.g., characteristics) when the cooler is running and in some cases respond to the occurrence of surge.

[0050] Advantageously, the machine learning model can be used to identify controllable cooler variable boundaries that enable the cooler to operate within a surge probability threshold. For example, once the machine learning model is generated, real-time values ​​of cooler operating data can be provided as inputs to the machine learning model. Based on the real-time values, the machine learning model can provide an output indicating the probability that such surge will occur based on the input. The surge probability can be compared with the probability threshold, and if the probability is close enough to the surge probability threshold (e.g., within a predetermined range), a controllable cooler variable boundary can be generated based on the input. The cooler controller can use the boundary and generate and transmit control signals to control the motor and / or the actuator of the compressor that runs the cooler. By running the compressor, the cooler controller can control the cooling capacity of the cooler and avoid surge. In addition, because the implementation of the systems and methods described herein can enable the cooler to operate without entering a surge state or otherwise minimizing the risk of entering a surge state, the cooler can operate more efficiently and require less maintenance.

[0051] In addition, by using the systems and methods described herein, the cooler controller can cause the cooler to operate close to a surge state without entering a surge state. Such operating points can cause the cooler to operate more efficiently. For example, in some cases, the closer the cooler operates to a surge state, the more efficiently the cooler can operate. Because systems that do not implement the systems and methods described herein cannot accurately predict the cooler operating point at which surge will occur, such systems often cause the cooler to operate in a safe area away from the surge state to ensure that the cooler does not enter a surge state. By using the systems and methods described herein, the cooler can operate closer to a surge state without entering a surge state, and therefore operate more efficiently.

[0052] Reference now Figure 1 , depicting an exemplary embodiment of a chiller 100. The chiller 100 shown includes a compressor 102 driven by a motor 104, a condenser 106, and an evaporator 108. A refrigerant may circulate through the chiller 100 in a vapor compression cycle. The chiller 100 may also include a control panel 114 to control the operation of the vapor compression cycle performed by the chiller 100.

[0053] The motor 104 may be powered by a variable speed drive (VSD) 110. The VSD 110 receives alternating current (AC) power having a specific fixed line voltage and a fixed line frequency from an AC power source (not shown), and provides power having a variable voltage and frequency to the motor 104. The motor 104 may be any type of motor that may be powered by the VSD 110. For example, the motor 104 may be a high-speed induction motor. The compressor 102 is driven by the motor 104 to compress the refrigerant vapor received from the evaporator 108 through the suction line 112, and deliver the refrigerant vapor to the condenser 106 through the discharge line 124. The compressor 102 may be a centrifugal compressor, a screw compressor, a scroll compressor, a turbine compressor, or any other type of suitable compressor.

[0054] The compressor 102 may be operated close to a surge state to operate efficiently. Whether the compressor 102 enters a surge state may be a function of the liquid flow through the compressor 102 and the pressure ratio between the inlet and the outlet of the compressor 102. In some cases, the higher the flow rate, the higher the pressure ratio may be before the cooler enters a surge state. For example, when the pressure ratio is too high for the corresponding amount of liquid flowing through the compressor 102, the compressor 102 may experience surge. Generally and in some cases, the closer the compressor 102 is operated to a surge state without entering a surge state, the more efficiently the compressor 102 can operate.

[0055] The evaporator 108 may include an internal tube bundle, a supply line 120, and a return line 122 for supplying and removing a process fluid to the internal tube bundle. The supply line 120 and the return line 122 may be in fluid communication with components within the HVAC system (e.g., an air handler) via a conduit that circulates the process fluid. The process fluid is a refrigerated liquid used to cool a building and may be, but is not limited to, water, ethylene glycol, calcium chloride brine, sodium chloride brine, or any other suitable liquid. The evaporator 108 is configured to reduce the temperature of the process fluid as the process fluid passes through the tube bundle of the evaporator 108 and exchanges heat with the refrigerant. Refrigerant vapor is formed in the evaporator 108 by the refrigerant liquid delivered to the evaporator 108 exchanging heat with the process fluid and undergoing a phase change to refrigerant vapor.

[0056] The refrigerant vapor delivered by the compressor 102 to the condenser 106 transfers heat to the fluid. Due to the heat transfer with the fluid, the refrigerant vapor condenses into a refrigerant liquid in the condenser 106. The refrigerant liquid from the condenser 106 flows through an expansion device (not shown) and returns to the evaporator 108 to complete the refrigerant cycle of the chiller 100. The condenser 106 includes a supply line 116 and a return line 118 for circulating the fluid between the condenser 106 and an external component of the HVAC system (e.g., a cooling tower). The fluid supplied to the condenser 106 via the return line 118 exchanges heat with the refrigerant in the condenser 106 and is removed from the condenser 106 via the supply line 116 to complete the cycle. The fluid circulating through the condenser 106 may be water or any other suitable liquid.

[0057] The refrigerant may have an operating pressure of less than 400 kPa or about 58 psi. For example, the refrigerant may be R1233zd. R1233zd is a non-flammable fluorinated gas that has a low global warming potential (GWP) relative to other refrigerants used in commercial chiller assemblies. GWP is a metric developed to allow comparison of the global warming impact of different gases by quantifying how much energy the emission of 1 ton of gas will absorb relative to the emission of 1 ton of carbon dioxide over a given period of time.

[0058] In some embodiments, a sensor (not shown) of the cooler 100 generates cooler operation data of the cooler 100 based on the operation of the cooler 100. The cooler operation data may also include a refrigerant type. The cooler operation data may be used to predict whether the cooler 100 may enter a surge state soon or immediately. As described herein, soon or immediately may be any amount of time in the future, such as, for example, 1 second, 5 seconds, 10 seconds, 1 minute, 5 minutes, 10 minutes, 30 minutes, 1 hour, etc. Examples of types of cooler operation data that the sensor may generate include, but are not limited to, refrigerant type, discharge pressure (Pd), suction pressure (Ps), stall voltage, stall current, input current, and bearing position. The discharge pressure may be the pressure generated at the output of the compressor 102. The refrigerant type may be the type of refrigerant currently flowing through the compressor 102. The suction pressure may be the pressure generated at the input of the compressor 102. A large difference between the discharge pressure and the suction pressure may cause surge to occur in the cooler because the pressure difference may cause the coolant to flow back into the compressor 102. The stall voltage and the stall current may be the voltage and the current across the motor 104, respectively, when the motor 104 is stalled. The stall voltage and the stall current may be fixed values ​​that may affect when the chiller 100 surges. The input current may be the amount of current currently driving the motor 104. A sudden increase or decrease in the input current may indicate that surge is occurring or will occur soon. The bearing position may be the position of various bearings on the shaft of the motor 104 driving the compressor 102. If the bearing position of the chiller 100 deviates too much from the position, the chiller 100 may surge.

[0059] In some embodiments, the chiller operating data can be used to determine that surge will occur if the VSD frequency is below or above a threshold, if the VGD position is above or below a threshold, and / or if the PRV position is above or below a threshold. Once any of these values ​​are determined, a control signal can be implemented to avoid causing the VSD frequency, VGD position, or PRV position to exceed or fall below the associated threshold. In some embodiments, the threshold is associated with the probability of surge occurring. For example, an administrator can enter a surge probability threshold of 40%. A threshold can be generated that enables the chiller 100 to operate with a surge probability of 40%.

[0060] Reference now Figure 2, a block diagram of a model-based chiller control system 200 according to some embodiments is shown. The chiller control system 200 shown includes a chiller controller 202, a VSD 207, a motor 209, an actuator 210, and a compressor 216. The compressor 216 can be operated in a chiller such as the chiller 100. The chiller controller 202 can receive input 204 from various sensors of the chiller, which indicates when the chiller is operating near a surge state and / or when the chiller is operating in a surge state. The input can include VSD frequency, pre-rotation vane (PRV) position / variable geometry diffuser (VGD) position, and / or values ​​of various other characteristics associated with the operation of the chiller, such as the pressure difference between the condenser and the evaporator of the chiller (e.g., the pressure difference between the inlet and outlet of the compressor 216). The PRV position can be a valve position at the inlet of the compressor 216. The VGD position can be a valve position at the outlet of the compressor 216. The chiller controller 202 may receive these inputs and generate a VSD control signal 206 to control a motor 209 driving a compressor 216, and / or generate a PRV / VGD control signal 208 to control an actuator 210 to control the outlet and / or inlet of the chiller's compressor 216 to avoid future surges or currently occurring surges.

[0061] To generate control signals 206 and 208 to control actuator 210, chiller controller 202 may use a mapping technique that maps motor speed, PRV position and / or VGD position, and / or pressure difference between the condenser and the evaporator to a predetermined value that indicates whether the chiller is in a surge state or whether surge is about to occur. For example, chiller controller 202 may maintain a database including a surge map, and chiller controller 202 may compare each of the values ​​described above with the surge map to determine whether the chiller is operating in a surge state or whether surge is about to occur. The surge map may include hard-coded values ​​of chiller variables that indicate when the chiller is experiencing surge. The surge map may include a control signal that corresponds to the value of the variable and should be based on these values ​​to get the chiller out of the surge state. If chiller controller 202 determines that the chiller is operating in a surge state or surge is about to occur, chiller controller 202 may identify a predetermined control signal from the surge map to send to actuator 210 to control compressor 216 to avoid or stop surge.

[0062] The actuator 210 may receive the signal and adjust the operation of the compressor 216 to avoid or stop any surge. The actuator 210 may include an actuator that controls the PRV position and / or VGD position of the compressor 216 (e.g., a control valve at the inlet and / or outlet of the compressor 216). The actuator 210 may operate based on the power and control signals they receive from the chiller controller 202. The motor 209 may receive a voltage from the VSD 207. The chiller controller 202 may control the voltage received by the motor 209 from the VSD 207 by adjusting the frequency of the VSD control signal 206 sent by the chiller controller 202 to the VSD 207. The motor 209 may operate the compressor 216. Similarly, the chiller controller 202 may send a PRV / VGD control signal 208 indicating the PRV position and / or VGD position to the actuator 210 to control the valve position at the inlet and outlet of the compressor 216. The actuator 210 may set a PRV / VGD position 214 based on the control signal 208 .

[0063] In some embodiments, the mapping technique used by the chiller controller 202 to generate the VSD, PRV, and / or VGD control signals 206 and 208 is a static process and must be manually updated by an administrator to adjust for any changes in the operation of the compressor 216. The chiller controller 202 may not be able to adjust the control signals based on chiller operating conditions that change over time. The mapping technique may be manually implemented based on existing knowledge and / or testing. The chiller controller 202 may not be able to take into account various other parameters of the chiller that may be important indicators of whether surge is about to occur or whether surge is occurring. Therefore, the chiller controller 202 may rely on limited parameters, require extensive testing, and may not accurately predict surge over time as conditions of the chiller change.

[0064] Reference now Figure 3 , showing reference according to some embodiments Figure 1A block diagram of a system 300 for predicting surge in a chiller 100 is shown and described. The system 300 is shown to include a chiller controller 302, a chiller 100, a VSD 110, a cloud database 328, and a user device 332. The chiller controller 302 is shown to include a model generator 312, a machine learning model 314, a cloud database 328, and a historical database 330. The chiller controller 302 can manage chiller operating data associated with the chiller 100 and other building equipment of a building management system. In short, the model generator 312 can receive or obtain chiller operating data from the cloud database 328, the historical database 330, and / or from the chiller 100 (not shown) to train one or more machine learning models of the machine learning model 314 to predict the probability of surge of the chiller 100 at various operating points. The machine learning model 314 can be implemented as a machine learning model that can dynamically predict the probability of surge of the operating point of the chiller 100 and send the prediction to the boundary generator 316. The boundary generator 316 can use the predicted surge probability to determine the operating boundaries of the chiller 100. In some cases, the boundary generator 316 generates a boundary that is predicted to keep the probability of surge occurring at or below a surge probability threshold. As described herein, the surge probability threshold can be provided by a user, retrieved from a memory, or obtained from any other data source. For example, the surge probability threshold can indicate a surge probability that an administrator is satisfied with. The variable adjuster 318 can use the boundary to adjust the value of the controllable chiller variable to ensure that the chiller 100 operates within the boundary.

[0065] The chiller controller 302 can typically be implemented using one or more servers associated with a building management system (BMS) (not shown). The chiller controller 302 can include a remote server (cloud), a local server (on-site), or any combination thereof. The chiller controller 302 can be implemented as a local controller with the chiller 100 and the variable regulator 318 installed on-site. In some embodiments, the chiller controller 302 is part of the chiller 100 itself. For example, the chiller 100 can be a smart chiller that includes its own controller. The chiller controller 302 can be part of the controller or implemented separately as part of the chiller 100.

[0066] The chiller controller 302 is shown to include a communication interface 304 and a processing circuit 306, which includes a processor 308 and a memory 310. The memory 310 may include instructions that, when executed by the processor 308, cause the processor 308 to perform one or more functions described herein. The processor 308 described herein may be implemented using one or more processors or processing devices, such as a tensor processing unit (TPU), which is designed to use machine learning techniques to determine the probability of whether surge will occur based on a set operating point of the chiller 100, a probability of an operating boundary of the chiller 100, or the operating boundary itself.

[0067] The memory 310 may be configured to store instructions and / or other data (e.g., data related to communication, configuration, and / or address data of nodes, etc.) that may be executed by the control circuit. The processing circuit 306 may be configured to implement any of the methods described herein and / or cause such methods to be executed by a processor (e.g., processor 305). The memory 310 is shown to include a model generator 312, a machine learning model 314, a boundary generator 316, and a historical database 330. In some embodiments, the memory 310 also includes a cloud database 328. The memory 310 may include any number of components and / or modules. The processing circuit 306 may implement any of the components 312, 316, 318, and / or 330 to process the cooler operating data 326, train the machine learning model 314 to generate a probability that surge will occur based on a set cooler operating point, generate boundaries, and operate the cooler 100. The processing circuit 306 may be considered to include or be connected or connectable to a memory 310 which may be configured to be accessible by the processing circuit 306 for reading and / or writing.

[0068] The communication interface 304 can be configured to facilitate communication with any device. In addition, the communication interface 304 can be configured to securely communicate with all devices and systems of the BMS, such as the chiller 100 and sensors (not shown) that collect chiller operation data of the chiller 100. In various embodiments, communication via the communication interface 304 can be direct (e.g., local wired or wireless communication) or via a communication network (e.g., WAN, Internet, cellular network, etc.).

[0069] The model generator 312 may receive chiller operation data from a cloud database 328 and / or a historical database 330. Each of the cloud database 328 and / or the historical database 330 may be a graph database, MySQL, Oracle, Microsoft SQL, PostgreSql, DB2, a document store, a search engine, a device identifier value store, etc. The cloud database 328 and / or the historical database 330 may be configured to store chiller operation data including any number of values. The cloud database 328 may be a cloud database associated with chiller operation data including data associated with a situation of surge occurring in the chiller 100 and / or chiller operation data from other chillers that provide chiller operation data to the cloud database 328 (e.g., a connected chiller from a building and / or a chiller from another building). If the chiller controller 302 is or includes a cloud server, the cloud database 328 may be part of the chiller controller 302, or if the chiller controller 302 does not include a cloud server, the cloud database may be a database separate from the chiller controller 302. If the cloud database 328 is separate from the chiller controller 302, the cloud database 328 can provide chiller operation data input from the cloud to the chiller controller 302. The historical database 330 can be a dedicated database for collecting and storing chiller operation data of the chiller 100. The historical database 330 can be a cloud database or a locally stored database corresponding to the chiller 100 at the chiller controller 302. Examples of chiller operation data that can be stored in the cloud database 328 and / or the historical database 330 include, but are not limited to, discharge pressure, suction pressure, stall voltage, bearing position, input current, coolant type, etc. Such values ​​can be associated with the time when they were collected or generated and / or an identifier that identifies the device or time series associated with these values.

[0070] The chiller operation data 326 may be communicated to each of the cloud database 328 and the historical database 330 from the sensor generating the data. For example, the sensor may detect information associated with the operation of the chiller and generate the chiller operation data 326 to be sent to a local gateway (not shown) within the same building as the chiller 100. The local gateway may collect, aggregate, and format the chiller operation data into a machine-readable format and send the chiller operation data to the cloud database 328 or the historical database 330 for storage. Multiple building management systems may communicate with the cloud database 328 to similarly send the chiller operation data locally to the building management system for storage in the cloud database 328. The chiller operation data 326 may be manually uploaded to the database 328 and / or 330 by a technician, transmitted to the database 328 and / or 330 by another controller (not shown), sent to the database 328 and / or 330 by a sensor having an Internet connection, and the like.

[0071] The model generator 312 may train the machine learning model 314 with a training data set including cooler operating data associated with a surge boundary. The surge boundary may be a threshold between a sample of operating data indicating that the cooler is operating in a surge state and a sample of cooler operating data indicating that the cooler is operating in a normal or non-surge state. The training data set may include cooler operating data values ​​when the cooler 100 enters a surge state (e.g., data values ​​when the cooler 100 is at a cooler boundary) and other cooler operating data associated with the surge boundary. The model generator 312 may generate a training data set including only such values ​​and apply it to the machine learning model 314 until the machine learning model 314 can accurately (e.g., accurately to a threshold) predict the probability of surge based on the cooler operating data. Advantageously, by training the model generator 312 using only data associated with a surge boundary, the model generator 312 may train the machine learning model 314 to predict the probability of whether the cooler 100 will enter a surge state, rather than predicting the probability of whether the cooler 100 will be completely in a surge state. However, the training data set may include any type of training data.

[0072] In some embodiments, to generate training data, the model generator 312 may identify cooler operating data in which the cooler 100 begins to operate in a surge state. The model generator 312 may identify the cooler operating data associated with the time when the cooler 100 begins to operate in a surge state as the cooler surge data. The model generator 312 may use the cooler surge data to train the machine learning model 314 to predict the probability of surge.

[0073] The model generator 312 can also use the cooler operation data associated with the surge boundary to train the machine learning model 314. In some embodiments, the model generator 312 can use the controllable cooler variable range to identify such cooler operation data. Each or a portion of the controllable cooler variables can be associated with a specific range. The specific range can be set by an administrator. The range can include any number of values ​​or within any percentage. When the model generator 312 detects the value of the cooler operation data at the surge boundary, the model generator 312 can identify other values ​​of the cooler controllable cooler variables within the specific range of each controllable cooler variable as being associated with the surge boundary. The model generator 312 can use the surge boundary data and the cooler operation data associated with the surge boundary to train the machine learning model 314.

[0074] For example, the model generator 312 can detect that the chiller 100 enters a surge state having an operating point with a VSD frequency of 49 Hz, a PRV position of 25% open, and a VGD position of 25% open. The VSD frequency can be associated with a range of 2 Hz; the PRV position can be associated with a range of 5%; and the VGD position can be associated with a range of 5%. The model generator 312 can identify these values ​​and generate a training data set that includes values ​​of the VSD frequency ranging between 47 Hz and 51 Hz, values ​​of the PRV position ranging from 20% to 30%, and values ​​of the VGD position ranging from 20% to 30%. The range of each chiller variable can be any value or percentage. The values ​​of other chiller operating data can remain constant for each value. The model generator 312 can use these values ​​and other chiller operating data to generate a training data set to train the machine learning model 314.

[0075] In some embodiments, the model generator 312 can detect the time when the cooler 100 enters a surge state, and use the cooler operating data associated with the time before, during and / or after the surge occurs to train the cooler 100. The model generator 312 can use the time range to determine which cooler operating data to include in the training data set. For example, the model generator 312 can obtain an indication that the cooler 100 experienced a 5-second surge at 1:00 p.m. on October 2. The model generator 312 can obtain a time series including values ​​associated with the time in one-second increments between 12:59 p.m. and 1:01 p.m. The model generator 312 can be configured to train the machine learning model 314 using training data, which includes cooler operating data associated with the cooler 100 from 5 seconds before the cooler 100 enters the surge state until the cooler 100 leaves the surge state. Thus, the model generator 312 may generate a training data set having data associated with values ​​from 5 seconds before 1:00 PM until 5 seconds after 1:00 PM (5 seconds before the chiller entered the surge state and the time the chiller was in the surge state). The model generator 312 may input the training data set into the machine learning model 314 for training. The model generator may use chiller operating data over any time range, including, in some embodiments, a time range of data associated with a time after the chiller 100 leaves the surge state.

[0076] In some embodiments, the model generator 312 can generate a training data set by formatting the cooler surge data into a feature vector including input-output pairs. The input can be cooler operation data, and the output can be whether the cooler operation data is associated with surge. The training data may include cooler operation data marked with an output indicating that the input is associated with surge and cooler operation data marked with an output indicating that the input is not associated with surge. The training data can be input into the machine learning model 314. The machine learning model 314 can predict the probability of surge based on the input output. The output can be compared with the output of the input pair of the training data. The machine learning model 314 can adjust its internal weight based on the difference between the output of the label and the output predicted by the machine learning model 314. The model generator 312 can perform a similar process using historical data stored in the cloud database 328 and / or the historical database 330. Therefore, the model generator 312 can train the machine learning model 314 to generate a surge probability.

[0077] In some cases, the model generator 312 may receive cooler operating data 326 from the cooler 100 in real time, and train the machine learning model 314 using the real-time cooler operating data. For example, the model generator 312 may receive cooler operating data generated by a sensor from a local gateway. The model generator 312 may use the generated cooler operating data to generate a training data set. In some cases, the model generator 312 may do so in response to detecting a surge occurring in the cooler 100. In some cases, in order to generate a training data set, the model generator 312 may determine whether the received cooler operating data is at or within a threshold of a surge boundary, and mark the training data set based on the determination. The model generator 312 may input the marked training data set into the machine learning model 314 for training. Therefore, the model generator 312 may train the machine learning model 314 to predict a value and / or a surge probability based on data samples generated when the cooler 100 enters a surge state.

[0078] The model generator 312 can train the machine learning model 314 in real time and / or before implementing the machine learning model 314. For example, the model generator 312 can train the machine learning model 314 to predict the probability of surge. In some cases, the administrator can input thousands of real-world training data sets consisting of data generated from the chiller 100 or from other chillers of similar types. For example, the administrator can input a data set that includes time series data in some cases. The machine learning model 314 can use this data and predict the probability that surge will occur based on this data. For any prediction that proves to be correct (for example, the machine learning model 314 predicted a high probability that surge will occur and surge occurred), the model generator 312 can create a labeled training data set indicating that the surge prediction is correct, and feed the labeled training data set back to the machine learning model 314 so that the machine learning model 314 can adjust its internal weights / parameters accordingly. This process can be repeated for any number of iterations until the model generator 312 determines that the machine learning model 314 is accurate to a threshold. Once such a determination is made, the machine learning model 314 can be implemented to make similar predictions using live data and continuously trained based on any correct predictions. In some cases, data from incorrect predictions may be used to train the machine learning model 314. The model generator 312 may identify the prediction as incorrect, label the data associated with the prediction as a correct prediction, and feed the labeled training data through the machine learning model 314 for training. A similar process may be performed to train the machine learning model of the machine learning model 314 to predict a threshold value for PRV / VGD position or VSD frequency to avoid exceeding or falling below the threshold value. Thus, the machine learning model 314 may be generated to predict surge probability and / or threshold values ​​for VSD frequency and / or PRV / VGD position.

[0079] In some embodiments, the chiller controller 302 receives chiller operation data from chillers located in different buildings and / or at different construction sites. The chiller operation data can be organized based on the chiller type (e.g., chiller model, chiller size, etc.). The chiller operation data can be used to train and / or generate a machine learning model associated with a specific chiller type. For example, the chiller operation data of the centrifugal chillers at various construction sites can be used to train the machine learning model of the machine learning model 314 to generate surge predictions and / or various control signal boundaries for the centrifugal chillers. Similarly, in another example, the chiller operation data of the screw-driven chillers at various construction sites can be used to train another machine learning model of the machine learning model 314 to generate surge predictions and / or various control signal boundaries for the screw-driven chillers. Any number of machine learning models can be generated (e.g., one machine learning model is generated for each different chiller type), and stored in the chiller controller 302, and used as needed.

[0080] The chiller controller 302 may automatically select which machine learning model to use based on the type of chiller providing real-time values ​​to the chiller controller 302. For example, the chiller 100 may send its model number to the chiller controller 302. The chiller controller 302 may select a machine learning model associated with the model number and use the selected machine learning model to predict surge in the chiller or generate a control signal to avoid surge.

[0081] The machine learning model 314 may include one or more machine learning models, such as, but not limited to, a neural network, a random forest, a support vector machine, etc., configured to automatically determine the probability of whether a surge will occur (e.g., in a binary classification system) and / or the VSD frequency and / or PRV / VGD position boundaries associated with a surge probability threshold. The surge probability threshold may be determined based on administrator input from the user device 332, or otherwise obtained as described above. For example, one machine learning model of the machine learning model 314 may receive a chiller operating point input and determine the probability of whether a surge will occur based on the chiller operating point. Depending on the embodiment, one or more other machine learning models of the machine learning model 314 may receive the same or similar input and the surge probability threshold, and determine the boundary VSD frequency and / or PRV / VGD position for the value of the controllable chiller variable to avoid exceeding or falling below when the chiller controller 302 operates the chiller 100 (e.g., via the variable regulator 318). The machine learning model 314 may transmit its output (e.g., the probability of surge or surge boundary) to the variable regulator 318.

[0082] For example, the machine learning model 314 can be trained to predict the probability of surge of the chiller 100 based on various operating points. The machine learning model 314 can receive a data set including chiller operating points, the data set including values ​​of current refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, variable speed drive frequency, current pre-rotation vane position, and / or current variable geometry diffuser position. The machine learning model 314 can use internal weights and / or parameters to determine the probability that surge will occur based on the received data set, and transmit the probability to the boundary generator 316. For example, the machine learning model 314 can receive a data set including chiller operating points, and based on the data set, determine that there is a 75% probability that surge will occur. The machine learning model 314 can transmit the probability to the boundary generator 316 for further processing.

[0083] In another example, the machine learning model 314 can be trained to automatically predict boundaries of controllable chiller variables, such as VSD frequency and / or VGD or PRV position. The machine learning model 314 can predict boundaries of any controllable chiller variable. The machine learning model 314 can receive a data set similar to the above and use internal weights and / or parameters to determine boundaries. In some embodiments, the machine learning model 314 further generates boundaries based on a surge probability threshold. An administrator can input the surge probability threshold, and the machine learning model 314 can be trained to predict boundaries using the surge probability as input. In some embodiments, the machine learning model 314 can be trained to use a data set that does not include values ​​of controllable chiller variables for predicting boundaries with it. For example, if the machine learning model 314 is trained to predict the VSD frequency based on the chiller operating point and the input probability of surge, the machine learning model 314 can be trained to make such predictions without using the value of the VSD frequency as input. The machine learning model 314 can be trained to predict boundaries of any controllable chiller variable and / or based on any surge probability threshold.

[0084] As the chiller 100 operates, the machine learning model 314 can adapt to operational changes of the chiller 100 (e.g., based on wear and tear of components of the chiller 100) via continuous training. Thus, the machine learning model 314 can provide updated feedback regarding the probability that surge will occur based on a given chiller operating point or controllable chiller variable boundaries, which the variable regulator 318 can implement to operate the chiller 100 at an operating point associated with a surge probability threshold.

[0085] The boundary generator 316 can be configured to generate boundaries of controllable cooler variables based on a surge probability threshold. The boundary generator 316 can obtain a surge probability threshold. The surge probability threshold can range from 0% to 100%. The boundary generator 316 can identify and use the surge probability to determine the boundaries of the controllable cooler variables, which the variable regulator 318 can use to avoid operating the cooler 100 at an operating point where the surge probability exceeds the surge probability threshold. The boundary generator 316 can do this by obtaining cooler operating data and generating a data set (e.g., a feature vector) to input into the machine learning model 314. Depending on how the machine learning model 314 is configured, the boundary generator 316 can generate a data set to include cooler operating data, which includes the value of the controllable cooler variable and / or the surge probability of the surge probability threshold. The boundary generator 316 can input the data set into the machine learning model 314, receive the output, and determine one or more controllable cooler variable boundaries based on the output.

[0086] As will be described in more detail below, in some cases, the boundary generator 316 may generate a data set to input into the machine learning model 314 until the boundary generator 316 can generate controllable cooler variable boundaries that meet the surge probability threshold. For example, the boundary generator 316 may receive the surge probability threshold from the user device 332 and compare the surge probability threshold to the surge probability output by the machine learning model 314. The boundary generator 316 may determine whether the surge probability is the same or within a threshold of the surge probability threshold, and if so, transmit the control signal used as input to the machine learning model 314 as a control signal boundary. However, if the boundary generator 316 determines that the probability is not within the threshold, the boundary generator 316 may adjust one or more control signals used as input to the machine learning model 314, apply the adjusted control signal to the machine learning model 314, and determine whether the output probability of the machine learning model 314 is within a threshold of the surge probability threshold. The boundary generator 316 may repeat this process until it identifies one or more control signal boundaries that meet the surge probability threshold.

[0087] The variable regulator 318 may be configured to operate the chiller 100 and / or various other building equipment of the building. The variable regulator 318 may be configured to receive signals from the boundary generator 316, such as VSD frequency and / or PRV / VGD position boundaries. Based on the boundaries, the variable regulator 318 may generate a VSD control signal 320 and / or a PRV / VGD control signal 322 to operate the chiller 100. For example, the variable regulator 318 may receive an output from the machine learning model 314 indicating that the minimum VSD frequency threshold is 51 Hz. In another example, the variable regulator 318 may receive a control signal threshold indicating that the PRV position should not exceed 80% of full open. The variable regulator 318 may use such predictions to ensure that it does not output a VSD frequency below 51 Hz or a PRV position that causes the PRV to open more than 80% of its maximum value. Therefore, the variable regulator 318 may control the capacity of the compressor 102 to avoid increasing the probability of surge to above the surge probability threshold.

[0088] In some embodiments, the variable regulator 318 can control the chiller 100 by adjusting the first value of the first controllable chiller variable to affect the cooling capacity of the chiller 100 until the desired cooling capacity is reached or the first value of the first controllable chiller variable reaches the boundary or threshold determined as above. If the first value of the first controllable chiller variable reaches the threshold before the desired cooling capacity is reached, the variable regulator 318 can adjust the second value of the second controllable chiller variable until the desired cooling capacity is reached. For example, the upper boundary of the PRV / VGD position can be determined to be fully open 70%. In order to operate the chiller 100, the variable regulator 318 can adjust the value of the PRV / VGD position control signal to be fully open 70%. If the chiller still cannot handle the new cooling load, the variable regulator 318 can increase the VSD frequency until the desired cooling load is reached or the VSD frequency reaches the determined boundary as described above. If the variable regulator 318 cannot operate the chiller 100 to handle the new cooling load, the variable regulator 318 can transmit a message to the user device 332 indicating that the desired cooling capacity cannot be reached.

[0089] Advantageously, by controlling the chiller 100 in the manner described above, the chiller 100 can operate more efficiently. In some embodiments, for example, when attempting to increase the cooling capacity of the chiller 100, increasing the PRV / VGD position may be more efficient than increasing the VSD frequency. Therefore, by determining and using the PRV / VGD position control upper boundary to control the chiller 100 before adjusting the VSD frequency, the variable regulator 318 can cause the chiller 100 to operate more efficiently. Similarly, when attempting to reduce the cooling capacity of the chiller 100, first determining and using the VSD frequency lower boundary to operate the chiller 100 can cause the chiller 100 to operate more efficiently than first reducing the PRV / VGD position.

[0090] The actuator 324 may be an actuator that operates the compressor 102 of the chiller 100. The actuator 324 may be similar to the actuator described in reference Figure 2 The actuator 210 is shown and described. The actuator 324 can operate the PRV position and / or VGD position of the compressor 102 based on the control signal received from the variable regulator 318. The actuator 324 can control the flow of liquid (e.g., coolant) into and out of the compressor 102 by opening or closing valves at each end of the compressor 102.

[0091] The motor 104 may drive the compressor 102 as described above. The motor 104 may receive voltage from the VSD 110 based on the frequency of the VSD control signal 320 sent to the VSD 110 by the variable regulator 318. The motor 104 may operate the compressor 102 based on the voltage provided by the VSD 110 via a shaft (not shown).

[0092] Advantageously, by controlling the VSD frequency, PRV position, and / or VGD position based on the boundaries generated by the boundary generator 316, the variable regulator 318 can actively control the capacity of the chiller 100 so that the compressor 102 can operate efficiently (e.g., approaching a surge state) within a range of values ​​associated with a surge probability threshold. The variable regulator 318 can receive the operating boundaries from the boundary generator 316. The boundary generator 316 can generate such boundaries based on the surge probability threshold. The variable regulator 318 can generate control signals to control the chiller 100 accordingly. Using the machine learning model 314 to predict the surge probability or generate controllable chiller variable boundaries to operate with a predetermined surge probability can enable the variable regulator 318 to proactively avoid surges in the chiller 100, improving the efficiency of the chiller 100 operation. Such predictions are superior to other systems that do not use the systems and methods described herein (e.g., systems and methods that use surge mapping), which often require the controller to react to surges that have already occurred and determine which control signals can stop the surge, if they will be applied. Such predictions or reactions are often made using static values ​​that are not adjusted based on changes in how the chiller operates. As a result, such systems may become more inaccurate over time. Thus, in previous approaches, more surges may occur, causing the surging chiller to lose an unpredictable amount of operating time.

[0093] Reference now Figure 4, shows a detailed block diagram of a chiller controller environment 400 according to some embodiments. The chiller controller environment 400 is shown to include a historical database 330, a machine learning model 314, a boundary generator 316, a variable regulator 318, and a user device 332. The boundary generator 316 is shown to include a data processor 402, a PRV / VGD boundary generator 404, a probability identifier 406, a VSD boundary generator 408, a cooling load identifier 410, a performance simulator 412, and a signal generator 414. In short, the data processor 402 can obtain chiller operating data from the historical database 330. The data processor 402 can generate a feature vector using the chiller operating data and input the feature vector into the machine learning model 314. The machine learning model 314 can determine the probability that surge will occur based on the chiller operating data and provide the probability to the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408. The PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 can determine whether the probabilities are the same or within a surge probability threshold. The surge probability threshold can be a threshold provided to the probability identifier 406 by an administrator via an input to the user device 332. If the probability is within the surge probability threshold, the performance simulator 412 can determine whether the chiller 100 can meet the expected new cooling load (e.g., the adjusted cooling load) based on the threshold. If one of the boundary generators 404 or 408 determines that the probability is not within the surge probability threshold or the performance simulator 412 determines that the expected cooling load may not be met, the data processor 402 can provide the adjusted input to the machine learning model 314 to repeat the process until the performance simulator 412 identifies a boundary that enables the chiller 100 to meet the cooling load. The signal generator 414 can transmit the determined boundary to the variable regulator 318 to control the chiller 100.

[0094] The data processor 402 may be configured to generate a feature vector including cooler operation data, and apply the feature vector to the machine learning model 314 to determine the probability that surge will occur based on various cooler operation data. In some cases, the data processor 402 may generate a feature vector in response to receiving an indication that a new cooling load will be provided for the cooler 100 to process. For example, the cooler controller 302 may receive an indication to process a higher cooling load, and the data processor 402 may generate a feature vector using the current cooler operation data of the cooler 100 to determine the boundary in which the cooler 100 operates. The boundary may enable the cooler 100 to avoid operating under operating conditions that cause the cooler 100 to have a probability of entering a surge state exceeding a surge probability threshold. The feature vector may include a digital identifier of the cooler operation data. The digital identifier may be a value of a cooler operation data set. The cooler operation data may include, but is not limited to, the refrigerant type, discharge pressure value, suction pressure value, stall voltage, bearing position, input current, VSD frequency, PRV position, VGD position, etc. of the cooler. Any non-digital chiller operating data can be associated with a number in a database (e.g., a historical database) of the chiller controller 302. The chiller operating data can include current values ​​(e.g., current values ​​of characteristics of the chiller). The formatted chiller operating data can be applied to one or more machine learning models for processing. As described below, the data processor 402 can adjust the values ​​of controllable chiller variables used as input to the machine learning model 314 to determine the boundaries of such variables.

[0095] The probability identifier 406 may be configured to receive or identify a surge probability threshold value to operate the chiller 100 and transmit the surge probability threshold value to the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408. In some embodiments, the probability identifier 406 may generate a user interface and transmit the user interface to the user device 332 so that an administrator may enter such a surge probability threshold value. The surge probability may be any value between 0% and 100%. The probability identifier 406 may receive the surge probability threshold value and transmit it to the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 for further processing.

[0096] The machine learning model 314 may receive a feature vector including chiller operating data and generate a probability that surge will occur based on the feature vector. In some cases, the machine learning model 314 may output the probability to the PRV / VGD boundary generator 404. The PRV / VGD boundary generator 404 may be configured to generate a PRV / VGD boundary for operating the chiller 100 based on the output probability of the machine learning model 314 and the probability identified by the probability identifier 406. The PRV / VGD boundary generator 404 may compare the probability from the machine learning model 314 with the surge probability threshold from the probability identifier 406 to determine whether the probability is the same or within a set range or threshold of the surge probability threshold. The range or threshold may be any value. If the PRV / VGD boundary generator 404 determines that the probability and the surge probability threshold are within a set range of each other, the PRV / VGD boundary generator 404 may identify the PRV / VGD position value used by the data processor 402 to generate the feature vector input into the machine learning model 314 to generate the probability. The PRV / VGD boundary generator 404 may identify such position values ​​as boundaries of the PRV / VGD positions.

[0097] For example, the PRV / VGD boundary generator 404 may receive a surge probability of 60% from the machine learning model 314 based on a feature vector including a PRV / VGD position value of 25% fully open. The PRV / VGD boundary generator 404 may also receive a surge probability threshold of 55% from the probability identifier 406. The PRV / VGD boundary generator 404 may be configured to determine that any surge probability within 5% of the surge probability threshold is associated with a valid variable boundary. The PRV / VGD boundary generator 404 may compare the predicted probability to the threshold and determine that the predicted surge probability from the machine learning model 314 is within an acceptable range of the surge probability threshold and generate a PRV / VGD position boundary with a value of 25% fully open. The PRV / VGD boundary generator 404 may output the determined position boundary to the performance simulator 412 to determine whether the generated boundary enables the chiller 100 to have sufficient capacity for the new cooling load.

[0098] If the PRV / VGD boundary generator 404 determines that the predicted probability from the machine learning model 314 is not within the threshold of the surge probability threshold, the PRV / VGD boundary generator 404 can transmit a signal to the data processor 402 indicating that the boundary cannot be generated. The data processor 402 can receive the signal and generate a new feature vector including a new PRV / VGD position value. The data processor 402 can generate a new PRV / VGD position value for the feature vector based on whether the predicted surge probability is above or below the surge probability threshold. For example, in some embodiments, the data processor 402 can increase the PRV / VGD position value from a value in the previous feature vector to have a higher value. A higher PRV / VGD position value can cause the valve to open to allow more liquid to flow through the compressor 102. On the other hand, if the predicted surge probability is too low, the data processor 402 can reduce the PRV / VGD position value to increase the predicted surge probability. Because the closer the chiller 100 operates to the surge boundary, the more efficiently the chiller 100 can operate, by increasing the probability of surge, the PRV / VGD boundary generator 404 can cause the chiller 100 to operate more efficiently while operating within the administrator-set parameters. The data processor 402 and the PRV / VGD boundary generator 404 can adjust the PRV / VGD values ​​until the PRV / VGD boundary generator 404 can generate a PRV / VGD boundary associated with a predicted surge probability within a surge probability threshold.

[0099] The cooling load identifier 410 can be configured to identify a change in the required cooling that the chiller 100 must provide. The cooling load identifier 410 can identify the desired cooling load based on a change in the configuration of the building equipment or based on a change in a set point (e.g., a zone air temperature set point) of the building. In some cases, such changes can be automatically performed based on a preset schedule or based on the conditions of the identified building. In some cases, such changes may need to be based on administrator input, such as a manual change in a thermostat. The change may be associated with an increase or decrease in the cooling load. The cooling load identifier 410 can identify the change and determine the new cooling load required to implement such a change. The cooling load identifier 410 can transmit the change, including the new cooling load requirement, to the performance simulator 412 to determine whether the boundaries generated by the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 enable the chiller 100 to have sufficient capacity to handle the new cooling load.

[0100] The performance simulation device 412 can be configured to receive a new cooling load from the cooling load identifier 410 and the boundaries generated by the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 (described below). The performance simulation device 412 can simulate how the chiller 100 will operate with the boundaries generated by the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 to determine whether the boundaries will enable the chiller 100 to have sufficient cooling capacity to handle the new cooling load identified by the cooling load identifier 410. The performance simulation device 412 can make such a determination based on the current operating conditions of the chiller and the boundaries of the control signal for either or both of the PRV / VGD boundary or the VSD boundary. For example, in some embodiments, the performance simulation device 412 can determine that a higher VSD frequency and / or PRV / VGD position can result in a higher cooling capacity. The performance simulation device 412 can use a linear function, where a higher VSD frequency and / or PRV / VGD position equals a higher cooling capacity in the chiller 100. However, the performance simulator 412 may determine the cooling capacity based on the VSD frequency and the PRV / VGD position.

[0101] In some cases, the performance simulator 412 may use the same values ​​to determine the maximum cooling capacity of the chiller for a set of operating conditions that the data processor 402 uses to generate the feature vector. For example, the performance simulator 412 may use the current chiller operating values ​​for the current refrigerant type, discharge pressure, suction pressure, stall voltage, input current, and bearing position, as well as boundary values ​​for any variable speed drive frequency, current pre-rotation vane position, and / or current variable geometry diffuser position to determine what the cooling load would be if the boundary values ​​were implemented as control values ​​for controlling the chiller 100. Such values ​​may be used in the feature vector that the data processor 402 inputs into the machine learning model 314, or they may be collected from the historical database 330 or the chiller 100. In some embodiments, the performance simulator 412 may also use characteristics of the chiller 100 to determine the cooling load, such as the size of the chiller 100 and / or the compressor 102. The performance simulator 412 may simulate how the chiller 100 will operate based on these inputs and determine whether the chiller 100 can handle the cooling load identified by the cooling load identifier 410 (e.g., determine whether the chiller 100 will have a cooling capacity that exceeds the cooling load identified by the cooling load identifier 410). If the performance simulator 412 determines that the inputs can handle the identified cooling load, the signal generator 414 may generate and transmit a signal to the variable regulator 318, the signal including the boundary values ​​generated by the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408. However, if the performance simulator 412 determines that the chiller 100 cannot handle the identified cooling load using the determined boundaries, the performance simulator 412 may transmit a signal to the data processor 402 to indicate that new boundaries need to be generated.

[0102] The data processor 402 may receive such a signal and generate a new feature vector that includes adjusted values ​​of controllable chiller variables that are different from the controllable chiller variables used to generate the previous boundaries. For example, if the performance simulator 412 determines that the generated boundaries for the PRV / VGD positions do not enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load, the data processor 402 may increase the VSD frequency of the previous features in the new feature vector. The data processor 402 may generate a new feature vector with an increased VSD frequency and input it into the machine learning model 314. The machine learning model 314 may generate a predicted surge probability based on the new feature vector and transmit the predicted surge probability to the VSD boundary generator 408.

[0103] The VSD boundary generator 408 may receive the predicted surge probability and the surge probability threshold from the probability identifier 406 and compare the two probabilities, similar to how the PRV / VGD boundary generator 404 compares the predicted surge probability to the surge probability threshold. The VSD boundary generator 408 may determine whether the predicted surge probability is within a threshold or range of the surge probability threshold. If the VSD boundary generator 408 determines that the probability is within a threshold of the surge probability threshold, the VSD boundary generator 408 may generate a VSD frequency boundary with a VSD frequency that is used in generating a feature vector for the predicted surge probability. The VSD boundary generator 408 may transmit the generated boundary to the performance simulator 412 to determine whether the VSD frequency boundary enables the chiller 100 to have a chiller capacity that meets a new cooling load similar to that described above. On the other hand, if the VSD boundary generator 408 determines that the probability is not within a threshold of the surge probability threshold, the data processor 402 may further adjust the VSD frequency in the new feature vector for application to the machine learning model 314. The data processor 402 may generate feature vectors with new VSD frequencies until the VSD boundary generator 408 generates a boundary associated with a predicted surge probability within a predetermined range of the surge probability threshold.

[0104] The performance simulator 412 may receive the boundaries generated by the VSD boundary generator 408 to determine whether the chiller 100 can operate at a sufficiently high cooling capacity to meet the new cooling load. The performance simulator 412 may make such a determination similar to how the performance simulator 412 determines the maximum cooling capacity based on the PRV / VGD boundaries generated by the PRV / VGD boundary generator 404. If the performance simulator 412 determines that the chiller 100 can provide sufficient cooling capacity to meet the new cooling load, the signal generator 414 may generate and / or transmit a signal to the variable regulator 318 with the generated VSD boundaries and / or PRV / VGD boundaries. In some embodiments, if the performance simulator 412 determines that the chiller 100 will not have a sufficiently high cooling capacity to meet the cooling load based on the generated boundaries, the performance simulator 412 may generate an alarm based on the surge probability threshold and transmit the alarm to the user device 332, the alarm indicating that the chiller 100 cannot meet the cooling load.

[0105] In some embodiments, instead of simulating whether the generated boundaries will enable the chiller to handle a given cooling load, the boundary generator 316 transmits the boundaries to the variable regulator 318, and the variable regulator 318 sends a control signal at the generated boundaries to the chiller 100. The variable regulator 318 determines the cooling capacity of the chiller 100 based on the control signal, and determines whether the expected cooling capacity is met. If the variable regulator 318 determines that the expected cooling is not met, the variable regulator 318 can transmit an indication to the data processor 402 indicating that a new boundary is needed, similar to the performance simulator 412. Otherwise, the variable regulator 318 can operate the chiller 100 based on the generated boundaries.

[0106] The boundary generator 316 may generate the control variable boundaries in a different order based on whether the change in cooling load is increasing or decreasing. For example, when the cooler cooling load needs to be increased, the boundary generator 316 may generate an upper boundary for the PRV / VGD position. If the boundary generator 316 determines that the cooler cooling load is not at the desired amount when the PRV / VGD position reaches its upper threshold, the boundary generator 316 may generate an upper boundary for the VSD frequency. On the other hand, when the cooler load needs to be reduced, the boundary generator 316 may first generate a lower boundary for the VSD frequency. If the boundary generator 316 determines that the cooler cooling load is not at the desired amount when the VSD frequency reaches the lower boundary, the boundary generator 316 may generate a lower boundary for the PRV / VGD position. The upper boundary may define to what extent a particular variable may be increased before the probability of surge becomes too high, and the lower boundary may define to what extent a particular variable may be reduced.

[0107] Advantageously, by implementing the systems and methods described herein, an administrator can input a surge probability threshold, and the chiller controller can automatically determine operating boundaries to implement based on the input surge probability threshold. The chiller can operate to avoid entering a surge state and therefore operate more efficiently.

[0108] Reference now Figure 5 , a diagram showing an exemplary neural network 500 for generating a surge probability according to some embodiments. The neural network 500 may be a reference Figure 3An exemplary embodiment of a machine learning model 314 is shown and described. The neural network 500 shown includes inputs associated with characteristics of the chiller 100. The inputs may include, for example, a discharge pressure 502 (e.g., the pressure at the outlet of the compressor 102), a suction pressure 504 (e.g., the pressure at the inlet of the compressor 102), a VSD frequency 506, a PRV / VGD position 508, a stall voltage 510, an input current 512, a refrigerant type (not shown), and a bearing position 514. In some embodiments, the neural network 500 is trained to generate surge predictions using each of the inputs 502 to 514 in addition to either the VSD frequency 506 or the PRV / VGD position 508. The inputs may include any type of input related to the operation of the chiller. For example, the inputs may also include a current control signal implemented by the controller to operate the chiller 100. The inputs 502 to 514 may be input nodes that provide outputs to nodes of the hidden layer 516. The nodes of the hidden layer 516 can perform various functions on the inputs and provide outputs (e.g., weighted signal outputs) to the output layer 518. The output layer 518 can include one or more output nodes and indicate the probability that surge will occur. The neural network 500 can include any number of components (e.g., any number of hidden layers, input nodes, output nodes, etc.).

[0109] The output of each of the input nodes 502 to 514 and the nodes of the hidden layer 516 can be a signal or a combination of signals of the neural network 500. Each of the signals or the combination of signals can be associated with a weight. For example, the input signal associated with the discharge pressure 502 can have a weight of 0.2, while the input signal associated with the bearing position 514 can have a weight of 0.7. The weight can indicate the importance of a particular input compared to other inputs. For example, the input associated with the stall voltage 510 can have a higher weight than the input associated with the input current 512. Other inputs that can be associated with higher weights can include the discharge pressure 502 and the suction pressure 504. In some embodiments, the input associated with the difference between the discharge pressure 502 and the suction pressure 504 can have a higher weight than other weights. The signal of each input can have any weight. The weight can be adjusted via training as described below. In addition, as described below, the nodes of the hidden layer 516 can perform one or more operations on the signal, and then output the weighted signal to the output layer 518.

[0110] Inputs 502 to 514 may be values ​​associated with the operation of the chiller 100. As the chiller 100 continues to operate, inputs 502 to 514 may be fed into the neural network 500 so that the neural network 500 may determine in real time whether a surge will occur. Inputs 502 to 514 may be continuously fed into the neural network 500 or fed into the neural network 500 as samples at periodic intervals (e.g., as time series data). The data of inputs 502 to 514 may be associated with a timestamp indicating when the chiller operation data was generated and / or an identifier associated with the chiller 100 or a time series of chiller operation data (e.g., a device identifier, a time series identifier, a MAC address, an IP address, etc.). For example, the neural network 500 may receive a time series of chiller operation data every minute. The time series may include values ​​of inputs 502 to 514 at intervals, such as every 5 milliseconds, 10 milliseconds, 20 milliseconds, etc. Each value may be associated with a timestamp indicating when it was generated by an associated sensor, when it was collected by a gateway that passed it to a building management system server, or when it was received by the chiller controller 302. The values ​​may also be associated with an identifier that indicates which chiller the values ​​are associated with and / or which device (e.g., a sensor) generated the chiller operating data. The timestamps and identifiers may be included as inputs in the neural network 500.

[0111] The input nodes associated with inputs 502 to 514 may be a first layer of the neural network 500 representing an input layer of the neural network 500. Each input may be a node of the input layer that sends a weighted signal to one or more nodes of the hidden layer 516. The inputs 502 to 514 may be values ​​from which values ​​collected by a server or processor associated with the neural network 500 are converted into numerical values, binary codes, matrices, vectors, etc. For example, the refrigerant type is converted into a numerical value based on a related number in a database of a server (not shown). The server may then normalize the numbers to values ​​between -1 and 1 using any technique so that operations may be performed on the numbers by the nodes of the hidden layer 516. The server may normalize the numbers to any range of values. After the server normalizes the numbers to values ​​between -1 and 1, the neural network 500 may transmit the values ​​as weighted signals to the nodes of the hidden layer 516.

[0112] Hidden layer 516 may be one or more node layers that receive input signals or combinations of input signals from inputs 502 to 514. The nodes of hidden layer 516 may perform one or more operations on the input signals or combinations of signals and provide signals or combinations of signals to output layer 518. Although two hidden layers are shown, any number of hidden layers may exist. The nodes of each hidden layer may provide output signals to the nodes of other hidden layers or to output layer 518. In some embodiments, the number of nodes or layers of hidden layer 516 is related to the number of input nodes and / or the number of output nodes of neural network 500. Neural network 500 may perform operations such as multiplication, linear operations, sigmoid / hyperbolic tangent or any other activation function at each layer of hidden layer 516 based on the values ​​associated with input nodes 502 to 514 and the weights associated with the signals or combinations of signals transmitted between input nodes 502 to 514 and nodes of hidden layer 516. The weighted signals or combinations of signals from hidden layer 516 may be sent to output layer 518.

[0113] The output layer 518 can be a layer of the neural network 500 dedicated to providing the probability of surge occurring in a binary classification system. In such a binary classification system, the output layer 518 is shown as including nodes representing the probability that surge will occur. The probability can be determined by the neural network 500 based on the inputs 502 to 514. If the inputs 502 to 514 indicate that surge will likely occur, the nodes of the output layer 518 can be associated with a high probability (e.g., a probability of more than 50%). If the inputs 502 to 514 indicate that surge will be unlikely to occur, the nodes of the output layer 518 can be associated with a low probability (e.g., a probability of less than 50%). As described above, the neural network 500 can send the determined probabilities to the PRV / VGD boundary generator 404 and / or the VSD boundary generator 408 to generate the cooler control signal boundaries.

[0114] The weights associated with the signals or combinations of signals propagating between the inputs 502 to 514 and the hidden layer 516 and then between the hidden layer 516 and the output layer 518 may be automatically determined based on training data provided by an administrator. The training data may include chiller operating data collected from a cloud database (e.g., cloud database 328). The data may include input data labeled with actual outputs (e.g., chiller characteristic values ​​and a flag indicating whether surge has occurred). The neural network 500 may initially have randomized weights associated with each of its signals or combinations of signals, but after a sufficient amount of training data has been input into the neural network 500, the weights may be determined to a degree of certainty that the administrator may identify as sufficient.

[0115] Reference now Figure 6, a flow chart of an exemplary process for generating and operating a chiller using boundaries of controllable chiller variables according to some embodiments is shown. Process 600 can be performed by any controller or processor of a building management system. For example, process 600 can be performed by one or more data processing systems (e.g., chiller controller 302). Process 600 can enable a chiller (e.g., chiller 100) including a compressor (e.g., compressor 102) to operate under different conditions over time within boundaries generated based on surge probability thresholds. Indications of changes in surge probability thresholds and cooling loads can be received. Controllable chiller variable boundaries can be generated based on indications of changes in surge probability thresholds and cooling loads. Thus, the chiller can be operated to avoid surges within surge probability thresholds. Process 600 can include any number of steps, and the steps can be performed in any order.

[0116] Process 600 is shown as including receiving input indicating a surge probability threshold and chiller operating data (step 602). For example, step 602 can be performed by boundary generator 316. An interface can be generated at a user device. The interface may include a field for an administrator to enter a surge probability threshold. The administrator can enter a surge probability between 0% and 100% of the surge probability threshold. In some embodiments, input can be automatically received from another controller or processor of the BMS. In addition, chiller operating data can be received. The chiller operating data may include values ​​for refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, current variable speed drive frequency, current pre-rotation vane position, or current variable geometry diffuser position. The chiller operating data may be current chiller operating data (e.g., values ​​associated with how the chiller is currently operating).

[0117] Process 600 is shown as including receiving an indication of a change in cooling load (step 604). For example, step 604 can be performed by boundary generator 316. An indication of a change in cooling load can be received based on a change in a set point of the BMS. The cooling load can be a cooling load required by the chiller. For example, an occupant of the BMS can manually change the zone air temperature set point of the building via a thermostat. Therefore, the thermostat can send an indication of a change in the zone air temperature set point to the BMS or to a processor within the BMS (e.g., chiller controller 302). The change can be identified, and a new cooling load can be determined based on the change in the zone air temperature set point. In another example, a building can host an event that accommodates a large number of people. The building's sensors sense an increase in temperature due to high occupancy and send an indication of the temperature increase to the processor of the BMS. A change in cooling load can be identified based on high occupancy. In yet another example, there can be a predetermined change in cooling load based on a predetermined event that occurs. The processor can determine when a predetermined event occurs and determine that there is a change in the cooling load of the chiller. For example, a planned increase in cooling load may be associated with an event that requires a large number of occupants.

[0118] Process 600 is shown as including determining whether the required cooling load has increased (step 606). For example, step 606 can be performed by boundary generator 316. An indication of a change in the cooling load can be identified. If the chiller is required to provide a higher cooling capacity, the required cooling load may increase, and if the chiller is required to provide a lower cooling capacity, the cooling load may decrease. Examples of situations in which the chiller may be required to provide a higher cooling capacity include, but are not limited to, events that may be attended by a large number of occupants, sudden reductions in zone air temperature set points, connection of new control systems that cause heat during operation, and the like. Examples of situations in which the required cooling load decreases include, but are not limited to, when there is a sudden decrease in the occupancy of a building or space, a predetermined decrease in the operation of building equipment, an increase in zone air temperature set points, and the like.

[0119] If it is determined at step 606 that the required cooling load is increased, process 600 is shown as including determining an upper boundary of the PRV / VGD position based on the surge probability threshold and the cooler operating data (step 608). For example, step 608 can be performed by the boundary generator 316. A feature vector including the cooler operating data can be generated. The feature vector can be applied to a machine learning model to determine the probability that surge will occur based on various cooler operating data. In some cases, the feature vector can be generated in response to an indication received that the cooling load to be handled by the cooler will increase. For example, an indication can be received that the cooler needs to handle an increased cooling load. A feature vector including current cooler operating data for the cooler can be generated and applied to the machine learning model, the current cooler operating data including the value of the PRV / VGD position.

[0120] The machine learning model can receive a feature vector including chiller operating data and generate a predicted probability that surge will occur based on the feature vector. The predicted surge probability can be compared to a surge probability threshold. It can be determined whether the predicted probability and the surge probability threshold are within a threshold or range of each other. If it is determined that the predicted surge probability and the surge probability threshold are within a threshold or range of each other, a PRV / VGD position value can be identified from the feature vector input into the machine learning model. The PRV / VGD position value can be generated as a boundary or candidate boundary for controlling the chiller.

[0121] For example, a feature vector may be generated and input into a machine learning model that includes cooler operating data and values ​​of the PRV / VGD position. The machine learning model may generate a predicted surge probability based on the feature vector. The predicted surge probability may be compared to a surge probability threshold to determine whether the probabilities are within each other's thresholds. If the probabilities are not within each other's thresholds, a new feature vector may be generated with a new PRV / VGD position value that depends on whether the predicted surge probability is above or below the surge probability threshold. For example, in some embodiments, if the predicted surge probability is higher, the new feature vector may include a higher PRV / VGD position value. However, if the predicted surge probability is lower, the new feature vector may include a lower PRV / VGD position value. New feature vectors may be generated until the machine learning model generates a predicted surge probability within a threshold of the surge probability threshold.

[0122] Process 600 is shown as including determining whether the upper boundary enables the cooler to reach the new required cooling (step 610). For example, step 606 can be performed by the boundary generator 316. The operation of the cooler can be simulated based on the generated PRV / VGD position upper boundary to determine whether the boundary enables the cooler to provide sufficient cooling capacity for the new cooling load. The determination can be made based on the current operating conditions of the cooler and the PRV / VGD position boundary. In some cases, the same or similar values ​​can be used to determine the maximum cooling capacity of the cooler for the same set of operating conditions for generating the feature vector that predicts the probability of surge. For example, if the PRV / VGD position boundary value is implemented as a control value for controlling the cooler, the current cooler operating values ​​of the current refrigerant type, discharge pressure, suction pressure, stall voltage, input current and bearing position and the boundary value of the PRV / VGD position can be used to determine what the cooling load will be. Such values ​​may have been used in the feature vectors that may have been input into the machine learning model or may have been collected from a database or cooler. In some embodiments, the characteristics of the cooler can be used to determine the cooling load, such as the size of the cooler or the compressor of the cooler. Operation of the chiller may be simulated based on these inputs, and a determination may be made as to whether the chiller is able to handle the new cooling load. If it is determined based on the PRV / VGD upper boundary that the chiller can handle the identified cooling load, then at step 614, an upper PRV / VGD position boundary may be generated for use in controlling the chiller. However, if it is determined based on a control signal below the generated PRV / VGD upper boundary that the chiller cannot handle the new cooling load, then at step 612, an upper boundary for the VSD frequency may be determined based on the same surge probability threshold.

[0123] For example, step 612 may be performed by the boundary generator 316. A new feature vector may be generated that includes a new value for the VSD frequency control signal. For example, if it is determined that the generated boundary for the PRV / VGD position does not enable the chiller to provide sufficient cooling capacity to handle the new cooling load, the new feature vector may include a value for the VSD frequency that is higher than the VSD frequency value of the previous feature vector. A new feature vector including a higher VSD frequency may be generated and used as an input to the machine learning model. The machine learning model may generate a new predicted surge probability based on the new feature vector. The new predicted surge probability and the new feature vector may be used to generate an upper boundary for the VSD frequency based on a surge probability threshold.

[0124] The upper boundary of the VSD frequency can be determined similarly to how the upper boundary of the PRV / VGD position is generated. For example, the machine learning model can generate a predicted surge probability based on an increased VSD frequency. The predicted surge can be compared to a surge probability threshold. It can be determined whether the two probabilities are within a threshold of each other. If it is determined that the two probabilities are within a threshold of each other, the VSD frequency upper boundary can be generated using the same VSD frequency value used in the feature vector used to generate the predicted surge probability. At step 614, the PRV / VGD position upper boundary can be used to control the cooler.

[0125] In some embodiments, it may be determined whether the VSD frequency boundaries, and in some cases the generated PRV / VGD position boundaries, enable the chiller to have chiller capacity to meet the new cooling load. Such determinations may be made using the VSD frequency boundaries and PRV / VGD position boundaries, as well as other chiller operating conditions, to simulate how the chiller would perform if the VSD frequency at the VSD frequency boundary was applied to the chiller. In some embodiments, the VSD frequency boundaries and PRV / VGD position boundaries may be used to simulate whether the use of each boundary will enable the chiller to provide sufficient cooling capacity to meet the change in cooling load. If it is determined that neither the VSD frequency boundary nor the signal at the PRV / VGD position boundary enables the chiller to provide sufficient cooling capacity to handle the change in cooling load, a signal indicating that the chiller cannot operate within the surge probability threshold may be sent to a user device. On the other hand, if it is determined that the boundary enables the chiller to provide the required cooling capacity, the chiller may be controlled with a control signal within the specified boundary.

[0126] If it is determined at step 606 that the required cooling load is reduced, process 600 may include determining a lower boundary of the VSD frequency based on the surge probability threshold and the chiller operating data (step 616). For example, step 616 may be performed by boundary generator 316. The lower boundary of the VSD frequency may be determined similarly to how the upper boundary of the PRV / VGD position is determined at step 608. For example, the feature vector may be input into a machine learning model that includes the chiller operating data and the value of the VSD frequency. The machine learning model may generate a predicted surge probability based on the feature vector. The predicted surge probability may be compared to the surge probability threshold to determine whether the probabilities are within each other's thresholds. If the probabilities are not within each other's thresholds, a new feature vector with a new VSD frequency may be generated, which may vary depending on whether the predicted surge probability is above or below the surge probability threshold. For example, in some embodiments, if the predicted surge probability is high, the new feature vector may include a higher VSD frequency. However, if the predicted surge probability is low, the new feature vector may include a lower VSD frequency. New feature vectors may be generated and applied to the machine learning model until the machine learning model generates a predicted surge probability that is within a threshold of the surge probability threshold.

[0127] Process 600 may include determining whether the predicted VSD frequency lower boundary can reach the cooling load (step 618). For example, step 618 may be performed by boundary generator 316. Step 618 may be performed similarly to how step 610 is performed. For example, the operation of the chiller may be simulated based on the generated lower VSD frequency to determine whether the boundary enables the chiller to provide sufficient cooling capacity for the new cooling load. The determination may be made based on the current operating conditions of the chiller and the VSD frequency value. In some cases, the same or similar values ​​may be used to determine the maximum cooling capacity of the chiller for the same set of operating conditions used to generate the feature vector that predicts the probability of surge. For example, if the VSD frequency lower boundary value is implemented as a control value for controlling the chiller, the current refrigerant type, discharge pressure, suction pressure, stall voltage, input current, and current chiller operating values ​​of the bearing position and the boundary value of the VSD frequency may be used to determine what the cooling load will be. Such values ​​may have been used in the feature vector that may have been input into the machine learning model or may have been collected from a database or chiller. In some embodiments, the characteristics of the chiller may be used to determine the cooling capacity, such as the size of the chiller or the compressor of the chiller. Operation of the chiller may be simulated based on these inputs, and a determination may be made as to whether the chiller is able to handle the new cooling load. If it is determined that the chiller can handle the identified cooling load based on the VSD frequency lower boundary, then at step 614, a VSD frequency lower boundary may be generated for use in controlling the chiller. However, if it is determined that the chiller cannot handle the new cooling load based on a control signal that is above the generated VSD frequency lower boundary, then at step 620, a PRV / VGD position lower boundary may be determined based on the same surge probability threshold.

[0128] For example, step 620 may be performed by the boundary generator 316. A new feature vector may be generated that includes new values ​​for the PRV / VGD position control signal. For example, if it is determined that the boundaries of the generated VSD frequency values ​​do not enable the chiller to provide sufficient cooling capacity to handle the new cooling load, the new feature vector may include a value for PRV / VGD position control that is lower than the PRV / VGD position value of the previous feature vector. A new feature vector including lower PRV / VGD position values ​​may be generated and used as input to the machine learning model. The machine learning model may generate a new predicted surge probability based on the new feature vector. The new predicted surge probability and the new feature vector may be used to generate a lower boundary for the PRV / VGD position based on a surge probability threshold.

[0129] The lower boundary of the PRV / VGD position can be determined similarly to how the lower boundary of the VSD frequency is generated. For example, the machine learning model can generate a predicted surge probability based on the reduced PRV / VGD position. The predicted surge can be compared to a surge probability threshold. It can be determined whether the two probabilities are within a threshold of each other. If it is determined that the two probabilities are within a threshold of each other, the PRV / VGD position lower boundary can be generated using the same PRV / VGD position value used in the feature vector used to generate the predicted surge probability. At step 614, the PRV / VGD position upper boundary can be used to control the cooler.

[0130] In some embodiments, it can be determined whether the generated PRV / VGD position lower boundary enables the cooler to have a cooler capacity that meets the new cooling load. Such a determination can be made using the PRV / VGD position lower boundary and other cooler operating conditions to simulate how the cooler will perform if the PRV / VGD position lower boundary is applied to the cooler. In some embodiments, the PRV / VGD position lower boundary can be used to simulate whether the use of the boundary will enable the cooler to provide sufficient cooling capacity to meet the change in cooling load. If it is determined that the signal at the PRV / VGD position lower boundary will not enable the cooler to provide sufficient cooling capacity to handle the change in cooling load, a signal indicating that the cooler cannot operate within the surge probability threshold can be sent to the user device. On the other hand, if it is determined that the PRV / VGD position lower boundary enables the cooler to provide the required cooling capacity, the cooler can be controlled with a control signal within the specified boundary.

[0131] In some embodiments, the cooler can be controlled by adjusting the first value of the first controllable cooler variable to affect the cooling capacity of the cooler until the desired cooling capacity is reached or the first value of the first controllable cooler variable reaches a boundary or threshold generated as above. If the first value of the first controllable cooler variable reaches the threshold before the desired cooling capacity is reached, the second value of the second controllable cooler variable can be adjusted until the desired cooling capacity is reached. For example, the upper boundary of the PRV / VGD position can be determined to be fully open at 70%. In order to operate the cooler, the value of the PRV / VGD position control signal can be adjusted to be fully open at 70%. If the cooler still cannot handle the new cooling load, the VSD frequency can be increased until the desired cooling load is reached or the VSD frequency reaches the determined boundary as described above. If it is determined based on the surge probability threshold that the cooler cannot be operated to handle the new cooling load, a message indicating that the desired cooling capacity cannot be reached can be transmitted to the user device.

[0132] Reference now Figure 7, a flow chart of an exemplary process 700 for generating and operating a chiller using boundaries of a controllable chiller variable according to some embodiments is shown. For example, the process 700 may be performed by one or more data processing systems (e.g., the chiller controller 302). The process 700 may enable a chiller (e.g., the chiller 100) including a compressor (e.g., the compressor 102) to operate under different conditions over time within boundaries generated based on a surge probability threshold. An indication of a change in a surge probability threshold (e.g., an acceptable probability of surge occurring) and a cooling load may be received. Controllable chiller variable boundaries may be generated based on an indication of a change in the surge probability threshold and the cooling load. Different values ​​of the controllable chiller variable may affect the surge probability. The process 700 accounts for such changes by incrementally adjusting the controllable chiller variable value until a boundary of another controllable chiller variable that can reach the cooling load is identified. Thus, by implementing the process 700, the effects that changing the controllable chiller variable may have when determining the boundaries of such variables may be identified. The process 700 may include any number of steps, and the steps may be performed in any order.

[0133] Process 700 is shown as including receiving input indicating a surge probability threshold and chiller operating data (step 702), receiving an indication of a change in cooling load (step 704), determining whether the cooling load has increased (step 706), determining an upper boundary of the PRV / VGD position based on the surge probability threshold and the chiller operating data (step 708), and determining whether the upper boundary of the PRV / VGD position can reach the cooling load (step 710). Each of steps 702 to 710 may be performed by a process similar to that described in reference to FIG. Figure 6 Steps 602 through 610 are shown and described as being performed by the boundary generator 316 .

[0134] Process 700 is shown as including adjusting the value of the VSD frequency if the PRV / VGD upper boundary cannot reach the cooling load (step 712). For example, step 712 can be performed by the boundary generator 316. In response to the PRV / VGD upper boundary being determined to not enable the chiller to provide sufficient cooling capacity to handle the new cooling load, the value of the VSD frequency can be adjusted. The value of the VSD frequency can be adjusted in the new feature vector according to the value of the VSD frequency of the previous feature vector used to generate the PRV / VGD upper boundary. The VSD frequency value can be increased or decreased from the previous VSD frequency value. For example, it can be determined at step 710 that the PRV / VGD upper boundary does not enable the chiller to have the cooling capacity to provide the new cooling load. A new feature vector can be generated that includes a VSD frequency that is higher than the VSD frequency used to generate the PRV / VGD upper boundary. At step 708, the new feature vector can be applied to the machine learning model to determine a new PRV / VGD upper boundary. Thus, steps 708 to 712 may be iteratively performed until a PRV / VGD upper boundary is generated that is determined to provide sufficient cooling capacity to meet the new cooling load. Once the PRV / VGD upper boundary is generated, the PRV / VGD upper boundary may be implemented together with the VSD frequency used to generate the PRV / VGD upper boundary to control the chiller.

[0135] In some embodiments, steps 708 to 712 may be performed until the value of the VSD frequency reaches a threshold value set by an administrator. For example, the administrator may set the VSD frequency threshold value to 60 Hz. If steps 708 to 712 are performed until a feature vector is generated based on a VSD frequency value of 60 Hz and / or a PRV / VGD upper boundary is generated using the feature vector, and it is determined based on the PRV / VGD upper boundary that the chiller is still inoperable, a signal may be generated and transmitted to the administrator indicating that the chiller cannot provide sufficient cooling capacity to meet the cooling load based on the surge probability threshold value.

[0136] Process 700 is shown to include determining a lower boundary of the VSD frequency based on the surge probability threshold and the chiller operating data (step 716), and determining whether the lower boundary can reach the new cooling load (step 718). If it is determined at step 706 that the cooling load has not increased, each of steps 716 and 718 may be performed. Steps 716 and 718 may be performed by boundary generator 316. Figure 6 Steps 616 and 618 are shown and described to perform steps 716 and 718.

[0137] Process 700 is shown as including adjusting the value of the PRV / VGD position (step 720). In some embodiments, step 720 can be performed by the boundary generator 316. The value of the PRV / VGD position can be adjusted similarly to how the value of the VSD frequency is adjusted in step 712. For example, step 720 can be performed by the boundary generator 316 or the variable adjuster 318. In response to the VSD lower boundary being determined to not enable the chiller to provide sufficient cooling capacity to handle the new cooling load, the value of the PRV / VGD position can be adjusted. The value of the PRV / VGD position can be adjusted in the new feature vector according to the value of the PRV / VGD position of the previous feature vector used to generate the VSD frequency lower boundary. The PRV / VGD position value can be increased or decreased from the previous PRV / VGD position value. For example, it can be determined at step 718 that the VSD frequency lower boundary does not enable the chiller to have the cooling capacity to provide the new cooling load. A new feature vector may be generated that includes a PRV / VGD position that is higher than the PRV / VGD position used to generate the VSD frequency lower boundary. At step 716, the new feature vector may be applied to the machine learning model to determine a new VSD frequency lower boundary. Thus, steps 716 to 720 may be iteratively performed until a VSD frequency lower boundary is generated that is determined to provide sufficient cooling capacity to meet the new cooling load. Once the VSD frequency lower boundary is generated, the VSD frequency lower boundary may be implemented along with the PRV / VGD position used to generate the VSD frequency lower boundary to control the chiller.

[0138] In some embodiments, steps 716 to 720 may be performed until the value of the PRV / VGD position reaches a threshold value set by an administrator. For example, the administrator may set the PRV / VGD position threshold value to 80% open. If steps 716 to 720 are performed until a feature vector is generated based on the PRV / VGD position value of 80% open and the VSD frequency lower boundary is generated using the feature vector, and it is determined based on the VSD frequency lower boundary that the chiller is still inoperable, a signal may be generated and transmitted to the administrator indicating that the chiller cannot provide sufficient cooling capacity to meet the cooling load based on the surge probability threshold value.

[0139] Reference now Figure 8, a flow chart of an exemplary process 800 for generating boundaries of controllable cooler variables according to some embodiments is shown. For example, process 800 can be performed by one or more data processing systems (e.g., cooler controller 302). Process 800 can be an exemplary process detailing how boundaries of controllable cooler variables can be generated in any of steps 608, 612, 616, 620, 708, and / or 716. Process 800 can also detail how boundaries of controllable cooler variables can be generated in any of steps 1210, 1214, 1218, and / or 1222. Advantageously, process 800 can be performed to automatically generate boundaries of controllable cooler variables based on surge probability thresholds. Therefore, surge can be avoided in the cooler, while the cooler can operate close to a surge state, improving the efficiency of the cooler. Process 800 can include any number of steps, and the steps can be performed in any order.

[0140] Process 800 is shown as including receiving input indicating a surge probability threshold and cooler operating data (step 802). For example, step 802 can be performed by boundary generator 316. Step 802 can be performed similar to how step 602 or 702 is performed. For example, input indicating a surge probability threshold can be received at a user interface. The input probability can be any value between 0% and 100%. The input can be received at a user device and transmitted to one or more data processing systems. The cooler operating data can be received or obtained from a database or from a sensor of the cooler. The database can be a cloud database or a local database of one or more processors. The cooler operating data can include values ​​of refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, current variable speed drive frequency, current pre-rotation blade position, or current variable geometry diffuser position. The cooler operating data can be current cooler operating data (e.g., values ​​associated with how the cooler is currently operating).

[0141] Process 800 is shown as including providing the cooler operation data to one or more machine learning models (step 804). For example, step 804 can be performed by the boundary generator 316. The cooler operation data can be formatted as a feature vector. The feature vector can include the numerical value of the cooler operation data. For any cooler operation data associated with a word, the corresponding value of the word can be obtained from the database. For example, different types of coolants can each be associated with a specific value. In some embodiments, when these values ​​are used in the feature vector, these values ​​can be normalized to values ​​between -1 and 1. Once the feature vector is generated, the feature vector can be applied to one or more machine learning models.

[0142] One or more machine learning models can each be any of a neural network, a support vector machine, a random forest, or any other type of machine learning model. Each of the machine learning models can be trained to predict the probability of surge occurrence based on a feature vector including cooler operating data. In some embodiments, one or more machine learning models can output the probability of surge to one output node. In some embodiments, one or more machine learning models can output the probability of surge to two output nodes in a binary classification system. For example, one or more machine learning models can output the probability that surge will occur based on the cooler operating data, and output the probability that surge will not occur based on the same cooler operating data. Each of the probabilities can be between 0% and 100%. For example, a machine learning model can output a 60% probability that surge will occur and a 40% probability that surge will not occur. The probabilities may or may not add up to 100%.

[0143] Process 800 is shown as including obtaining an output from the machine learning model indicating a probability that surge will occur based on the chiller operating data (step 806). For example, step 806 can be performed by the boundary generator 316. The output can be obtained from the machine learning model based on a feature vector including the chiller operating data. The output can include a probability that surge will occur. In some embodiments, the output can also include a probability that surge will not occur.

[0144] Process 800 is shown as including determining whether the output probability is within a threshold of a surge probability threshold (step 808). For example, step 806 can be performed by boundary generator 316. The output probability can be compared to the surge probability threshold obtained in step 802. A difference between the two probabilities can be determined. The difference can be compared to a threshold. The threshold can be a predetermined threshold set by an administrator. For example, the threshold can be 7%. The threshold can be any value.

[0145] Process 800 is shown as including updating a controllable cooler variable (step 810). For example, step 810 can be performed by the boundary generator 316. If it is determined that the difference between the two probabilities is higher than the threshold, then at step 810, the controllable cooler variable of the cooler can be updated. The updated controllable cooler variable can be a controllable cooler variable for which a boundary (upper or lower) is determined or generated. Examples of controllable cooler variables include VSD frequency, PRV / VGD position, or any other controllable cooler variable that can be adjusted to control the cooler. In some embodiments, the PRV position or VGD position can be a separate controllable cooler variable (and therefore controlled separately). The controllable cooler variable can be adjusted by adjusting the value of the new feature vector according to the value in the previous feature vector provided to the machine learning model at step 804. The controllable cooler variable can be adjusted to have a value higher or lower than its previous value. For example, if the difference between the probabilities is too high, the controllable cooler variable can be adjusted to have a higher value, and if the difference between the probabilities is too low, the controllable cooler variable can be adjusted to have a lower value. Once the controllable chiller variables are adjusted in the new feature vector, at step 804, the new feature vector including otherwise identical values ​​may be provided to the machine learning model.

[0146] Process 800 is shown as including generating boundaries of controllable cooler variables based on surge probability thresholds and cooler operating data (step 812). For example, step 812 can be performed by boundary generator 316. If it is determined that the difference between the two probabilities is lower than the threshold, then at step 812, boundaries of controllable cooler variables can be generated. The boundary can be an upper boundary or a lower boundary of any controllable cooler variable. For example, the boundary can be an upper boundary or a lower boundary of a VSD frequency or a PRV / VGD position. The boundary can be used to operate the cooler. As described above, the boundary can only be used if it is determined that the boundary enables the cooler to provide sufficient cooling capacity to meet the new cooling capacity.

[0147] Reference now Fig. 9 , a flow chart of an exemplary process 900 for generating boundaries of controllable chiller variables according to some embodiments is shown. For example, process 900 can be performed by one or more data processing systems (e.g., chiller controller 302). Process 900 can be an exemplary process detailing how boundaries of controllable chiller variables can be generated in any of steps 608, 616, 708, and / or 716. Advantageously, process 900 can enable one or more data processing systems to avoid generating boundaries outside of chiller operation. Process 900 can include any number of steps, and the steps can be performed in any order.

[0148] Process 900 is shown as including determining the boundaries of a controllable cooler variable (step 902). For example, step 902 can be performed by boundary generator 902. Step 902 can be performed using the steps provided in process 800. Process 900 is shown as including determining whether the boundary is less than a maximum boundary (step 904). Step 904 can be performed by boundary generator 316. A maximum boundary can be received as an administrator input. The maximum boundary can indicate the limit at which the cooler may operate or the limit at which the cooler can operate. The maximum boundary can be the lowest value that the controllable cooler variable boundary can be, or the highest value that the controllable cooler variable boundary can be. For example, the maximum boundary can be the lowest that the VSD frequency lower boundary can be, or the highest that the VSD frequency upper boundary can be. In some cases, the maximum boundary can be input for both the lowest controllable cooler variable boundary that can be and the highest controllable cooler variable boundary that can be. The maximum boundary can be set for any number of controllable cooler variables. In some embodiments, the maximum boundary can indicate 80% of the limit at which the cooler can operate. The boundary determined in step 902 can be compared with the maximum boundary. If it is determined at step 904 that the determined boundary is less than the maximum boundary, then at step 906, the controllable chiller variable boundary may be generated at the determined boundary. On the other hand, if it is determined at step 904 that the determined boundary is greater than the maximum boundary, then at step 908, the controllable chiller variable boundary may be generated at the maximum boundary. Thus, in some embodiments, the controllable chiller variable boundary may never be generated outside of the possible operating boundaries of the chiller or the boundaries set by the administrator.

[0149] Reference now Fig.10 , shows a detailed block diagram of a chiller controller environment 1000 according to some embodiments. The chiller controller environment 1000 may include a historical database 330, a machine learning model 314, a boundary generator 316, a variable regulator 318, and a user device 332. The boundary generator 316 is shown as including a time series generator 1002, a data processor 1004, a PRV / VGD boundary generator 1006, a probability identifier 1008, a VSD boundary generator 1010, a cooling load identifier 1012, a performance simulator 1014, and a signal generator 1016. Each of the components 1004 to 1016 may be similar to the reference Figure 4 Components 402 through 414 are shown and described. Chiller controller environment 1000 may include more or fewer components, depending on its configuration.

[0150] The time series generator 1002 can be configured to obtain chiller data, such as chiller data from the historical database 330, and format the data into a time series. The time series generator 1002 can transmit the time series to the historical database 330 for storage and / or transmission to the data processor 1004. The time series can be a collection of values ​​of a specific point (e.g., an exhaust air temperature point of an air handling unit, an exhaust air temperature, an intake fan state, a zone air temperature, humidity, pressure, etc.), or a value of chiller operation data as described herein generated at different times (e.g., at periodic intervals). These values ​​may include identifiers of building devices (e.g., air handlers, VAV boxes, controllers, chillers, boilers, vents, dampers, etc.) associated with these points or be associated with them. Each time series may include a series of values ​​of the same point or chiller operation data and a timestamp for each of the data values. For example, a time series of points provided by a temperature sensor (e.g., provided by a local gateway) may include a series of temperature values ​​measured by the temperature sensor and the corresponding time of the measured temperature values. The time series may be generated by the data processor 1004. Examples of time series that may be generated by data processor 1004 are as follows:

[0151] [<key, timestamp 1 ,value 1 >,<key, timestamp 2 ,value 2 >,<key, timestamp 3 ,value 3 >]

[0152] The key is the identifier of the source of the original data sample (e.g., time series ID, sensor ID, device ID, etc.), and the timestamp i It can identify the time when the i-th sample is collected, and the value i may indicate the value of the i-th sample.

[0153] The data processor 1004 can receive and use the time series to generate a feature vector for input into the machine learning model 314. As will be described in more detail below, the machine learning model 314 can be trained to predict the probability of surge based on the time series data of the feature vector received as input. The machine learning model 314 can receive any feature vector including the time series and output such surge probability to the PRV / VGD boundary generator 1006 or the VSD boundary generator 1010 depending on the boundary being generated by the boundary generator 316.

[0154] Advantageously, by using time series data and training the machine learning model 314 to use the time series data to determine the probability of surge based on the chiller operating data, more data can be used to make such predictions. The time series can include a series of values ​​that the machine learning model 314 can use, rather than just one value for each input of the chiller operating data. Because the time series can include more data, the machine learning model 314 can have more data when it processes it to more accurately predict the probability of surge.

[0155] Reference now Fig.11 , a diagram showing an exemplary neural network 1100 for generating surge probabilities using time series data according to some embodiments. The neural network 1100 may be similar to that of reference Figure 5 The neural network 500 shown and described, except that a time series of values ​​is used as input for each input of the chiller operating data, rather than a single value. For example, instead of using the current value of the discharge pressure as an input to the neural network, a time series including multiple values ​​before the current value, and in some cases, the current value of the discharge pressure can be used as an input to the neural network 1100. The neural network 1100 may include a discharge pressure time series 1102, a suction pressure time series 1104, a variable speed drive frequency time series 1106, a pre-rotation vane position / variable geometry diffuser position time series 1108, a stall voltage time series 1110, an input current time series 1112, and a bearing position time series 1114. The time series inputs 1102 to 1114 may represent multiple values ​​of each input. Each value may be a value of a respective time series and is input into a separate input node. In some embodiments, time may also be used as an input to the neural network 1100. Each of the components 1102 to 1114 may represent an input to the neural network 1100. The neural network 1100 is further shown to include a hidden layer 1116 and an output layer 1118. The inputs 1102 to 1114 may be converted into weighted signals or combinations of signals and passed to the hidden layer 1116. The nodes of the hidden layer 1116 may perform various operations on the weighted signals, similar to the reference Figure 5 Nodes of hidden layer 516 are shown and described. Output layer 1118 may provide nodes indicating the probability that surge will occur based on inputs 1102 to 1114.

[0156] In some embodiments, the values ​​of each time series of time series inputs 1102 to 1114 can be weighted according to their proximity in time to the current value. For example, if the suction pressure time series 1104 includes 10 values, each one second apart until the current value, the current suction pressure value may be associated with the highest weight of the 10 values, and the weight associated with each value may decrease as the values ​​are further away in time from the current suction pressure value. A similar weighting system is implemented on all time series inputs 1102 to 1114. Each value can be weighted in any manner according to the training of the neural network 1100.

[0157] For example, similar to reference Figure 5 How to train neural network 500 is shown and described to train neural network 1100. For example, labeled feature vectors including time series data may be fed into neural network 1100. Feature vectors may be labeled with a flag indicating whether they are associated with surge. In some embodiments, to avoid any errors in training, any training data associated with surge may be associated with a cooler only when it is operating at or near a surge threshold. Thus, neural network 1100 may predict the probability of surge based on the time at which the cooler may enter a surge state rather than the time at which the cooler goes deep into a surge state.

[0158] Reference now Fig.12 , a flowchart of an exemplary process 1200 for generating and using boundaries of controllable cooler variables to operate a cooler is shown according to some embodiments. For example, process 1200 can be performed by one or more data processing systems (e.g., cooler controller 302). Process 1200 can enable a cooler (e.g., cooler 100) including a compressor (e.g., compressor 102) to operate under different conditions over time within boundaries generated based on surge probability thresholds. Process 1200 can be performed similar to process 600. Time series data can be generated from cooler operating data. Time series data can be used to generate controllable cooler boundaries. Advantageously, by using time series to generate controllable cooler boundaries, any predictions made by a machine learning model can be more accurate because the time series data provides more values ​​to make such predictions. Process 600 can include any number of steps, and the steps can be performed in any order.

[0159] Process 1200 is shown as including receiving input indicating a surge probability threshold and cooler operating data (step 1202). For example, step 1202 may be performed by boundary generator 316. Figure 6Step 1202 is performed by performing the step 602 shown and described. Process 1200 is shown as including generating a time series based on the chiller operation data (step 1204). The chiller data can be obtained and formatted as a time series. The time series can be transmitted to a database for storage. The time series can be a collection of values ​​of a specific point (e.g., the exhaust air temperature point of the air handling unit, the exhaust air temperature, the intake fan state, the zone air temperature, humidity, pressure, etc.), or the value of the chiller operation data as described herein generated at different times (e.g., at periodic intervals). These values ​​may include identifiers of building devices (e.g., air handlers, VAV boxes, controllers, chillers, boilers, vents, dampers, etc.) associated with these points or are associated with them. Each time series may include a series of values ​​of the same point or chiller operation data and a timestamp for each of the data values. For example, a time series of points provided by a temperature sensor (e.g., provided by a local gateway) may include a series of temperature values ​​measured by the temperature sensor and the corresponding time of the measured temperature values. A time series can be generated. An example of a time series may be as follows:

[0160] [<key, timestamp 1 ,value 1 >,<key, timestamp 2 ,value 2 >,<key, timestamp 3 ,value 3 >]

[0161] The key is the identifier of the source of the original data sample (e.g., time series ID, sensor ID, device ID, etc.), and the timestamp i It can identify the time when the i-th sample is collected, and the value i may indicate the value of the i-th sample.

[0162] The time series can be used to generate a feature vector for input into a machine learning model. The machine learning model can be trained to predict the surge probability based on the time series data of the feature vector received as input. The machine learning model can receive any feature vector including a time series and output a surge probability. The surge probability can be used to generate a PRV / VGD position upper boundary and / or a PRV / VGD position lower boundary and / or a VSD frequency upper boundary and / or a VSD frequency lower boundary.

[0163] Process 1200 is shown to include receiving an indication of a change in cooling load (step 1206). Process 1200 is shown to include determining whether the required cooling load has increased (step 1208). Process 1200 is shown to include determining an upper boundary of the PRV / VGD position based on a surge probability threshold and chiller operating data (step 1210). Process 1200 is shown to include determining whether the upper boundary can be achieved to cool the new cooling load (step 1212). Process 1200 is shown to include determining an upper boundary VSD frequency based on a surge probability threshold (step 1214). Process 1200 is shown to include determining a lower boundary of the VSD frequency based on a surge probability threshold and chiller operating data (step 1218). Process 1200 is shown to include determining whether the lower boundary can be achieved to cool the new cooling load (step 1220). Process 1222 is shown to include determining whether the lower boundary of the PRV / VGD position is based on a surge probability threshold (step 1222). Each of steps 1206 to 1214 and 1218 to 1222 may be performed by, for example, boundary generator 316. Process 1200 is shown as including operating the chiller based on the determined boundaries (step 1216). For example, step 1216 may be performed by variable adjuster 318. Figure 6 Steps 602 through 614 are shown and described to perform steps 1202 and 1206 through 1216.

[0164] Reference now Fig.13 , shows a detailed block diagram of a chiller controller environment 1300 according to some embodiments. The chiller controller environment 1300 may include a historical database 330, a machine learning model 314, a boundary generator 316, a variable regulator 318, and a user device 332. The boundary generator 316 is shown to include a data processor 1302, a probability identifier 1304, a PRV / VGD boundary generator 1306, a VSD boundary generator 1308, a cooling load identifier 1310, a performance simulator 1312, and a signal generator 1314. The cooling load identifier 1310, the performance simulator 1312, and the signal generator 1314 may be similar to those of reference 1302, 1304, 1306, 1308, 1310, 1312, and 1314, respectively. Figure 4 Shown and described are cooling load identifier 410, performance simulator 412, and signal generator 414. Chiller controller environment 1300 may include more or fewer components, depending on its configuration.

[0165] The data processor 1302 may be configured to format and generate a feature vector for input into the machine learning model 314. The data processor 1302 may format and generate a feature vector for input into the machine learning model 314 using chiller operation data from a historical database and / or live chiller operation data as described herein. The data processor 1302 may format the feature vector using a numeric identifier for each value. In some embodiments, the numeric identifier may be normalized to a value between -1 and 1. In addition to using the chiller operation data to generate a feature vector, the data processor 1302 may use a surge probability threshold value that an administrator inputs into the user device 332 or otherwise obtains in the feature vector. The surge probability threshold value may be between 0% and 100%. The probability identifier 1304 may identify the surge probability threshold value and transmit it to the data processor 1302. The data processor 1302 may receive the surge probability threshold value and use it together with the chiller operation data to generate a feature vector for input into the machine learning model 314 for processing.

[0166] The machine learning model 314 may include a PRV / VGD boundary predictor 1316 and a VSD boundary predictor 1318. The PRV / VGD boundary predictor 1316 and the VSD boundary predictor 1318 may be or include machine learning models, such as neural networks, support vector machines, random forests, etc. The PRV / VGD boundary predictor 1316 may be a machine learning model configured to predict a PRV / VGD boundary that may enable the chiller 100 (not shown) to operate with a surge probability that does not exceed a surge probability threshold. The PRV / VGD boundary predictor 1316 may be trained with training data generated from surge data (e.g., chiller operating data associated with the chiller when the chiller experiences surge). In some embodiments, the PRV / VGD boundary predictor 1316 may be trained to make predictions using the chiller operating data described herein, in addition to the control signals including the PRV / VGD signal. In some embodiments, the training data may be associated with the cooler operating data when the cooler has just crossed a surge threshold (e.g., has just entered a surge state). Once the PRV / VGD boundary predictor 1316 is trained, the PRV / VGD boundary predictor 1316 may predict a PRV / VGD boundary that enables the cooler 100 to operate within a surge probability threshold. For example, the PRV / VGD boundary predictor 1316 may receive the cooler operating data and a 75% surge probability as inputs in a feature vector. The PRV / VGD boundary predictor 1316 may predict a PRV / VGD boundary that is 60% open based on the inputs. The PRV / VGD boundary predictor 1316 may transmit the PRV / VGD boundary to the PRV / VGD boundary generator 1306 for further processing.

[0167] In addition, in some embodiments, the PRV / VGD boundary predictor 1316 may include two machine learning models. One machine learning model may predict an upper boundary of the PRV / VGD position, and another machine learning model may predict a lower boundary of the PRV / VGD position. Each machine learning model may be trained with similar data, except with markers specific to its respective boundary. The machine learning models may receive the same input (e.g., the same cooler operating data and surge probability) and output upper and lower boundaries of the PRV / VGD position, respectively. For example, each of the machine learning models may receive inputs including cooler operating data and surge probability. The machine learning model may output PRV / VGD positions of 25% open and 60% open as operating boundaries so that the cooler 100 operates without having an excessively high surge probability. Finally, in some embodiments, the PRV / VGD boundary predictor 1316 may include different machine learning models that are trained to predict PRV boundaries and VGD boundaries, respectively. Such machine learning models may include machine learning models that predict upper and lower boundaries of each variable or only predict one boundary. Such a machine learning model can be trained as described above for a machine learning model that predicts PRV / VGD location boundaries together.

[0168] The PRV / VGD boundary predictor 1316 may transmit the predicted boundaries to the PRV / VGD boundary generator 1306. The PRV / VGD boundary generator 1306 may generate the predicted boundaries and transmit the generated boundaries to the performance simulator 1312 to determine whether the generated boundaries will enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load. If the performance simulator 1312 determines that one or more boundaries will not enable the chiller 100 to provide sufficient cooling capacity, the data processor 1302 may generate a feature vector including similar chiller operating data and the same surge probability threshold value, and apply the feature vector to the VSD boundary predictor 1318 to generate one or more VSD boundaries for the variable regulator 318 to use to control the chiller 100, in addition to the generated PRV / VGD boundaries. However, if a corresponding VSD frequency boundary has been generated and it is determined that it does not enable the chiller 100 to provide sufficient cooling capacity, a message may be transmitted to the user device 332 indicating that the chiller 100 cannot provide sufficient cooling capacity with the surge probability threshold value.

[0169] Similarly, the VSD boundary predictor 1318 may be a machine learning model that is trained to predict a VSD frequency boundary based on a surge probability threshold and cooler operating data. In some cases, the VSD boundary predictor 1318 may be trained using a training method similar to the PRV / VGD boundary predictor 1316 with cooler operating data that occurs at or near a surge threshold. The VSD boundary predictor 1318 may be trained to predict a VSD boundary based on the same or similar cooler data as the PRV / VGD boundary predictor 1316, except that the VSD boundary predictor 1318 may use the PRV / VGD control signal data as input while excluding input related to VSD frequency. Therefore, any predictions made by the VSD boundary predictor 1318 may not be affected by the VSD control signal (which may not have any correlation with the predicted VSD frequency boundary). For example, the VSD boundary predictor 1318 may receive inputs including cooler operating data excluding the VSD frequency and a surge probability threshold. The VSD boundary predictor 1318 can output a VSD frequency that is predicted as a boundary for the variable regulator 318 to avoid crossing when controlling the chiller 100. In some embodiments, the VSD boundary predictor 1318 includes a machine learning model for an upper boundary of the VSD frequency and a machine learning model for a lower boundary of the VSD frequency. For example, each of the machine learning models can receive inputs including chiller operating data and a surge probability. The machine learning model can output frequencies of 45 Hz and 55 Hz as operating boundaries to operate the chiller 100 without having an excessively high probability of surge. Each of these machine learning models can be trained using similar training data, except for the labels associated with its corresponding boundary.

[0170] The VSD boundary predictor 1318 may communicate the predicted boundary to the VSD boundary generator 1308. The VSD boundary generator 1308 may generate the predicted boundary and communicate the generated boundary to the performance simulator 1312 to determine whether the generated boundary will enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load. If the performance simulator 1312 determines that the boundary will not enable the chiller 100 to provide sufficient cooling capacity, the data processor 1302 may generate a feature vector including similar chiller operating data and the same surge probability threshold value and apply the feature vector to the PRV / VGD boundary predictor 1316 to generate a PRV / VGD boundary for the variable regulator 318 to use to control the chiller 100 in addition to the generated VSD boundary. However, if the corresponding PRV / VGD position boundary has been generated and it is determined that it does not enable the chiller 100 to provide sufficient cooling capacity, a message may be transmitted to the user device 332 indicating that the chiller 100 cannot provide sufficient cooling capacity with the surge probability threshold value.

[0171] Advantageously, by using the chiller operating data and the surge probability prediction boundaries, controllable chiller operating boundaries of the chiller 100 can be accurately predicted. Instead of using an iterative process, a machine learning model can automatically predict such boundaries, which can speed up processing and enable the variable regulator 318 to more quickly adjust the control of the chiller 100 in real time.

[0172] Reference now Fig.14 , a diagram showing an exemplary neural network 1400 for predicting VSD frequency boundaries according to some embodiments. The neural network 1400 may be similar to the reference Figure 5 The neural network 500 is shown and described, except that the neural network 1400 uses a surge probability threshold to predict the value of the VSD frequency boundary. In some embodiments, each possible value of the VSD frequency boundary can be treated as a different class in a multi-class classification system. The neural network 1400 can be representative of other machine learning models that can receive the same or similar inputs to predict VSD frequency boundaries. The neural network 1400 can include input nodes for discharge pressure 1402, suction pressure 1404, probability that surge will occur 1406, pre-rotation vane position / variable geometry diffuser position 1408, stall voltage 1410, input current 1412, and bearing position 1414. The components 1402, 1404, 1408, 1410, 1412, and 1414 can be similar to those described in reference to FIG. Figure 5 1402, 1404, 1408, 1410, 1412, and 1414 are shown and described. Each of the components 1402 to 1414 may represent an input to the neural network 1400. The neural network 1400 is further shown to include a hidden layer 1416 and an output layer 1418. The inputs 1402 to 1414 may be converted into weighted signals or combinations of signals and passed to the hidden layer 1416. The nodes of the hidden layer 1416 may perform various operations on the weighted signals, similar to those described in reference to FIG. Figure 5 Nodes of hidden layer 516 are shown and described. The nodes of hidden layer 1416 may transmit further weighted signals to output layer 1418. Output layer 1418 may include one or more nodes and / or layers indicating that frequency boundaries are implemented based on inputs 1402-1414.

[0173] In some embodiments, the weights of the weighted signals may be generated or determined based on training data provided to the neural network 1400. The neural network 1400 may be trained using a method similar to that described above with reference to Figure 5The neural network 1400 may be trained using supervised, semi-supervised, or unsupervised methods of training as described. For example, the neural network 1400 may be trained using labeled training data including values ​​for each of the inputs 1402 to 1414. For example, a training data set including values ​​for each of the inputs 1402 to 1414 may be input to the neural network 1400. The training data set may be labeled with a correct value of 60 Hz. The neural network 1400 may output a VSD frequency boundary of 55 Hz based on the training data set. The correct value may be fed back into the neural network 1400 to adjust its internal weights and / or parameters to make more accurate predictions for similar data sets in the future. The accuracy percentage may be determined based on how close the neural network 1400 predicts the labeled value on multiple training data sets. For example, in some embodiments, the closer the neural network 1400 predicts the correct frequency, the higher the accuracy percentage. The training data set may be input to the neural network 1400 for training until the neural network 1400 is accurate to a threshold. In some embodiments, the training data set may be input until the neural network 1400 is accurate to a threshold for a preset number of training data sets.

[0174] Reference now Fig.15 , a diagram showing an exemplary neural network 1500 for generating PRV / VGD position boundaries according to some embodiments. The neural network 1500 may be similar to the reference Fig.14 The neural network 1400 is shown and described, except that the neural network 1500 uses the VSD frequency input 1506 as an input instead of the PRV / VGD position input to predict the value of the PRV / VGD boundary. In some embodiments, each possible value of the PRV / VGD position boundary output can be considered a different class in a multi-class classification system. The neural network 1500 can be representative of other machine learning models that can receive the same or similar inputs to predict VSD frequency boundaries. The neural network 1500 can include input nodes for discharge pressure 1502, suction pressure 1504, VSD frequency 1506, probability that surge will occur 1508, stall voltage 1510, input current 1512, and bearing position 1514. The components 1502, 1504, 1506, 1510, 1512, and 1514 can be similar to those of reference 1504. Fig.14 1402, 1404, 1408, 1410, 1412, and 1414 are shown and described. Each of the components 1502 to 1514 may represent an input to the neural network 1500. The neural network 1500 is further shown to include a hidden layer 1516 and an output layer 1518. The inputs 1502 to 1514 may be converted into weighted signals or combinations of signals and passed to the hidden layer 1516. The nodes of the hidden layer 1516 may perform various operations on the weighted signals, similar to those described in reference to FIG. Fig.14Nodes of hidden layer 1416 are shown and described. The nodes of hidden layer 1516 may transmit further weighted signals to output layer 1518. Output layer 1518 may include one or more nodes and / or layers indicating implementation of frequency boundaries based on inputs 1402-1414.

[0175] In some embodiments, the weights of the weighted signals may be generated or determined based on training data provided to the neural network 1500. The neural network 1500 may be generated or determined based on training data provided to the neural network 1500. Figure 5 The neural network 1500 may be trained using supervised, semi-supervised, or unsupervised methods of training as described. For example, the neural network 1500 may be trained using labeled training data including values ​​for each of the inputs 1502 to 1514. For example, a training data set including values ​​for each of the inputs 1502 to 1514 may be input to the neural network 1500. The training data set may be labeled with a 50% correct value. The neural network 1500 may output a 45% PRV / VGD position boundary based on the training data set. The 50% correct value may be fed back to the neural network 1500 to adjust its internal weights and / or parameters to make more accurate predictions for similar data sets in the future. The accuracy percentage may be determined based on how close the neural network 1500 predicts the labeled values ​​on multiple training data sets. For example, in some embodiments, the closer the neural network 1500 is to predicting the correct PRV / VGD position of the training data set, the higher the accuracy percentage. The training data set may be input to the neural network 1500 for training until the neural network 1500 is accurate to a threshold. In some embodiments, training data sets may be input until neural network 1500 is accurate to a threshold for a preset number of training data sets.

[0176] Reference now Fig.16 , a flow chart of an exemplary process 1600 for generating and setting boundaries of controllable chiller variables according to some embodiments is shown. For example, process 1600 may be performed by one or more data processing systems (e.g., chiller controller 302). Process 1600 may be a flowchart detailing how the process may be performed in Figure 6 , 7An exemplary process for generating boundaries of controllable cooler variables in any of steps 608, 612, 616, 620, 708, 716, 1210, 1214, 1218, and / or 1222 of process 1600 or 12. Advantageously, process 1600 can be performed to automatically generate boundaries of controllable cooler variables based on surge probability thresholds. As a result, surge can be avoided in the cooler, and the cooler can be operated close to surge conditions, improving the efficiency of the cooler. Instead of using an iterative process that may require the use of multiple inputs, a single input to a machine learning model can be used to determine the boundaries. Process 1600 can include any number of steps, and the steps can be performed in any order.

[0177] Process 1600 is shown as including receiving input indicating a surge probability threshold and cooler operating data (step 1602). For example, step 1602 can be performed by boundary generator 316. Step 1602 can be performed similarly to how steps 602, 702, or 802 are performed. For example, an input indicating a surge probability threshold can be received at a user interface. The input probability can be any value between 0% and 100%. The input can be received at a user device and transmitted to one or more data processing systems. The cooler operating data can be received or obtained from a database or from a sensor of the cooler. The database can be a cloud database or a local database of one or more processors. The cooler operating data can include values ​​for refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, current variable speed drive frequency, current pre-rotation blade position, or current variable geometry diffuser position. The cooler operating data can be current cooler operating data (e.g., values ​​associated with how the cooler is currently operating).

[0178] Process 1600 is shown as including providing cooler operating data including surge probability to one or more machine learning models (step 1604). For example, step 1604 can be performed by boundary generator 316. The cooler operating data and surge probability can be formatted as a feature vector. The feature vector can include numerical values ​​of the cooler operating data and the surge probability. For any cooler operating data associated with a word, the corresponding value of the word can be obtained from a database. For example, different types of coolants can each be associated with a specific value. In some embodiments, when these values ​​are used in the feature vector, the values ​​can be normalized to values ​​between -1 and 1. Once the feature vector is generated, the feature vector can be applied to one or more machine learning models.

[0179] One or more machine learning models may include a PRV / VGD boundary prediction model and / or a VSD boundary prediction model. The PRV / VGD boundary prediction model and the VSD boundary prediction model may be or may include machine learning models such as neural networks, support vector machines, random forests, etc. The PRV / VGD boundary prediction model may be a machine learning model configured to predict a PRV / VGD boundary that may enable the cooler to operate with a surge probability that does not exceed a surge probability threshold. The PRV / VGD boundary prediction model may be trained with training data generated from surge data (e.g., cooler operating data associated with the cooler when the cooler experiences surge). In some embodiments, the PRV / VGD boundary prediction model may be trained to make predictions using the cooler operating data described herein, in addition to control signals including PRV / VGD signals. In some embodiments, the training data may be associated with the cooler operating data when the cooler has just crossed a surge threshold (e.g., has just entered a surge state). Once the PRV / VGD boundary prediction model is trained, the PRV / VGD boundary prediction model can predict a PRV / VGD boundary that causes the cooler to operate with a surge probability below a threshold. For example, the PRV / VGD boundary prediction model can receive cooler operating data and a surge probability of 75% as input in a feature vector. The PRV / VGD boundary prediction model can predict a 60% open PRV / VGD boundary based on the input.

[0180] In addition, in some embodiments, the PRV / VGD boundary prediction model may include two machine learning models. One machine learning model may predict the upper boundary of the PRV / VGD position, and the machine learning model may predict the lower boundary of the PRV / VGD position. Each machine learning model can be trained with similar data, except with markers specific to its respective boundary. The machine learning model may receive the same input (e.g., the same cooler operating data and surge probability) and output the upper and lower boundaries of the PRV / VGD position, respectively. For example, each of the machine learning models may receive inputs including cooler operating data and surge probability. The machine learning model may output PRV / VGD positions of 25% open and 60% open as operating boundaries so that the cooler operates without having an excessively high probability of surge. Finally, in some embodiments, the PRV / VGD boundary prediction model may include different machine learning models that are trained to predict PRV boundaries and VGD boundaries, respectively. Such machine learning models may include machine learning models that predict the upper and lower boundaries of each variable or only predict one boundary. Such a machine learning model can be trained as described above for a machine learning model that predicts PRV / VGD location boundaries together.

[0181] Predicted boundaries for the PRV / VGD positions may be generated. It may be determined whether the generated boundaries will enable the chiller to provide sufficient cooling capacity to handle the new cooling load. If it is determined that the predicted boundaries will not enable the chiller to provide sufficient cooling capacity, a new feature vector may be generated that includes similar chiller operating data and the same surge probability threshold. However, the new feature vector may not include values ​​for the VSD frequency and may instead include values ​​for the PRV / VGD positions. The feature vector may be applied to a VSD boundary prediction model to generate a VSD boundary. The VSD boundary may be used to control the chiller in conjunction with the generated PRV / VGD boundary.

[0182] The VSD prediction model can be a machine learning model that is trained to predict the VSD frequency boundary based on the surge probability threshold and the cooler operation data. In some cases, the VSD prediction model can be trained with cooler operation data that occurs at or near the surge threshold using a training method similar to the PRV / VGD prediction model. The VSD prediction model can be trained to predict the VSD boundary based on the same or similar cooler data as the PRV / VGD prediction model, except that the VSD prediction model can use the PRV / VGD control signal data as input while excluding input related to the VSD frequency. Therefore, any prediction made by the VSD prediction model may not be affected by the VSD control signal (which may not have any correlation with the predicted VSD frequency boundary). For example, the VSD prediction model may receive inputs including cooler operation data excluding the VSD frequency and the surge probability threshold. The VSD prediction model can output the VSD frequency predicted as the boundary to avoid crossing when controlling the cooler. In some embodiments, the VSD prediction model includes a machine learning model for the upper boundary of the VSD frequency and a machine learning model for the lower boundary of the VSD frequency. For example, each of the machine learning models may receive inputs including chiller operating data and a surge probability. The machine learning model may output frequencies of 45 Hz and 55 Hz as operating boundaries to operate the chiller without having an excessively high probability of surge. Each of these machine learning models may be trained using similar training data, except for the labels associated with its corresponding boundaries.

[0183] Predicted boundaries can be generated. It can be determined whether the generated boundaries will enable the chiller to provide sufficient cooling capacity to handle the new cooling load. If it is determined that the boundaries are preventing the chiller from providing sufficient cooling capacity, a feature vector including similar chiller operating data and the same surge probability threshold can be generated and applied to the PRV / VGD boundary prediction model. In some cases, a PRV / VGD boundary control chiller can be generated in addition to the generated VSD boundary. However, if it has been determined that there is no PRV / VGD boundary that enables the chiller to meet the cooling load. A message can be transmitted to a user device indicating that the chiller cannot provide the necessary cooling load based on the surge probability threshold.

[0184] Process 1600 is shown as including obtaining output indicating controllable cooler variable boundaries from one or more machine learning models (step 1606). For example, step 1606 can be performed by boundary generator 316. Output including values ​​of controllable cooler variables can be received. Boundaries can be generated based on the values ​​of the controllable cooler variables.

[0185] Process 1600 is shown as including setting boundaries of controllable chiller variables based on output indicating thresholds of controllable chiller variables (step 1608). For example, step 1608 may be performed by boundary generator 316. The boundaries may be upper and / or lower boundaries of any controllable chiller variable. For example, the boundaries may be upper and / or lower boundaries of VSD frequency or PRV / VGD position control signals. The boundaries may be used to operate the chiller. As described above, the boundaries may be used only if it is determined that the boundaries enable the chiller to provide sufficient cooling capacity to meet the new cooling capacity.

[0186] Reference now Fig.17 , showing a schematic diagram of a method according to some embodiments Figure 3 Detailed block diagram of a chiller controller environment of FIG. 1700. The chiller controller environment 1700 may include a historical database 330, a machine learning model 314, a boundary generator 316, a variable regulator 318, and a user device 332. The boundary generator 316 is shown to include a data processor 1702, a probability identifier 1704, a PRV / VGD boundary selector 1706, a VSD boundary selector 1708, a cooling load identifier 1710, a performance simulator 1712, and a signal generator 1714. The data processor 1702, the probability identifier 1704, the cooling load identifier 1710, the performance simulator 1712, and the signal generator 1714 may be similar to those of reference 1, respectively. Fig.13 The data processor 1302, probability identifier 1304, cooling load identifier 1310, performance simulator 1312, and signal generator 1314 are shown and described. Chiller controller environment 1700 may include more or fewer components, depending on its configuration.

[0187] The machine learning model 314 may include a PRV / VGD boundary predictor 1716 and a VSD boundary predictor 1718. The PRV / VGD boundary predictor 1716 and the VSD boundary predictor 1718 may be or include machine learning models such as neural networks, support vector machines, random forests, etc. The PRV / VGD boundary predictor 1716 may be a machine learning model configured to predict the probability of multiple PRV / VGD boundary values. The probability may be a confidence score indicating the likelihood that the boundary value is the correct value generated, so the cooler 100 may operate at a surge probability threshold. Each node of the output layer of the PRV / VGD boundary predictor 1716 may be associated with a position value. The position value may range from 0% to 100%. The PRV / VGD boundary predictor 1716 may include any number of values. The PRV / VGD boundary predictor 1716 may predict such probabilities based on an input feature vector including cooler operation data other than the PRV / VGD position value in some cases. The feature vector may also include a surge probability threshold. The PRV / VGD boundary predictor 1716 may receive a feature vector comprising such data and output probabilities for any number of PRV / VGD locations.

[0188] The PRV / VGD boundary predictor 1716 can be trained with training data generated from surge data (e.g., cooler operating data associated with the cooler when the cooler experiences surge). In some embodiments, the training data can be associated with the cooler operating data when the cooler has just crossed a surge threshold (e.g., has just entered a surge state). Once the PRV / VGD boundary predictor 1716 is trained, the PRV / VGD boundary predictor 1716 can predict the probability of a PRV / VGD position boundary based on a surge probability threshold. For example, the PRV / VGD boundary predictor 1716 can receive the cooler operating data and a surge probability of 75% as inputs in a feature vector. The PRV / VGD boundary predictor 1716 can predict on input that the probability of a 25% open PRV / VGD position is 60%, the probability of a 50% open PRV / VGD position is 26%, and the probability of a 24% open PRV / VGD position is 24. The PRV / VGD boundary predictor 1716 may predict any probability for any number of PRV / VGD positions. The PRV / VGD boundary predictor 1716 may pass the probabilities to the PRV / VGD boundary selector 1706 for further processing.

[0189] Additionally, in some embodiments, the PRV / VGD boundary predictor 1716 may include two machine learning models. One machine learning model may predict the probability of a value of an upper boundary of the PRV / VGD position, and the other machine learning model may predict the probability of a value of a lower boundary of the PRV / VGD position. Each machine learning model may be trained with similar data, except with markers specific to its respective boundary. The machine learning models may receive the same input (e.g., the same cooler operating data and surge probability) and output the probabilities of the upper and lower boundaries of the PRV / VGD position, respectively.

[0190] In some embodiments, the PRV / VGD boundary predictor 1716 may include different machine learning models that are trained to predict the probability of PRV boundaries and VGD boundaries, respectively. Such machine learning models may include machine learning models that predict the probability of an upper boundary and a lower boundary of each of the PRV or VGD or just one of the variables. Such machine learning models may be trained as described above for a machine learning model that predicts PRV / VGD location boundaries together.

[0191] The PRV / VGD boundary predictor 1716 may transmit the probability of the boundary to the PRV / VGD boundary selector 1706. The PRV / VGD boundary selector 1706 may select the boundary associated with the highest probability and generate the boundary based on the selected boundary. In some embodiments, if the boundary is associated with a probability exceeding a threshold, the PRV / VGD boundary selector 1706 only selects the boundary associated with the highest probability. The PRV / VGD boundary selector 1706 may transmit the generated boundary to the performance simulator 1712 to determine whether the generated boundary will enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load. If the performance simulator 1712 determines that the boundary will not enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load, the data processor 1702 may generate a feature vector including similar chiller operating data and the same surge probability threshold, and apply the feature vector to the VSD boundary predictor 1718 to generate one or more VSD boundaries for the variable regulator 318 to use to control the chiller 100, in addition to the generated PRV / VGD boundaries. However, if a corresponding VSD frequency boundary has been generated and determined to be such that chiller 100 cannot provide sufficient cooling capacity, a message may be transmitted to user device 332 indicating that chiller 100 cannot provide sufficient cooling capacity with a surge probability threshold.

[0192] The VSD boundary predictor 1718 may be a machine learning model that is trained to predict the probability of a value of a VSD frequency boundary based on a surge probability threshold and cooler operating data. In some cases, the VSD boundary predictor 1718 may be trained using a training method similar to the PRV / VGD boundary predictor 1716 with cooler operating data that occurs at or near a surge threshold. The VSD boundary predictor 1718 may be trained to predict a VSD boundary based on the same or similar cooler data as the PRV / VGD boundary predictor 1716, except that the VSD boundary predictor 1718 may use the PRV / VGD control signal data as input while excluding input related to the VSD frequency. Therefore, any predictions made by the VSD boundary predictor 1718 may not be affected by the VSD control signal (which may not have any correlation with the predicted VSD frequency boundary). For example, the VSD boundary predictor 1718 may receive inputs including cooler operating data excluding the VSD frequency and a surge probability threshold. The VSD boundary predictor 1718 may output a probability of 70% for a frequency of 48 Hz, a probability of 10% for a frequency of 52 Hz, and a probability of 20% for a frequency of 55 Hz. The VSD boundary predictor 1718 may output any probability for any number of VSD frequencies. The VSD boundary predictor 1718 may output the VSD frequencies to the VSD boundary selector 1708 for further processing. In some embodiments, the VSD boundary predictor 1718 includes a machine learning model for an upper VSD frequency boundary and a machine learning model for a lower VSD frequency boundary. Each machine learning model may similarly output the frequency of its corresponding VSD frequency boundary. Each of these machine learning models may be trained using similar training data, except for the labels associated with its corresponding boundary.

[0193] The VSD boundary predictor 1718 may transmit the predicted probability of the VSD frequency value to the VSD boundary selector 1708. The VSD boundary selector 1708 may select the frequency value associated with the highest probability and generate a boundary at the selected frequency. In some embodiments, the VSD boundary selector 1708 may generate a VSD frequency boundary only if the probability associated therewith exceeds a threshold. The VSD boundary selector may transmit the generated boundary to the performance simulator 1712 to determine whether the generated boundary will enable the chiller 100 to provide sufficient cooling capacity to handle the new cooling load. If the performance simulator 1712 determines that the boundary will not enable the chiller 100 to provide sufficient cooling capacity, the data processor 1702 may generate a feature vector including similar chiller operating data and the same surge probability threshold, and apply the feature vector to the PRV / VGD boundary predictor 1716 to generate a PRV / VGD boundary for the variable regulator 318 to use to control the chiller 100, in addition to the generated VSD boundary. However, if a corresponding PRV / VGD position boundary has been generated and is determined to prevent chiller 100 from providing sufficient cooling capacity, a message may be transmitted to user device 332 indicating that chiller 100 cannot provide sufficient cooling capacity with a surge probability threshold.

[0194] Advantageously, by using the chiller operating data and the surge probability prediction boundaries, controllable chiller operating boundaries of the chiller 100 can be accurately predicted. Instead of using an iterative process, a machine learning model can automatically predict such boundaries, which can speed up processing and enable the variable regulator 318 to more quickly adjust the control of the chiller 100 in real time.

[0195] Reference now Fig.18 , shows a diagram of a neural network 1800 for predicting the probability of a PRV / VGD location boundary according to some embodiments. The neural network 1800 may be used with reference Fig.15 The neural network 1800 may include input nodes for discharge pressure 1802, suction pressure 1804, VSD frequency 1806, probability that surge will occur 1808, stall voltage 1810, input current 1812, and bearing position 1814. The components 1802, 1804, 1806, 1808, 1810, 1812, and 1814 may be similar to those of reference 1800. Fig.151502, 1504, 1506, 1508, 1510, 1512, and 1514 are shown and described. Each of the components 1802 to 1814 can represent an input to the neural network 1800. The neural network 1800 is further shown to include a hidden layer 1816 and an output layer 1818. The inputs 1802 to 1814 can be converted into weighted signals or combinations of signals and passed to the hidden layer 1816. The nodes of the hidden layer 1816 can perform various operations on the weighted signals, similar to those of reference 1800. Fig.15 Nodes of hidden layer 1516 are shown and described. The nodes of hidden layer 1816 may transmit further weighted signals to output layer 1818. Output layer 1818 may include one or more nodes and / or layers that indicate the probabilities of potential values ​​of the frequency boundaries implemented based on inputs 1802-1814.

[0196] In some embodiments, the weights of the weighted signals may be generated or determined based on training data provided to the neural network 1800. The neural network 1800 may be generated or determined using a training data set similar to that described above with reference to FIG. Fig.15 The neural network 1800 may be trained using a supervised, semi-supervised, or unsupervised method for the training described. For example, the neural network 1800 may be trained using labeled training data including the values ​​of each of the inputs 1802 to 1814. For example, a training data set including the values ​​of each of the inputs 1802 to 1814 may be input to the neural network 1800. The training data set may be labeled with a correct value of 70% open. The neural network 1800 may output the probability of a position value frequency value of 50%, 60%, 70%, and 80% open. The neural network 1800 may adjust its internal weights and / or parameters proportional to the closeness of each probability to the correct value. Continuing with the above example, the neural network 1800 may adjust its internal weights so that if a similar data set is input into the neural network 1800, it will predict a higher probability of 70% and a lower probability of each of 50%, 60%, and 80% open. The accuracy percentage may be determined based on the closeness of the neural network 1800 in predicting the labeled values ​​on multiple training data sets. For example, in some embodiments, the closer the neural network 1800 predicts the correct frequency, the higher the accuracy percentage. The training data set can be input into the neural network 1800 for training until the neural network 1800 is accurate to a threshold. In some embodiments, the training data set can be input until the neural network 1800 is accurate to a threshold for a preset number of training data sets.

[0197] Reference now Fig.19 , shows a diagram of an exemplary neural network 1900 for predicting the probability of a VSD frequency boundary according to some embodiments. The neural network 1900 may be used with reference Fig.14The neural network 1900 may include input nodes for discharge pressure 1902, suction pressure 1904, pre-rotation vane position / variable geometry diffuser position 1906, probability that surge will occur 1908, stall voltage 1910, input current 1912, and bearing position 1914. The components 1902, 1904, 1906, 1908, 1910, 1912, and 1914 may be similar to those described with reference to FIG. Fig.14 1402, 1404, 1406, 1408, 1410, 1412, and 1414 are shown and described. Each of the components 1902 to 1914 may represent an input to the neural network 1900. The neural network 1900 is further shown to include a hidden layer 1916 and an output layer 1918. The inputs 1902 to 1914 may be converted into weighted signals or combinations of signals and passed to the hidden layer 1916. The nodes of the hidden layer 1916 may perform various operations on the weighted signals, similar to those described in reference to FIG. Fig.14 Nodes of hidden layer 1416 are shown and described. The nodes of hidden layer 1916 may transmit further weighted signals to output layer 1918. Output layer 1918 may include one or more nodes and / or layers that indicate the probabilities of potential values ​​of the frequency boundaries implemented based on inputs 1902-1914.

[0198] In some embodiments, the weights of the weighted signals may be generated or determined based on training data provided to the neural network 1900. The neural network 1900 may be generated or determined using a training data set similar to that described above with reference to FIG. Fig.14The neural network 1900 may be trained using supervised, semi-supervised, or unsupervised methods of the training described. For example, the neural network 1900 may be trained using labeled training data including values ​​for each of the inputs 1902 to 1914. For example, a training data set including values ​​for each of the inputs 1902 to 1914 may be input to the neural network 1900. The training data set may be labeled with the correct value of 60 Hz. The neural network 1900 may output probabilities of frequency values ​​of 55 Hz, 60 Hz, 65 Hz, and 70 Hz. The neural network 1900 may adjust its internal weights and / or parameters proportional to the closeness of each probability to the correct value. Continuing with the above example, the neural network 1900 may adjust its internal weights so that if a similar data set is input into the neural network 1900, it will predict a higher probability of 60 Hz and a lower probability of 55 Hz and 65 Hz. The accuracy percentage may be determined based on how close the neural network 1900 predicts the labeled values ​​on multiple training data sets. For example, in some embodiments, the closer the neural network 1900 predicts the correct frequency, the higher the accuracy percentage. The training data set may be input into the neural network 1900 for training until the neural network 1900 is accurate to a threshold. In some embodiments, the training data set may be input until the neural network 1900 is accurate to a threshold for a preset number of training data sets.

[0199] Reference now Fig. 20 , a flowchart of an exemplary process for generating and setting boundaries of controllable chiller variables according to some embodiments is shown. For example, process 2000 may be performed by one or more data processing systems (e.g., chiller controller 302). Process 2000 may be a flowchart detailing how the controllable chiller variables may be generated and set boundaries of controllable chiller variables according to some embodiments. Figure 6 , 7 An exemplary process for generating boundaries of controllable cooler variables in any of steps 608, 612, 616, 620, 708, 716, 1210, 1214, 1218, and / or 1222 of process 2000 or 12. Advantageously, process 2000 can be performed to automatically generate boundaries of controllable cooler variables based on surge probability thresholds. As a result, surge can be avoided in the cooler, and the cooler can be operated close to surge conditions, improving the efficiency of the cooler. Instead of using an iterative process that may require the use of multiple inputs, a single input to a machine learning model can be used to determine the boundaries. Process 2000 can include any number of steps, and the steps can be performed in any order.

[0200] Process 2000 is shown as including receiving input indicating a surge probability threshold and chiller operating data (step 2002). For example, step 2002 can be performed by boundary generator 316. Step 2002 can be performed similar to how step 1602 is performed. Process 2000 is also shown as including providing the chiller operating data including the surge probability to one or more machine learning models (step 2004). For example, step 2004 can be performed by boundary generator 316.

[0201] Process 2000 is shown as including obtaining an output indicating the probability of multiple values ​​of a controllable cooler variable from a machine learning model (step 2006). For example, step 2006 can be performed by the boundary generator 316. The PRV / VGD boundary prediction model and the VSD boundary prediction model can be or include machine learning models such as neural networks, support vector machines, random forests, etc. The PRV / VGD boundary prediction model can be a machine learning model configured to predict the probability of multiple PRV / VGD boundary values. The probability can be a confidence score indicating the possibility that the boundary value is the correct value generated, so the cooler can operate at a surge probability threshold. Each node of the output layer of the PRV / VGD boundary prediction model can be associated with a position value. The position value can range from 0% to 100%. The PRV / VGD boundary prediction model can include any number of values. The PRV / VGD boundary prediction model can predict such probabilities based on an input feature vector including cooler operation data other than the PRV / VGD position value in some cases. The feature vector can also include a surge probability threshold. The PRV / VGD boundary prediction model can receive a feature vector comprising such data and output the probabilities of any number of PRV / VGD locations.

[0202] The PRV / VGD boundary prediction model can be trained with training data generated from surge data (e.g., cooler operating data associated with the cooler when the cooler experiences surge). In some embodiments, the training data can be associated with the cooler operating data when the cooler has just crossed a surge threshold (e.g., has just entered a surge state). Once the PRV / VGD boundary prediction model is trained, the PRV / VGD boundary prediction model can predict the probability of a PRV / VGD position boundary based on a surge probability threshold. For example, the PRV / VGD boundary prediction model can receive cooler operating data and a surge probability of 75% as input in a feature vector. The PRV / VGD boundary prediction model can predict on input a probability of 60% for a 25% open PRV / VGD position, 26% for a 50% open PRV / VGD position, and 24% for a 24% open PRV / VGD position. The PRV / VGD boundary prediction model can predict any probability of any number of PRV / VGD positions.

[0203] In addition, in some embodiments, the PRV / VGD boundary prediction model may include two machine learning models. One machine learning model may predict the probability of a value of an upper boundary of the PRV / VGD position, and another machine learning model may predict the probability of a value of a lower boundary of the PRV / VGD position. Each machine learning model may be trained with similar data, except with markers specific to its respective boundary. The machine learning models may receive the same input (e.g., the same cooler operating data and surge probability) and output the probabilities of the upper and lower boundaries of the PRV / VGD position, respectively.

[0204] In some embodiments, the PRV / VGD boundary prediction model may include different machine learning models that are trained to predict the probability of PRV boundaries and VGD boundaries, respectively. Such machine learning models may include machine learning models that predict the probability of an upper boundary and a lower boundary of each of PRV or VGD or just one of the variables. Such machine learning models may be trained as described above for a machine learning model that predicts PRV / VGD position boundaries together.

[0205] The boundary associated with the highest probability can be selected and generated. In some embodiments, the boundary associated with the highest probability can be selected only when the boundary is associated with a probability of exceeding a threshold. It can be determined whether the generated boundary will enable the chiller to provide sufficient cooling capacity to handle the new cooling load. If it is determined that the boundary will not enable the chiller to provide sufficient cooling capacity to handle the new cooling load, a new feature vector can be generated that includes similar chiller operation and the same surge probability threshold. The new feature vector can be an input to a VSD boundary prediction model to generate a VSD boundary to control the chiller, in addition to the generated PRV / VGD boundary. However, if a corresponding VSD frequency boundary has been generated and it is determined that it does not enable the chiller to provide sufficient cooling capacity, a message can be transmitted to a user device indicating that the chiller cannot provide sufficient cooling capacity with the input surge probability threshold.

[0206] The VSD boundary prediction model can be a machine learning model that is trained to predict the probability of the value of the VSD frequency boundary based on the surge probability threshold and the cooler operation data. In some cases, the VSD boundary prediction model can be trained with cooler operation data that occurs at or near the surge threshold using a training method similar to the PRV / VGD boundary prediction model. The VSD boundary prediction model can be trained to predict the VSD boundary based on the same or similar cooler data as the PRV / VGD boundary prediction model, except that the VSD boundary prediction model can use the PRV / VGD control signal data as input while excluding input related to the VSD frequency. Therefore, any prediction made by the VSD boundary prediction model may not be affected by the VSD control signal (which may not have any correlation with the predicted VSD frequency boundary). For example, the VSD boundary prediction model may receive inputs including cooler operation data excluding the VSD frequency and the surge probability threshold. The VSD boundary prediction model can output a probability of 70% for a frequency of 48 Hz, a probability of 10% for a frequency of 52 Hz, and a probability of 20% for a frequency of 55 Hz. The VSD boundary prediction model can output any probability of any number of VSD frequencies. In some embodiments, the VSD boundary prediction model includes a machine learning model for the VSD upper frequency boundary and a machine learning model for the VSD lower frequency boundary. Each machine learning model can similarly output the probability of the value of its corresponding VSD frequency boundary. Each of these machine learning models can be trained using similar training data, except for the labels associated with its corresponding boundary.

[0207] Process 2000 is shown as including identifying boundaries based on probabilities of multiple values ​​of a controllable cooler variable (step 2008). For example, step 2008 can be performed by boundary generator 316. Probabilities of values ​​of controllable cooler variables generated by the machine learning model can be obtained. The probabilities can be compared to each other. The value of the controllable cooler variable associated with the highest probability can be selected. For example, a probability of a VSD frequency value can be generated. The probability of 47 Hz can be 30%, 50 Hz can be 40%, and 52 Hz can be 30%. The 50 Hz value can be selected because it is associated with the highest probability. In some embodiments, the highest probability can be compared to a threshold. If the probability does not exceed the threshold, the value can not be selected and an error can be generated. Continuing with the above example, if the threshold probability is 50%, an error may be generated because the probability associated with 50 Hz does not exceed the threshold probability of 50%.

[0208] Process 2000 is shown as including setting boundaries of controllable chiller variables (step 2010). For example, step 2010 can be performed by boundary generator 316. The output can include boundaries of controllable chiller variables to avoid crossing during operation. The boundaries can be upper boundaries and / or lower boundaries of any controllable chiller variable. For example, the boundaries can be upper boundaries and / or lower boundaries of VSD frequency or PRV / VGD position control signals. The boundaries can be used to operate the chiller. As described above, the boundaries can only be used if it is determined that the boundaries enable the chiller to provide sufficient cooling capacity to meet the new cooling capacity.

[0209] Reference now Fig.21 , a flowchart of an exemplary process 2100 for generating controllable chiller variable boundaries based on the output of a machine learning model according to some embodiments is shown. Process 2100 can be performed by any controller or processor of a building management system. For example, process 2100 can be performed by one or more data processing systems (e.g., chiller controller 302). Process 2100 can enable a chiller (e.g., chiller 100) including a compressor (e.g., compressor 102) to operate under different conditions over time within boundaries set based on surge probability thresholds. The machine learning model can be trained with training data generated from chiller operating data in a cloud database, so that the machine learning model can provide accurate predicted surge probabilities when connected to the chiller. The machine learning model can be trained over time when surge occurs, so that the machine learning model can dynamically adapt and adjust to changing conditions of the chiller. In some embodiments, the machine learning model is trained to generate VSD frequency and / or VGD and PRV position boundaries. Process 2100 can include any number of steps, and the steps can be performed in any order.

[0210] Process 2100 is shown as including collecting chiller operation data associated with one or more chillers (step 2102). For example, step 2102 can be performed by chiller controller 302. Chiller operation data can be collected from chillers associated with a first building management system or chillers associated with other building management systems. For example, a cloud database (e.g., cloud database 328) including chiller operation data generated by chillers of other building management systems and / or chillers of the first building management system can be accessed via a network. Chiller operation data can be collected from a cloud database. In some embodiments, the chiller operation data in the cloud database is generated by a sensor and sent to the chiller from which the chiller operation data is collected via a local gateway. Chiller operation data may include operating parameters of the chiller, which indicate various characteristics of each chiller when the chiller is operating correctly and / or characteristics of each chiller when the chiller experiences surge. The value of the parameter can be marked with a label indicating whether the chiller is experiencing surge or otherwise operating at a surge threshold (e.g., just beginning or about to begin experiencing surge). For example, if the cooler experiences surge within 10 milliseconds, 20 milliseconds, 50 milliseconds, 100 milliseconds, one second, etc. when the cooler operation data is generated or collected, the cooler operation data can be marked as being associated with surge. Any time period can be used to associate the cooler operation data with surge or surge threshold. The cooler operation data can also be marked as being associated with surge or associated with the normal operation of the cooler. The characteristics of the cooler can include the type of refrigerant, discharge pressure (Pd), suction pressure (Ps), stall voltage, stall current, bearing position, input current, etc. of the cooler at various time points. These features can also include variable speed drive frequency, pre-rotation vane (PRV) position and / or variable geometry diffuser (VGD) position control signal, which makes the compressor of the cooler run and corresponds to the time point of other characteristics. Therefore, the cooler operation data can indicate how the cooler operates under various control signals. The cooler operation data can include a timestamp indicating when the cooler operation data is generated or collected, and in some cases, an identifier indicating which cooler, sensor or time series the cooler operation data is associated with.

[0211] In some embodiments, the chiller operating data may be tagged with a VSD frequency, PRV position, and / or VGD position control signal that is implemented and causes the chiller to operate in a normal state, a surge state, or near a surge threshold. In addition, in some embodiments, the chiller operating data may be tagged or otherwise associated with the type of refrigerant used, and associated with various data points. Thus, the chiller operating data may indicate how the chiller operates under various control signals and / or when using different types of refrigerants. Such data may be used to determine minimum and / or maximum values ​​for one or more of the VSD frequency and / or PRV / VGD position control signals under various operating conditions, and in some cases, determine the best refrigerant to use (e.g., finding the refrigerant associated with the least amount of surge).

[0212] Advantageously, by collecting cooler operating data from multiple coolers from a cloud database, more information can be used for training to identify the probability of surge occurrence and / or the boundaries of various control signals to avoid crossing. Although coolers can operate differently from each other, they can have some similarities in how they operate. When cooler operating data is collected to avoid surge in a cooler that the data processing system is newly connected to or configured to operate, the cooler operating data from other coolers can be used to train a machine learning model to predict the probability of surge in the cooler and / or control signal value (e.g., VSD frequency, PRV position, and / or VGD position) thresholds to avoid crossing or causing the risk of surge.

[0213] In addition, chiller operating data can be collected and added to a database (e.g., a historical database 330) dedicated to storing chiller operating data for a particular chiller and / or a first building management system to which the chiller is connected. When the chiller is operating within the first building management system, chiller operating data associated with the operation of the chiller can be automatically added to the dedicated database. The chiller operating data can be added to the database periodically at various time intervals, or in some cases, when surge occurs. For example, an indication of surge occurring in the chiller can be obtained (e.g., received as a signal or determined based on the chiller operating data). Various sensors of the chiller can be polled in response to obtaining an identification to obtain characteristic values ​​of the chiller before, during, and / or after the surge occurs. The chiller operating data can be stored in a database. These values ​​can be marked to indicate whether the chiller operating data is associated with surge.

[0214] Cooler operation data can be collected from the cooler in real time. For example, for cooler operation data associated with various characteristics, the cooler and / or sensor that generates the cooler operation data can be continuously polled. In another example, the sensor can generate cooler operation data based on the operation of the cooler, and automatically send the cooler operation data via the gateway at various intervals (e.g., every 5 milliseconds, every 10 milliseconds, every 20 milliseconds, etc.). Therefore, rapid pressure changes can be monitored. In another example, a technician can manually upload the cooler operation data. By collecting such cooler operation data, the cooler operation data can be used to generate boundaries of controllable cooler variables based on surge probability thresholds.

[0215] Process 2100 is shown as including training one or more machine learning models with collected chiller operating data (step 2104). For example, step 2104 can be performed by the model generator 312 or the machine learning model 314 as described above. One or more machine learning models can be machine learning models such as, but not limited to, neural networks, random forests, or support vector machines. Each machine learning model can be associated with the same or different inputs and outputs. One or more machine learning models can be trained to associate outputs with given inputs (e.g., one machine learning model can determine the probability that surge will occur based on current chiller characteristics, and / or one machine learning model can determine the VSD frequency and / or VGD or PRV position control signal boundaries to avoid crossing based on the input surge probability). For example, a first machine learning model can receive inputs associated with the characteristics of a chiller during operation and output the probability of whether the chiller will experience surge, and a second machine learning model can receive inputs similar to the first machine learning model, in addition to controllable chiller variables, and additionally include a surge probability threshold, and predict the boundaries of lost controllable chiller variables based on similar inputs and surge probability thresholds.

[0216] Collected chiller operating data (e.g., historical data collected from a cloud database or a local database, live chiller operating data collected directly from the chiller's sensors, etc.) can be used to train one or more machine learning models. For example, chiller operating data associated with multiple chillers can be retrieved from a cloud database. A training data set can be generated using the retrieved chiller operating data, which includes characteristic values ​​of the chiller during operation and a timestamp associated with the value. The training data set can be associated with values ​​from the chiller and can include values ​​at one or more time points within a time frame determined by an administrator. Each value can be marked with an indication of whether the value is associated with surge. For example, a value generated or collected within a time frame selected by an administrator or during surge can be marked as associated with surge. Otherwise, the value can be marked as not associated with surge. The training data set can be fed into one or more machine learning models, and an output can be obtained including a probability that a surge is about to occur in the chiller based on the training data set. The output can be fed back into one or more machine learning models for training as described herein.

[0217] One or more machine learning models may be trained using supervised, semi-supervised, and / or unsupervised training methods. One or more machine learning models may be trained in a supervised system that implements feedback (e.g., back propagation in a neural network). During supervised training, one or more machine learning models may be trained with training data having input and output pairs. The output may be a labeled correct output that one or more machine learning models should predict based on the input of the pair. In this case, the input may be a characteristic value of the cooler at different time points. For example, the values ​​of multiple characteristics of the cooler may be associated with a timestamp as an input. Each value and timestamp may be associated with an output tag indicating whether they are associated with surge. In some cases, the input may be a time series of values ​​associated with multiple timestamps, each timestamp being labeled with an associated output indicating whether they are associated with surge. One or more machine learning models may internally adjust their weights or parameters based on the input and output pairs. In some embodiments, after the training data is used as input and the one or more machine learning models identify the probability of the output, the one or more machine learning models identify the expected output from the training data, and identify the difference between the actual output and the expected output, and update the internal process of the one or more machine learning models based on the difference. For example, if the one or more machine learning models include a neural network, the neural network can identify the difference as a delta and modify its internal process to more accurately predict the probability of surge and / or control signals for future training data sets or live data. As more training data is fed into the one or more machine learning models, the one or more machine learning models can become more accurate.

[0218] In a semi-supervised system, one or more machine learning models may receive a mixture of labeled and unlabeled cooler operating data. Labeled and unlabeled data may be implemented into one or more machine learning models for training. For example, one or more machine learning models may be trained with cooler operating data detected and collected when the cooler is in a surge state. A portion of the cooler operating data may be labeled with a label indicating that the cooler operating data was collected when the cooler was in a surge state, another portion of the cooler operating data may be labeled with a label indicating that the cooler operating data was collected when the cooler was not in a surge state, and the remainder of the cooler operating data may not be labeled. All cooler operating data may be fed into one or more machine learning models for semi-supervised training.

[0219] In an unsupervised system, unlabeled chiller operation data can be applied to one or more machine learning models for training. In such a system, one or more machine learning models can group the training data together based on data with similar values ​​or other characteristics, and provide outputs based on these groups. For example, chiller operation data associated with surge can be grouped into one cluster, while chiller operation data associated with normally operating chillers can be grouped into another cluster. Using unsupervised learning techniques, one or more machine learning models can group the chiller operation data into any group based on the chiller operation data that has common characteristics with these groups.

[0220] In some embodiments, one or more machine learning models are trained with cooler operation data collected from the cooler when the cooler experiences surge. For example, cooler operation data indicating that the cooler is experiencing surge can be obtained from a sensor of the cooler. The cooler operation data may include cooler operation data for the time before, during, and after the surge is detected. The time before (reaching a threshold) and during the surge can be marked as associated with surge, while the time after the surge ends can be associated with normal operation. In some embodiments, only the time before and after the surge is detected (until the time threshold) is marked as associated with approaching the surge threshold. It can be determined based on the value of the cooler operation data (e.g., a high pressure value of the compressor of the cooler, an indication of coolant flowing backward through the compressor, etc.) that the cooler is experiencing surge. The cooler operation data can be marked with a mark (e.g., an expected output) indicating that the cooler operation data associated with surge (e.g., cooler operation data associated with the time before and during the surge) or approaching the surge threshold is surge data, while the data not associated with surge is not surge data. The labeling may be binary, such as "1" for chiller operation data associated with surge and "0" for chiller operation data not associated with surge. The chiller operation data may be formatted into a format recognizable by one or more machine learning models (e.g., a feature vector as described below) and provided to one or more machine learning models as supervised training. A similar process may be used to label and train one or more machine learning models using chiller operation data not associated with surge.

[0221] In some embodiments, one or more machine learning models can predict VSD frequency and / or VGD / PRV position boundaries based on surge probability thresholds to avoid crossing. Labeled training data with correct boundaries can be generated based on surge probability thresholds and current cooler operating parameters. The labeled training data can be input into one or more machine learning models that predict VSD frequency and / or VGD / PRV position boundaries. Such one or more machine learning models can output associated predicted boundaries. The expected output can be compared with the actual output. One or more machine learning models can adjust their internal weights and / or parameters in proportion to the difference between the actual output and the expected output.

[0222] Advantageously, by continuously providing surge data and non-surge data to one or more machine learning models for training, the chiller operation data specific to the chiller can be used to continuously train the one or more machine learning models to make them more accurate. The one or more machine learning models can automatically adapt to changes in chiller operation. Thus, the one or more machine learning models may be able to accurately predict the probability of surge for the boundaries of the chiller controllable chiller variables that vary over time, so that the boundaries of operation can be identified, and control signals can be generated to operate the chiller within such boundaries and avoid any undesirable surge probabilities.

[0223] Process 2100 is shown as including applying the live cooler operation data associated with the cooler as input to one or more machine learning models (step 2106). For example, step 2106 can be performed by the boundary generator 316 as described above. For cooler operation data, the sensor of the cooler can be polled at various intervals, or the cooler operation data can be obtained from the sensor described above in other ways. For example, the sensor can be polled once every 5 seconds. The cooler operation data can be received and processed into a form that can be used as an input for one or more machine learning models. For example, a feature vector can be generated from a cooler operation data set. The feature vector can include a digital identifier of the data. The digital identifier can be the value of the cooler operation data set. The live cooler operation data can include but is not limited to the refrigerant type, discharge pressure value, suction pressure value, stall voltage, bearing position, input current, VSD frequency, PRV position, VGD position, etc. of the cooler. The live cooler operation data can include current values ​​(e.g., current values ​​of cooler characteristics). The formatted cooler operation data can be applied to one or more machine learning models for processing.

[0224] Process 2100 is shown as including determining boundaries of controllable chiller variables for a surge probability threshold based on outputs of one or more machine learning models (step 2108). For example, step 2108 may be performed by boundary generator 316. One or more machine learning models may generate an output indicating a probability of whether a surge is likely to occur soon (e.g., within 10 milliseconds, 50 milliseconds, 1 second, 1 minute, 5 minutes, 10 minutes, 30 minutes, 1 hour, etc. as determined by an administrator) or otherwise indicating a probability of whether a surge will occur. A surge probability threshold may be obtained. Probabilities from one or more machine learning models may be obtained that indicate a probability that a surge will occur based on set operating conditions for the chiller. The probabilities may be compared. If the probabilities are within a threshold of each other, the controllable chiller variable values ​​used as inputs to the machine learning models may be used to generate operating boundaries for the chiller. If the probabilities are not within a threshold of each other, the controllable chiller variable values ​​input to the one or more machine learning models may be adjusted until the one or more machine learning models generate probabilities within a threshold of the surge probability threshold.

[0225] In some embodiments, if the probabilities are within thresholds of each other, it is determined based on the boundaries whether the cooler is able to provide sufficient cooling capacity to meet the desired cooling load. The simulator can simulate cooler performance based on the current operating conditions of the cooler, the characteristics of the cooler itself, and the generated boundaries of the controllable cooler variables. If it is determined based on such inputs that the cooler is able to provide the desired cooling load, boundaries can be generated and used to control the cooler. Otherwise, new boundaries can be generated based on other controllable cooler variables (e.g., another one of the PRV / VGD position or VSD frequency) until boundaries are generated that enable the cooler to provide the desired cooling load. If boundaries are determined for each controllable variable for a given surge probability threshold, and it is determined based on the surge probability threshold that the cooler cannot provide the desired cooling load, an alarm can be transmitted to a user device indicating that the desired cooling load cannot be met based on the surge probability threshold.

[0226] In some embodiments, one or more machine learning models may generate outputs indicating boundaries of the VSD frequency and / or VGD / PRV position control signals based on the surge probability threshold to avoid crossing. One or more machine learning models may generate boundaries. In some embodiments, examples of VSD frequency control signal thresholds include 40Hz, 45Hz, 50Hz, 55Hz, 60Hz, etc. Depending on how one or more machine learning models are configured, they may predict boundaries to avoid exceeding or falling below the boundaries. For example, in one configuration, one or more machine learning models may predict that the VSD frequency control signal should not be lower than 50Hz. In another configuration, one or more learning models may predict that the VSD frequency control signal should not exceed 60Hz. One or more machine learning models may predict any boundaries of any configuration. In addition, in some embodiments, examples of VGD / PRV position control signal boundaries include fully open 20%, fully open 40%, fully open 60%, fully open 70%, fully open 90%, etc. These positions may be percentages of any valve position at the inlet and / or outlet of the compressor. Similar to VSD frequency, depending on the configuration of the one or more machine learning models, the one or more machine learning models can predict boundaries for PRV and / or VGD positions to avoid exceeding or falling below the boundaries. The one or more machine learning models can predict any boundaries for the valve position control signal and / or VSD frequency.

[0227] In some embodiments, one or more machine learning models generate a probability distribution including probabilities of multiple boundaries based on inputs including actual cooler characteristics. For example, one or more machine learning models can generate a probability distribution including probabilities of VGD position boundaries to avoid the risk of exceeding or causing surge in the cooler. The probability distribution can include probabilities such as 10% for fully open 60%, 20% for fully open 70%, and 70% for fully open 80%. One or more machine learning models can predict any probability of any control signal boundary. The predicted probabilities can be compared to each other and / or to a threshold. If the probability is the highest and / or exceeds the threshold, the control signal boundary associated with the probability can be identified and used as a boundary for controlling the cooler. Therefore, the output of the machine learning model can be used to create a boundary of values ​​within which the cooler should operate to avoid surge and, in some cases, operate efficiently.

[0228] Process 2100 is shown as including affecting the operation of the cooler based on the output of the determined boundary (step 2110). For example, step 2110 can be performed by the variable regulator 318 as described above. For example, the boundary can be determined based on the surge probability and surge probability threshold generated by the machine learning model. Using the boundary, a new control signal can be generated to apply to the compressor to operate at the surge probability threshold. The control signal can be generated based on the boundary and transmitted to the cooler. For example, because operation close to the surge state may make the cooler operate more efficiently, a control signal at the controllable cooler variable boundary can be generated. For example, it can be determined that the valve position of the inlet and outlet of the compressor can be increased according to the newly generated PRV and VGD boundaries. Therefore, a control signal for controlling the actuator can be generated to open each of the valves, and the actuator repositions the valve at the inlet and outlet. A control signal can be generated so that the probability of surge occurring is still within the surge probability threshold. In another example, it can be determined that the VSD frequency can be increased based on the new cooling load according to the newly generated VSD frequency boundary. Therefore, a control signal to a VSD controlling a motor controlling the compressor may be generated at a higher frequency to the determined boundary so that the VSD will increase the voltage to control the motor. A control signal may be generated to generate a probability of a surge probability threshold.

[0229] Advantageously, by implementing one or more machine learning models to accurately identify the probability of whether surge will occur or the boundary of the surge probability threshold, the systems and methods described herein can enable the server and / or controller to operate the cooler at high efficiency without causing the cooler to enter a surge state. When the cooler experiences surge, the machine learning model can be continuously trained, so the machine learning model can continue to generate the probability of indicating whether surge will occur or the boundary of the control signal to avoid crossing. When training data is applied to the machine learning model, the probability can become more accurate. The server and / or controller can identify when the cooler may soon experience surge or the boundary of the control signal to avoid crossing, and therefore, the controller can adjust the control signal for driving the motor and the actuator running the cooler. By predicting an accurate surge probability, the use of a machine learning model can enable the cooler to operate near the surge boundary without entering a surge state. Because operating near the surge boundary enables the cooler to operate more efficiently. Therefore, using the method described herein can enable the cooler to operate more efficiently. In addition, because the machine learning model can be continuously trained, the machine learning model can automatically adapt to changes in the operation of the cooler to determine or generate the boundary of the control signal to avoid crossing.

[0230] Configuration of the exemplary embodiment

[0231] The construction and arrangement of the systems and methods shown in each exemplary embodiment are illustrative only. Although only a few embodiments are described in detail in the present disclosure, many modifications are possible (e.g., changes in the size, scale, structure, shape and ratio, parameter values, installation arrangements, use of materials, colors, orientations, etc. of various elements). For example, the position of the element can be reversed or otherwise changed, and the nature or quantity of the discrete element or position can be changed or changed. Therefore, all such modifications are intended to be included in the scope of the present disclosure. The order or sequence of any process or method step can be changed or reordered according to alternative embodiments. Without departing from the scope of the present disclosure, other replacements, modifications, changes and omissions can be made in the design, operating conditions and arrangement of the exemplary embodiments.

[0232] The present disclosure contemplates methods, systems, and program products on any machine-readable medium for completing various operations. The embodiments of the present disclosure may be implemented using an existing computer processor, or implemented by a dedicated computer processor for an appropriate system, merged for this or another purpose, or implemented by a hard-wired system. Embodiments within the scope of the present disclosure include program products, which include machine-readable media for carrying or having machine-executable instructions or data structures stored thereon. Such machine-readable media may be any available medium that can be accessed by a general or special-purpose computer or other machine with a processor. As an example, such machine-readable media may include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of machine-executable instructions or data structures and can be accessed by a general or special-purpose computer or other machine with a processor. When information is transmitted or provided to a machine via a network or another communication connection (hard-wired, wireless, or a combination of hard-wired or wireless), the machine appropriately regards the connection as a machine-readable medium. Therefore, any such connection is appropriately referred to as a machine-readable medium. The above combination is also included in the scope of machine-readable media. Machine-executable instructions comprise, for example, instructions and data which cause a general purpose computer, special purpose computer, or special purpose processing machines to perform a certain function or group of functions.

[0233] Although the accompanying drawings show a specific order of method steps, the order of steps may be different from that depicted. Two or more steps may also be performed simultaneously or partially simultaneously. Such variations will depend on the selected software and hardware systems and the designer's choice. All such variations are within the scope of the present disclosure. Similarly, software implementations may be completed with standard programming techniques having rule-based logic and other logic to complete various connection steps, processing steps, comparison steps, and decision steps.

Claims

1. A system, comprising: processing circuitry; and a memory having stored thereon instructions which, when executed by the processing circuit, cause the processing circuit to: generating a bound on the value of the controllable cooler variable to prevent the future cooler surge event from occurring based on a probability that a future cooler surge event will occur if the value of the controllable cooler variable is used to operate the cooler, wherein the instructions, when executed by the processing circuit, cause the processing circuit to: comparing the probability that the future cooler surge event will occur to a probability threshold; as well as responsive to determining that the probability that the future cooler surge event will occur does not violate the probability threshold, setting the value of the controllable cooler variable to the boundary of the controllable cooler variable; as well as Operation of the chiller is affected based on the boundary to prevent the future chiller surge event from occurring. 2 . The system of claim 1 , wherein the instructions, when executed by the processing circuit, cause the processing circuit to receive the probability threshold via a user device.

3. The system of claim 1 , wherein the instructions, when executed by the processing circuit, cause the processing circuit to: Detecting past surge events in historical chiller operating data; marking a portion of the historical cooler operation data as surge cooler operation data; and The surge cooler operating data is applied to one or more machine learning models to generate the boundaries. 4 . The system of claim 3 , wherein the surge cooler operating data comprises a time series including values ​​of one or more variables of the historical cooler operating data associated with times before and during the past surge event.

5. The system of claim 1 , wherein the instructions, when executed by the processing circuit, cause the processing circuit to affect operation of the chiller by determining values ​​of one or more controllable chiller variables predicted to prevent the future surge event, the one or more controllable chiller variables comprising at least one of: variable speed drive frequency; Pre-rotation blade position; or Variable geometry diffuser position.

6. The system of claim 1 , wherein the controllable cooler variable is a first controllable cooler variable, and wherein the instructions, when executed by the processing circuit, cause the processing circuit to: The first controllable chiller variable is adjusted to affect the operating capacity of the chiller until: The operating capacity of the cooler reaches the desired operating capacity; or the first controllable cooler variable reaches the boundary; as well as In response to the first controllable chiller variable reaching the boundary before the operating capacity of the chiller reaches the desired operating capacity, a second controllable chiller variable is adjusted until the operating capacity of the chiller reaches the desired operating capacity.

7. The system of claim 1, wherein the instructions, when executed by the processing circuit, cause the processing circuit to refrain from applying a control signal having an additional value of the controllable cooler variable that exceeds the boundary.

8. The system of claim 1 , wherein the instructions, when executed by the processing circuit, cause the processing circuit to: The chiller operating data is applied as an input to one or more machine learning models, wherein the input includes the value of the controllable chiller variable, and wherein the output of the one or more machine learning models corresponds to the probability that the future chiller surge event will occur if the value of the controllable chiller variable is used to operate the chiller.

9. The system of claim 8, wherein the instructions, when executed by the processing circuit, cause the processing circuit to: The one or more machine learning models are generated by the following steps: Create a training dataset by following these steps: detecting past surge events in historical chiller operating data; and marking a portion of the historical cooler operating data as surge cooler operating data associated with the past surge event; and The one or more machine learning models are trained using the training dataset. 10 . The system of claim 9 , wherein the surge cooler operating data comprises a time series including values ​​of one or more variables of the historical cooler operating data associated with times before and during the past surge event.

11. The system of claim 8, wherein the chiller operating data comprises at least one of: refrigerant type, discharge pressure, suction pressure, stall voltage, input current, bearing position, current variable speed drive frequency, current pre-rotation vane position, current variable geometry diffuser position, or any combination thereof.

12. A controller configured to control operation of a chiller, wherein the controller comprises: a memory including instructions stored thereon, wherein the instructions, when executed by one or more processors, cause the controller to: determining a probability that a future chiller surge event will occur if values ​​of controllable chiller variables are used to operate the chiller; comparing the probability that the future cooler surge event will occur to a probability threshold; In response to determining that the probability that the future cooler surge event will occur does not violate the probability threshold, setting the value of the controllable cooler variable to a boundary of the controllable cooler variable; and Operation of the chiller is affected based on the boundary to prevent the future chiller surge event from occurring.

13. The controller of claim 12, wherein the controllable cooler variable is a first controllable cooler variable, and wherein the instructions, when executed by the one or more processors, cause the controller to: The first controllable chiller variable is adjusted to affect the operating capacity of the chiller until: The operating capacity of the cooler reaches a desired operating capacity; or the first controllable chiller variable reaches the boundary; and In response to the first controllable chiller variable reaching the boundary before the operating capacity of the chiller reaches the desired operating capacity, a second controllable chiller variable is adjusted until the operating capacity of the chiller reaches the desired operating capacity.

14. The controller of claim 13, wherein the first controllable cooler variable is pre-rotation vane position or variable geometry diffuser position, and wherein the second controllable cooler variable is variable speed drive frequency.

15. The controller of claim 13, wherein the first controllable cooler variable is a variable speed drive frequency, and wherein the second controllable cooler variable is a pre-rotation vane position or a variable geometry diffuser position.

16. The controller of claim 13, wherein the instructions, when executed by the one or more processors, cause the controller to: determining a second boundary for the second controllable chiller variable; adjusting the second controllable chiller variable to affect the operating capacity of the chiller until: The operating capacity of the cooler reaches the desired operating capacity; or the second controllable cooler variable reaches the second boundary; as well as In response to the second controllable chiller variable reaching the second boundary before the operating capacity of the chiller reaches the desired operating capacity, sending a notification to a user device indicating that the desired operating capacity cannot be reached.

17. A system, comprising: processing circuitry; and a memory having stored thereon instructions which, when executed by the processing circuit, cause the processing circuit to: determining an upper bound, a lower bound, or both, on the value of the controllable cooler variable based on a probability that a future cooler surge event will occur if the value of the controllable cooler variable is used to operate the cooler to prevent the future cooler surge event from occurring, wherein the instructions, when executed by the processing circuit, cause the processing circuit to: comparing the probability that the future cooler surge event will occur to a probability threshold; and In response to determining that the probability that the future cooler surge event will occur does not violate the probability threshold, setting the value of the controllable cooler variable to the upper boundary or the lower boundary of the controllable cooler variable; and Operation of the cooler is affected based on the upper boundary or the lower boundary. The system of claim 17 , wherein the probability threshold is received via a user device.

19. The system of claim 17, wherein the instructions, when executed by the processing circuit, cause the processing circuit to refrain from outputting a control signal having an additional value of the controllable cooler variable that exceeds the upper boundary or the lower boundary.

20. The system of claim 19, wherein the controllable cooler variable is a first controllable cooler variable, and wherein the instructions, when executed by the processing circuit, cause the processing circuit to: The first controllable chiller variable is adjusted to affect the operating capacity of the chiller until: The operating capacity of the cooler reaches a desired operating capacity; or The first controllable chiller variable reaches the upper boundary or the lower boundary; and In response to the first controllable chiller variable reaching the upper boundary or the lower boundary before the operating capacity of the chiller reaches the desired operating capacity, a second controllable chiller variable is adjusted until the operating capacity of the chiller reaches the desired operating capacity.

Images