Air conditioning control system

Abstract

To solve the following problem: an environment of a target space cannot be adjusted sufficiently close to a desired environmental condition since operation condition of an air conditioner is decided based on an existing database.SOLUTION: An air-conditioning control system 1 performs air conditioning in a target space using an air conditioner 10. The air-conditioning control system 1 includes a generation unit 34, a prediction unit 33, an evaluation unit 35, an extraction unit 36, and an air-conditioning controller 19. The generation unit 34 generates multiple candidates of an air-conditioning setting value of the air conditioner 10. The prediction unit 33 predicts an environmental condition using an NN model 41. The NN model 41 predicts an environmental condition in the target space based on the air-conditioning setting value. The evaluation unit 35 evaluates an environmental condition corresponding to the air-conditioning setting value. The extraction unit 36 extracts an air-conditioning setting value evaluated as the highest by the evaluation unit 35 from among the candidates of the air-conditioning setting value. The air-conditioning controller 19 controls the air conditioner 10 with the air-conditioning setting value extracted by the extraction unit 36.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] Regarding air conditioning control systems. [Background technology]

[0002] As disclosed in Patent Document 1 (Japanese Patent Laid-Open No. 2010-15192), there is a technique for determining the optimal operating conditions of an air conditioner so that a target space is in a desired environmental state. Summary of the Invention [Problem to be solved by the invention]

[0003] In Patent Document 1, the operating conditions of the air conditioner are determined based on an existing database, which poses a problem in that the target space cannot be brought sufficiently close to the desired environmental state. [Means for solving the problem]

[0004] An air conditioning control system according to a first aspect conditions a target space with an air conditioning device. The air conditioning control system includes a generation unit, a prediction unit, an evaluation unit, an extraction unit, and a control unit. The generation unit generates a plurality of candidate operating conditions for the air conditioning device. The prediction unit predicts an environmental state using a learning model. The learning model predicts the environmental state of the target space based on the operating conditions. The evaluation unit evaluates the environmental state corresponding to the operating conditions. The extraction unit extracts, from the plurality of candidate operating conditions, operating conditions whose evaluation by the evaluation unit satisfies predetermined conditions. The control unit controls the air conditioning device according to the operating conditions extracted by the extraction unit.

[0005] The air conditioning control system of the first aspect extracts operating conditions for the air conditioner based on the environmental state of the target space. As a result, the air conditioning control system can determine operating conditions for the air conditioner that bring the target space closer to a desired environmental state.

[0006] An air conditioning control system according to a second aspect is the air conditioning control system according to the first aspect, further comprising a learning unit, which generates a learning model.

[0007] An air conditioning control system according to a third aspect is the air conditioning control system according to either the first or second aspect, further comprising a calculation unit. The calculation unit performs a CFD (Computational Fluid Dynamics) simulation of the environmental state using the operating conditions as input. The learning model predicts the environmental state, which is the output of the CFD simulation, using the operating conditions as input.

[0008] With this configuration, the air conditioning control system of the third aspect can determine the operating conditions of the air conditioner for each operating condition, so that the target space approaches the desired environmental state, without actually controlling the air conditioner.

[0009] The air conditioning control system of the fourth aspect is the air conditioning control system of the third aspect, and the learning model is a model that has been trained using the operating conditions as explanatory variables and the environmental state, which is the output of the CFD simulation, as the objective variable.

[0010] An air conditioning control system according to a fifth aspect is the air conditioning control system according to either the first or second aspect, further comprising a calculation unit. The calculation unit performs a CFD simulation of the environmental state using operating conditions as input. The calculation unit calculates a first environmental state and a second environmental state that has lower simulation accuracy than the first environmental state. The learning model predicts the first environmental state using the second environmental state as input.

[0011] An air conditioning control system according to a fifth aspect performs a CFD simulation of the environmental state using operating conditions as input. The air conditioning control system calculates a second environmental state, which has low accuracy in the CFD simulation, from the operating conditions. The air conditioning control system predicts a first environmental state from the second environmental state using a learning model. Therefore, the air conditioning control system can reduce the time required for the CFD simulation compared to when the first environmental state, which has high accuracy in the CFD simulation, is calculated from the operating conditions. As a result, the air conditioning control system can determine, in a realistic amount of time, the operating conditions of the air conditioning device that will bring the target space closer to the desired environmental state.

[0012] An air conditioning control system according to a sixth aspect is the air conditioning control system according to the fifth aspect, in which the learning model is a model that has been learned using the second environmental state as an explanatory variable and the first environmental state as a target variable.

[0013] An air conditioning control system according to a seventh aspect is the air conditioning control system according to either the first or second aspect, wherein the environmental state is the value of an environmental parameter at a plurality of locations including a first location and a second location within the target space. The air conditioning control system further includes an acquisition unit. The acquisition unit acquires the value of the environmental parameter at the first location by actual measurement. The learning model predicts the value of the environmental parameter at the second location based on the value of the environmental parameter at the first location acquired by the acquisition unit when the air conditioning device is controlled according to the operating conditions.

[0014] In the air conditioning control system of the seventh aspect, the learning model predicts the value of an environmental parameter at a second location based on the actual measured value of the environmental parameter at a first location. As a result, the air conditioning control system can predict the value of the environmental parameter at all locations from the actual measured values ​​of the environmental parameter at some locations.

[0015] An air conditioning control system of an eighth aspect is an air conditioning control system of the seventh aspect, in which the learning model is a model that has been trained using the values ​​of the environmental parameters at a first location acquired by an acquisition unit when the air conditioning device is controlled according to the operating conditions as explanatory variables and the values ​​of the environmental parameters at a second location as objective variables.

[0016] An air conditioning control system according to a ninth aspect is the air conditioning control system according to the seventh aspect, further comprising a calculation unit. The calculation unit performs a CFD simulation of the values ​​of the environmental parameters at the first location and the second location using the operating conditions as input. The learning model predicts the value of the environmental parameter at the second location based on the value of the environmental parameter at the first location acquired by the acquisition unit and the values ​​of the environmental parameters at the first location and the second location calculated by the calculation unit.

[0017] An air conditioning control system of a tenth aspect is the air conditioning control system of the ninth aspect, wherein the learning model is a model that has been trained using the values ​​of the environmental parameters at the first location acquired by the acquisition unit as explanatory variables and the values ​​of the environmental parameters at the second location as objective variables. The values ​​of the environmental parameters at the second location are predicted by inputting the values ​​of the environmental parameters at the second location calculated by the calculation unit into a first learning model that has been trained using the values ​​of the environmental parameters at the first location calculated by the calculation unit as explanatory variables and the values ​​of the environmental parameters at the first location acquired by the acquisition unit as objective variables.

[0018] An air conditioning control system according to an eleventh aspect is the air conditioning control system according to any one of the first to tenth aspects, wherein the generation unit generates candidates for operating conditions by reinforcement learning that uses the evaluation of the evaluation unit as a value.

[0019] With this configuration, the air conditioning control system of the eleventh aspect can determine the operating conditions of the air conditioning unit so that the target space approaches the desired environmental state, without relying on the operator's experience and without generating all candidate operating conditions.

[0020] An air conditioning control system of a twelfth aspect is the air conditioning control system of the eleventh aspect, in which the generation unit generates candidate operating conditions using a learning model that determines candidates for the next operating conditions based on the current operating conditions.

[0021] The air conditioning control system of the twelfth aspect can easily generate candidates for operating conditions by reinforcement learning using a learning model, even if there are a large number of elements of the operating conditions.

[0022] An air conditioning control system according to a thirteenth aspect is the air conditioning control system according to any one of the first to twelfth aspects, wherein the target space includes an aisle between server racks in a data center.

[0023] The air conditioning control system of the thirteenth aspect can realize reduction in power consumption of servers by bringing the environmental state of the aisle between server racks in a data center closer to a desired environmental state.

[0024] A placement determination device of a fourteenth aspect determines the placement of air conditioning units in a target space. The placement determination device includes a generation unit, a prediction unit, an evaluation unit, and an extraction unit. The generation unit generates a plurality of candidate placement conditions for the air conditioning units. The prediction unit predicts the environmental state using a learning model that predicts the environmental state of the target space based on the placement conditions. The evaluation unit evaluates the environmental state corresponding to the placement conditions. The extraction unit extracts placement conditions from the plurality of candidate placement conditions whose evaluation by the evaluation unit satisfies predetermined conditions.

[0025] A placement determination device according to a fourteenth aspect extracts placement conditions for air conditioners based on the environmental state of a target space. As a result, the placement determination device can determine placement conditions for air conditioners that bring the target space closer to a desired environmental state.

[0026] The arrangement determination device according to a fifteenth aspect is the arrangement determination device according to the fourteenth aspect, further comprising a learning unit, which generates a learning model.

[0027] A placement determination device according to a sixteenth aspect is the placement determination device according to either the fourteenth or fifteenth aspect, further comprising a calculation unit. The calculation unit receives placement conditions as input and performs a CFD simulation of the environmental state. The learning model receives placement conditions as input and predicts the environmental state, which is the output of the CFD simulation.

[0028] With this configuration, the placement determination device of the sixteenth aspect can determine placement conditions for air conditioners for each placement condition, such that the target space approaches the desired environmental state, without actually operating the air conditioners.

[0029] A placement determination device according to a seventeenth aspect is the placement determination device according to either the fourteenth or fifteenth aspect, further comprising a calculation unit. The calculation unit performs a CFD simulation of an environmental state using placement conditions as input. The calculation unit calculates a first environmental state and a second environmental state having a lower simulation accuracy than the first environmental state. The learning model predicts the first environmental state using the second environmental state as input.

[0030] A seventeenth aspect of the placement determination device performs a CFD simulation of an environmental state using placement conditions as input. The placement determination device calculates a second environmental state, which has low CFD simulation accuracy, from the placement conditions. The placement determination device predicts a first environmental state from the second environmental state using a learning model. Therefore, the placement determination device can reduce the time required for the CFD simulation compared to when the first environmental state, which has high CFD simulation accuracy, is calculated from the placement conditions. As a result, the placement determination device can determine, in a realistic amount of time, placement conditions for air conditioning units that bring the target space closer to the desired environmental state.

[0031] The placement determination device of an eighteenth aspect is the placement determination device of any one of the fourteenth to seventeenth aspects, in which the generation unit generates candidates for placement conditions by reinforcement learning that uses the evaluation of the evaluation unit as a value.

[0032] With this configuration, the placement determination device of the 18th aspect can determine placement conditions for air conditioning units that will bring the target space closer to the desired environmental state, without relying on the operator's experience and without generating candidates for all placement conditions.

[0033] The placement determination device of a 19th aspect is the placement determination device of the 18th aspect, in which the generation unit generates candidates for placement conditions using a learning model that determines candidates for the next placement conditions based on the current placement conditions.

[0034] The placement determination device of the nineteenth aspect can easily generate placement condition candidates by reinforcement learning using a learning model, even if there are a large number of placement condition elements.

[0035] The placement determination device according to a twentieth aspect is the placement determination device according to any one of the fourteenth to nineteenth aspects, wherein the target space includes an aisle between server racks in a data center.

[0036] The placement determination apparatus according to the twentieth aspect can reduce the power consumption of servers by bringing the environmental state of the aisle between server racks in a data center closer to a desired environmental state. [Brief explanation of the drawings]

[0037] [Figure 1] FIG. 2 is a schematic plan view of the data center according to the first embodiment. [Figure 2] 1 is a schematic cross-sectional view of a data center according to a first embodiment. [Figure 3] FIG. 2 is a functional block diagram of the air conditioning control system according to the first embodiment. [Figure 4] 1 is a graph showing the relationship between the number of spatial meshes and the CFD execution time in the first embodiment. [Figure 5] 4 is a graph showing the accuracy of the NN model in the first embodiment. [Figure 6] 4 is a heat map showing the wind speed distribution in the server room in the first embodiment. [Figure 7] 4 is a flowchart of a process for controlling an air conditioner using optimal air conditioning setting values ​​in the first embodiment. [Figure 8] FIG. 10 is a schematic plan view of a data center according to a second embodiment. [Figure 9]FIG. 10 is a functional block diagram of an air conditioning control system according to a second embodiment. [Figure 10] 10 is a flowchart of a process for controlling an air conditioner using optimal air conditioning setting values ​​in a second embodiment. [Figure 11] FIG. 10 is a schematic plan view of a data center according to a third embodiment. [Figure 12] FIG. 10 is a schematic cross-sectional view of a data center according to a third embodiment. [Figure 13] FIG. 11 is a functional block diagram of a placement determination device according to a third embodiment. [Figure 14] 11 is a flowchart of a process for determining optimal placement conditions in the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0038] First Embodiment (1) Background In recent years, the Internet of Things (IoT) has been developing, connecting various devices, such as automobiles, electronic devices, and sensors, to the Internet and transmitting and receiving information. The development of the IoT is expected to lead to a rapid increase in the number of M2M (Machine-to-Machine) connections, which represents the scale of machine-to-machine communications. The number of M2M connections worldwide is expected to increase 2.4-fold between 2018 and 2023. Furthermore, by 2023, the number of M2M connections is expected to reach 14.7 billion, or 1.8 devices per capita. As the number of M2M connections increases, the amount of data transmitted and received also increases, increasing the amount of computing resources required at data centers to process this data. As a result, data centers are expected to expand in size and consume more and more power. Therefore, reducing power consumption at data centers has become a challenge.

[0039] Two methods have been proposed to reduce power consumption in data centers. The first is to reduce server power consumption by improving the efficiency of task allocation to servers. The second is to optimize the settings of the air conditioning units (hereafter referred to as air conditioning settings) that ventilate the server rooms in data centers. Existing research has focused primarily on optimizing the cooling efficiency of air conditioning units. Server power consumption is said to account for approximately 30% of a data center's total power consumption, and air conditioning settings have a significant impact on server power consumption. Server power consumption depends on the rotation speed of the server's internal fan, which in turn depends on the intake air speed and ambient temperature. Therefore, server power consumption is heavily dependent on the wind speed distribution within the data center, in addition to the temperature within the data center. Generally, a uniform wind speed distribution in the server room allows the air conditioning units to deliver cool air to the servers evenly, resulting in efficient cooling. Therefore, achieving a uniform wind speed distribution in the server room is important for reducing server power consumption.

[0040] (2) Overall structure The air conditioning control system 1 mainly comprises an air conditioner 10 and an optimization device 20. The air conditioner 10 and the optimization device 20 are communicatively connected via a network NW. The air conditioning control system 1 uses the optimization device 20 to determine optimal air conditioning setting values ​​that bring the target space closer to a desired environmental state. The air conditioning control system 1 then cools (air conditions) the target space using the air conditioner 10 based on the optimal air conditioning setting values.

[0041] In this embodiment, the target space is a server room SP in a data center DC. FIG. 1 is a schematic plan view of the data center DC. FIG. 2 is a schematic cross-sectional view of the data center DC. As shown in FIGS. 1 and 2, the server room SP has a server installation area SP1 and an air conditioner installation area SP2. In the server installation area SP1, six server racks RK are lined up in a row (hereinafter, a group of server racks RK may be referred to as a module). On both sides of the module, there are a cold aisle CA and a hot aisle HA. Each server rack RK houses 22 servers SV. Therefore, a total of 528 servers SV are present in the server installation area SP1. In the air conditioner installation area SP2, three indoor units 11 to 13, which are part of an air conditioning device 10, are installed. Two aisle meshes AM, with adjustable air permeability, are stretched between the air conditioner installation area SP2 and the cold aisle CA of the server installation area SP1. Additionally, a rack mesh RM with adjustable air permeability is stretched on the cold aisle CA side of each server rack RK. As shown by the arrows in Figures 1 and 2, the cool air blown out from the indoor units 11 to 13 is rectified once in the space in front of the indoor units 11 to 13. The rectified cool air passes through the aisle mesh AM and rack mesh RM and is supplied to the servers SV. The air is heated inside the servers SV, passes through the hot aisle HA and the ceiling AC, and is again drawn into the indoor units 11 to 13.

[0042] The environmental state of the server room SP is, for example, the wind speed distribution and temperature distribution of the server room SP. Furthermore, the desired environmental state of the server room SP is a state in which the wind speed distribution of the server room SP is uniform, or a state in which the temperature distribution of the server room SP is biased to a specific location. In this embodiment, the desired environmental state of the server room SP is a state in which the wind speed distribution of the hot aisle HA of the server room SP is uniform. The optimization device 20 determines optimal air conditioning setting values ​​that will result in a uniform wind speed distribution in the hot aisle HA of the server room SP. Therefore, the optimization device 20 needs to predict the wind speed distribution of the server room SP from the air conditioning setting values. The optimization device 20 uses CFD simulation to predict the wind speed distribution of the server room SP from the air conditioning setting values.

[0043] In a CFD simulation, the server room SP is divided into many small spaces (hereinafter sometimes referred to as spatial meshes), and environmental parameters such as wind speed and temperature are predicted for each spatial mesh. In this embodiment, the environmental parameter is wind speed. The more spatial meshes there are, the more precise the results that can be obtained, but a CFD simulation with a large number of spatial meshes requires an enormous amount of time for each calculation. However, in a data center DC, servers SV are densely installed in server racks RK, so precise results must be obtained and the number of spatial meshes cannot be reduced.

[0044] Therefore, in this embodiment, the optimization device 20 performs a low-precision CFD simulation (hereinafter, sometimes referred to as low-precision CFD) (with a relatively small number of spatial meshes). Then, the optimization device 20 performs regression prediction of the results of a high-precision CFD simulation (hereinafter, sometimes referred to as high-precision CFD) (with a relatively large number of spatial meshes) from the results of the low-precision CFD. A neural network model (hereinafter, sometimes referred to as an NN (Neural Network) model) is used for the regression prediction. This makes it possible to reduce the execution time of the high-precision CFD.

[0045] Furthermore, in this embodiment, optimization device 20 uses a reinforcement learning model (hereinafter sometimes referred to as a DRL (Deep Reinforcement Learning) model) to reduce the number of times the CFD simulation is performed. Optimization device 20 can determine optimal air conditioning setting values ​​in a realistic amount of time without searching for all combinations of air conditioning setting values.

[0046] (3) Detailed configuration (3-1) Air conditioning equipment Figure 3 is a functional block diagram of the air conditioning control system 1. As shown in Figures 1 to 3, the air conditioning apparatus 10 mainly has indoor units 11 to 13 and an air conditioning controller 19. In this embodiment, the air conditioning apparatus 10 includes an aisle mesh AM and a rack mesh RM as components.

[0047] The indoor units 11 to 13 are so-called packaged devices. The indoor units 11 to 13 are installed on the floor of the air conditioner installation area SP2. Each of the indoor units 11 to 13 mainly includes a compressor, a condenser, an expansion valve, an evaporator, and a fan. The compressor, condenser, expansion valve, and evaporator are connected by a refrigerant circuit.

[0048] The compressor draws in low-pressure refrigerant, compresses it using a compression mechanism, and discharges the compressed refrigerant. The compression mechanism of the compressor is driven by a compressor motor, and the rotation speed of the compressor motor can be controlled by an inverter.

[0049] In the condenser, heat is exchanged between the refrigerant flowing inside the condenser and water cooled by a heat source unit installed on the roof of a data center DC, for example. At this time, the refrigerant flowing inside the condenser condenses into liquid refrigerant.

[0050] The expansion valve adjusts the pressure and flow rate of the refrigerant flowing through the refrigerant circuit. The expansion valve is an electronic expansion valve with adjustable opening.

[0051] The fan supplies air from the air conditioner installation area SP2 to the evaporator. The fan is driven by a fan motor. The rotation speed of the fan motor can be controlled by an inverter.

[0052] In the evaporator, heat is exchanged between the refrigerant flowing through the evaporator and the air in the air conditioner installation area SP2 supplied by the fan. As shown in FIG. 2, the indoor units 11-13 drive their fans to draw in air from the air conditioner installation area SP2 through the air inlets at their tops. The drawn air from the air conditioner installation area SP2 passes through the evaporator. At this time, the refrigerant flowing inside the evaporator evaporates and becomes gas refrigerant. The air that has passed through the evaporator is cooled, and cool air is blown out from the air outlets of the indoor units 11-13.

[0053] The air conditioning controller 19 controls the operations of the various devices that make up the air conditioner 10.

[0054] The air conditioning controller 19 has a control arithmetic device and a storage device. The control arithmetic device is a processor such as a CPU or GPU. The storage device is a storage medium such as a RAM, a ROM, or a flash memory. The control arithmetic device reads out programs stored in the storage device and performs predetermined arithmetic processing in accordance with the programs, thereby controlling the operation of the various devices that make up the air conditioning apparatus 10. The control arithmetic device can also write calculation results to the storage device and read out information stored in the storage device in accordance with the programs.

[0055] The air conditioning controller 19 is electrically connected to the various devices that make up the air conditioner 10 so as to be able to exchange control signals and information. The air conditioning controller 19 is also connected so as to be able to communicate with the various sensors provided in the air conditioner 10. The air conditioning controller 19 also exchanges various information, including air conditioning set values, and various signals with the control unit 29 of the optimization device 20 via the network NW. The air conditioning controller 19 receives optimal air conditioning set values ​​from the optimization device 20 and controls the various devices that make up the air conditioner 10 using these optimal air conditioning set values. The air conditioning set values ​​include, for example, the blowing air speed, set temperature, set humidity, etc. of each of the indoor units 11 to 13, the transmittance of each of the aisle mesh AM and rack mesh RM, etc.

[0056] (3-2) Optimization device The optimization device 20 of this embodiment is a computer installed on the cloud. The optimization device 20 may be installed, for example, in a data center DC. As shown in FIG. 3 , the optimization device 20 mainly includes a storage unit 21, an input unit 22, a display unit 23, a communication unit 24, and a control unit 29.

[0057] (3-2-1) Storage section The storage unit 21 is a storage device such as a RAM, a ROM, and an HDD (hard disk drive), etc. The storage unit 21 stores programs executed by the control unit 29, data necessary for executing the programs, and the like.

[0058] (3-2-2) Input section The input unit 22 is a keyboard and a mouse. Various commands and information for the optimization device 20, such as the initial values ​​of the air conditioning setting values ​​to be searched for, can be input using the input unit 22.

[0059] (3-2-3) Display section The display unit 23 is a monitor, and can display the results of the CFD simulation, etc.

[0060] (3-2-4) Communications Department The communication unit 24 is a network interface device for communicating with the air conditioner 10 via the network NW.

[0061] (3-2-5) Control Unit The control unit 29 is a processor such as a CPU or a GPU. In this embodiment, the processor is an Intel Core i7-based processor. TM The optimization device 20 was run on a 3.7GHz i7-8700K processor. The control unit 29 reads and executes programs stored in the storage unit 21 to implement various functions of the optimization device 20. The control unit 29 can also write calculation results to the storage unit 21 and read information stored in the storage unit 21 in accordance with the programs.

[0062] The control unit 29 exchanges various information including air conditioning setting values ​​and various signals with the air conditioning controller 19 of the air conditioner 10 via the network NW.

[0063] 3, the control unit 29 has, as functional blocks, a calculation unit 31, a first learning unit 32, a prediction unit 33, a generation unit 34, an evaluation unit 35, an extraction unit 36, and a second learning unit 37. The calculation unit 31, the first learning unit 32, the prediction unit 33, the generation unit 34, the evaluation unit 35, the extraction unit 36, and the second learning unit 37 are functional blocks for realizing the DRL models 51 to 53.

[0064] (3-2-5-1) Arithmetic section The calculation unit 31 receives the air conditioning setting values ​​(operating conditions) as input and performs a CFD simulation of the wind speed distribution (environmental state) in the server room SP.

[0065] The air conditioning set value in this embodiment has five elements: the blowing air speed of each of the indoor units 11 to 13, and the transmittance of each of the aisle mesh AM and rack mesh RM. Each element of the air conditioning set value satisfies the condition of the following equation 1.

[0066]

number

[0067] Here, S is the air conditioning setting value, and M aisle is the transmittance of the air mesh AM, and M rack indicates the transmittance of the rack mesh RM, and P1 to P3 indicate the blowing air speeds of the indoor units 11 to 13, respectively.

[0068] In this embodiment, Flow Designer provided by Advance Knowledge Laboratory is used as software for performing the CFD simulation. The results of the CFD simulation (wind speed for each spatial mesh) are output to a CSV file or the like.

[0069] The calculation unit 31 calculates a high-precision wind speed distribution (first environmental state; hereinafter, sometimes referred to as high-precision wind speed distribution) that is the result of high-precision CFD for the server room SP, and a low-precision wind speed distribution (second environmental state; hereinafter, sometimes referred to as low-precision wind speed distribution) that is the result of low-precision CFD for the server room SP.

[0070] When searching for optimal air conditioning setting values, calculation unit 31 calculates a low-accuracy air speed distribution for each air conditioning setting value to be searched for, in order to input it into NN model 41. Furthermore, calculation unit 31 calculates a pair of a low-accuracy air speed distribution and a high-accuracy air speed distribution (training data set for NN model 41) for each of a plurality of air conditioning setting values, in order to pre-train NN model 41.

[0071] (3-2-5-2) First Learning Section The first learning unit 32 generates the NN model 41 using the learning dataset for the NN model 41 calculated by the calculation unit 31. At this time, the first learning unit 32 learns the NN model 41 using the low-accuracy wind speed distribution as an explanatory variable and the high-accuracy wind speed distribution as a target variable.

[0072] In this embodiment, in order to create a training dataset for the NN model 41, 896 air conditioning setting values ​​that satisfy the following equation 2 under the constraint of equation 1 were used.

[0073]

number

[0074] The calculation unit 31 performed three types of low-precision CFD and one type of high-precision CFD, as shown in Table 1 below, for each of these air conditioning setting values.

[0075] [Table 1]

[0076] Figure 4 is a graph showing the relationship between the number of spatial meshes and the execution time of a CFD simulation (hereinafter sometimes referred to as CFD execution time) for one air conditioning setting value. As shown in Figure 4, as the number of spatial meshes increases, the CFD execution time increases. For example, when the number of spatial meshes is 23,750, the CFD execution time is approximately 20 seconds, and when the number of spatial meshes is 1,000,000, the CFD execution time is approximately 780 seconds.

[0077] In this embodiment, 80% of the training data set for the NN model 41 was used as training data and 20% as test data. FIG. 5 is a graph showing the accuracy of the NN model 41. The accuracy evaluation index is RMSE (Root Mean Square Error). The solid line graph in FIG. 5 shows the accuracy of the NN model 41 trained by associating low-accuracy wind speed distributions with high-accuracy wind speed distributions. The dashed line graph in FIG. 5 shows the accuracy of a training model trained by associating air conditioning setting values ​​with high-accuracy wind speed distributions. For example, when the number of spatial meshes is 23,750, the RMSE of the solid line graph is approximately 0.009 m / s, and the RMSE of the dashed line graph is approximately 0.07 m / s. Therefore, when the number of spatial meshes is 23,750, it can be seen that the NN model 41 has significantly higher accuracy than a training model that predicts high-accuracy wind speed distributions from air conditioning setting values. It is also apparent that the accuracy of the NN model 41 when the number of spatial meshes is 23,750 is almost the same as the accuracy when the number of spatial meshes is 500,000.

[0078] In other words, the time required to calculate a high-accuracy wind speed distribution from a low-accuracy wind speed distribution using the NN model 41 is approximately 20 times shorter than the time required to calculate a high-accuracy wind speed distribution from an air conditioning setting value using high-accuracy CFD.

[0079] From consideration of FIGS. 4 and 5, in this embodiment, when predicting a high-accuracy wind speed distribution, a low-accuracy wind speed distribution with 23,750 spatial meshes is used.

[0080] (3-2-5-3) Prediction section The prediction unit 33 predicts a high-accuracy wind speed distribution using the NN model 41 generated by the first learning unit 32. Specifically, the prediction unit 33 inputs the low-accuracy wind speed distribution calculated by the calculation unit 31 based on the air conditioning setting values ​​into the NN model 41, and predicts a high-accuracy wind speed distribution.

[0081] (3-2-5-4) Generation part There are tens of thousands to hundreds of thousands of possible combinations of air conditioning settings, so it is not realistic to search for all of them. Therefore, a DRL model is used as a method for searching for optimal air conditioning settings. In this embodiment, the learning model used to select the next action in the DRL model is a Dueling Network model (hereinafter sometimes referred to as a DN model).

[0082] When a search is performed using a single DRL model, the number of combinations of air conditioning setting values ​​remains in the tens of thousands to hundreds of thousands. Therefore, in this embodiment, three DRL models 51 to 53 are used. The search range and search width of the air conditioning setting values ​​are gradually changed for each DRL model 51 to 53, thereby reducing the number of combinations of air conditioning setting values. In this way, the number of combinations for each DRL model 51 to 53 is kept to around one thousand and several hundred, significantly reducing the number of times the CFD simulation needs to be run. The search range and search width of the air conditioning setting values ​​for each DRL model 51 to 53 are shown in Table 2 below.

[0083] [Table 2]

[0084] By using the DRL models 51 to 53, it is possible to finally determine the optimum air conditioning setting values ​​for the transmittance of the aisle mesh AM and the rack mesh RM in 1% increments and for the blowing air speed of the indoor units 11 to 13 in 0.1 m / s increments.

[0085] The states (air conditioning setting values) and actions in the DRL models 51 to 53 are shown in the following equation 3.

[0086]

number

[0087] where S is the state, a is the action, and M aisle + etc. show the behavior of increasing or decreasing the corresponding air conditioning setting value by the search width. For example, M in DRL model 51 aisle + is M aisle This means increasing it by 10%.

[0088] In this embodiment, the total number of searches to find the optimal air conditioning setting values ​​is 5,000 steps across all of the DRL models 51 to 53.

[0089] The generation unit 34 generates candidates for air conditioning setting values ​​to be searched for using DRL models 51-53, which use the evaluation by the evaluation unit 35 as a value (reward). Specifically, the generation unit 34 generates candidates for air conditioning setting values ​​with a high expected reward using DN models 61-63, which determine candidates for the next air conditioning setting value based on the current air conditioning setting value. The DN models 61-63 are learning models corresponding to the DRL models 51-53, respectively.

[0090] (3-2-5-5) Evaluation Section The evaluation unit 35 evaluates the high-precision air velocity distribution corresponding to the air conditioning setting value.

[0091] The evaluation unit 35 first calculates an evaluation value for the air conditioning setting value based on the air conditioning setting value and the high-precision air speed distribution corresponding to the air conditioning setting value.

[0092] In this embodiment, the evaluation value is defined by the following equation 4.

[0093]

number

[0094] The evaluation value is calculated by averaging the variance score, mesh score, and air conditioning score with weights α, β, and γ. The variance score, mesh score, and air conditioning score are each configured to take values ​​between 0 and 100 (as a result, the evaluation value also takes values ​​between 0 and 100). The larger the evaluation value, variance score, mesh score, and air conditioning score, the higher the evaluation.

[0095] The dispersion score depends on the dispersion of the set of wind speeds in specific spatial meshes that exist behind each server SV (hot aisle HA side), which are included in the high-precision wind speed distribution. The dispersion score is configured so that the smaller the dispersion, the larger the value. The dispersion score represents the uniformity of the wind speed distribution in the hot aisle HA of the server room SP. The larger the dispersion score, the more uniform the wind speed distribution in the hot aisle HA of the server room SP.

[0096] The mesh score depends on the sum of the permeability of the aisle mesh AM and the permeability of the rack mesh RM. The mesh score is configured so that the greater the permeability of the aisle mesh AM or the rack mesh RM, the larger the value. The higher the mesh score, the more wind passes through the aisle mesh AM or the rack mesh RM, and the smaller the dispersion score.

[0097] The air conditioning score depends on the sum of the cubes of the blown air speeds of the indoor units 11 to 13. The air conditioning score is configured so that the smaller the exhaust air speeds of the indoor units 11 to 13 (the smaller the power consumption of the air conditioner 10), the larger the value. The air conditioning score is defined in this way because the power consumption of the air conditioner 10 is proportional to the cube of the blown air speed.

[0098] In this embodiment, since emphasis is placed on making the wind speed distribution in the hot aisle HA uniform, α is set to 0.6 (a relatively large value), β is set to 0.1, and γ is set to 0.3.

[0099] After calculating the evaluation values, the evaluation unit 35 calculates rewards for the DRL models 51 to 53 based on the evaluation values, as shown in Table 3 below.

[0100] [Table 3]

[0101] In this embodiment, the current evaluation value is compared with the average of the evaluation values ​​of the past 10 steps to stabilize the DRL models 51 to 53.

[0102] (3-2-5-6) Extraction part The extraction unit 36 ​​extracts the optimal air conditioning setting value with the highest evaluation value from the evaluation unit 35 for each of the DRL models 51 to 53 from among the multiple air conditioning setting value candidates generated by the generation unit 34. The control unit 29 ultimately transmits the optimal air conditioning setting value for the DRL model 53 to the air conditioner 10.

[0103] (3-2-5-7) Second Learning Section During the search using the DRL models 51 to 53, the second learning unit 37 associates a state with a reward expected when each action is taken from the state, and stores the associations in the storage unit 21 for each of the DRL models 51 to 53.

[0104] The second learning unit 37 uses the above data stored in the storage unit 21 to learn and update the DN models 61 to 63 as needed.

[0105] (4) Verification results In this embodiment, the air conditioning setting values ​​of the following equation 5 are used as the initial values ​​of the air conditioning setting values ​​in the DRL model 51.

[0106]

number

[0107] The air conditioning setting value is a value at which the mesh score and air conditioning score are 100.

[0108] Table 4 below shows the optimal air conditioning setting values ​​extracted in the DRL models 51 to 53 and their evaluation values.

[0109] [Table 4]

[0110] As shown in Table 4, DRL models 51 to 53 extract air conditioning setting values ​​with relatively high evaluation values.

[0111] Fig. 6 is a heat map showing the air speed distribution in the server room SP when the optimal air conditioning setting values ​​in the DRL model 53 are set for the air conditioner 10. As shown in Fig. 6, the air speed distribution in the hot aisle HA is almost uniform (the degree of shading is the same).

[0112] Table 5 below shows the number of searches and the number of times the CFD simulation was performed for each of the DRL models 51 to 53.

[0113] [Table 5]

[0114] As shown in Table 5, by performing a stepwise search using DRL models 51 to 53, it was possible to significantly reduce the number of times the CFD simulation was run. In this verification, the air conditioning setting value that appeared for the first time in the process of searching for the air conditioning setting value and the corresponding high-precision air speed distribution were stored in memory unit 21. If the air conditioning setting value stored in memory unit 21 appeared again in the process of searching for the air conditioning setting value, the CFD simulation was not performed, and the corresponding high-precision air speed distribution was read out from memory unit 21, thereby reducing the number of times the simulation was run.

[0115] Each search by the DRL models 51 to 53 requires 0.3 seconds. Calculating a high-precision air speed distribution requires 20.3 seconds (10 seconds for inputting air conditioning setting values, 5 seconds for low-precision CFD, 5 seconds for output to a CSV file or the like, and 0.3 seconds for calculating the high-precision air speed distribution using the NN model 60). Therefore, the air conditioning control system 1 in this embodiment was able to determine the optimal air conditioning setting values ​​in the DRL model 53 in approximately 3.77 hours (= (5000 × 0.3 seconds + 596 × 20.3 seconds) ÷ 3600).

[0116] (5) Processing An example of processing for controlling the air conditioner 10 using optimal air conditioning setting values ​​will be described using the flowchart in Figure 7. In this flowchart, steps S1 to S8 are repeated three times in response to the DRL models 51 to 53 being executed in sequence.

[0117] The first processing from step S1 to step S8 is processing related to the DRL model 51.

[0118] As shown in step S1, the initial values ​​of the air conditioning setting values ​​in the DRL model 51 are input to the optimization device 20.

[0119] After step S1 is completed, the air conditioning control system 1 calculates a low-accuracy air speed distribution from the air conditioning setting values, as shown in step S2.

[0120] After step S2 is completed, the air conditioning control system 1 predicts a high-accuracy wind speed distribution from the low-accuracy wind speed distribution using the NN model 41, as shown in step S3.

[0121] After step S3, as shown in step S4, the air conditioning control system 1 calculates an evaluation value and a reward for the air conditioning setting value from the air conditioning setting value and the high-precision air speed distribution.

[0122] After step S4, the air conditioning control system 1 uses the DN model 61 to generate candidates for air conditioning setting values ​​with high expected rewards, as shown in step S5.

[0123] After step S5, as shown in step S6, the air conditioning control system 1 determines whether the number of searches by the DRL model 51 is equal to or less than a predetermined number (a total number of searches by the DRL model 51 set in advance). If the number of searches by the DRL model 51 is equal to or less than the predetermined number, the process proceeds to step S2. If the number of searches by the DRL model 51 is greater than the predetermined number, the process proceeds to step S7.

[0124] When proceeding from step S6 to step S2, the air conditioning control system 1 calculates a low-accuracy air speed distribution from the air conditioning setting values ​​generated in step S5. In other words, the air conditioning control system 1 repeats steps S2 to S6 until the number of searches by the DRL model 51 becomes greater than a predetermined number.

[0125] When the process proceeds from step S6 to step S7, the air conditioning control system 1 extracts the optimal air conditioning setting value with the highest evaluation value from among the air conditioning setting values ​​found in the DRL model 51.

[0126] After step S7, as shown in step S8, the air conditioning control system 1 determines whether or not processing has been completed for all of the DRL models 51 to 53. At this stage, only DRL model 51 has been completed, so the process proceeds to step S1.

[0127] The second processing from step S1 to step S8 is processing related to the DRL model 52. Basically, it is the same as the first processing from step S1 to step S8, so only the different parts will be explained.

[0128] In step S1, the air conditioning control system 1 inputs the optimal air conditioning setting values ​​of the DRL model 51 extracted in the first step S7 to the optimization device 20 as the initial values ​​of the air conditioning setting values ​​in the DRL model 52.

[0129] In step S5, the air conditioning control system 1 uses the DN model 62 to generate candidates for air conditioning setting values ​​with high expected rewards.

[0130] In step S6, the air conditioning control system 1 determines whether the number of searches by the DRL model 52 is equal to or less than a predetermined number (a total number of searches by the DRL model 52 set in advance).

[0131] In step S7, the air conditioning control system 1 extracts the optimal air conditioning setting value with the highest evaluation value from the air conditioning setting values ​​found in the DRL model 52.

[0132] In step S8, the air conditioning control system 1 determines whether or not processing has been completed for all of the DRL models 51 to 53. At this stage, only the DRL models 51 and 52 have been completed, so the process proceeds to step S1.

[0133] The processing from step S1 to step S8 for the third time is processing related to the DRL model 53. Since it is basically the same as the processing from step S1 to step S8 for the first time, only the different parts will be explained.

[0134] In step S1, the air conditioning control system 1 inputs the optimal air conditioning setting values ​​of the DRL model 52 extracted in the second step S7 to the optimization device 20 as the initial values ​​of the air conditioning setting values ​​in the DRL model 53.

[0135] In step S5, the air conditioning control system 1 uses the DN model 63 to generate candidates for air conditioning setting values ​​with high expected rewards.

[0136] In step S6, the air conditioning control system 1 determines whether the number of searches by the DRL model 53 is equal to or less than a predetermined number (a total number of searches by the DRL model 53 set in advance).

[0137] In step S7, the air conditioning control system 1 extracts the optimal air conditioning setting value with the highest evaluation value from the air conditioning setting values ​​found in the DRL model 53.

[0138] In step S8, the air conditioning control system 1 determines whether or not the processing has been completed for all of the DRL models 51 to 53. At this stage, the processing has been completed for all of the DRL models 51 to 53, so the process proceeds to step S9.

[0139] When proceeding from step S8 to step S9, the air conditioning control system 1 transmits the optimal air conditioning setting values ​​of the DRL model 53 extracted in the third step S7 from the optimization device 20 to the air conditioning device 10, and controls the air conditioning device 10 using the optimal air conditioning setting values.

[0140] The air conditioning control system 1 learns and updates the DN models 61 to 63 as needed while processing the DRL models 51 to 53.

[0141] (6) Features (6-1) 2. Description of the Related Art Conventionally, there are techniques for determining optimal operating conditions for an air conditioning device so that a target space is in a desired environmental state.

[0142] However, with conventional technologies, the operating conditions of the air conditioning unit are determined based on an existing database, which means that the target space cannot be brought sufficiently close to the desired environmental state.Furthermore, with conventional technologies, there is also the problem that the entire target space cannot be brought sufficiently close to the desired environmental state at the same time.

[0143] The air conditioning control system 1 of this embodiment conditions the air in a server room SP using an air conditioner 10. The air conditioning control system 1 includes a generation unit 34, a first learning unit 32, a prediction unit 33, an evaluation unit 35, an extraction unit 36, and an air conditioning controller 19. The generation unit 34 generates multiple candidates for air conditioning setting values ​​for the air conditioner 10. The first learning unit 32 generates an NN model 41. The prediction unit 33 predicts an air speed distribution using the NN model 41. The NN model 41 predicts the air speed distribution in the server room SP based on the air conditioning setting values. The evaluation unit 35 evaluates the air speed distribution corresponding to the air conditioning setting values. The extraction unit 36 ​​extracts the air conditioning setting value with the highest evaluation by the evaluation unit 35 from the multiple candidate air conditioning setting values. The air conditioning controller 19 controls the air conditioning device 10 using the air conditioning setting values ​​extracted by the extraction unit 36.

[0144] The air conditioning control system 1 extracts air conditioning setting values ​​based on the air velocity distribution in the server room SP. As a result, the air conditioning control system 1 can determine optimal air conditioning setting values ​​that bring the server room SP closer to the desired air velocity distribution. The air conditioning control system 1 can also bring the entire server room SP closer to the desired air velocity distribution at the same time.

[0145] (6-2) The air conditioning control system 1 of this embodiment further includes a calculation unit 31. The calculation unit 31 performs a CFD simulation of wind speed distribution using air conditioning setting values ​​as input. The calculation unit 31 calculates a high-precision wind speed distribution and a low-precision wind speed distribution. The NN model 41 predicts the high-precision wind speed distribution using the low-precision wind speed distribution as input. The NN model 41 is a model that has undergone learning using the low-precision wind speed distribution as an explanatory variable and the high-precision wind speed distribution as a target variable.

[0146] The air conditioning control system 1 performs a CFD simulation of the wind speed distribution in the server room SP using the air conditioning setting values ​​as input. The air conditioning control system 1 calculates the low-accuracy wind speed distribution from the air conditioning setting values. The air conditioning control system 1 predicts the high-accuracy wind speed distribution from the low-accuracy wind speed distribution using the NN model 41.

[0147] Therefore, the air conditioning control system 1 can reduce the time required for CFD simulation compared to when calculating a highly accurate air speed distribution from air conditioning setting values. As a result, the air conditioning control system 1 can find, in a realistic amount of time, optimal air conditioning setting values ​​that bring the server room SP closer to the desired air speed distribution.

[0148] (6-3) In the air conditioning control system 1 of this embodiment, the generation unit generates candidates for air conditioning setting values ​​through reinforcement learning that uses the evaluation by the evaluation unit 35 as a value.

[0149] As a result, the air conditioning control system 1 can determine the optimal air conditioning setting values ​​that bring the server room SP closer to the desired air speed distribution, without relying on the operator's experience and without generating all possible air conditioning setting values.

[0150] (6-4) In the air conditioning control system 1 of this embodiment, the generation unit 34 generates candidates for air conditioning setting values ​​using DN models 61 to 63 that determine candidates for the next air conditioning setting value based on the current air conditioning setting value.

[0151] As a result, the air conditioning control system 1 can easily generate optimal air conditioning setting value candidates using the DRL models 51-53 that use the DN models 61-63, even if there are a large number of elements for the air conditioning setting values.

[0152] (6-5) In the air conditioning control system 1 of this embodiment, the target space includes an aisle between server racks RK in a data center DC.

[0153] As a result, the air conditioning control system 1 can reduce the power consumption of the servers SV by bringing the air velocity distribution in the cold aisle CA and the hot aisle HA of the server room SP closer to a desired air velocity distribution.

[0154] (7) Variations (7-1) Variation 1A In this embodiment, the air conditioning apparatus 10 air-conditioned the server room SP using floor-standing, packaged indoor units 11-13. However, the air conditioning apparatus 10 may also air-condition the server room SP using a so-called building multi-type or central air-conditioning type. The indoor units 11-13 may also be ceiling-suspended or wall-mounted.

[0155] (7-2) In this embodiment, the target space of the air conditioning control system 1 is a server room SP in a data center DC. However, this is not limited to this, and the target space of the air conditioning control system 1 may be a space in a commercial facility, an office space, etc.

[0156] (7-3) In this embodiment, the air conditioning control system 1 calculates a low-accuracy wind speed distribution from the air conditioning setting values ​​and predicts a high-accuracy wind speed distribution by inputting the low-accuracy wind speed distribution into the NN model 41. However, the air conditioning control system 1 may also use a learning model 42 that predicts a high-accuracy wind speed distribution using the air conditioning setting values ​​as input. The learning model 42 is, for example, a neural network.

[0157] At this time, in order to pre-train the learning model 42, the calculation unit 31 calculates a high-precision wind speed distribution (a learning data set for the learning model 42) for each of a plurality of air conditioning setting values.

[0158] The first learning unit 32 generates the learning model 42 using the learning dataset for the learning model 42 calculated by the calculation unit 31. At this time, the first learning unit 32 learns the learning model 42 using the air conditioning setting values ​​as explanatory variables and the high-precision air speed distribution as a target variable.

[0159] The prediction unit 33 predicts a high-accuracy wind speed distribution using the learning model 42 generated by the first learning unit 32. Specifically, the prediction unit 33 inputs the air conditioning setting values ​​into the learning model 42 and predicts a high-accuracy wind speed distribution.

[0160] As a result, the air conditioning control system 1 can find optimal air conditioning setting values ​​that bring the server room SP closer to the desired air speed distribution for each air conditioning setting value, without actually controlling the air conditioner 10.

[0161] (7-4) In this embodiment, the environmental condition is the wind speed distribution in the server room SP. The evaluation value of the air conditioning setting value is defined so that when the wind speed distribution in the server room SP becomes a desired wind speed distribution, the evaluation value of the corresponding air conditioning setting value becomes high (Equation 4).

[0162] However, the environmental condition may also be the temperature distribution in the server room SP. In this case, the evaluation value of the air conditioning setting value is defined so that the evaluation of the corresponding air conditioning setting value is high when the temperature distribution in the server room SP becomes a desired temperature distribution. The desired temperature distribution is, for example, a uniform temperature distribution or a temperature distribution that eliminates hot spots where the temperature in a specific part becomes high.

[0163] In this way, a desired environmental state can be achieved by arbitrarily setting the evaluation value of the air conditioning setting value.

[0164] (7-5) Although the embodiments of the present disclosure have been described above, it will be understood that various changes in form and details can be made without departing from the spirit and scope of the present disclosure as defined in the claims.

[0165] Second Embodiment The following description will focus on the differences from the first embodiment. Everything other than what will be described in this embodiment is basically the same as the first embodiment.

[0166] (1) Overall structure The air conditioning control system 2 mainly includes an air conditioner 10 and an optimization device 120. The air conditioning control system 2 uses the optimization device 120 to find optimal air conditioning setting values ​​that bring the server room SP closer to a desired environmental state.

[0167] FIG. 8 is a schematic plan view of a data center DC in this embodiment. The difference from the first embodiment is that environmental sensors for measuring the surrounding wind speed, temperature, and the like are installed on the rear (hot aisle HA side) of some server racks RK. In this embodiment, the environmental sensor is a wind speed sensor SS1. Also, in FIG. 8, a virtual wind speed sensor SS2 (hereinafter, sometimes referred to as a virtual wind speed sensor SS2) that is not actually installed is shown on the rear of the server rack RK where the wind speed sensor SS1 is not installed. In FIG. 8, the virtual wind speed sensor SS2 is distinguished from the virtual wind speed sensor SS2 by shading it in. In the air-conditioning control system 2, the environmental state of the server room SP is the value of the wind speed (environmental parameter) at the installation location (first location) of the wind speed sensor SS1 and the installation location (second location) of the virtual wind speed sensor SS2 in the server room SP. In other words, the environmental state of the server room SP is the measurement values ​​of the wind speed sensor SS1 and the virtual wind speed sensor SS2. The desired environmental state of the server room SP is a state in which the measured values ​​of the wind speed sensor SS1 and the virtual wind speed sensor SS2 are uniform. In other words, the desired environmental state of the server room SP is a state in which the wind speed distribution in the hot aisle HA is uniform, as in the first embodiment.

[0168] (2) Detailed configuration (2-1) Air conditioning equipment In this embodiment, the air conditioning apparatus 10 further includes an air velocity sensor SS1 as a component. The air conditioning controller 19 is communicably connected to the air velocity sensor SS1. The air conditioning controller 19 can obtain the measurement values ​​of the air velocity sensor SS1 by actual measurement (hereinafter, the measurement values ​​obtained by actual measurement of the air velocity sensor SS1 may be referred to as actual sensor values).

[0169] (2-2) Optimization device 9 is a functional block diagram of the air conditioning control system 2. As shown in FIG. 9, the optimization device 120 mainly includes a storage unit 21, an input unit 22, a display unit 23, a communication unit 24, and a control unit 129.

[0170] 9, the control unit 129 has, as functional blocks, an acquisition unit 138, a calculation unit 131, a first learning unit 132, a prediction unit 133, a generation unit 34, an evaluation unit 35, an extraction unit 36, and a second learning unit 37. The acquisition unit 138, the calculation unit 131, the first learning unit 132, the prediction unit 133, the generation unit 34, the evaluation unit 35, the extraction unit 36, and the second learning unit 37 are functional blocks for realizing the DRL models 51 to 53.

[0171] (2-2-1) Acquisition department The acquisition unit 138 acquires the measured sensor values ​​via the air conditioning controller 19.

[0172] (2-2-2) Arithmetic section The calculation unit 131 uses the air conditioning setting values ​​as input and performs a CFD simulation of the measurement values ​​of the wind speed sensor SS1 and the virtual wind speed sensor SS2 (hereinafter, the measurement values ​​of the wind speed sensor SS1 that have been subjected to the CFD simulation may be referred to as the CFD sensor values, and the measurement values ​​of the virtual wind speed sensor SS2 that have been subjected to the CFD simulation may be referred to as the CFD virtual sensor values.) Here, the measurement values ​​of the wind speed sensor SS1 and the virtual wind speed sensor SS2 that have been subjected to the CFD simulation refer to the simulated wind speed values ​​in the spatial mesh that corresponds to the installation locations of the wind speed sensor SS1 and the virtual wind speed sensor SS2.

[0173] (2-2-3) First Learning Section The first learning unit 132 generates a learning model 43. The learning model 43 predicts the measurement value of the virtual wind speed sensor SS2 using as input the actually measured sensor value when the air conditioner 10 is controlled by a certain air conditioning setting value.

[0174] The first learning unit 132 pre-learns the learning model 43 in three stages.

[0175] In the first stage, the first learning unit 132 prepares a plurality of air conditioning setting values. For each of the prepared plurality of air conditioning setting values, the first learning unit 132 controls the air conditioner 10 using the prepared plurality of air conditioning setting values ​​and acquires the measured sensor values ​​at that time. The first learning unit 132 also performs a CFD simulation and acquires CFD sensor values ​​and CFD virtual sensor values ​​for each of the prepared plurality of air conditioning setting values. The first learning unit 132 associates the CFD sensor values ​​with the measured sensor values ​​and trains the learning model 43a. At this time, the first learning unit 132 trains the learning model 43a (first learning model) using the CFD sensor values ​​as explanatory variables and the measured sensor values ​​as target variables.

[0176] In the second stage, the first learning unit 132 inputs the CFD virtual sensor values ​​acquired in the first stage into the learning model 43a, and predicts the measurement values ​​of the virtual wind speed sensor SS2.

[0177] In the third stage, the first learning unit 132 associates the actual sensor values ​​acquired in the first stage with the measurement values ​​of the virtual wind speed sensor SS2 predicted in the second stage to learn the learning model 43. At this time, the first learning unit 132 learns the learning model 43 using the actual sensor values ​​as explanatory variables and the measurement values ​​of the virtual wind speed sensor SS2 as objective variables. The learning model 43 is used as a learning model that predicts the measurement values ​​of the virtual wind speed sensor SS2 using the actual sensor values ​​as input.

[0178] The learning models 43 and 43a are, for example, neural networks.

[0179] (2-2-4) Prediction section The prediction unit 133 inputs the actually measured sensor values ​​when the air conditioner 10 is controlled using a certain air conditioning setting value into the learning model 43, and predicts the measurement values ​​of the virtual wind speed sensor SS2.

[0180] (2-2-5) Evaluation Section The variance score calculated by the evaluation unit 35 depends on the variance of the set consisting of the actual sensor value and the measurement value of the virtual wind speed sensor SS2 predicted by the learning model 43.

[0181] (3) Processing An example of the process for controlling the air conditioner 10 using optimal air conditioning setting values ​​will be described using the flowchart in Figure 10. The process of steps S1 and S5 to S9 will not be described as they are the same as those in Figure 7. The process of steps S10 to S12 described below is common to the DRL models 51 to 53.

[0182] After step S1 is completed, as shown in step S10, the air conditioning control system 2 transmits the air conditioning setting values ​​from the optimization device 120 to the air conditioner 10, and controls the air conditioner 10 using the air conditioning setting values.

[0183] After step S10, the air conditioning control system 2 acquires an actual measured sensor value from the air velocity sensor SS1 as shown in step S11.

[0184] After step S11 is completed, as shown in step S12, the air conditioning control system 2 uses the learning model 43 to predict the measurement value of the virtual wind speed sensor SS2 from the actually measured sensor value.

[0185] After completing step S12, as shown in step S13, the air conditioning control system 2 calculates an evaluation value and reward for the air conditioning setting value from the actual measured sensor value and the measurement value of the virtual wind speed sensor SS2 predicted by the learning model 43.

[0186] (4) Features (4-1) In the air conditioning control system 2 of this embodiment, the environmental state is the wind speed values ​​at multiple locations in the server room SP, including the installation location of the wind speed sensor SS1 and the installation location of the virtual wind speed sensor SS2. The air conditioning control system 2 further includes an acquisition unit 138. The acquisition unit 138 acquires the wind speed value at the installation location of the wind speed sensor SS1 by actual measurement. The learning model 43 predicts the wind speed value at the installation location of the virtual wind speed sensor SS2 based on the wind speed value at the installation location of the wind speed sensor SS1 acquired by the acquisition unit 138 when the air conditioning apparatus 10 is controlled by the air conditioning setting value. The learning model 43 is a model that has been trained using the wind speed value at the installation location of the wind speed sensor SS1 acquired by the acquisition unit 138 when the air conditioning apparatus 10 is controlled by the air conditioning setting value as an explanatory variable and the wind speed value at the installation location of the virtual wind speed sensor SS2 as a target variable.

[0187] In the air conditioning control system 2, the learning model 43 predicts the measurement value of the virtual wind speed sensor SS2 based on the actually measured sensor value. As a result, the air conditioning control system 2 can predict the sensor values ​​of all locations from the actually measured sensor values ​​of some locations.

[0188] (4-2) The air conditioning control system 2 of this embodiment further includes a calculation unit 131. The calculation unit 131 uses air conditioning setting values ​​as input and performs a CFD simulation of the wind speed values ​​at the installation location of the wind speed sensor SS1 and the installation location of the virtual wind speed sensor SS2. The learning model 43 predicts the wind speed value at the installation location of the virtual wind speed sensor SS2 based on the wind speed value at the installation location of the wind speed sensor SS1 acquired by the acquisition unit 138 and the wind speed values ​​at the installation locations of the wind speed sensor SS1 and the virtual wind speed sensor SS2 calculated by the calculation unit 131. The learning model 43 is a model that has undergone learning using the wind speed value at the installation location of the wind speed sensor SS1 acquired by the acquisition unit 138 as an explanatory variable and the wind speed value at the installation location of the virtual wind speed sensor SS2 as a target variable. The value of the wind speed at the installation location of the virtual wind speed sensor SS2 is predicted by inputting the value of the wind speed at the installation location of the virtual wind speed sensor SS2 calculated by the calculation unit 131 into the learning model 43a that has been trained, using the value of the wind speed at the installation location of the wind speed sensor SS1 calculated by the calculation unit 131 as the explanatory variable and the value of the wind speed at the installation location of the wind speed sensor SS1 acquired by the acquisition unit 138 as the target variable.

[0189] (5) Variations Although the embodiments of the present disclosure have been described above, it will be understood that various changes in form and details can be made without departing from the spirit and scope of the present disclosure as defined in the claims.

[0190] Third Embodiment (1) Overall structure The placement determination device 220 determines the placement of the air conditioning devices 210 in the target space. The placement determination device 220 finds the optimal conditions for placing the air conditioning devices 210 (hereinafter sometimes referred to as placement conditions) that bring the target space closer to the desired environmental state.

[0191] In this embodiment, the target space is a server room SP in a data center DC. Fig. 11 is a schematic plan view of the data center DC. Fig. 12 is a schematic cross-sectional view of the data center DC. As shown in Figs. 11 and 12, the server room SP has a server installation area SP1 and an air conditioner installation area SP2.

[0192] In the server installation area SP1, six server racks RK are lined up in a row (hereinafter, a group of server racks RK may be referred to as a module). On either side of the module, there is a cold aisle CA and a hot aisle HA. Each server rack RK houses a server SV.

[0193] Three indoor units 211-213, which are part of the air conditioner 210, are provisionally installed in the air conditioner installation area SP2. The placement condition for the air conditioner 210 in this embodiment is the position of each of the indoor units 211-213 in the X-axis direction in the air conditioner installation area SP2.

[0194] Two aisle meshes AM, with adjustable airflow transmittance, are stretched between the air conditioner installation area SP2 and the cold aisle CA of the server installation area SP1. Furthermore, a rack mesh RM, with adjustable airflow transmittance, is stretched on the cold aisle CA side of each server rack RK. As indicated by the arrows in Figures 11 and 12, the cool air blown out from the indoor units 211-213 is rectified once in the space in front of the indoor units 211-213. The rectified cool air passes through the aisle meshes AM and rack mesh RM and is supplied to the servers SV. The air heated inside the servers SV and discharged from the servers SV is drawn back into the indoor units 211-213 via the hot aisle HA and the attic AC.

[0195] The environmental state of the server room SP is, for example, the wind speed distribution and temperature distribution of the server room SP. Furthermore, the desired environmental state of the server room SP is a state in which the wind speed distribution of the server room SP is uniform, or a state in which the temperature distribution of the server room SP is biased to a specific location. In this embodiment, the desired environmental state of the server room SP is a state in which the wind speed distribution of the hot aisle HA of the server room SP is uniform. The placement determination device 220 determines the optimal placement conditions that will result in a uniform wind speed distribution in the hot aisle HA of the server room SP. Therefore, the placement determination device 220 needs to predict the wind speed distribution of the server room SP from the placement conditions. The placement determination device 220 uses CFD simulation to predict the wind speed distribution of the server room SP from the placement conditions.

[0196] In a CFD simulation, the server room SP is divided into many small spaces (hereinafter sometimes referred to as spatial meshes), and environmental parameters such as wind speed and temperature are predicted for each spatial mesh. In this embodiment, the environmental parameter is wind speed. The more spatial meshes there are, the more precise the results that can be obtained, but a CFD simulation with a large number of spatial meshes requires an enormous amount of time for each calculation. However, in a data center DC, servers SV are densely installed in server racks RK, so precise results must be obtained and the number of spatial meshes cannot be reduced.

[0197] Therefore, in this embodiment, the placement determination device 220 performs a low-precision CFD simulation (hereinafter, sometimes referred to as low-precision CFD) (with a relatively small number of spatial meshes). Then, the placement determination device 220 performs regression prediction of the results of a high-precision CFD simulation (hereinafter, sometimes referred to as high-precision CFD) (with a relatively large number of spatial meshes) from the results of the low-precision CFD. A neural network model (hereinafter, sometimes referred to as an NN (Neural Network) model) is used for the regression prediction. This makes it possible to reduce the execution time of the high-precision CFD.

[0198] In this embodiment, the placement determination device 220 uses a reinforcement learning model (hereinafter, may be referred to as a DRL (Deep Reinforcement Learning) model) to reduce the number of times that the CFD simulation is executed. The placement determination device 220 can determine optimal placement conditions in a realistic amount of time without searching for all combinations of placement conditions.

[0199] (2) Detailed configuration (2-1) Air conditioning equipment The air conditioner 210 cools (air conditions) the server room SP.

[0200] The air conditioner 210 mainly includes indoor units 211 to 213 and an air conditioning controller 219.

[0201] The indoor units 211-213 are so-called packaged devices. The indoor units 211-213 are installed on the floor of the air conditioner installation area SP2. Each of the indoor units 211-213 mainly includes a compressor, a condenser, an expansion valve, an evaporator, and a fan. The compressor, condenser, expansion valve, and evaporator are connected by a refrigerant circuit.

[0202] The compressor draws in low-pressure refrigerant, compresses it using a compression mechanism, and discharges the compressed refrigerant. The compression mechanism of the compressor is driven by a compressor motor, and the rotation speed of the compressor motor can be controlled by an inverter.

[0203] In the condenser, heat is exchanged between the refrigerant flowing inside the condenser and water cooled by a heat source unit installed on the roof of a data center DC, for example. At this time, the refrigerant flowing inside the condenser condenses into liquid refrigerant.

[0204] The expansion valve adjusts the pressure and flow rate of the refrigerant flowing through the refrigerant circuit. The expansion valve is an electronic expansion valve with adjustable opening.

[0205] The fan supplies air from the air conditioner installation area SP2 to the evaporator. The fan is driven by a fan motor. The rotation speed of the fan motor can be controlled by an inverter.

[0206] In the evaporator, heat is exchanged between the refrigerant flowing through the evaporator and the air in the air conditioner installation area SP2 supplied by the fan. As shown in FIG. 12, the indoor units 211-213 drive their fans to draw in air from the air conditioner installation area SP2 through the air inlets at their tops. The drawn air from the air conditioner installation area SP2 passes through the evaporator. At this time, the refrigerant flowing inside the evaporator evaporates and becomes gas refrigerant. The air that has passed through the evaporator is cooled, and cool air is blown out from the air outlets of the indoor units 211-213.

[0207] The air conditioning controller 219 controls the operations of the various devices that make up the air conditioner 210 .

[0208] The air conditioning controller 219 has a control arithmetic device and a storage device. The control arithmetic device is a processor such as a CPU or GPU. The storage device is a storage medium such as a RAM, a ROM, or a flash memory. The control arithmetic device reads out programs stored in the storage device and performs predetermined arithmetic processing in accordance with the programs, thereby controlling the operation of the various devices that make up the air conditioning apparatus 210. The control arithmetic device can also write calculation results to the storage device and read out information stored in the storage device in accordance with the programs.

[0209] The air conditioning controller 219 is electrically connected to the various devices that make up the air conditioner 210 so as to be able to exchange control signals and information with them. The air conditioning controller 219 is also connected so as to be able to communicate with the various sensors provided in the air conditioner 210. The air conditioning controller 219 controls the various devices that make up the air conditioner 210 using predetermined air conditioning set values. The air conditioning set values ​​include, for example, the blown air speed, set temperature, set humidity, etc. of each of the indoor units 211 to 213.

[0210] (2-2) Placement determination device 13 is a functional block diagram of the arrangement determination device 220. As shown in FIG.

[0211] (2-2-1) Storage section The storage unit 221 is a storage device such as a RAM, a ROM, and an HDD (hard disk drive), etc. The storage unit 221 stores programs executed by the control unit 229, data necessary for executing the programs, and the like.

[0212] (2-2-2) Input section The input unit 222 is a keyboard and a mouse. Various commands and information for the placement determination device 220, such as initial values ​​of placement conditions to be searched for, can be input using the input unit 222.

[0213] (2-2-3) Display section The display unit 223 is a monitor, and can display the results of the CFD simulation, etc.

[0214] (2-2-4) Control Unit The control unit 229 is a processor such as a CPU or a GPU. The control unit 229 reads and executes programs stored in the storage unit 221 to realize various functions of the arrangement determination device 220. The control unit 229 can also write calculation results to the storage unit 221 and read information stored in the storage unit 221 according to the programs.

[0215] 13, the control unit 229 has, as functional blocks, a calculation unit 231, a first learning unit 232, a prediction unit 233, a generation unit 234, an evaluation unit 235, an extraction unit 236, and a second learning unit 237. The calculation unit 231, the first learning unit 232, the prediction unit 233, the generation unit 234, the evaluation unit 235, the extraction unit 236, and the second learning unit 237 are functional blocks for realizing the DRL model 251.

[0216] (2-2-4-1) Arithmetic section The calculation unit 231 performs a CFD simulation of the wind speed distribution in the server room SP using the arrangement conditions, the transmittances of the aisle mesh AM and the rack mesh RM, and the air conditioning setting values ​​as input. In this embodiment, the transmittances of the aisle mesh AM and the rack mesh RM and the air conditioning setting values ​​are fixed to predetermined values. Therefore, the calculation unit 231 essentially performs a CFD simulation of the wind speed distribution in the server room SP using the arrangement conditions as input.

[0217] The placement condition in this embodiment is the position in the X-axis direction of each of the indoor units 211 to 213. Each element of the placement condition satisfies the condition of the following equation 6.

[0218]

number

[0219] Here, S indicates the arrangement condition, X1 to X3 indicate the X-axis coordinates of the indoor units 211 to 213, respectively, and L indicates the length of the server room SP in the X-axis direction.

[0220] An example of software for performing CFD simulations is Flow Designer provided by Advanced Knowledge Research Institute. The results of the CFD simulation (wind speed for each spatial mesh) are output to a CSV file or similar.

[0221] The calculation unit 231 calculates a high-precision wind speed distribution (hereinafter sometimes referred to as high-precision wind speed distribution) that is the result of high-precision CFD for the server room SP, and a low-precision wind speed distribution (hereinafter sometimes referred to as low-precision wind speed distribution) that is the result of low-precision CFD for the server room SP.

[0222] When searching for optimal placement conditions, the calculation unit 231 calculates a low-accuracy wind speed distribution for each placement condition to be searched for, in order to input the distribution to the NN model 241. Furthermore, in order to pre-train the NN model 241, the calculation unit 231 calculates a pair of a low-accuracy wind speed distribution and a high-accuracy wind speed distribution (a learning data set for the NN model 241) for each of a plurality of placement conditions.

[0223] (2-2-4-2) First Learning Section The first learning unit 232 generates the NN model 241 using the learning dataset of the NN model 241 calculated by the calculation unit 231. At this time, the first learning unit 232 learns the NN model 241 using the low-accuracy wind speed distribution as an explanatory variable and the high-accuracy wind speed distribution as a target variable.

[0224] (2-2-4-3) Prediction section The prediction unit 233 predicts a high-accuracy wind speed distribution using the NN model 241 generated by the first learning unit 232. Specifically, the prediction unit 233 inputs the low-accuracy wind speed distribution calculated by the calculation unit 231 based on the placement conditions into the NN model 241, and predicts a high-accuracy wind speed distribution.

[0225] (2-2-4-4) Generation part In this embodiment, a DRL model 251 is used as a method for searching for optimal placement conditions. In this embodiment, a learning model used to select the next action in the DRL model 251 is a Dueling Network model 261 (hereinafter, sometimes referred to as the DN model 261).

[0226] The states (placement conditions) and actions in the DRL model 251 are shown in the following equation 7.

[0227]

number

[0228] where S is the state, a is the action, and X1 + etc. indicate actions to increase or decrease the corresponding placement conditions by the search width.

[0229] The generation unit 234 generates candidates for placement conditions to be searched for using the DRL model 251, which uses the evaluation by the evaluation unit 235 as a value (reward). Specifically, the generation unit 234 generates candidates for placement conditions with a high expected reward using the DN model 61, which determines candidates for the next placement conditions based on the current placement conditions.

[0230] (2-2-4-5) Evaluation Section The evaluation unit 235 evaluates the high-precision wind speed distribution corresponding to the arrangement conditions.

[0231] First, the evaluation unit 235 calculates an evaluation value of the arrangement condition based on the arrangement condition and the high-precision wind speed distribution corresponding to the arrangement condition.

[0232] The evaluation value is defined, for example, as shown in the following equation 8.

[0233]

number

[0234] The evaluation value is calculated by averaging the variance score, mesh score, and air conditioning score with weights α, β, and γ. The variance score, mesh score, and air conditioning score are each configured to take values ​​between 0 and 100 (as a result, the evaluation value also takes values ​​between 0 and 100). The larger the evaluation value, variance score, mesh score, and air conditioning score, the higher the evaluation.

[0235] The dispersion score depends on the dispersion of the set of wind speeds in specific spatial meshes that exist behind each server SV (hot aisle HA side), which are included in the high-precision wind speed distribution. The dispersion score is configured so that the smaller the dispersion, the larger the value. The dispersion score represents the uniformity of the wind speed distribution in the hot aisle HA of the server room SP. The larger the dispersion score, the more uniform the wind speed distribution in the hot aisle HA of the server room SP.

[0236] The mesh score depends on the sum of the permeability of the aisle mesh AM and the permeability of the rack mesh RM. The mesh score is configured so that the greater the permeability of the aisle mesh AM or the rack mesh RM, the larger the value. The higher the mesh score, the more wind passes through the aisle mesh AM or the rack mesh RM, and the smaller the dispersion score.

[0237] The air conditioning score depends on the sum of the cubes of the blown air speeds of the indoor units 211 to 213. The air conditioning score is configured so that the smaller the exhaust air speed of the indoor units 211 to 213 (the smaller the power consumption of the air conditioner 210), the larger the value. The air conditioning score is defined in this way because the power consumption of the air conditioner 210 is proportional to the cube of the blown air speed.

[0238] In this embodiment, since emphasis is placed on making the wind speed distribution in the hot aisle HA uniform, α is set to 0.6 (a relatively large value), β is set to 0.1, and γ is set to 0.3.

[0239] After calculating the evaluation value, the evaluation unit 235 calculates the reward for the DRL model 251 based on the evaluation value, for example, as shown in Table 6 below.

[0240] [Table 6]

[0241] In this embodiment, the DRL model 251 is stabilized by comparing the current evaluation value with the average of the evaluation values ​​of the past 10 steps.

[0242] (2-2-4-6) Extraction part The extraction unit 236 extracts the optimum placement condition having the highest evaluation value from the evaluation unit 235 from among the multiple placement condition candidates generated by the generation unit 234.

[0243] (2-2-4-7) Second Learning Section During the search using the DRL model 251, the second learning unit 237 associates a state with a reward expected when each action is taken from that state, and stores the associations in the storage unit 221.

[0244] The second learning unit 237 uses the above data stored in the storage unit 221 to learn and update the DN model 261 as needed.

[0245] (3) Processing An example of the process for determining the optimum placement conditions will be described with reference to the flowchart of FIG.

[0246] As shown in step S21, the initial values ​​of the placement conditions are input to the placement determination device 220.

[0247] After completing step S21, the placement determination device 220 calculates a low-accuracy wind speed distribution from the placement conditions, as shown in step S22.

[0248] After completing step S22, the arrangement determination device 220 predicts a high-accuracy wind speed distribution from the low-accuracy wind speed distribution using the NN model 241, as shown in step S23.

[0249] After completing step S23, as shown in step S24, the placement determination device 220 calculates an evaluation value and a reward for the placement condition from the placement condition and the high-accuracy wind speed distribution.

[0250] After completing step S24, the placement determination device 220 generates placement condition candidates with high expected rewards using the DN model 261, as shown in step S25.

[0251] After step S25, as shown in step S26, the placement determination device 220 determines whether the number of searches by the DRL model 251 is equal to or less than a predetermined number (a total number of searches by the DRL model 251 set in advance). If the number of searches by the DRL model 251 is equal to or less than the predetermined number, the process proceeds to step S22. If the number of searches by the DRL model 251 is greater than the predetermined number, the process proceeds to step S27.

[0252] When the process proceeds from step S26 to step S22, the placement determination device 220 calculates a low-accuracy wind speed distribution from the placement conditions generated in step S25. In other words, the placement determination device 220 repeats steps S22 to S26 until the number of searches by the DRL model 251 becomes greater than a predetermined number.

[0253] When the process proceeds from step S26 to step S27, the placement determination device 220 extracts the optimal placement condition with the highest evaluation value from among the placement conditions searched for in the DRL model 251.

[0254] The placement determination device 220 learns and updates the DN model 261 as needed while processing the DRL model 251.

[0255] (4) Features (4-1) When designing a data center or adding modules to a server room, it is necessary to determine the optimal placement conditions for air conditioning devices so that the server room will have the desired environmental conditions.

[0256] The placement determination device 220 of this embodiment determines the placement of air conditioning devices 210 in a server room SP. The placement determination device 220 includes a generation unit 234, a first learning unit 232, a prediction unit 233, an evaluation unit 235, and an extraction unit 236. The generation unit 234 generates multiple candidate placement conditions for the air conditioning devices 210. The first learning unit 232 generates an NN model 241. The prediction unit 233 predicts a wind speed distribution using the NN model 241. The NN model 241 predicts the wind speed distribution in the server room SP based on the placement conditions. The evaluation unit 235 evaluates the wind speed distribution corresponding to the placement conditions. The extraction unit 236 extracts the placement condition that has the highest evaluation by the evaluation unit 235 from the multiple candidate placement conditions.

[0257] The placement determination device 220 extracts placement conditions based on the wind speed distribution in the server room SP. As a result, the placement determination device 220 can obtain optimal placement conditions that bring the server room SP closer to a desired wind speed distribution.

[0258] (4-2) The arrangement determination device 220 of this embodiment further includes a calculation unit 231. The calculation unit 231 receives arrangement conditions as input and performs a CFD simulation of wind speed distribution. The calculation unit 231 calculates a high-accuracy wind speed distribution and a low-accuracy wind speed distribution. The NN model 241 receives the low-accuracy wind speed distribution as input and predicts the high-accuracy wind speed distribution.

[0259] The placement determination device 220 performs a CFD simulation of the wind speed distribution in the server room SP using the placement conditions as input. The placement determination device 220 calculates a low-accuracy wind speed distribution from the placement conditions. The placement determination device 220 predicts a high-accuracy wind speed distribution from the low-accuracy wind speed distribution using the NN model 241.

[0260] Therefore, the placement determination device 220 can reduce the time required for the CFD simulation compared to when calculating a highly accurate wind speed distribution from the placement conditions. As a result, the placement determination device 220 can obtain, in a realistic amount of time, the optimal placement conditions that bring the server room SP closer to a desired wind speed distribution.

[0261] (4-3) In the placement determination device 220 of this embodiment, the generation unit 234 generates candidates for placement conditions by reinforcement learning that values ​​the evaluation of the evaluation unit 235.

[0262] As a result, the placement determination device 220 can determine optimal air conditioning setting values ​​that bring the server room SP closer to the desired air velocity distribution, without relying on the experience of the operator and without generating candidates for all placement conditions.

[0263] (4-4) In the arrangement determination device 220 of this embodiment, the generation unit 234 generates candidates for arrangement conditions using a DN model 261 that determines candidates for the next arrangement condition based on the current arrangement condition.

[0264] As a result, the placement determination device 220 can easily generate optimal placement condition candidates by using the DRL model 251 that uses the DN model 261, even if there are many elements of the placement conditions.

[0265] (4-5) In the placement determination device 220 of this embodiment, the target space includes an aisle between server racks RK in a data center DC.

[0266] As a result, the placement determination device 220 can reduce the power consumption of the servers SV by bringing the wind speed distribution in the cold aisle CA and the hot aisle HA of the server room SP closer to a desired wind speed distribution.

[0267] (5) Variations (5-1) Variation 1A In this embodiment, the air conditioner 210 is a floor-standing type that conditions the air in the server room SP using packaged indoor units 11 to 13. However, the air conditioner 210 may also be a so-called building multi-type or central air conditioning type that conditions the air in the server room SP. Furthermore, the indoor units 211 to 213 may be ceiling-suspended or wall-mounted.

[0268] (5-2) In this embodiment, the target space of the placement determination device 220 is a server room SP in a data center DC. However, the target space of the placement determination device 220 may be a space in a commercial facility, an office space, or the like.

[0269] (5-3) In this embodiment, the placement determination device 220 fixed the orientation of the air outlets of the indoor units 211 to 213 and searched for placement conditions one-dimensionally. However, the placement determination device 220 may search for placement conditions two-dimensionally or three-dimensionally, taking into account the translational degree of freedom and the rotational degree of freedom.

[0270] (5-4) In this embodiment, the placement determination device 220 calculates a low-accuracy wind speed distribution from the placement conditions and inputs the low-accuracy wind speed distribution into the NN model 241 to predict a high-accuracy wind speed distribution. However, the placement determination device 220 may also use a learning model 242 that predicts a high-accuracy wind speed distribution using the placement conditions as input. The learning model 242 is, for example, a neural network.

[0271] At this time, in order to pre-learn the learning model 242, the calculation unit 231 calculates a high-precision wind speed distribution (a learning data set for the learning model 242) for each of a plurality of arrangement conditions.

[0272] The first learning unit 232 generates the learning model 42 using the learning dataset of the learning model 242 calculated by the calculation unit 231. At this time, the first learning unit 232 learns the learning model 242 using the placement conditions as explanatory variables and the high-precision wind speed distribution as a target variable.

[0273] The prediction unit 233 predicts a high-accuracy wind speed distribution using the learning model 242 generated by the first learning unit 232. Specifically, the prediction unit 233 inputs the placement conditions into the learning model 242 and predicts a high-accuracy wind speed distribution.

[0274] As a result, the placement determination device 220 can determine the optimal placement conditions for each placement condition, such that the server room SP approaches the desired wind speed distribution, without actually operating the air conditioners 210.

[0275] (5-5) Although the embodiments of the present disclosure have been described above, it will be understood that various changes in form and details can be made without departing from the spirit and scope of the present disclosure as defined in the claims. [Explanation of symbols]

[0276] 1,2 Air conditioning control system 10 Air conditioning equipment 19 Air conditioning controller (control unit) 31,131 Arithmetic unit 32,132 First Learning Section (Learning Section) 33,133 Forecasting Department 34 Generation part 35 Evaluation Department 36 Extraction part 41 NN model (learning model) 42,43 Learning Model 43a Learning Model (First Learning Model) 61~63 DN model (learning model) 138 Acquisition Department CA Cold Aisle (Aisle) DC Data Center HA Hot Aisle (Aisle) RK Server Rack SP server room (target space) [Prior art documents] [Patent documents]

[0277] [Patent Document 1] Japanese Patent Application Laid-Open No. 2010-15192

Claims

1. An air conditioning control system (1, 2) for conditioning a target space (SP1) using an air conditioning device (10), a generation unit (34) that generates candidates for the operating conditions when searching for operating conditions of the air conditioning device; a prediction unit (33, 133) that predicts the environmental state corresponding to the candidate operating conditions using a learning model (41 to 43) that predicts the environmental state of the target space based on the candidate operating conditions; an evaluation unit (35) that evaluates the candidate operating conditions using the environmental conditions corresponding to the candidate operating conditions; an extraction unit (36) that extracts, from the candidates for the operating condition generated by the generation unit, candidates for the operating condition for which the evaluation by the evaluation unit satisfies a predetermined condition; a control unit (19) that controls the air conditioning device based on the candidate operating conditions extracted by the extraction unit; a calculation unit (31, 131) that performs a CFD (Computational Fluid Dynamics) simulation of the environmental state and calculates the environmental state as a result of the CFD simulation; Equipped with the prediction unit predicts the environmental state as a result of the CFD simulation using the candidate operating conditions as input, or predicts the environmental state when the air conditioning apparatus is controlled by the candidate operating conditions using the learning model learned using the environmental state as a result of the CFD simulation; the operating conditions are air conditioning settings, The environmental state is a value of an environmental parameter at a plurality of locations including a first location and a second location in the target space. Air conditioning control system (1, 2).

2. a learning unit (32, 132) that generates the learning model; Further provided with An air conditioning control system (1, 2) according to claim 1.

3. the prediction unit uses the candidate operating conditions as input to the learning model and predicts the environmental state as a result of the CFD simulation using the candidate operating conditions as input. An air conditioning control system (1) according to claim 1 or 2.

4. The learning model is a model that has been learned using the operating conditions as explanatory variables and the environmental state as a result of the CFD simulation as a target variable. An air conditioning control system (1) according to claim 3.

5. the prediction unit predicts a first environmental state, which is the environmental state as a result of the CFD simulation using the candidate operating conditions as input, as an input to the learning model, and The second environmental state is simulated with less accuracy than the first environmental state. An air conditioning control system (1) according to claim 1 or 2.

6. the learning model is a model that has been learned using the second environmental state as an explanatory variable and the first environmental state as a target variable; An air conditioning control system (1) according to claim 5.

7. an acquisition unit (138) that acquires the value of the environmental parameter at the first location by actual measurement; Furthermore, the prediction unit predicts the value of the environmental parameter at the second location when the air conditioning apparatus is controlled by the candidate operating condition, based on the value of the environmental parameter at the first location acquired by the acquisition unit when the air conditioning apparatus is controlled by the candidate operating condition, using the learning model learned using the values ​​of the environmental parameter at the first location and the second location calculated by the calculation unit; An air conditioning control system (2) according to claim 1 or 2.

8. The learning model is a model that has been trained using values ​​of the environmental parameters at the first location acquired by the acquisition unit when the air conditioning device is controlled under the operating conditions as explanatory variables, and values ​​of the environmental parameters at the second location when the air conditioning device is controlled under the operating conditions as objective variables. An air conditioning control system (1) according to claim 7.

9. The learning model is the value of the environmental parameter at the first location acquired by the acquisition unit when the air conditioning apparatus is controlled according to the operating condition is used as an explanatory variable; a model that has been trained using the values ​​of the environmental parameters at the first location calculated by the calculation unit as explanatory variables and the values ​​of the environmental parameters at the first location acquired by the acquisition unit when the air conditioning device is controlled under the operating conditions as objective variables, predicted by inputting the values ​​of the environmental parameters at the second location calculated by the calculation unit into a first learning model (43a) that has been trained using the values ​​of the environmental parameters at the second location when the air conditioning device is controlled under the operating conditions as objective variables; An air conditioning control system (1) according to claim 8.

10. The generation unit generates the candidate operating conditions through reinforcement learning that values ​​the evaluation of the candidate operating conditions by the evaluation unit. An air conditioning control system (1, 2) according to claim 1 or 2.

11. The generation unit generates the candidate operating conditions using a reinforcement learning model (61 to 63) that determines the next candidate operating condition based on the current candidate operating condition. An air conditioning control system (1, 2) according to claim 10.

12. The target space includes an aisle (CA, HA) between server racks (RK) in a data center (DC), An air conditioning control system (1, 2) according to claim 1 or 2.

Images