JP5511698B2 - Air conditioner linkage control system, air conditioner linkage control method, and air conditioner linkage control program - Google Patents

Description

  The present invention relates to an air conditioner linkage control system, an air conditioner linkage control method, and an air conditioner linkage control program. The present invention relates to an air conditioner installed in a data center and an information processing device such as a server. An air conditioner linkage control system, an air conditioner linkage control method for controlling the power consumption of the information processing device and the air conditioner to an optimum state within a predetermined temperature and humidity range using humidity and load information of the information processing device, and The present invention relates to an air conditioner linkage control program.

  In a data center where information processing apparatuses such as a large number of servers are installed, temperature / humidity sensors are used so that the information processing apparatus operates normally in a temperature / humidity environment within a predetermined range. It is necessary to adjust the air conditioner setting conditions based on the temperature and humidity measurement results. Here, as a method for estimating whether or not the temperature / humidity environment is within a specified range even after a predetermined time has elapsed from a certain time (for example, the current time), heat generally used in air conditioning performance evaluation or the like is used. There is a method of estimation by using fluid analysis software, but since it takes a long time to calculate until the evaluation result is obtained, it cannot be used as a mechanism for controlling the setting conditions of the air conditioner online.

  On the other hand, as one of the load distribution methods for reducing the power consumption of information processing devices such as a large number of servers in a data center, A virtualization technology that consolidates processing on a server is being studied. However, in such a virtualization technology, heat generation is also concentrated locally, making it difficult to effectively cool the data center room without waste.

  Under these circumstances, the cooling temperature in the data center is excessive in order to avoid the risk of high-temperature failures in information processing equipment such as servers due to air conditioner failure or a rise in the data center room temperature during a power failure. It has become common to set it to a low value.

  In order to improve the energy saving performance while avoiding the risk of high temperature failure of the information processing apparatus at the time of failure of the air conditioner or power failure, Japanese Patent Application Laid-Open No. 2010-133626 “Air conditioning equipment and ICT equipment” In the “linkage control method”, the set temperature of the air conditioner installed in the data center and the processing load (power consumption) of the information processing device such as a server are linked, and the server or the like is linked every predetermined time span. Even if the air conditioner stops operating for the processing load allocated to the information processing device, the cooling target zone of the air conditioner should be set below the temperature that does not exceed the allowable limit temperature within the standard recovery time. A technique to be controlled is disclosed.

JP 2010-133626 A

  However, in the conventional technique, even if the technique as in Patent Document 1 is used, the temperature and humidity environment of the information processing apparatus such as the server is changed by changing the setting condition of the air conditioner installed in the data center. It is difficult to estimate in detail in real time whether or not it will change. Therefore, as a result, in the data center, even if the air conditioner is shut down, within the standard time required for recovery, the temperature and humidity environment of the information processing device such as a server is set as a predetermined allowable limit value. However, there remains a problem that efficient air-conditioning operation is difficult. For example, it is frequently operated at a temperature lower than necessary in order to ensure that the temperature does not exceed.

  In addition, with regard to virtualization technology that attempts to reduce power consumption by consolidating processing to some servers, it only takes into account the general method of reducing power consumption of information processing devices such as servers. As described above, not only is it difficult to effectively cool the room of the data center without waste, but also the setting conditions of the air conditioner of the data center where the information processing device is installed and the information It is difficult to control the power consumption required for both the air conditioner and the information processing device to an optimum state because it does not consider efficient cooling by linking with the load of the processing device. .

  The present invention has been made in view of the problems as described above, and the temperature and humidity at one or more locations at the time when a predetermined time has elapsed from an arbitrarily designated time (for example, the current time), that is, In an environment in which the temperature and humidity of one or more predetermined locations in the data center where the information processing device is installed are estimated at high speed, and each temperature and humidity is within a predetermined range. While satisfying the constraint of operating the information processing device, the optimum combination of the load distribution of the information processing device and the setting conditions of the air conditioner to minimize the total power consumption of the information processing device and the air conditioner To provide an air conditioner linkage control system, an air conditioner linkage control method, and an air conditioner linkage control program that can be extracted and controlled in real time. That.

  The present invention comprises the following technical means in order to solve the above-mentioned problems.

The first technical means includes a power consumption measuring unit that measures power consumption of one or more air conditioners and one or more information processing apparatuses installed in the data center, and one or more predetermined locations. An air conditioner linkage control system that includes at least a temperature and humidity measurement unit that measures temperature and humidity, and controls both the setting conditions of the air conditioner and load distribution of a CPU (Central Processing Unit) of the information processing apparatus in a linked manner A power consumption estimation model for estimating power consumption between the air conditioner and the information processing device within a period from a time point arbitrarily designated until a predetermined time elapses, and outputting the estimated power consumption, The temperature / humidity is estimated by estimating the temperature / humidity of one or more of the locations after the predetermined time has elapsed since the arbitrarily designated time, and outputting the estimated temperature / humidity of each location. A machine model for the power consumption estimation model and the temperature / humidity estimation model, the constant model, the power consumption measurement information measured by the power consumption measurement unit, and the temperature / humidity measurement information measured by the temperature / humidity measurement unit. A learning database unit that stores the training data, and the measurement information of the power consumption and the measurement information of the temperature and humidity stored in the learning database unit as the training data for machine learning. A machine learning unit for adjusting parameters and parameters of the temperature / humidity estimation model in advance, and the machine learning unit for processing a load in the information processing apparatus in a temperature / humidity environment at an arbitrarily designated time. The estimated temperature / humidity output from the temperature / humidity estimation model with adjusted parameters is within a predetermined range. The condition of setting the air conditioner and the CPU of the information processing device that minimize the estimated power consumption output by the power consumption estimation model whose parameters have been adjusted in the machine learning unit Optimal search unit for searching for a combination with the load distribution of the air conditioner as an optimal combination state and controlling the setting condition of the air conditioner and the load distribution of the CPU of the information processing apparatus so as to be the optimal combination state searched In the air conditioner linkage control system according to the first technical means, the power consumption estimation model includes temperature / humidity measurement information measured by the temperature / humidity measurement unit and setting of the air conditioner. The information and the CPU load information in each information processing device are used as inputs, and within a period from the time point arbitrarily specified until the predetermined time elapses. The value of power consumption between the air conditioner and the information processing device or the probability distribution of the power consumption is estimated and output as the estimated power consumption, and the temperature and humidity estimation model is measured by the temperature and humidity measurement unit One or more after a predetermined time elapses from an arbitrarily designated time point, using as input the measured temperature and humidity measurement information, the setting information of the air conditioner, and the CPU load information in each information processing device A temperature / humidity value or a probability distribution of the temperature / humidity at each location is estimated and output as the estimated temperature / humidity .

According to a second technical means, in the air conditioner linkage control system according to the first technical means, in addition to the temperature / humidity measurement information measured by the temperature / humidity measurement unit, heat using thermal fluid analysis software is used. The conditions and simulation results of fluid simulation are stored in the learning database unit, and the stored conditions and simulation results of the thermal fluid simulation are used as the training data for machine learning of the temperature and humidity estimation model in the machine learning unit. It is characterized by.

According to a third technical means, in the air conditioner linkage control system according to the first or second technical means, the estimated power consumption and the estimated temperature are integrated by integrating the estimated power consumption model and the temperature / humidity estimation model. It is configured as a multi-output neural network that outputs humidity.

The fourth technical means includes a power consumption measuring unit that measures power consumption of one or more air conditioners and one or more information processing apparatuses installed in the data center, and one or more predetermined locations. An air conditioner linkage control method comprising at least a temperature / humidity measuring unit for measuring temperature and humidity, and controlling both the setting condition of the air conditioner and load distribution of a CPU (Central Processing Unit) of the information processing apparatus in a linked manner A power consumption estimation model for estimating power consumption between the air conditioner and the information processing device within a period from a time point arbitrarily designated until a predetermined time elapses, and outputting the estimated power consumption, Estimating the temperature and humidity of one or more of the locations after the predetermined time has elapsed from the arbitrarily specified time, and outputting the estimated temperature and humidity as the estimated temperature and humidity of the locations The power consumption measurement information measured by the power consumption measurement unit and the temperature / humidity measurement information measured by the temperature / humidity measurement unit, the power consumption estimation model and the temperature / humidity estimation model Using the training data for machine learning of the power consumption estimation model and the temperature / humidity estimation model stored as training data for machine learning, the parameters of the power consumption estimation model and the parameters of the temperature / humidity estimation model are stored. The estimated temperature / humidity output by the temperature / humidity estimation model whose parameters have been adjusted when the load is to be processed by the information processing apparatus in a temperature / humidity environment at an arbitrarily designated time is determined in advance. The power consumption estimation model with adjusted parameters is output under the condition that it is within the specified range. The combination of the setting condition of the air conditioner and the load distribution of the CPU of the information processing device that minimizes the estimated power consumption is searched as an optimal combination state, and the searched optimal combination state is obtained. The setting condition of the air conditioner and the load distribution of the CPU of the information processing device are controlled, and the power consumption estimation model includes temperature / humidity measurement information measured by the temperature / humidity measurement unit and setting information of the air conditioner. And CPU load information in each of the information processing devices as an input, or a value of power consumption between the air conditioner and the information processing device in a period until a predetermined time elapses from an arbitrarily designated time point or The probability distribution of the power consumption is estimated and output as the estimated power consumption, and the temperature / humidity estimation model is the temperature / humidity measurement information measured by the temperature / humidity measurement unit. And the setting information of the air conditioner and the CPU load information in each information processing apparatus as inputs, and the temperature and humidity of each of the one or more locations after the predetermined time elapses from an arbitrarily designated time point. A value or a probability distribution of the temperature and humidity is estimated and output as the estimated temperature and humidity .

According to a fifth technical means, in the air conditioner linkage control method according to the fourth technical means, in addition to the temperature / humidity measurement information measured by the temperature / humidity measurement unit, heat using thermal fluid analysis software is used. The fluid simulation conditions and simulation results are stored, and the stored thermal fluid simulation conditions and simulation results are used as the training data for machine learning of the temperature and humidity estimation model.

A sixth technical means is for calculating power consumption in the air conditioner linkage control method according to the fourth or fifth technical means, in addition to the power consumption measurement information measured by the power consumption measuring unit. A calculation result derived using an approximate expression or a table is stored, and the stored calculation result is used as the training data for machine learning of the power consumption estimation model.

A seventh technical means integrates the power consumption estimation model and the temperature / humidity estimation model in the air conditioner linkage control method according to the fourth, fifth, or sixth technical means, so that the estimated power consumption and It is configured as a multi-output neural network that outputs the estimated temperature and humidity.

An eighth technical means is an air conditioner linkage control program that implements the air conditioner linkage control method according to any of the fourth to seventh technical means as a program executable by a computer. And

  According to the air conditioner linkage control system, the air conditioner linkage control method, and the air conditioner linkage control program of the present invention, the following effects can be obtained.

  One or more air conditioners and one when controlling setting conditions of one or more air conditioners installed in a data center and load distribution of a central processing unit (CPU) of one or more information processing apparatuses Minimize the total power consumption of the air conditioner and the information processing device under the constraint that the temperature and humidity environment in the data center where the information processing device is installed is within a predetermined range. By treating this as a combination optimization problem for the purpose, it is possible to optimize the operating state of the CPU load distribution of the information processing apparatus and the setting conditions of the air conditioner.

  In addition to the calculation results of power consumption derived using approximate equations and tables and thermal fluid simulation data using thermal fluid analysis software, actually measured data is used as training data for machine learning, By performing machine learning in advance in the power consumption estimation model and temperature / humidity estimation model, it is possible to calculate the CPU load distribution of the information processing device and the setting conditions of the air conditioner relatively easily at high speed. The information processing apparatus and the air conditioner can be controlled in real time.

It is a system configuration diagram for explaining an example of a system configuration of an air conditioner linkage control system according to the present invention. It is explanatory drawing for demonstrating an example of the power consumption estimation model of the air-conditioner linkage control system shown in FIG. 1, and a temperature / humidity estimation model. It is explanatory drawing for demonstrating an example of the production method of the training data used for the machine learning of the power consumption estimation model and temperature / humidity estimation model shown in FIG. It is a flowchart for demonstrating an example of the optimal search method of the optimal search part of the air conditioning machine linkage control system shown in FIG.

  Hereinafter, an example of a preferred embodiment of an air conditioner linkage control system, an air conditioner linkage control method, and an air conditioner linkage control program according to the present invention will be described in detail with reference to the drawings. In the following description, an air conditioner linkage control system and an air conditioner linkage control method according to the present invention will be described. However, such an air conditioner linkage control method is implemented as an air conditioner linkage control program executable by a computer. Of course, the air-conditioner linkage control program may be recorded on a computer-readable recording medium.

(Features of the present invention)
Prior to the description of the embodiments of the present invention, an outline of the features of the present invention will be described first. The present invention relates to a CPU (Central Processing Unit) load control of an information processing device and air conditioning control in the data center in a data center in which one or more air conditioners and one or more information processing devices such as servers are set. Are integrated and integrated, the temperature and humidity environment in the data center can be changed to a specified temperature (for example, when the load on the CPU of the information processing device fluctuates at this time, power failure, etc. From the time when the fluctuation occurs, every predetermined period, etc.) until the predetermined time elapses, the temperature and humidity conditions within which the information processing apparatus can operate normally fall within a predetermined range. It is also related to a mechanism that can obtain the optimum conditions for minimizing the total power consumption of the information processing device and the air conditioner at a high speed under the constraint condition By controlling the air conditioner setting conditions and the CPU load distribution of the information processing device in association with each other, the total power consumption is minimized while maintaining the temperature and humidity within the specified range. It is characterized in that the air conditioner and the information processing apparatus can be operated under optimum operating conditions.

  In the conventional technology, in order to search for the optimum condition that minimizes power consumption and to perform future prediction of required temperature and humidity changes, complex phenomena including thermal fluids in the data center are performed at high speed. Even if an approximate model capable of computation is to be created, as described above, there is a problem that it is extremely difficult to realize.

  In contrast, the present invention solves the problems in the prior art by creating each estimation model related to power consumption, temperature and humidity in advance using machine learning based on training data such as actual measurement data and simulation data. Is the main feature.

  More specifically, in the present invention, the following mechanism is adopted. That is, power consumption, temperature / humidity measurement data and simulation data in a data center where one or more information processing devices and one or more air conditioners are installed are stored in advance as training data for machine learning in the learning database unit. Store it. Then, by inputting the training data for machine learning stored in the learning database unit to the power consumption estimation model and the temperature / humidity estimation model created in advance, the actual measurement data stored in the learning database unit and Machine learning is performed in the machine learning unit using simulation data or the like, and parameters of the power consumption estimation model and the temperature / humidity estimation model are adjusted in advance.

  After that, after entering into the state of full-scale operation, when trying to process the load applied to the information processing device in a certain temperature and humidity environment, the power consumption estimation model with adjusted parameters and Using the temperature / humidity estimation model, under the constraint that the temperature / humidity environment is within the specified range until a predetermined time elapses from an arbitrarily specified time (for example, the current time) , Quickly search for the optimum condition that can minimize the total power consumption of the information processing device and the air conditioner, and control the information processing device and the air conditioner to match the searched optimum condition .

(Embodiment of the present invention)
Next, an example of the system configuration of the air conditioner linkage control system according to the present invention will be described with reference to FIG. FIG. 1 is a system configuration diagram for explaining an example of a system configuration of an air conditioner linkage control system according to the present invention. A data center in which a plurality of information processing apparatuses and air conditioners to be controlled are installed. This case is illustrated. However, the air conditioner linkage control system according to the present invention is not limited to such a case. For example, the information processing apparatus may be one or the air conditioner may be one. .

  In the system configuration shown in FIG. 1, the control object controlled by the air conditioner linkage control system 100 is a plurality of information processing apparatuses 211, 212,..., A plurality of information processing stored in a plurality of racks 210, 220,. .. And each of the information processing apparatuses 211, 212,..., And a plurality of air conditioners 310, 320,... For maintaining the temperature and humidity environment required by each of the information processing apparatuses 221, 222,. It is. Here, electric power is supplied from the commercial power supply 400 to each of the information processing apparatuses 211, 212,..., Each of the information processing apparatuses 221, 222,.

  The air conditioner linkage control system 100 includes at least a power consumption measurement unit 110, a temperature / humidity measurement unit 120, a learning database unit 130, a machine learning unit 140, a power consumption estimation model 150, a temperature / humidity estimation model 160, and an optimum search unit 170. It is prepared for.

  In the air-conditioner linkage control system 100, meters and sensors are arranged at one or more predetermined locations arbitrarily determined as means for measuring power consumption and temperature / humidity, and at predetermined timings, respectively. Actual power consumption data and temperature / humidity measurement data are collected by the power consumption measurement unit 110 and the temperature / humidity measurement unit 120.

  Here, the predetermined locations where the meters and sensors for measuring temperature and humidity are arranged are the information processing apparatuses 211, 212,..., The racks 210, 220,. .., Each information processing device 221, 222,... Is an arbitrary location inside the information processing device 221, 222,..., And one or more arbitrary locations that are necessary for determining the optimum operating state in the data center. . For example, the cooling air inlets and air outlets of the racks 210, 220,... In which the information processing apparatuses 211, 212,. Meters and sensors are installed at various locations where heating elements such as CPUs (Central Processing Units) and motherboards of the processing devices 221, 222,.

  In addition, in the present embodiment, the predetermined place where the meter or sensor for measuring the power consumption is arranged is arranged at a power supply position to which a power supply line from the commercial power supply 400 is connected, and the information processing apparatuses 211, 212, ..., the total amount of power supplied to both the information processing apparatuses 221, 222, ... and the air conditioners 310, 320, ... is obtained by integrating the power consumption consumed by each information processing apparatus and each air conditioner. The case where it acquires as power consumption is shown. However, the present invention is not limited to such an arrangement form, and power supply lines that supply power to the information processing apparatuses 211, 212,..., The information processing apparatuses 221, 222,. It is also possible to disperse them appropriately.

  In the air conditioner linkage control system 100, actual data and simulation data relating to power consumption and temperature / humidity collected at predetermined timings by the power consumption measurement unit 110 and temperature / humidity measurement unit 120 are stored in the learning database unit 130, respectively. It is stored and used as training data for machine learning of the power consumption estimation model 150 and the temperature / humidity estimation model 160.

  The machine learning unit 140 performs machine learning using actual measurement data, simulation data, and the like regarding power consumption and temperature / humidity stored as training data in the learning database unit 130, so that a power consumption estimation model can be obtained in a state where high-speed computation is possible. 150 and parameters of the temperature / humidity estimation model 160 are adjusted in advance.

  In the power consumption estimation model 150, parameters are adjusted in advance by the machine learning unit 140, and each information processing device 211, 212,. .., Information processing devices 211, 212,..., Information within a period until a predetermined time elapses from an arbitrarily designated time point (for example, the current time) when trying to process a certain load at 221, 222,. The total power consumption of the processing devices 221, 222, ... and the air conditioners 310, 320, ... is estimated and output as estimated power consumption.

  The temperature / humidity estimation model 160 also has parameters adjusted in advance by the machine learning unit 140, and each information processing device 211, 212,..., Each information processing device in a temperature / humidity environment at an arbitrarily designated time (for example, the current time). When a certain load is to be processed at 221, 222,..., A meter or sensor for actually measuring the temperature and humidity is measured when the predetermined time elapses from an arbitrarily designated time (for example, current time). The temperature / humidity of each of one or more predetermined locations is estimated and output as the estimated temperature / humidity.

  In the optimum search unit 170, an optimum combination of the CPU load distribution of each information processing device 211, 212,..., Each information processing device 221, 222,. To do. That is, the predetermined specified range even when a predetermined time elapses from the time when the estimated temperature / humidity output by the temperature / humidity estimation model 160 whose parameters have been adjusted by the machine learning unit 140 is arbitrarily specified (for example, the current time). The predetermined time elapses in advance from the time point when the estimated power consumption output by the power consumption estimation model 150 whose parameters have been adjusted by the machine learning unit 140 is arbitrarily specified (for example, the current time) .., The CPU load distribution of each information processing device 221, 222,... And the setting conditions (settings) of each air conditioner 310, 320,. A combination of temperature, blowing temperature, air volume, wind direction, etc.) is searched for as an optimal combination. Then, the information processing devices 211, 212,..., The information processing devices 221, 222,..., The air conditioners 310, 320,.

  Next, machine learning models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 will be described with reference to FIG. FIG. 2 is an explanatory diagram for explaining an example of the power consumption estimation model 150 and the temperature / humidity estimation model 160 of the air conditioner linkage control system 100 shown in FIG. 1. Here, in the present embodiment shown in FIG. 2, both the power consumption estimation model 150 and the temperature / humidity estimation model 160 are integrated into one estimation model, and a predetermined time has elapsed from an arbitrarily designated time (for example, the current time). In the example shown, it is configured as a multi-output neural network that outputs both the total estimated power consumption during the period until the time and the estimated temperature and humidity at each location after the predetermined time has elapsed. Of course, it may be configured to be separated into different models as estimation models for estimating the estimated temperature and humidity.

  As shown in FIG. 2, the power consumption estimation model 150 and the temperature / humidity estimation model 160 both include temperature / humidity measurement information 181 that is actual temperature / humidity measurement data measured by the temperature / humidity measurement unit 120 and each air conditioner. The setting information 182 indicating the setting conditions (setting temperature, air volume, etc.) of 310, 320,..., And information on the load to be processed by the CPU in each information processing device 211, 212,. Is input information.

  The temperature / humidity measurement information 181 is information indicating temperature / humidity measurement data at one or more locations arbitrarily determined in advance as locations where sensors and meters are installed and should be adjusted to the optimum temperature / humidity. Are not only temperature information generated by the information processing apparatuses 221, 222,..., But also temperature and humidity measurement results measured at arbitrary locations in the data center, data not only in the data center, and data Also includes temperature and humidity measurement information of outside air outside the center.

  The setting information 182 of the setting information of each air conditioner 310, 320,... Is information indicating setting conditions such as a set temperature, a blowing temperature, an air volume, and a wind direction set for each air conditioner 310, 320,. Further, the CPU load information 183 includes information on the load that each information processing device 211, 212,..., And the information processing devices 221, 222,. For example, the information regarding the processing load of each information processing device 211, 212,..., And the information processing devices 221, 222,. .., 212 can be directly collected from the CPUs of the information processing apparatuses 221, 222,.

  In the power consumption estimation model 150, based on the input information as described above, the information processing apparatuses 211, 212,... Within a period until a predetermined time elapses from an arbitrarily designated time (for example, the current time). , And the total power consumption of the air conditioners 310, 320,... Are estimated and output as estimated power consumption 191. Further, in the temperature / humidity estimation model 160, sensors and meters are installed when the predetermined time elapses from an arbitrarily designated time (for example, current time) based on the input information as described above. The temperature and humidity of one or more of the locations are estimated and output as the estimated temperature and humidity 192 of each location.

  The estimated power consumption 191 output from the power consumption estimation model 150 integrates the power consumption of all the information processing apparatuses 211, 212, ..., the information processing apparatuses 221, 222, ... and all the air conditioners 310, 320, .... The estimated power value or its probability distribution is obtained by estimating the amount of power consumption within a predetermined period of time after a predetermined time (for example, the current time).

  Further, the estimated temperature / humidity 192 output from the temperature / humidity estimation model 160 is an estimated value of the temperature / humidity or a probability distribution in each of the predetermined locations where the sensors and meters are installed, and is arbitrarily designated. This is an estimate of the temperature and humidity at each predetermined location after a predetermined time has elapsed from the time (for example, the current time).

  Here, if each of the neural networks constituting the power consumption estimation model 150 and the temperature / humidity estimation model 160 is constituted by a Bayesian model, each of the power consumption estimation model 150 and the temperature / humidity estimation model 160 will be described. The estimated power consumption 191 and the estimated temperature / humidity 192 output from are output as probability distributions of the power consumption and temperature / humidity, respectively, until a predetermined time elapses from an arbitrarily designated time point (for example, the current time). It is possible to output a probability distribution of power consumption within a period of time, a probability distribution of temperature and humidity at each predetermined location after a predetermined time has elapsed from an arbitrarily designated time point (for example, the present time).

  As described above, the learning database unit 130 of the air conditioner linkage control system 100 shown in FIG. 1 has input values as training data for machine learning of the estimation models of the power consumption estimation model 150 and the temperature / humidity estimation model 160. And a set of output values. The machine learning unit 140 outputs the output values of the estimated models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 to the input value of the training data for machine learning stored in the learning database unit 130. The parameters in the estimation models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 are adjusted so as to approach the output value of the training data for machine learning stored in 130. Here, as in the example shown in FIG. 2, when the estimation models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 are configured by a three-layer perceptron having a hierarchical structure of an input layer, an intermediate layer, and an output layer. Can learn parameters in each estimation model by using Backward Propagation of Errors.

  Next, a method of creating training data for machine learning of the power consumption estimation model 150 and the temperature / humidity estimation model 160 will be described with reference to FIG. FIG. 3 is an explanatory diagram for explaining an example of a method for creating training data used for machine learning of the power consumption estimation model 150 and the temperature / humidity estimation model 160 shown in FIG.

  Training data of each estimation model of the power consumption estimation model 150 and the temperature / humidity estimation model 160 can be created at each stage before and during operation of the air conditioner linkage control system 100 as shown in FIG. .

  In the stage before operation, as shown in step ST11 of FIG. 3, the power consumption estimation model 150 is calculated based on the calculation result derived for the input information (input value) using an approximate expression or table for power consumption calculation. Training data is obtained, and, as shown in step ST12 of FIG. 3, temperature / humidity simulation values (simulation conditions, simulation results) with respect to input information (input values) are acquired using thermal fluid analysis software. Thus, training data of the temperature / humidity estimation model 160 is created. As described above, in the pre-operation stage, using the training data obtained by simulation or the like in advance, about the estimated models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 is roughly learned by the machine learning unit 140. By performing in advance, stable control can be performed in the initial operation stage.

  At the stage of operation, as shown in step ST21 of FIG. 3, the measured data measured by a sensor or meter installed in the target data center is used as the power consumption measuring unit 110, the temperature / humidity measuring unit. By collecting at 120, training data of each estimation model of the power consumption estimation model 150 and the temperature / humidity estimation model 160 is created. By using actual measurement data such as this, it is possible to realize the estimation models of the power consumption estimation model 150 and the temperature / humidity estimation model 160 peculiar to the structure and the environment as relatively simple models. Arithmetic can be performed.

  However, there are cases where it is effective to use simulation results using thermal fluid analysis software as training data even during operation. For example, temperature / humidity data outside the specified range required for an information processing apparatus installed in a data center is hardly obtained as measured data. Therefore, the air conditioner shown in FIG. 1 is supplemented by performing simulation using thermal fluid analysis software as shown in step ST22 of FIG. The determination as to whether or not the linkage control system 100 can perform control so as to maintain the temperature and humidity within a specified range can be estimated more accurately.

  Next, an optimal search method of the optimal search unit 170 of the air conditioner linkage control system 100 shown in FIG. 1 will be described using the flowchart of FIG. FIG. 4 is a flowchart for explaining an example of the optimum search method of the optimum search unit 170 of the air conditioner linkage control system 100 shown in FIG. 1, and the setting conditions of the air conditioners 310, 320,. An example of a control method of the information processing apparatuses 211, 212,... And the CPU load distribution of the information processing apparatuses 221, 222,.

  The control shown in the flowchart of FIG. 4 is periodically activated every time a predetermined predetermined control time is reached (step S1). When the predetermined control time is reached and the flowchart of FIG. 4 is started, the information processing apparatuses 211, 210 are considered in consideration of the current setting conditions of the air conditioners 310, 320,. In order to adjust the setting conditions of the air conditioners 310, 320,... For the CPU load distribution candidates of the information processing apparatuses 221, 222,. , 212,..., Information processing apparatuses 221, 222,..., 212,. When the temperature / humidity measured at each location and the setting conditions of each air conditioner are input, the output of the learned temperature / humidity estimation model 160, that is, the estimated temperature / humidity is arbitrarily designated ( Even after a predetermined time has passed a predetermined from example, if the present time) to check whether within the specified range determined in advance (step S2).

  When the output of the temperature / humidity estimation model 160, that is, the estimated temperature / humidity is not within the specified range (NG in step S2), the setting of the air conditioners 310, 320,. Is adjusted / changed (step S3), the process returns to step S2, and the determination process of step S2 is repeated.

  In step S3, when determining the next setting condition of each air conditioner 310, 320,..., That is, when adjusting / changing the setting of each air conditioner using the optimization method, tabu search Metaheuristic techniques such as (tabu search) and genetic algorithms are suitable because they can calculate practical approximate solutions relatively quickly even for large-scale optimization problems.

  On the other hand, if the output of the temperature / humidity estimation model 160, that is, the estimated temperature / humidity is within the specified range (OK in step S2), the output of the learned power consumption estimation model 150, that is, the estimated power consumption, is determined in advance. It is then evaluated whether or not it is minimum within the predetermined time period (step S4). If the selected candidate does not satisfy the selection candidate end condition (no in step S5), after adjusting / changing the settings of the air conditioners 310, 320,. In step S3), the process returns to step S2, and the determination process in step S2 and the evaluation in step S3 are repeated. In addition, what is necessary is just to make it prescribe | regulate as the completion | finish conditions of the selection candidate in step S5 using the repetition frequency of the selected candidate, calculation time, etc., for example.

  If the selected candidate satisfies the selection candidate end condition in step S5 (yes in step S5), the information processing apparatuses 211, 212,..., The information processing apparatuses 221, 222,. It is checked whether or not the end conditions for all the CPU load distribution candidates are satisfied (step S6). In addition, what is necessary is just to make it prescribe | regulate as the completion conditions of all the candidates in step S6, for example using the repetition frequency of the selected candidate, calculation time, etc. When the completion conditions for all candidates are not yet satisfied (no in step S6), the information processing apparatuses 211, 212,..., The information processing apparatuses 221, 222,. After adjusting / changing the CPU load distribution (step S7), the process returns to step S2, and the determination process in step S2, the evaluation in step S3, and the determination process in steps S5 and S6 are repeated.

  In step S7, when determining CPU load distribution candidates for the information processing apparatuses 211, 212,..., The information processing apparatuses 221, 222,. In the case of adjusting / changing the CPU load distribution of the device, the metaheuristic method such as tabu search (taboo search) or genetic algorithm is very similar to the case of determining the next setting condition of each air conditioner described above. Even in a large-scale optimization problem, a practical approximate solution can be calculated at a relatively high speed.

  When the completion conditions for all candidates are satisfied (Yes in step S6), the restriction condition that the temperature / humidity is within a predetermined range is satisfied even after a predetermined time has elapsed. , The information processing apparatuses 211, 212,..., And the information processing apparatuses 221, 222,... That the power consumption in the period from the arbitrarily specified time (for example, the current time) until the predetermined time elapses is minimized. A combination of the CPU load distribution of each information processing apparatus and the setting conditions of each air conditioner 310, 320,... Is extracted as an optimal solution (step S8). Thereafter, a control signal based on the extracted optimum solution is created and transmitted to the air conditioners 310, 320,..., The information processing apparatuses 211, 212,. And the CPU load distribution of the information processing apparatuses 221, 222,... Are controlled (step S9).

  The power consumption estimation model 150 and the temperature / humidity estimation model 160 are configured using the Bayesian learning model, and the estimated power consumption 191 and the estimated temperature / humidity 192 from the power consumption estimation model 150 and the temperature / humidity estimation model 160 are used as the power consumption. When outputting as a probability distribution or a probability distribution of temperature / humidity, the output of the temperature / humidity estimation model 160 falls within the specified range when extracting the setting condition regarding the minimum power consumption in step S8 or in step S2. When checking whether or not, the probability that the probability of deviating from the minimum power consumption and the probability that the temperature and humidity deviate from the specified range are smaller than a predetermined threshold may be used as a constraint. If the threshold is determined according to the user's policy or the like, the error of the power consumption estimation model 150 and the temperature / humidity estimation model 160 can be adjusted flexibly.

DESCRIPTION OF SYMBOLS 100 ... Air-conditioner linkage control system, 110 ... Power consumption measuring part, 120 ... Temperature / humidity measuring part, 130 ... Learning database part, 140 ... Machine learning part, 150 ... Power consumption estimation model, 160 ... Temperature / humidity estimation model, 170 ... Optimal search unit, 181 ... temperature / humidity measurement information, 182 ... setting information of each air conditioner, 183 ... CPU load information in each information processing apparatus, 191 ... estimated power consumption, 192 ... estimated temperature / humidity, 210, 220 ... rack , 211, 212, 221, 222 ... information processing device, 310, 320 ... air conditioner, 400 ... commercial power supply.

Claims (8)

A power consumption measuring unit that measures the power consumption of one or more air conditioners and one or more information processing devices installed in the data center, and a temperature and humidity that measures the temperature and humidity of one or more predetermined locations An air conditioner linkage control system that includes at least a measurement unit and controls both the setting condition of the air conditioner and the load distribution of the CPU (Central Processing Unit) of the information processing apparatus,
A power consumption estimation model that estimates power consumption of the air conditioner and the information processing device within a period from a time point arbitrarily designated until a predetermined time elapses, and outputs the estimated power consumption;
A temperature / humidity estimation model that estimates the temperature / humidity of one or more of the locations after the predetermined time has elapsed from an arbitrarily designated time point and outputs the estimated temperature / humidity of each location;
The power consumption measurement information measured by the power consumption measurement unit and the temperature / humidity measurement information measured by the temperature / humidity measurement unit are used as training data for machine learning of the power consumption estimation model and the temperature / humidity estimation model. A learning database section to store;
Using the power consumption measurement information and the temperature / humidity measurement information stored in the learning database unit as the training data for machine learning, the power consumption estimation model parameters and the temperature / humidity estimation model parameters are set in advance. A machine learning unit to adjust;
The estimated temperature / humidity output by the temperature / humidity estimation model whose parameters have been adjusted by the machine learning unit when the load is to be processed by the information processing apparatus in a temperature / humidity environment at an arbitrarily designated time point. A condition for setting the air conditioner that minimizes the estimated power consumption output from the power consumption estimation model whose parameters have been adjusted by the machine learning unit under the condition that the power consumption is within a predetermined range. The combination with the load distribution of the CPU of the information processing device is searched as an optimal combination state, and the setting condition of the air conditioner and the load distribution of the CPU of the information processing device are set so as to obtain the searched optimal combination state. An optimal search unit for controlling
With at least,
The power consumption estimation model is arbitrarily specified by using as input the measurement information of the temperature and humidity measured by the temperature and humidity measurement unit, the setting information of the air conditioner, and the CPU load information in each of the information processing devices. Estimating a power consumption value or a probability distribution of the power consumption between the air conditioner and the information processing device in a period until the predetermined time elapses, and outputting the estimated power consumption, The temperature / humidity estimation model is a time point arbitrarily designated as input of measurement information of temperature / humidity measured by the temperature / humidity measurement unit, setting information of the air conditioner, and CPU load information in each information processing apparatus. Estimating the temperature / humidity value or probability distribution of the temperature / humidity at each of the one or more locations after the predetermined time has passed in advance as the estimated temperature / humidity Air conditioner coordinated control system characterized by force. 2. The air conditioner linkage control system according to claim 1 , wherein, in addition to the temperature and humidity measurement information measured by the temperature and humidity measurement unit, conditions and simulation results of thermal fluid simulation using thermal fluid analysis software are learned. An air conditioner linkage control system which is stored in a database unit and uses the stored thermal fluid simulation conditions and simulation results as the training data for machine learning of the temperature and humidity estimation model in the machine learning unit. The air conditioner linkage control system according to claim 1 or 2 , wherein the power consumption estimation model and the temperature / humidity estimation model are integrated to form a multi-output neural network that outputs the estimated power consumption and the estimated temperature / humidity. An air conditioner linkage control system. A power consumption measuring unit that measures the power consumption of one or more air conditioners and one or more information processing devices installed in the data center, and a temperature and humidity that measures the temperature and humidity of one or more predetermined locations An air-conditioner linkage control method that includes at least a measurement unit and controls both the setting condition of the air-conditioner and the load distribution of the CPU (Central Processing Unit) of the information processing apparatus,
A power consumption estimation model that estimates power consumption of the air conditioner and the information processing device within a period from a time point arbitrarily designated until a predetermined time elapses, and outputs the estimated power consumption;
A temperature / humidity estimation model that estimates the temperature / humidity of one or more of the locations after the predetermined time has elapsed from an arbitrarily designated time point and outputs the estimated temperature / humidity of each location;
Have
The power consumption measurement information measured by the power consumption measurement unit and the temperature / humidity measurement information measured by the temperature / humidity measurement unit are used as training data for machine learning of the power consumption estimation model and the temperature / humidity estimation model. Using the training data for machine learning of the stored power consumption estimation model and the temperature / humidity estimation model stored, the parameters of the power consumption estimation model and the parameters of the temperature / humidity estimation model are adjusted in advance,
The estimated temperature / humidity output by the temperature / humidity estimation model whose parameters are adjusted when the load is to be processed by the information processing device in a temperature / humidity environment at an arbitrarily specified time is within a predetermined specified range. A combination of the setting condition of the air conditioner and the load distribution of the CPU of the information processing apparatus that minimizes the estimated power consumption that is output by the power consumption estimation model whose parameters have been adjusted Is determined as an optimal combination state, and the setting condition of the air conditioner and the load distribution of the CPU of the information processing device are controlled so as to be the optimal combination state searched for,
The power consumption estimation model is arbitrarily specified by using as input the measurement information of the temperature and humidity measured by the temperature and humidity measurement unit, the setting information of the air conditioner, and the CPU load information in each of the information processing devices. Estimating a power consumption value or a probability distribution of the power consumption between the air conditioner and the information processing device in a period until the predetermined time elapses, and outputting the estimated power consumption, The temperature / humidity estimation model is a time point arbitrarily designated as input of measurement information of temperature / humidity measured by the temperature / humidity measurement unit, setting information of the air conditioner, and CPU load information in each information processing apparatus. Estimating the temperature / humidity value or probability distribution of the temperature / humidity at each of the one or more locations after the predetermined time has passed in advance as the estimated temperature / humidity Air conditioner linkage control method characterized by force. 5. The air conditioner linkage control method according to claim 4 , wherein, in addition to the temperature / humidity measurement information measured by the temperature / humidity measurement unit, conditions and simulation results of thermal fluid simulation using thermal fluid analysis software are stored. And storing the thermal fluid simulation conditions and simulation results as the training data for machine learning of the temperature and humidity estimation model. 6. The calculation result derived by using an approximate expression for power consumption calculation or a table in addition to the power consumption measurement information measured by the power consumption measurement unit in the air conditioner linkage control method according to claim 4 or 5. And the stored calculation result is used as the training data for machine learning of the power consumption estimation model. 7. The multi-output neural network according to claim 4, 5, or 6 , wherein the power consumption estimation model and the temperature and humidity estimation model are integrated to output the estimated power consumption and the estimated temperature and humidity. An air conditioner linkage control method characterized by comprising: 8. An air conditioner linkage control program, wherein the air conditioner linkage control method according to claim 4 is implemented as a program executable by a computer.

Images