ML model management device and ML model management method

Abstract

To appropriately determine the requirement of a relearning of an ML model.SOLUTION: In a management calculator 5000, a control plan generation part 5120 generates a first control plan based on a RE power generation amount predicted by an estimation using an ML model 5320 and a prediction value of a power consumption. An apparatus control part 5130 controls a control object apparatus by using the first control plan. Further, the control plan generation part 5120 generates a second control plan on the basis of the RE power generation amount or an actual value of the power consumption in the case where an evaluation that an accuracy of the estimation is deteriorated is made by a relearning requirement determination part 5140. The relearning requirement determination part 5140 determines whether or not a negative difference occurs against a second control result when the control object apparatus is controlled by using the second control plan in the first control result when controlling the control object apparatus by using the first control plan, and determines that a relearning of the ML model 5320 is required when it is determined that the negative difference occurs.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to an apparatus and method for managing an updatable machine learning (ML) model through ML. [Background technology]

[0002] In recent years, there has been an increasing demand for reducing greenhouse gas emissions to protect the global environment, and as a result, the use of electricity produced by renewable energy (RE) power generation, such as solar power generation and wind power generation, is increasing. However, since the amount of electricity generated by RE power generation fluctuates significantly depending on the weather, etc., in order to achieve a stable supply of electricity to electricity consumers, it is necessary to accurately predict the amount of RE power generation and the amount of electricity consumed by electricity consumers, and to control generators other than RE power generation and to control the charging and discharging of storage batteries that temporarily store surplus electricity.

[0003] Regarding prediction of RE power generation and power consumption, the technology described in Patent Document 1 is known. Patent Document 1 discloses a method for predicting the power generation and power consumption of various devices using a machine learning-based prediction model (ML model) and controlling the amount of power stored in a storage battery.

[0004] Furthermore, regarding re-learning of an ML model, a technology described in Patent Document 2 is known. Patent Document 2 discloses a method for evaluating a machine learning model and determining whether or not to re-learn the machine learning model based on the evaluation result. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] US Patent Application Publication No. 2021 / 0003974 [Patent Document 2] US Patent Application Publication No. 2016 / 0371601 Summary of the Invention [Problem to be solved by the invention]

[0006] In the technology described in Patent Document 1, by applying the technology of Patent Document 2, it is possible to re-learn the ML model used to predict the RE power generation amount and the power consumption amount, thereby improving the prediction accuracy. However, the technology of Patent Document 2 does not take into consideration the degree of influence on each device due to the deterioration of the prediction accuracy of the RE power generation amount and the power consumption amount, so there is a risk that the ML model will be re-learned at an unnecessary timing.

[0007] The present invention has been made in consideration of the above, and has an object to appropriately determine whether or not re-learning of an ML model is necessary. [Means for solving the problem]

[0008] The ML model management device according to the present invention includes a control plan generation unit that generates a first control plan for a controlled device based on a predicted value of a physical quantity predicted by inference using an ML model, an equipment control unit that controls the controlled device using the first control plan, and a re-learning necessity determination unit that determines whether or not re-learning of the ML model is necessary based on the predicted value and an actual measurement value of the physical quantity when the controlled device is controlled by the equipment control unit, wherein the re-learning necessity determination unit evaluates the accuracy of the inference based on the predicted value and the actual measurement value, and when the re-learning necessity determination unit evaluates that the accuracy of the inference has deteriorated, the control plan generation unit generates a second control plan for the controlled device based on the actual measurement value, and the re-learning necessity determination unit determines whether or not a negative difference occurs in a first control result when the controlled device is controlled using the first control plan compared to a second control result when the controlled device is controlled using the second control plan, and when it is determined that a negative difference occurs, determines that re-learning of the ML model is necessary. The ML model management method according to the present invention includes: A method using a computer, the method comprising: generating a first control plan for the controlled device based on a predicted value of a physical quantity predicted by inference using the ML model; The computer Controlling the control target device using the first control plan; The computer evaluating accuracy of the inference based on the predicted value and an actual measurement value of the physical quantity when the controlled device is controlled; The computer generating a second control plan for the control target device based on the actual measurement value when it is evaluated that the accuracy of the inference has deteriorated; The computer determining whether a first control result obtained by controlling the controlled device using the first control plan has a negative difference from a second control result obtained by controlling the controlled device using the second control plan; The computer If it is determined that a negative difference occurs, it is determined that re-learning of the ML model is necessary. Effect of the Invention

[0009] According to the present invention, it is possible to appropriately determine whether or not re-learning of an ML model is necessary. [Brief description of the drawings]

[0010] [Figure 1] 1 is an overall configuration diagram of a basic type of a computer system according to a first embodiment of the present invention; [Diagram 2] FIG. 2 is a diagram showing the overall configuration of a modified example of the computer system according to the first embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram showing the internal configuration of a management computer. [Figure 4] A diagram of the internal structure of the data store. [Diagram 5] A diagram of the ML model management table. [Figure 6] FIG. 4 is a diagram showing the configuration of a storage battery management table. [Figure 7] FIG. 13 is a diagram showing the configuration of a computer management table. [Figure 8] FIG. 13 is a diagram showing the configuration of an RE rate management table. [Figure 9] FIG. [Figure 10] FIG. [Figure 11] FIG. 4 is a diagram showing a flow of a process for determining whether or not re-learning is necessary according to the first embodiment of the present invention. [Figure 12] FIG. 4 is a diagram showing an example of a display screen of a re-learning necessity determination result according to the first embodiment of the present invention. [Figure 13] FIG. 11 is a diagram showing another example of the display screen of the re-learning necessity determination result according to the first embodiment of the present invention. [Figure 14] FIG. 11 is a diagram showing a flow of a process for determining whether or not re-learning is necessary according to a second embodiment of the present invention. [Figure 15] FIG. 13 is a diagram showing a flow of a re-learning necessity determination process according to the third embodiment of the present invention. [Figure 16] FIG. 13 is a diagram showing an example of a display screen for a re-learning necessity determination result according to the third embodiment of the present invention. [Figure 17] FIG. 13 is a diagram showing a flow of a re-learning necessity determination process according to a fourth embodiment of the present invention. [Figure 18] FIG. 13 is a diagram showing an example of a display screen for a re-learning necessity determination result according to the fourth embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] (First embodiment) A computer system according to a first embodiment of the present invention will be described below. In this embodiment, a system that manages an updatable ML model by machine learning, predicts the amount of power generated by renewable energy and the amount of power consumed by equipment by inference using the ML model, and aims to improve the utilization rate of renewable energy will be described.

[0012] Fig. 1 is a diagram showing the overall configuration of a basic form of a computer system 10 according to a first embodiment of the present invention. The computer system 10 shown in Fig. 1 is configured by connecting one or a plurality of computing environments 100 and a computing environment 200 having a management function for each computing environment 100 via a wide area network 400 such as the Internet. The computing environments 100 and 200 each correspond to, for example, a data center base.

[0013] The computing environment 100 has one or more computers 3000. Each computer 3000 is connected to a wide area network 400 via a management network 8000 such as a LAN (Local Area Network). The computers 3000 are information devices capable of executing any arithmetic processing, and correspond to power consuming devices that consume power to operate. The operation of the computers 3000 is controlled by a management computer 5000 possessed by the computing environment 200, and the computers 3000 in the same computing environment 100 can communicate with each other via a data network 7000 such as a LAN, and exchange various data used in the arithmetic processing. Each computing environment 100 may include, in addition to the computers 3000, devices that consume power, such as air conditioners and lighting equipment.

[0014] The computer 3000 is connected to one or more storage batteries 4100 via a charging / discharging device 4000. The charging / discharging device 4000 is connected to an RE power source 300 provided corresponding to each computing environment 100, and supplies the power supplied from the RE power source 300 to the computer 3000, and charges and stores the surplus power in the storage battery 4100. When the power supplied from the RE power source 300 is insufficient, the storage battery 4100 is discharged, and the power stored in the storage battery 4100 is supplied to the computer 3000. The storage battery 4100 corresponds to a power storage device capable of storing the surplus power for the computer 3000, and is charged and discharged by the charging / discharging device 4000, and the amount of stored power is managed by the management computer 5000. The charging / discharging device 4000 is connected to a management network 8000, and its operation is controlled by the management computer 5000 of the computing environment 200.

[0015] The RE power supply 300 generates power by RE power generation using renewable energy and supplies it to the computation environment 100. In the computation environment 100, the power from the RE power supply 300 is supplied to the computer 3000 via the charging / discharging device 4000 as described above, and the surplus power is stored in the storage battery 4100. The RE power supply 300 may be realized by directly purchasing RE-generated power from a power company, for example, or may be realized by private power generation using RE power generation equipment such as a solar power generation system installed in each computation environment 100. The RE power supply 300 may also be realized by purchasing an environmental value certificate (such as a green power certificate) issued by a power generation company that generates RE power, and thereby substantially procuring RE-generated power from a power company.

[0016] The computing environment 200 includes a management computer 5000 and a data store 6000. The management computer 5000 is connected to the wide area network 400 via a management network 8000. As a result, the management computer 5000 and the computers 3000 and the charging / discharging devices 4000 of each computing environment 100 are connected to each other, and the management computer 5000 controls the operation of each computer 3000 and each charging / discharging device 4000 to manage the power consumption of each computer 3000 and the amount of stored power of each storage battery 4100. The data store 6000 is a storage device capable of storing and holding various data, and is connected to the wide area network 400 via the management network 8000, similar to the management computer 5000. In the computing environment 200, the data collected by the management computer 5000 from the computers 3000 and the charging / discharging devices 4000 of each computing environment 100 is stored and held in the data store 6000. Furthermore, the management computer 5000 and the data store 6000 may be configured as an integrated computer.

[0017] The computer 3000 and the management computer 5000 can be configured using any information device. For example, the computer 3000 and the management computer 5000 may be realized using a server or the like, or information devices other than a server, such as a data storage device or a network switch, may be used as the computer 3000 and the management computer 5000. Furthermore, the computer 3000 and the management computer 5000 are not limited to physical computers (bare metal), but may be virtual computers (VMs, containers, etc.), or may be realized using a cloud service. The cloud service may be provided using a server, or may be a serverless cloud service. In addition to this, the computer system 10 of this embodiment can be realized by using any information device or information service as the computer 3000 and the management computer 5000.

[0018] Fig. 2 is an overall configuration diagram of a modified example of the computer system 10 according to the first embodiment of the present invention. The computer system 10 shown in Fig. 2 is operated in an on-premise environment without being connected to a wide area network 400, and differs from that described in Fig. 1 in that the computing environment 200 does not exist, and one computing environment 100 has a computer 3000, a charging / discharging device 4000, and a storage battery 4100, as well as a management computer 5000 and a data store 6000.

[0019] 2, a management computer 5000 and a data store 6000 are connected to each of the computers 3000 and the charging / discharging devices 4000 in the same computing environment 100 via a management network 8000. This enables the management of the computers 3000 and the storage battery 4100 to be performed from the management computer 5000, and management data for the computer 3000 and the storage battery 4100 collected by the management computer 5000 from the computers 3000 and the charging / discharging devices 4000 is stored and held in the data store 6000.

[0020] Next, the internal configuration of the management computer 5000 and the data store 6000 will be explained below with reference to FIG. 3 and FIG.

[0021] 3 is a diagram showing the internal configuration of the management computer 5000. The management computer 5000 comprises a processor 5100, a memory 5200, a storage device 5300, a management network interface 5400, and an I / O device 5500.

[0022] The processor 5100 executes various programs deployed in the memory 5200 to function as an inference unit 5110, a control plan generating unit 5120, an equipment control unit 5130, a relearning necessity determining unit 5140, and a learning unit 5150. The memory 5200 is used as a working area for the processor 5100, and stores therein a data collection program 5210, a learning program 5220, an inference program 5225, a battery charge / discharge planning and control program 5230, an IT workload planning and control program 5240, a relearning necessity determining program 5250, and an ML model management table 5260. The storage device 5300 is used as an auxiliary storage device for the processor 5100, and stores therein learning data 5310, an ML model 5320, and inference result data 5330. Some or all of the programs and data stored in the memory 5200 may be stored in the storage device 5300, and the processor 5100 may read them from the storage device 5300 and load them in the memory 5200 as necessary.

[0023] The management network interface 5400 performs communication interface processing with the management network 8000. The management computer 5000 communicates with the computer 3000 and the charging / discharging device 4000 through the communication interface processing of the management network interface 5400, and can control these to manage the computer 3000 and the storage battery 4100. The I / O device 5500 is a device that provides information to a user and receives operational input from the user, and is realized by combining, for example, a display device such as a liquid crystal display, an input device such as a keyboard or a mouse, or a touch panel that serves both as a display device and an input device.

[0024] The inference unit 5110 performs inference using the ML model 5320 to predict physical quantities related to power consumption in each computing environment 100. Specifically, for each computing environment 100, the amount of power generated by the RE power source 300 (hereinafter referred to as "RE power generation amount") and the amount of power consumed by the computer 3000 are predicted as physical quantities related to power consumption. The prediction results of these physical quantities by the inference unit 51110 are stored in the storage device 5300 as inference result data 5330. Note that the amount of power consumption predicted by the inference unit 5110 may include the amount of power consumed by power consumption devices other than the computer 3000, such as air conditioners and lighting.

[0025] The control plan generating unit 5120 generates a control plan for the management computer 5000 to control the computer 3000 and the charging / discharging device 4000 based on the inference result data 5330 obtained by the inference unit 5110 .

[0026] The device control unit 5130 controls the computer 3000 and the charging / discharging device 4000 using the control plan generated by the control plan generating unit 5120. As a result, the amount of power consumed by the computer 3000 is adjusted, and the amount of stored power in the storage battery 4100 is adjusted.

[0027] The relearning necessity determination unit 5140 compares the predicted values ​​of the RE power generation amount and the power consumption amount by the inference unit 5110 with the actual measured values ​​of the RE power generation amount and the power consumption amount when the computer 3000 and the charging / discharging device 4000 are controlled using a control plan based on these predicted values, and evaluates the accuracy of the inference performed by the inference unit 5110 based on the comparison result. If it is evaluated as a result that the accuracy of the inference has deteriorated, it determines whether or not relearning of the ML model 5320 used for the inference of the inference unit 5110 is required. Details of a method for determining whether or not relearning of the ML model 5320 by the relearning necessity determination unit 5140 is required will be described later.

[0028] When the re-learning necessity determination unit 5140 determines that re-learning of the ML model 5320 is necessary, the learning unit 5150 re-learns the ML model 5320 by using the learning data 5310. This updates the ML model 5320, and the inference accuracy by the inference unit 5110 is improved.

[0029] The data collection program 5210 is a program for acquiring data on the operating state from the computer 3000 and the charging / discharging device 4000 of each computing environment 100, and storing the data in the data store 6000. The data collected by the data collection program 5210 includes, for example, data on the operating state and power consumption of the computer 3000, the amount of RE power generation input to the charging / discharging device 4000, and the amount of power stored in the storage battery 4100 that is charged and discharged by the charging / discharging device 4000. The collected data is converted into a data format suitable for each table in Figs. 6 to 8, which will be described later, and then stored in the data store 6000. The management computer 5000 can collect data from the computer 3000 and the charging / discharging device 4000 at regular time intervals and store the data in the data store 6000 by periodically executing the data collection program 5210.

[0030] The learning program 5220 is a program for learning or relearning the ML model 5320, and causes the processor 5100 to function as a learning unit 5150. The learning program 5220 may use a known learning algorithm, such as Random Forest. The learning program 5220 is executed by the processor 5100 when the relearning necessity determination unit 5140 determines that the ML model 5320 needs to be relearned, as described above. Alternatively, the learning program 5220 may be manually executed at any timing by a user performing a predetermined operation using an I / O device 5500 in the management computer 5000.

[0031] The inference program 5225 is a program for performing inference using the ML model 5320 and obtaining predicted values ​​of the RE power generation amount and the power consumption amount of the computer 3000 as the inference result, and causes the processor 5100 to function as the inference unit 5110. The inference program 5225 is periodically executed by the processor 5100 during a predetermined control period, for example, every hour, to obtain predicted values ​​of the RE power generation amount and the power consumption amount in the next control period. In addition, for example, the inference program 5225 may be executed manually at any timing by a user performing a predetermined operation using the I / O device 5500 in the management computer 5000. The ML model 5320 includes a model for predicting the RE power generation amount and a model for predicting the power consumption amount of the computer 3000, and the inference program 5225 can obtain predicted values ​​of the RE power generation amount and the power consumption amount by using these models. As described above, when the inference unit 5110 predicts the power consumption amount of power consumption devices other than the computer 3000, such as air conditioners and lighting, the ML model 5320 may include models corresponding to these devices.

[0032] The battery charge / discharge planning and control program 5230 is a program for generating a control plan for charging / discharging each storage battery 4100 by the charge / discharge device 4000 and controlling the operation of the charge / discharge device 4000 according to the control plan. The IT workload planning and control program 5240 is a program for generating a control plan for each computer 3000 and controlling the operation of each computer 3000 according to the control plan. These programs are called by the inference program 5225 and executed periodically by the processor 5100 executing the inference program 5225, and generate a control plan for the next control period. After that, when the time set in the control plan arrives, the control of the charge / discharge device 4000 and the computer 3000 are executed according to the control plan, and the amount of stored power in the storage battery 4100 and the amount of power consumed by the computer 3000 are adjusted. The processor 5100 executes the battery charge / discharge planning and control program 5230 and the IT workload planning and control program 5240 , so that the processor 5100 can function as a control plan generating unit 5120 and an equipment control unit 5130 .

[0033] The relearning necessity determination program 5250 is a program for determining whether or not relearning of the ML model 5320 is necessary, and causes the processor 5100 to function as a relearning necessity determination unit 5140. The relearning necessity determination program 5250 is periodically executed by the processor 5100 at a predetermined control period, for example, every hour, after the inference program 5225 is executed. As a result, whether or not relearning of the ML model 5320 is necessary is periodically determined, and if it is determined that relearning is necessary, the learning program 5220 is called and executed by the processor 5100. As a result, relearning of the ML model 5320 is performed.

[0034] The ML model management table 5260 is a table that stores data used for managing the ML model 5320. The management computer 5000 can manage the learning status of each model constituting the ML model 5320, the correspondence between inferences performed using each model and controlled devices, and the like, by referring to the ML model management table 5260. Note that when re-learning of the ML model 5320 is performed, the contents of the ML model management table 5260 are updated.

[0035] 4 is a diagram showing the internal configuration of a data store 6000. The data store 6000 includes a processor 6100, a memory 6200, a storage device 6300, a management network interface 6400, and an I / O device 6500.

[0036] The processor 6100 uses the memory 6200 as a working area to execute various programs. A data input / output program 6210 is stored in the memory 6200. A part or all of the data input / output program 6210 may be stored in the storage device 6300, and the processor 6100 may read it from the storage device 6300 and load it in the memory 6200 as necessary. The storage device 6300 is used as an auxiliary storage device for the processor 6100, and stores a storage battery management table 6310, a computer management table 6320, and a RE rate management table 6330. The storage battery management table 6310, the computer management table 6320, and the RE rate management table 6330 may each be held in the memory 6200.

[0037] The management network interface 6400 and the I / O device 6500 are similar to the management network interface 5400 and the I / O device 5500, respectively, of the management computer 5000 in FIG.

[0038] The data input / output program 6210 is a program for inputting and outputting data to and from the management computer 5000 via the management network 8000. For example, the data collected from the computers 3000 and the charging / discharging devices 4000 of each computing environment 100 by executing the data collection program 5210 in the management computer 5000 is input from the management computer 5000 to the data store 6000 by the data input / output program 6210, and the data is appropriately converted and stored in the storage battery management table 6310, the computer management table 6320, and the RE rate management table 6330. The data stored in each of these tables is read from the storage device 6300 by the data input / output program 6210 and output to the management computer 5000. A known data input / output method (protocol) can be used for the data input / output method (protocol) in the data input / output program 6210. For example, the data input / output program 6210 can be realized using APIs such as REST (Representational State Transfer) and GraphQL, and data transmission / reception protocols such as NFS (Network File System), iSCSI (Internet Small Computer System Interface), and FTP (File Transfer Protocol).

[0039] The storage battery management table 6310 is a table for managing the storage batteries 4100 in each computing environment 100, and stores data related to the performance and power storage state of each storage battery 4100. The computer management table 6320 is a table for managing the computers 3000 in each computing environment 100, and stores data related to the operating state and power consumption of each computer 3000. The RE rate management table 6330 is a table for managing the utilization rate of renewable energy (RE rate) in each computing environment 100.

[0040] Next, the ML model management table 5260, the storage battery management table 6310, the computer management table 6320, and the RE rate management table 6330 will be described in detail.

[0041] 5 is a diagram showing the configuration of an ML model management table 5260. In the ML model management table 5260 held in the memory 5200 of the management computer 5000, records are set corresponding to each model of the ML model 5320. Each record in the ML model management table 5260 has data such as an ML model ID 5261, an ML model version 5262, a learning data Uniform Resource Identifier (URI) 5263, an inference target ID 5264, a computing environment ID 5265, re-learning power consumption 5266, and a re-learning cost 5267.

[0042] The ID of the ML model is stored in the ML model ID 5261. Any information may be used as the ID as long as it can uniquely identify the ML model.

[0043] A number indicating the version of the ML model corresponding to the record is stored in the ML model version 5262. The value of the ML model version 5262 is incremented every time training (retraining) of the ML model is performed.

[0044] In the ML model management table 5260 of this embodiment, the combination of the ML model ID 5261 and the ML model version 5262 makes it possible to uniquely identify the ML model corresponding to the record.

[0045] Information for uniquely identifying the training data 5310 used in training the ML model corresponding to the record is stored in the training data URI 5263. For example, when data made public by the data store 6000 is used as the training data 5310, the URI of the made-public data is stored in the training data URI 5263.

[0046] The inference target ID 5264 stores the ID of a device (computer 3000 or storage battery 4000) related to the physical quantity that is the target of inference by the ML model corresponding to the record. For example, in the case of an ML model used to predict the amount of RE power generation, the ID of the storage battery 4100 that stores the amount of RE power generation is stored here, and in the case of an ML model used to predict the amount of power consumption, the ID of the computer 3000 that consumes the amount of power consumption is stored here.

[0047] The computation environment ID 5265 stores the ID of the computation environment 100 related to the physical quantity to be inferred by the ML model corresponding to the record. In the case of an ML model used to predict the RE power generation amount, the ID of the computation environment 100 to which the RE power generation amount is supplied from the RE power source 300 is stored, and in the case of an ML model used to predict the power consumption amount, the ID of the computation environment 100 including the computer 3000 that consumes the power consumption amount is stored.

[0048] Data indicating the power consumption [W] and monetary cost [yen] required for re-learning the ML model corresponding to the record are stored in the re-learning power consumption 5266 and the re-learning cost 5267. Note that these data are required in the third and fourth embodiments described later, but are not essential elements in the ML model management table 5260 in this embodiment, and may be omitted.

[0049] 6 is a diagram showing the configuration of a storage battery management table 6310. In the storage battery management table 6310 held in the storage device 6300 of the data store 6000, records are set corresponding to each of the storage batteries 4100 of each computing environment 100. Each record in the storage battery management table 6310 has data such as a storage battery ID 6311, a maximum capacity 6312, a capacity 6313, and an excess storage amount increment 6314.

[0050] The ID of the storage battery corresponding to the record is stored in the storage battery ID 6311. Here, an ID common to the ID stored in the inference target ID 5264 of any record in the ML model management table 5260 is stored. In this way, each record of the ML model used for inferring the RE power generation amount in the ML model management table 5260 is linked to each record in the storage battery management table 6310.

[0051] The maximum capacity of the storage battery corresponding to the record is stored in the maximum capacity 6312. A fixed value determined in advance for each storage battery is stored here as the maximum capacity.

[0052] The capacity (amount of stored electricity) of the storage battery corresponding to the record for each time is stored in the capacity 6313. For example, the average amount of stored electricity of each storage battery for each hour calculated based on data collected from the charging / discharging device 4000 at a predetermined time interval is stored here in combination with time information.

[0053] The surplus increment of stored power amount 6314 stores the surplus increment of stored power amount according to the decrease in the RE rate in the storage battery corresponding to the record. In the computer system 10 of this embodiment, if the estimation accuracy of the RE power generation amount and the power consumption amount by the ML model is low, the power generated by normal power generation (thermal power generation, etc.) other than RE power generation is purchased from the power company and supplied to the computer 3000, and the power exceeds the original required amount, and the RE rate of the power supplied to the computer 3000 decreases while the stored power amount of the storage battery 4100 increases. In such a case, the computer system 10 of this embodiment calculates the increased stored power amount for each storage battery as the surplus increment and records it in the storage battery management table 6310 as the surplus increment of stored power amount 6314. Note that the surplus increment of stored power amount 6314 is necessary in a third embodiment described later, but is not an essential element in the storage battery management table 6310 in this embodiment, and may be omitted.

[0054] The units of the above maximum capacity 6312, capacity 6313, and surplus storage increment 6314 may be, for example, the amount of power (kWh) per hour, or the amount of current (mAh) per hour. Alternatively, they may be expressed using other physical quantities.

[0055] 7 is a diagram showing the configuration of a computer management table 6320. In the computer management table 6320 held in the storage device 6300 of the data store 6000, records are set corresponding to each of the computers 3000 included in each computing environment 100. Each record in the computer management table 6320 has data such as a computer ID 6321 and metrics 6322, for example.

[0056] The ID of the computer corresponding to the record is stored in the computer ID 6321. An ID common to the ID stored in the inference target ID 5264 of any record in the ML model management table 5260 is stored here. This allows each record of the ML model used for inferring the power consumption amount in the ML model management table 5260 to be linked to each record in the computer management table 6320.

[0057] Metrics 6322 stores information indicating the operating state of the computer corresponding to the record for each unit time. For example, information such as hourly CPU utilization, RAM (memory) usage, IOPS, and power consumption calculated based on data collected from each computer at a specified time interval is stored here in combination with time information. Furthermore, metrics 6322 may include information such as data transfer speed, network bandwidth usage, and various metrics of applications and middleware running on the computer.

[0058] 8 is a diagram showing the configuration of the RE rate management table 6330. In the RE rate management table 6330 held in the storage device 6300 of the data store 6000, records are set corresponding to each computing environment 100. Each record in the RE rate management table 6330 has data such as a computing environment ID 6331, an RE rate 6332, and an RE shortage amount 6333, for example.

[0059] The computation environment ID 6331 stores the ID of the computation environment corresponding to the record. The ID stored here is the same as the ID stored in the computation environment ID 5265 of any record in the ML model management table 5260. This links each record of the ML model used for estimating the RE power generation amount and power consumption amount in the ML model management table 5260 with each record in the RE rate management table 6330. However, as in the configuration of FIG. 2, when a management computer 5000 exists for each computation environment, the management computer 5000 only needs to manage the RE rate of that computation environment. Therefore, the ML model management table 5260 and the RE rate management table 6330 do not need to include the computation environment IDs 5265 and 6331, respectively. On the other hand, as in the configuration of FIG. 1, when one management computer 5000 manages multiple computation environments, the management computer 5000 needs to manage the RE rate of each computation environment individually. Therefore, in the ML model management table 5260 and the RE rate management table 6330, the computation environment IDs 5265 and 6331 are required, respectively.

[0060] The RE rate 6332 stores the RE rate for each unit time of the computation environment corresponding to the record. As described above, the RE rate indicates the utilization rate of renewable energy in the computation environment, and is calculated as the percentage [%] of RE power generation in the total power consumption of the computation environment. The total power consumption of the computation environment is calculated as the total value of the power consumption of all the computers 3000 included in the computation environment. However, if the computation environment includes other power consuming devices such as air conditioners and lighting, the power consumption of those devices may also be added together to calculate the total power consumption.

[0061] The RE shortage 6333 stores the shortage of the RE power generation amount with respect to the total power consumption of the computing environment corresponding to the record. For example, when the total power consumption of the computing environment cannot be covered by the RE power generation amount alone and the shortage is covered by purchasing power from a power company by normal power generation, the purchased amount of power [kWh] is recorded as the RE shortage 6333 in the RE rate management table 6330. Note that the RE shortage 6333 is necessary in a fourth embodiment described later, but is not an essential element in the RE rate management table 6330 in this embodiment, and may be omitted.

[0062] Next, the operations of the data collection program 5210, the inference program 5225, and the relearning necessity determination program 5250 will be described.

[0063] 9 is a diagram showing the flow of data collection processing by the data collection program 5210. In the management computer 5000, the processor 5100 periodically executes processing by the data collection program 5210 in accordance with a data collection processing flow 10000 in FIG.

[0064] When the data collection process flow 10000 is started in step 10010, in step 10020, the management computer 5000 acquires metrics data of devices that are managed by the management computer 5000. Here, the metrics data of the computers 3000 and the storage batteries 4100 of each computing environment 100 are acquired as devices to be managed by the management computer 5000. Specifically, information such as the amount of stored power of each storage battery 4100 is acquired from the charging / discharging device 4000 as metrics data of the storage battery 4100, and information on the operating status and power consumption of each computer 3000 is acquired as metrics data of the computer 3000. In addition to the metrics data of the computers 3000 and the storage batteries 4100, metrics data of the computing environment itself, such as the room temperature inside the computing environment 100 and the outside temperature, may be acquired.

[0065] In step 10030, the metrics data acquired in step 10020 is stored in a management table. Here, the metrics data acquired from the management computer 5000 is output to the data store 6000, and the data store 6000 is controlled so that the metrics data of the storage battery 4100 is stored in each record of the storage battery management table 6310, and the metrics data of the computer 3000 is stored in each record of the computer management table 6320. At this time, it is preferable to process the metrics data as necessary and convert it into a data format suitable for these management tables.

[0066] In step 10040, it is determined whether or not there are any managed devices for which metrics data has not been acquired in step 10020 in each computing environment 100 connected to the management computer 5000 via the management network 8000. If the result shows that there are still managed devices for which data has not been acquired, the process returns to step 10020 to continue acquiring metrics data, and if metrics data has been acquired for all managed devices, the process proceeds to step 10050.

[0067] In step 10050, the amount of RE power generation in each computation environment 100 is acquired. Here, the amount of RE power generation in each computation environment 100 is acquired by acquiring the amount of power (RE power generation amount) supplied from the RE power source 300 to the charging / discharging device 4000 in each computation environment 100. Information on the amount of RE power generation may be acquired from the charging / discharging device 4000 or from other devices. For example, when RE-generated power is purchased from a power company and used as the RE power source 300, the amount of RE power generation can be acquired by communication from the power company's system. Also, when an RE power generation facility such as a solar power generation system is used as the RE power source 300, the amount of RE power generation can be acquired from its monitoring system (DEMS: Data center Energy Management System). Furthermore, other information that affects the amount of RE power generation, such as information on the amount of solar radiation in a solar power generation system, may be acquired in addition to the amount of RE power generation.

[0068] In step 10060, the RE rate of each computing environment 100 is calculated based on the RE power generation amount acquired in step 10050. Then, the RE rate calculation result is output from the management computer 5000 to the data store 6000, and the data store 6000 is controlled to store the RE rate in each record of the RE rate management table 6330, thereby storing the RE rate of each computing environment 100 in the management table.

[0069] After the process of step 10060 is completed, the data collection process flow 10000 ends in step 10070.

[0070] In the data collection process flow 10000 of Fig. 9, data other than those described above may be collected and stored in the management table. The type of data to be acquired and stored can be determined arbitrarily depending on the type of data to be used for learning the ML model. In other words, the data input / output program 6210 can execute the data collection process flow 10000 according to any procedure as long as it can acquire data necessary for learning the ML model and evaluating the accuracy.

[0071] Fig. 10 is a diagram showing an inference processing flow by the inference program 5225. In the management computer 5000, the processor 5100 periodically executes processing according to the inference processing flow 15000 of Fig. 10 by the inference program 5225. Note that, when the inference processing flow 15000 is executed, it is assumed that the learning processing of the ML model 5320 has been executed in advance by the learning program 5220, and the learned ML model 5320 has been stored in the storage device 5300.

[0072] When inference processing flow 15000 is started in step 15010, in step 15020, the ML model management table 5260 is referenced to obtain an ML model corresponding to the physical quantity that is the current target of inference. Here, by selecting one of the unselected records in the ML model management table 5260 and referring to the values ​​of the ML model ID 5261 and ML model version 5262 of that record, one of the models can be selected from the ML models 5320 stored in the storage device 5300 and obtained as the ML model corresponding to the physical quantity that is the current target of inference.

[0073] In step 15030, inference is executed using the ML model acquired in step 15020, and the obtained predicted value is output as inference result data. Through this process, a predicted value of the RE power generation amount or the power consumption amount of the computer 3000 corresponding to the acquired ML model is obtained, and output as inference result data. The inference result data output in step 15030 is stored as inference result data 5330 in the storage device 5300.

[0074] In step 15040, it is determined whether or not there is an ML model for which inference in step 15030 has not been performed among the ML models 5320 stored in the storage device 5300. If an unselected record exists in the ML model management table 5260 in step 15020 and inference has not yet been performed for the ML model corresponding to that record, the process returns to step 15020 to continue selection of an ML model and inference, and if inference has been performed for all ML models, the process proceeds to step 15050.

[0075] In step 15050, the storage battery charge / discharge plan / control program 5230 is called, and a first charge / discharge plan is generated for each storage battery 4100 using the predicted value represented by the inference result data output in step 15030. Here, a target value of the amount of stored power of each storage battery 4100 is determined based on the predicted value of the amount of RE power generation by the RE power source 300 and the predicted value of the amount of power consumption by the computer 3000 so that the RE rate in each computing environment 100 is as high as possible. Then, a charge / discharge plan is generated for each storage battery 4100 so that the amount of stored power of each storage battery 4100 becomes the target value, and this is set as the first charge / discharge plan.

[0076] In step 15060, the IT workload planning and control program 5240 is called, and a first IT workload plan is generated for each computer 3000 using the predicted value represented by the inference result data output in step 15030. Here, a workload plan for each computer 3000 is generated based on the predicted value of the RE power generation amount by the RE power source 300 and the predicted value of the power consumption amount of the computer 3000 so that the RE rate in each computing environment 100 is as high as possible, and this is set as the first IT workload plan.

[0077] After the processing of step 15060 is completed, the inference processing flow 15000 ends in step 15070.

[0078] After the inference process flow 15000 in FIG. 10 is executed, the battery charge / discharge planning and control program 5230 and the IT workload planning and control program 5240 control the operation of the charge / discharge device 4000 and the computer 3000 based on the first charge / discharge plan and the first IT workload plan generated in steps 15050 and 15060, respectively. Specifically, for example, the battery charge / discharge planning and control program 5230 switches the operation state of the charge / discharge device 4000 in each computing environment 100 at predetermined time intervals according to the first charge / discharge plan. As a result, the optimal operation mode for each computing environment 100 is selected and switched at predetermined timing from among an RE charge mode in which the RE power source 300 charges the storage battery 4100 to increase the amount of stored power, a discharge mode in which the storage battery 4100 discharges power to the computer 3000, and a pass-through mode in which power is supplied from the RE power source 300 to the computer 3000 via the charge / discharge device 4000. On the other hand, for example, the IT workload planning and control program 5240 determines the load distribution of the computers 3000 in each computing environment 100 according to the first IT workload plan, and controls the execution, suspension, migration, resumption, etc. of the computational processing by each computer 3000 accordingly. In this way, the management computer 5000 can control the charging / discharging device 4000 and the computers 3000, which are the devices to be controlled.

[0079] In the inference process flow 15000 of FIG. 10, only one of the process of step 15050 or the process of step 15060 may be executed, and the other may not be executed. When the process of step 15050 is executed but the process of step 15060 is not executed, the charging / discharging device 4000 may be controlled according to the first charging / discharging plan generated in step 15050, and the computer 3000 may be controlled without using the prediction result using the ML model. Similarly, when the process of step 15060 is executed but the process of step 15050 is not executed, the computer 3000 may be controlled according to the first IT workload plan generated in step 15060, and the charging / discharging device 4000 may be controlled without using the prediction result using the ML model.

[0080] 11 is a diagram showing a relearning necessity determination process flow according to the relearning necessity determination program 5250 according to the first embodiment of the present invention. In the management computer 5000 of this embodiment, the processor 5100 periodically executes processing according to the relearning necessity determination program 5250 in accordance with the relearning necessity determination process flow 20000 in FIG.

[0081] When the re-learning necessity determination process flow 20000 is started in step 20010, in step 20020, the predicted values ​​of the RE power generation amount and the power consumption amount of the computer 3000 in each computing environment 100 are obtained from the previous inference result data. Here, in the inference process flow 15000 of FIG. 10 executed in the previous control period, the predicted values ​​of the RE power generation amount and the power consumption amount represented by the inference result data 5330 output in step 15030 and stored in the storage device 5300 are obtained. As a result, the predicted values ​​of the RE power generation amount and the power consumption amount used in generating the control plan (first charge / discharge plan, first IT workload plan) used for controlling each controlled device (charge / discharge device 4000, computer 3000) in the current control period can be obtained from the previous inference result data.

[0082] In step 20030, actual measurements of the amount of power generated by RE and the amount of power consumed by the computer 3000 in each computing environment 100 are obtained. Here, the amount of power generated by RE and the amount of power consumed by the computer 3000 obtained in steps 10050 and 10020, respectively, in the data collection process flow 10000 of FIG. 9 executed during the current control period can be obtained as these actual measurements.

[0083] In step 20040, the accuracy of inference by the ML model 5320 is calculated based on the predicted value acquired in step 20020 and the actual value acquired in step 20030. For example, for each model included in the ML model 5320, the difference between the actual value and the predicted value of the corresponding RE power generation amount or power consumption is calculated, and the accuracy of inference for each ML model can be calculated from the difference. Specifically, if the actual value is Vr and the predicted value is Ve, the inference accuracy Q [%] for each ML model can be calculated, for example, by the following formula (1). Q=100×(Vr-|Vr-Ve|) / Vr ··(1)

[0084] In step 20050, the inference accuracy of each ML model calculated in step 20040 is evaluated. Specifically, for example, the inference accuracy Q calculated by the above formula (1) is compared with a predetermined threshold (for example, 80%), and if Q is equal to or less than the threshold, it is evaluated that the inference accuracy of the ML model has decreased. On the other hand, if Q is greater than the threshold, it is evaluated that the inference accuracy of the ML model has not decreased and is sufficient. In step 20050, such a determination process is performed for each of the multiple ML models stored in the storage device 5300 as the ML model 5320, so that an evaluation result of the inference accuracy for each ML model can be obtained. As a result, if it is evaluated that the inference accuracy of at least one ML model has decreased, the process proceeds to step 20060, and if it is evaluated that the inference accuracy of all ML models has not decreased, the process proceeds to step 20100.

[0085] In step 20060, the storage battery charge / discharge plan / control program 5230 is called, and a second charge / discharge plan is generated for each storage battery 4100 using the actual measured values ​​of the RE power generation amount and the power consumption amount of the computer 3000 acquired in step 20030. Here, similar to step 15050 in the inference process flow 15000 of FIG. 10, a target value of the power storage amount of each storage battery 4100 is determined based on the actual measured values ​​of the RE power generation amount and the power consumption amount of the computer 3000 so that the RE rate in each computing environment 100 is as high as possible. Then, a charge / discharge plan is generated for each storage battery 4100 so that the power storage amount of each storage battery 4100 becomes the target value, and this is set as the second charge / discharge plan.

[0086] In step 20070, the IT workload planning and control program 5240 is called, and a second IT workload plan is generated for each computer 3000 using the actual measured values ​​of the RE power generation amount and the power consumption amount of the computer 3000 acquired in step 20030. Here, similar to step 15060 in the inference process flow 15000 of Fig. 10, a workload plan for each computer 3000 is generated based on the actual measured values ​​of the RE power generation amount and the actual measured values ​​of the power consumption amount of the computer 3000 so that the RE rate in each computing environment 100 is as high as possible, and this is set as the second IT workload plan.

[0087] As described above, in the inference processing flow 15000 of FIG. 10, only one of the processing of step 15050 or the processing of step 15060 is executed, and when one of the first charge / discharge plan or the first IT workload plan exists but the other does not exist, only one of the processing of steps 20060 and 20070 may be executed. That is, when the first charge / discharge plan exists but the first IT workload plan does not exist, the second charge / discharge plan may be generated in step 20060, but step 20070 may not be executed. Also, when the first IT workload plan exists but the first charge / discharge plan does not exist, the second IT workload plan may be generated in step 20070, but step 20060 may not be executed.

[0088] In step 20080, it is determined whether or not there is a negative difference between the first charge / discharge plan or the first IT workload plan used to control the charge / discharge device 4000 or the computer 3000 in the previous control period and the second charge / discharge plan generated in step 20060 or the second IT workload plan generated in step 20070. The negative difference here means that the value of the stored power amount or the RE rate of the storage battery 4100 when the charge / discharge device 4000 or the computer 3000 was actually controlled according to the first charge / discharge plan or the first IT workload plan in the previous control period is reduced compared to the value of the stored power amount or the RE rate of the storage battery 4100 when it is assumed that the charge / discharge device 4000 or the computer 3000 is controlled according to the second charge / discharge plan or the second IT workload plan.

[0089] Specifically, the storage battery management table 6310 and the RE rate management table 6330 are referenced to acquire the amount of stored power and the RE rate when the first charge / discharge plan and the first IT workload plan are implemented, and the storage amount and the RE rate of the storage battery 4100 when these are implemented are acquired from the target value of the amount of stored power when the second charge / discharge plan is generated and the target value of the RE rate when the second IT workload plan is generated. Then, these values ​​are compared, and as a result of controlling the charging / discharging device 4000 according to the first charge / discharge plan and charging / discharging the storage battery 4100, if the amount of stored power or the RE rate of the storage battery 4100 decreases compared to when the charging / discharging device 4000 is controlled according to the second charge / discharge plan and the storage battery 4100 is charged / discharged, it is determined that there is a negative difference, and the process proceeds to step 20090. On the other hand, if neither the amount of stored power nor the RE rate of the storage battery 4100 decreases, it is determined that there is no negative difference, and the process proceeds to step 20100. Furthermore, if the result of controlling the computer 3000 according to the first IT workload plan is that the amount of stored power in the storage battery 4100 or the RE rate is reduced compared to when the computer 3000 is controlled according to the second IT workload plan, it is determined that there is a negative difference, and the process proceeds to step 20090. On the other hand, if neither the amount of stored power in the storage battery 4100 nor the RE rate is reduced, it is determined that there is no negative difference, and the process proceeds to step 20100.

[0090] In step 20090, it is determined that re-learning is necessary for each ML model that was evaluated in step 20050 as having a decline in inference accuracy.

[0091] In step 20100, it is determined that re-learning is not necessary for all ML models.

[0092] Once the process of either step 20090 or 20100 has been completed, the re-learning necessity determination process flow 20000 ends in step 20110.

[0093] After execution of the re-learning necessity determination process flow 20000 in FIG. 11, the learning program 5220 performs re-learning of the ML model determined to require re-learning in step 20090 at any timing, and updates the contents of the ML model 5320.

[0094] Fig. 12 is a diagram showing an example of a display screen of the relearning necessity determination result displayed after execution of the relearning necessity determination program 5250 according to the first embodiment of the present invention. A display screen 5250A shown in Fig. 12 is a screen for explaining the determination result of whether or not relearning is necessary and the basis for it, and is displayed on a display device included in an I / O device 5500 by, for example, the processor 5100 in the management computer 5000 executing the relearning necessity determination program 5250.

[0095] Display screen 5250A includes display frames 5250A1, 5250A2, 5250A3, and 5250A4, a confirmation button 5250A5, and a re-learning execution button 5250A6.

[0096] Display frame 5250A1 displays information indicating which ML model the determination result of whether or not relearning is necessary currently displayed on display screen 5250A applies to. Specifically, display frame 5250A1 displays the managed device corresponding to the physical quantity that is the subject of inference by the ML model, the ID and version of the ML model, and the inference accuracy of the ML model. This information can be acquired from ML model management table 5260 or the processing result of step 20040 in relearning necessity determination processing flow 20000 in FIG. 11.

[0097] The display frame 5250A2 shows the time change of the decrease in the amount of stored power of the storage battery 4100 caused by the difference between the predicted value and the actual value based on the inference result of the ML model shown in the display frame 5250A1. That is, the difference between the amount of stored power of the storage battery 4100 obtained as a result of actually controlling the charging / discharging device 4000 and the computer 3000 according to the first charging / discharging plan and the first IT workload plan generated in steps 15050 and 15060, respectively, in the inference processing flow 15000 of Fig. 10 and the amount of stored power of the storage battery 4100 assumed when the charging / discharging device 4000 and the computer 3000 are respectively controlled according to the second charging / discharging plan and the second IT workload plan generated in steps 20060 and 20070, respectively, in the re-learning necessity determination processing flow 20000 of Fig. 11 is shown at predetermined time intervals in the display frame 5250A2.

[0098] In addition, when the decrease in the amount of stored power shown in the display frame 5250A2 is increasing, it indicates that the influence of the first charge / discharge plan and the first IT workload plan being deviated from the original planned value by the amount of increase is large. Here, the cause of the first charge / discharge plan and the first IT workload plan being deviated from the original planned value is an error in the predicted value of the RE power generation amount and the power consumption amount of the computer 3000 obtained by inference using the ML model. Therefore, when the decrease in the amount of stored power exceeds a predetermined threshold, that is, when it is determined that there is a negative difference in step 20080 in the re-learning necessity determination process flow 20000 of FIG. 11, it is determined that the ML model with the deteriorated inference accuracy should be re-learned.

[0099] The display frame 5250A3 shows the time change of the reduction amount of the RE rate caused by the difference between the predicted value and the actual value based on the inference result of the ML model shown in the display frame 5250A1. That is, the difference between the RE rate obtained as a result of actually controlling the charging / discharging device 4000 and the computer 3000 according to the first charging / discharging plan and the first IT workload plan generated in steps 15050 and 15060, respectively, in the inference processing flow 15000 of Fig. 10 and the RE rate assumed when the charging / discharging device 4000 and the computer 3000 are hypothetically controlled according to the second charging / discharging plan and the second IT workload plan generated in steps 20060 and 20070, respectively, in the re-learning necessity determination processing flow 20000 of Fig. 11 is shown in the display frame 5250A3 at predetermined time intervals.

[0100] Note that, similar to the decrease in the amount of stored power in display frame 5250A2, when the decrease in the RE rate shown in display frame 5250A3 is increasing, it indicates that the influence of the first charge / discharge plan or the first IT workload plan deviating from the original planned value by the amount of the increase was large. Therefore, when the decrease in the RE rate exceeds a predetermined threshold, that is, when it is determined that there is a negative difference in step 20080 in the re-learning necessity determination process flow 20000 in Fig. 11, it is determined that the ML model with reduced inference accuracy should be re-learned.

[0101] In display frame 5250A4, an explanation of the determination result of whether or not re-learning is necessary is displayed. In display screen 5250A in Fig. 12, although the inference accuracy of the ML model shown in display frame 5250A1 is below a threshold, the amount of stored power and the amount of decrease in the RE rate shown in display frames 5250A2 and 5250A3, respectively, do not exceed the threshold, so that it is determined that re-learning is unnecessary, and this is explained in display frame 5250A4.

[0102] When the user presses the confirmation button 5250A5 on the display screen 5250A, the management computer 5000 determines that the user has confirmed and ends the display of the display screen 5250A. Furthermore, when the user presses the re-learning execution button 5250A6, the management computer 5000 forcibly executes re-learning of the ML model regardless of the determination result of whether or not re-learning is necessary. However, the re-learning execution button 5250A6 is not essential on the display screen 5250A and may be omitted.

[0103] The management computer 5000 may output to another computer the same information as that displayed on the display screen 5250A as a log of the relearning necessity determination program 5250. Furthermore, the relearning necessity determination result may be displayed using a screen configuration other than the display screen 5250A.

[0104] Fig. 13 is a diagram showing another example of a display screen of the relearning necessity determination result displayed after execution of the relearning necessity determination program 5250 according to the first embodiment of the present invention. Like the display screen 5250A shown in Fig. 12, the display screen 5250B shown in Fig. 13 is a screen for explaining the determination result of whether or not relearning is necessary and the basis for it, and is displayed on a display device included in the I / O device 5500, for example, in the management computer 5000, as a result of the processor 5100 executing the relearning necessity determination program 5250.

[0105] Display screen 5250B includes display frames 5250B1, 5250B2, 5250B3, and 5250B4, a confirmation button 5250B5, and a re-learning execution button 5250B6. The contents of information shown in display frames 5250B1-5250B4 and the functions of confirmation button 5250B5 and re-learning execution button 5250B6 are similar to display frames 5250A1-5250A4, confirmation button 5250A5, and re-learning execution button 5250A6 in display screen 5250A in FIG.

[0106] Display screen 5250B in Figure 13 shows an example of a screen in which it is determined that re-learning is necessary because the inference accuracy of the ML model shown in display frame 5250B1 is below a threshold and the amount of decrease in the stored power amount shown in display frame 5250B2 exceeds a threshold.

[0107] According to the first embodiment of the present invention described above, the following advantageous effects can be obtained.

[0108] (1) The management computer 5000 includes a control plan generating unit 5120 that generates a first control plan (a first charge / discharge plan and a first IT workload plan) for the control target devices (the charging / discharging device 4000 and the computer 3000) based on a predicted value of a physical quantity predicted by inference using the ML model 5320, an equipment control unit 5130 that controls the charging / discharging device 4000 and the computer 3000 using the first charge / discharge plan and the first IT workload plan, and a relearning necessity determining unit 5140 that determines whether or not relearning of the ML model 5320 is necessary based on the predicted value of the physical quantity and an actual measurement value of the physical quantity when the charging / discharging device 4000 and the computer 3000 are controlled by the equipment control unit 5130. In this management computer 5000, the relearning necessity determining unit 5140 evaluates the accuracy of inference based on the predicted value and the actual measurement value (steps 20040, 20050). When the re-learning necessity determination unit 5140 evaluates that the accuracy of inference has deteriorated (step 20050: Yes), the control plan generation unit 5120 generates a second control plan (a second charge / discharge plan and a second IT workload plan) for the charging / discharging equipment 4000 and the computer 3000 based on the actual measurement value (steps 20060, 20070). In addition, the re-learning necessity determination unit 5140 determines whether or not a negative difference occurs in a first control result when the charging / discharging equipment 4000 and the computer 3000 are controlled using the first charge / discharge plan and the first IT workload plan with respect to a second control result when the charging / discharging equipment 4000 and the computer 3000 are controlled using the second charge / discharge plan and the second IT workload plan (step 20080), and when it is determined that a negative difference occurs (step 20080: Yes), it determines that re-learning of the ML model 5320 is necessary (step 20090). This makes it possible to appropriately determine whether or not the ML model needs to be retrained.

[0109] (2) The physical quantity includes at least one of the amount of power generated by renewable energy (RE power generation amount) that can supply power to a power consumption device (computer 3000) that operates by consuming power, and the amount of power consumed by computer 3000. In addition, the negative difference includes at least one of a decrease in the amount of power stored in a power storage device (storage battery 4100) that can store surplus power for computer 3000, and a decrease in the utilization rate of renewable energy (RE rate) in the amount of power consumed by computer 3000. In this manner, when the accuracy of inference using an ML model deteriorates, it is possible to appropriately determine whether or not re-learning of the ML model is necessary.

[0110] (3) The controlled device is the charging / discharging device 4000 that charges / discharges the storage battery 4100, and the device control unit 5130 controls the charging / discharging device 4000. The controlled device is also the computer 3000, and the device control unit 5130 controls the workload of the computer 3000. In this manner, the amount of stored power in the storage battery 4100 and the amount of power consumed by the computer 3000 can be appropriately adjusted according to the control plan generated by the control plan generating unit 5120.

[0111] Second Embodiment A computer system according to the second embodiment of the present invention will be described. The computer system of this embodiment has a similar configuration to the computer system 10 of Figures 1 and 2 described in the first embodiment. Therefore, the computer system of this embodiment will be described below using the configuration of the computer system 10 of Figures 1 and 2.

[0112] Fig. 14 is a diagram showing a relearning necessity determination process flow by the relearning necessity determination program 5250 according to the second embodiment of the present invention. In the management computer 5000 of this embodiment, the processor 5100 periodically executes processing in accordance with the relearning necessity determination process flow 30000 in Fig. 14 by the relearning necessity determination program 5250. Note that in the following explanation, it is assumed that the data collection process flow 10000 in Fig. 9 and the inference process flow 15000 in Fig. 10 have each been executed before the relearning necessity determination process flow 30000 is executed.

[0113] When re-learning necessity determination process flow 30000 is started in step 30010, steps 30020 to 30080 perform the same processes as steps 20020 to 20080 in Fig. 11 described in the first embodiment. However, if it is evaluated in step 30050 that the inference accuracy has not decreased for all ML models, the process proceeds to step 30120. Also, if it is determined in step 30080 that there is a negative difference, the process proceeds to step 30090, and if it is determined that there is no negative difference, the process proceeds to step 30100.

[0114] In step 30100, it is determined whether or not there is a positive difference between the first charge / discharge plan or the first IT workload plan used to control the charge / discharge device 4000 or the computer 3000 in the previous control period and the second charge / discharge plan generated in step 30060 or the second IT workload plan generated in step 30070. In contrast to the negative difference described above, the positive difference here means that the value of the stored power amount or the RE rate of the storage battery 4100 when the charge / discharge device 4000 or the computer 3000 is actually controlled according to the first charge / discharge plan or the first IT workload plan in the previous control period is increased compared to the value of the stored power amount or the RE rate of the storage battery 4100 when it is assumed that the charge / discharge device 4000 or the computer 3000 is controlled according to the second charge / discharge plan or the second IT workload plan.

[0115] Specifically, a comparison similar to that in step 30080 is performed using the amount of stored power and the RE rate when the first charge / discharge plan and the first IT workload plan are implemented, which are acquired at the time of the determination in step 30080, and the amount of stored power and the RE rate when the second charge / discharge plan and the second IT workload plan are implemented. As a result, if the amount of stored power or the RE rate of the storage battery 4100 increases by controlling the charging / discharging device 4000 according to the first charge / discharge plan and charging / discharging the storage battery 4100 compared to the case where the charging / discharging device 4000 is controlled according to the second charge / discharge plan and charging / discharging the storage battery 4100, it is determined that there is a positive difference, and the process proceeds to step 30110. On the other hand, if neither the amount of stored power nor the RE rate of the storage battery 4100 increases, it is determined that there is no positive difference, and the process proceeds to step 30120. Furthermore, if controlling the computer 3000 according to the first IT workload plan increases the amount of stored power or the RE rate of the storage battery 4100 compared to when the computer 3000 is controlled according to the second IT workload plan, it is determined that there is a positive difference, and the process proceeds to step 30110. On the other hand, if neither the amount of stored power nor the RE rate of the storage battery 4100 increases, it is determined that there is no positive difference, and the process proceeds to step 30120.

[0116] In step 30110, it is determined whether the amount of stored power in any of the storage batteries 4100 has reached maximum capacity due to an increase in the amount of stored power or the RE rate of the storage batteries 4100 caused by a positive difference. Here, it is determined whether or not there is a storage battery whose amount of stored power has reached maximum capacity from the reference result of the storage battery management table 6310 made at the time of the determination in step 30080. As a result, if there is even one storage battery whose amount of stored power has reached maximum capacity, it is determined that the amount of stored power in that storage battery has reached saturation and the process proceeds to step 30090, and if there is no storage battery whose amount of stored power has reached maximum capacity, the process proceeds to step 30120.

[0117] In step 30090, it is determined that re-learning is required for each ML model that was evaluated in step 30050 as having a decline in inference accuracy.

[0118] In step 30120, it is determined that re-learning is not required for all ML models.

[0119] Once the process of either step 30090 or 30120 has been completed, the re-learning necessity determination process flow 30000 ends in step 30130.

[0120] After execution of the re-learning necessity determination process flow 30000 in FIG. 14, the learning program 5220 performs re-learning of the ML model determined to require re-learning in step 30090 at any timing, and updates the contents of the ML model 5320.

[0121] According to the second embodiment of the present invention described above, even if a positive difference occurs between the first control result when the charging / discharging device 4000 and the computer 3000 are controlled using the first charge / discharge plan and the first IT workload plan and the second control result when the charging / discharging device 4000 and the computer 3000 are controlled using the second charge / discharge plan and the second IT workload plan (step 30100: Yes) and the amount of stored power in the storage battery 4100 is saturated (step 30110: Yes), the relearning necessity determination unit 5140 determines that relearning of the ML model is necessary (step 30090). As a result, when power waste occurs due to saturation of the amount of stored power, it is possible to determine that relearning of the ML model is necessary and reduce the waste.

[0122] (Third embodiment) A computer system according to the third embodiment of the present invention will be described. The computer system of this embodiment also has a similar configuration to the computer system 10 of Figures 1 and 2 described in the first embodiment. Therefore, the computer system of this embodiment will be described below using the configuration of the computer system 10 of Figures 1 and 2.

[0123] Fig. 15 is a diagram showing a relearning necessity determination process flow by the relearning necessity determination program 5250 according to the third embodiment of the present invention. In the management computer 5000 of this embodiment, the processor 5100 periodically executes processing in accordance with the relearning necessity determination process flow 40000 in Fig. 15 by the relearning necessity determination program 5250. Note that in the following explanation, it is assumed that the data collection process flow 10000 in Fig. 9 and the inference process flow 15000 in Fig. 10 have each been executed before the relearning necessity determination process flow 40000 is executed.

[0124] When the re-learning necessity determination process flow 40000 is started in step 40010, steps 40020 to 40080 perform the same processes as steps 20020 to 20080 in Fig. 11 described in the first embodiment. However, if it is evaluated in step 40050 that the inference accuracy has not decreased for all ML models, the process proceeds to step 40120. Also, if it is determined in step 40080 that there is a negative difference, the process proceeds to step 40090, and if it is determined that there is no negative difference, the process proceeds to step 40120.

[0125] In step 40090, an excess increment of the amount of stored power caused by a difference (plan difference) between the first charge / discharge plan and the first IT workload plan used for controlling the charge / discharge device 4000 and the computer 3000 in the previous control period and the second charge / discharge plan and the second IT workload plan generated in steps 40060 and 40070 is calculated. Specifically, the difference between the amount of stored power when the first charge / discharge plan and the first IT workload plan acquired at the time of the determination in step 40080 are implemented and the amount of stored power when the second charge / discharge plan and the second IT workload plan are implemented is calculated as the excess increment of the amount of stored power. Then, the calculated excess increment is accumulated for each storage battery and recorded in the excess increment of stored power 6314 in the storage battery management table 6310, thereby updating the storage battery management table 6310.

[0126] As described in the first embodiment, when the above-mentioned plan difference occurs due to low accuracy of inferring the RE power generation amount and the power consumption amount by the ML model, the power generated by normal power generation (thermal power generation, etc.) other than RE power generation is purchased from the power company and supplied to the computer 3000, and the power exceeds the amount actually required. As a result, the RE rate of the power supplied to the computer 3000 decreases while the amount of stored power in the storage battery 4100 increases. Therefore, in this embodiment, the increased amount of stored power for each storage battery is calculated as an excess increment and recorded as the storage amount surplus increment 6314 in the storage battery management table 6310, so that the excess increment of the amount of stored power due to the plan difference can be managed. In this case, the amount of stored power increases while the RE rate decreases, so that in step 40080, the decrease in the RE rate is determined as a negative difference.

[0127] In step 40100, for each ML model evaluated in step 40050 as having a decreased inference accuracy, the power consumption required for re-learning the ML model is compared with the surplus increment of the stored power amount. If the power consumption required for re-learning is equal to or less than the surplus increment of the stored power amount, the process proceeds to step 40110, and if the power consumption required for re-learning exceeds the surplus increment of the stored power amount, the process proceeds to step 40130. The power consumption required for re-learning of each ML model can be acquired by referring to the value of the re-learning power consumption 5266 in the ML model management table 5260 in FIG. 5. The surplus increment of the stored power amount can be acquired by identifying a record corresponding to the ML model in the storage battery management table 6310 in FIG. 6 based on the storage battery ID 6311 and referring to the value of the surplus increment of the stored power amount 6314 of the record.

[0128] In step 40110, it is determined that re-learning is necessary for each ML model for which the inference accuracy was evaluated to have decreased in step 40050 and for which it was determined in step 40100 that the power consumption required for re-learning is equal to or less than the surplus increment of stored power.

[0129] In step 40120, it is determined that re-learning is not necessary for all ML models.

[0130] In step 40130, it is determined that re-learning is necessary for each ML model for which it has been evaluated in step 40050 that the inference accuracy has decreased and it has been determined in step 40100 that the power consumption required for re-learning exceeds the surplus increment of the amount of stored power. However, in this case, it is not determined that re-learning is necessary immediately, but rather the surplus increment of the amount of stored power is continuously monitored, and it is determined that re-learning is necessary when the value becomes equal to or greater than the power consumption required for re-learning.

[0131] Once any one of steps 40110, 40120, and 40130 has been completed, the re-learning necessity determination process flow 40000 ends in step 40140.

[0132] 15, the learning program 5220 executes, at any timing, relearning of the ML models determined to require relearning in steps 40110 and 40130, and updates the contents of the ML model 5320. However, for the ML models determined to require relearning in step 40130, relearning is not executed immediately, but is executed when the surplus increase in the amount of stored power becomes equal to or exceeds the power consumption required for relearning, as described above.

[0133] Fig. 16 is a diagram showing an example of a display screen of a relearning necessity determination result displayed after execution of a relearning necessity determination program 5250 according to the third embodiment of the present invention. Like the display screens 5250A and 5250B in Fig. 12 and 13 described in the first embodiment, the display screen 5250C shown in Fig. 16 is a screen for explaining the determination result of whether or not relearning is necessary and the basis for it, and is displayed on a display device included in an I / O device 5500, for example, in the management computer 5000, as a result of the processor 5100 executing the relearning necessity determination program 5250.

[0134] The display screen 5250C includes display frames 5250C1, 5250C2, 5250C3, and 5250C4, a confirmation button 5250C5, and a re-learning execution button 5250C6. The contents of information displayed in the display frames 5250C1 to 5250C4 and the functions of the confirmation button 5250C5 and the re-learning execution button 5250C6 are similar to the display frames 5250A1 to 5250A4, the confirmation button 5250A5, and the re-learning execution button 5250A6 in the display screen 5250A of Fig. 12. However, in addition to the information in the display frame 5250A1, the display frame 5250C1 further displays the surplus increase in the amount of stored power and the power consumption required for re-learning the ML model.

[0135] 16 shows an example of a screen in which it is determined that re-learning is necessary because the inference accuracy of the ML model shown in display frame 5250C1 is below a threshold and the decrease in the amount of stored power shown in display frame 5250C2 exceeds a threshold. However, in this case, the surplus increase in the amount of stored power is below the power consumption required for re-learning the ML model, so display frame 5250C4 explains that re-learning will be performed when the surplus increase in the amount of stored power becomes equal to or exceeds the power consumption required for re-learning the ML model.

[0136] According to the third embodiment of the present invention described above, the negative difference is a decrease in the RE rate, and the re-learning necessity determination unit 5140 calculates the surplus increase in the amount of stored power due to the decrease in the RE rate (step 40090), and determines that re-learning of the ML model is necessary when the accumulated value of the surplus increase becomes equal to or greater than the power consumption required for re-learning the ML model 5320 (steps 40110, 40130). In this manner, it is possible to determine whether or not re-learning of the ML model is necessary so that re-learning of the ML model is performed at optimal timing while minimizing waste of power.

[0137] (Fourth embodiment) A computer system according to a fourth embodiment of the present invention will be described. The computer system of this embodiment also has a similar configuration to the computer system 10 of Figures 1 and 2 described in the first embodiment. Therefore, the computer system of this embodiment will be described below using the configuration of the computer system 10 of Figures 1 and 2.

[0138] Fig. 17 is a diagram showing a relearning necessity determination process flow by the relearning necessity determination program 5250 according to the fourth embodiment of the present invention. In the management computer 5000 of this embodiment, the processor 5100 periodically executes processing in accordance with the relearning necessity determination process flow 50000 in Fig. 17 by the relearning necessity determination program 5250. Note that in the following explanation, it is assumed that the data collection process flow 10000 in Fig. 9 and the inference process flow 15000 in Fig. 10 have each been executed before the relearning necessity determination process flow 50000 is executed.

[0139] When the re-learning necessity determination process flow 50000 is started in step 50010, steps 50020 to 50080 perform the same processes as steps 20020 to 20080 in Fig. 11 described in the first embodiment. However, if it is evaluated in step 50050 that the inference accuracy has not decreased for all ML models, the process proceeds to step 50120. Also, if it is determined in step 50080 that there is a negative difference, the process proceeds to step 50090, and if it is determined that there is no negative difference, the process proceeds to step 50120.

[0140] In step 50090, the RE shortage caused by the first charge / discharge plan and the first IT workload plan used in the control of the charge / discharge device 4000 and the computer 3000 in the previous control period is calculated. Specifically, the difference between the power consumption and the RE power generation of the computer 3000 in each computing environment 100 acquired in step 50030 is calculated as the RE shortage. Then, the calculated RE shortage is accumulated for each computing environment and recorded in the RE shortage 6333 in the RE ratio management table 6330, thereby updating the RE ratio management table 6330.

[0141] If the accuracy of inferring the amount of RE power generation and the amount of power consumption by the ML model is low, the RE rate of the power supplied to the computer 3000 may decrease as a result of purchasing power from a power company that is not RE power generation but normal power generation (thermal power generation, etc.) and supplying it to the computer 3000. Therefore, in this embodiment, the amount of power supplied by power generation other than RE power generation is calculated as the RE shortage amount and recorded as the RE shortage amount 6333 in the RE rate management table 6330, thereby making it possible to manage the cost increase accompanying the decrease in the RE rate. Note that in this case, since the RE rate decreases, in step 50080, the decrease in the RE rate will be determined as a negative difference.

[0142] In step 50100, for each ML model evaluated in step 50050 as having a decreased inference accuracy, the cost required for re-learning the ML model is compared with the additional procurement cost required to cover the shortage of RE power generation (additional RE procurement cost). If the cost required for re-learning is equal to or less than the additional RE procurement cost, the process proceeds to step 50110, and if it exceeds the additional RE procurement cost, the process proceeds to step 50130. The cost required for re-learning each ML model can be obtained by referring to the value of the re-learning cost 5267 in the ML model management table 5260 in FIG. 5. The additional RE procurement cost can be obtained by identifying a record corresponding to the ML model in the RE rate management table 6330 in FIG. 8 based on the computing environment ID 6331, and multiplying the value of the RE shortage amount 6333 of the record by a predetermined coefficient (for example, the purchase price of a green power certificate).

[0143] In step 50110, it is determined that re-learning is necessary for each ML model for which the inference accuracy was evaluated to have decreased in step 50050 and the cost required for re-learning was determined to be equal to or less than the additional RE procurement cost in step 50100.

[0144] In step 50120, it is determined that re-learning is not necessary for all ML models.

[0145] In step 50130, it is determined that re-learning is necessary for each ML model for which it has been evaluated in step 50050 that the inference accuracy has decreased and it has been determined in step 50100 that the cost required for re-learning exceeds the additional RE procurement cost. However, in this case, it is not determined that re-learning is necessary immediately, but rather the additional RE procurement cost is continuously monitored, and when the value becomes equal to or greater than the cost required for re-learning, it is determined that re-learning is necessary.

[0146] After the processing of any one of steps 50110, 50120, and 50130 is completed, the re-learning necessity determination processing flow 50000 ends in step 50140.

[0147] 17, the learning program 5220 executes, at any timing, re-learning of the ML models determined to require re-learning in steps 50110 and 50130, and updates the contents of the ML model 5320. However, for ML models determined to require re-learning in step 50130, re-learning is not executed immediately, but is executed when the additional RE procurement cost becomes equal to or exceeds the cost required for re-learning, as described above.

[0148] Fig. 18 is a diagram showing an example of a display screen of a relearning necessity determination result displayed after execution of a relearning necessity determination program 5250 according to the fourth embodiment of the present invention. Like the display screens 5250A and 5250B in Fig. 12 and 13 described in the first embodiment, the display screen 5250D shown in Fig. 18 is a screen for explaining the determination result of whether or not relearning is necessary and the basis for it, and is displayed on a display device included in an I / O device 5500, for example, in the management computer 5000, as a result of the processor 5100 executing the relearning necessity determination program 5250.

[0149] The display screen 5250D includes display frames 5250D1, 5250D2, 5250D3, and 5250D4, a confirmation button 5250D5, and a re-learning execution button 5250D6. The contents of the information displayed in the display frames 5250D1 to 5250D4 and the functions of the confirmation button 5250D5 and the re-learning execution button 5250D6 are similar to the display frames 5250A1 to 5250A4, the confirmation button 5250A5, and the re-learning execution button 5250A6 in the display screen 5250A of Fig. 12. However, in addition to the information in the display frame 5250A1, the display frame 5250D1 further displays the RE additional procurement cost and the cost required for re-learning the ML model.

[0150] 18 shows an example of a screen in which it is determined that re-learning is necessary because the inference accuracy of the ML model shown in display frame 5250D1 is below a threshold and the amount of decrease in the amount of stored power shown in display frame 5250D2 exceeds a threshold. However, in this case, since the additional RE procurement cost is below the cost required for re-learning the ML model, display frame 5250D4 explains that re-learning will be performed when the additional RE procurement cost becomes equal to or exceeds the cost required for re-learning the ML model.

[0151] According to the fourth embodiment of the present invention described above, the negative difference is a decrease in the RE rate, and the re-learning necessity determination unit 5140 calculates the additional cost due to the decrease in the RE rate (steps 50090, 50100), and determines that re-learning of the ML model is necessary when the accumulated value of the additional cost becomes equal to or greater than the cost required for re-learning the ML model 5320 (steps 50110, 50130). In this manner, it is possible to determine whether or not to re-learn the ML model so that unnecessary cost increases are suppressed and the ML model is re-learned at optimal timing.

[0152] (Modification) Assume that the first IT workload plan generated by the process of step 15060 in the inference process flow 15000 of FIG. 10 is a plan for shifting the workload of the computer 3000 by changing the operation schedule. In this case, a deviation due to the workload shift will inevitably occur between the predicted value of the power consumption obtained in step 15030 and the actual value of the power consumption when the computer 3000 is controlled according to the first IT workload plan. Therefore, in the first to fourth embodiments, when a plan for shifting the workload of the computer 3000 is generated as the first IT workload plan, a predicted value of the power consumption when the computer 3000 is controlled according to the first IT workload plan may be obtained, and the accuracy of the inference by the ML model 5320 may be calculated based on this predicted value.

[0153] Specifically, in the first embodiment, in step 20020 of the relearning necessity determination process flow 20000 in FIG. 11, instead of using the inference result data 5330, inference reflecting the first IT workload plan is performed to calculate a predicted value of the power consumption when the workload of the computer 3000 is shifted. In steps 20040 to 20050, the inference accuracy for each ML model is calculated and evaluated based on the difference between the predicted value and the actual measurement value. Similarly, in the second to fourth embodiments, in step 30020 of the relearning necessity determination process flow 30000 in FIG. 14, step 40020 of the relearning necessity determination process flow 40000 in FIG. 15, and step 50020 of the relearning necessity determination process flow 50000 in FIG. 17, instead of using the inference result data 5330, inference reflecting the first IT workload plan is performed to calculate a predicted value of the power consumption when the workload of the computer 3000 is shifted. In addition, in steps 30040 to 30050 of re-learning necessity determination process flow 30000, steps 40040 to 40050 of re-learning necessity determination process flow 40000, and steps 50040 to 50050 of re-learning necessity determination process flow 50000, the inference accuracy for each ML model is calculated and evaluated based on the difference between the predicted value and the actual measured value.

[0154] According to the modified example described above, the control plan generating unit 5120 generates a plan for shifting the workload of the computer 3000 as the first IT workload plan. The relearning necessity determining unit 5140 evaluates the accuracy of inference based on the predicted value and the actual value of the power consumption when the computer 3000 is controlled using the first IT workload plan (steps 20040 to 20050, 30040 to 30050, 40040 to 40050, 50040 to 50050). As a result, even if the workload of the computer 3000 is shifted, the accuracy of inference using the ML model 5320 can be accurately evaluated.

[0155] The present invention is not limited to the above-described embodiment and modifications, and may be implemented using any components without departing from the spirit of the present invention. In addition, the embodiments and modifications may be implemented in any combination.

[0156] The above-mentioned embodiments and modifications are merely examples, and the present invention is not limited to these contents as long as the characteristics of the invention are not impaired. In addition, although various embodiments and modifications have been described above, the present invention is not limited to these contents. Other aspects that are conceivable within the scope of the technical idea of ​​the present invention are also included in the scope of the present invention. [Explanation of symbols]

[0157] 10: computer system, 100, 200: computing environment, 300: RE power source, 400: wide area network, 3000: computer, 4000: charging / discharging equipment, 4100: storage battery, 5000: management computer, 5100: processor, 5110: inference unit, 5120: control plan generation unit, 5130: equipment control unit, 5140: re-learning necessity determination unit, 5150: learning unit, 5200: memory, 5210: data collection program, 5220: learning program, 5225: inference program, 5230: storage battery charge / discharge planning / control program, 5240: IT workload planning / control program, 5250: re-learning necessity determination program RAM, 5260: ML model management table, 5300: storage device, 5310: learning data, 5320: ML model, 5330: inference result data, 5400: management network interface, 5500: I / O device, 6000: data store, 6100: processor, 6200: memory, 6210: data input / output program, 6300: storage device, 6310: battery management table, 6320: computer management table, 6330: RE rate management table, 6400: management network interface, 6500: I / O device, 7000: data network, 8000: management network

Claims

1. a control plan generation unit that generates a first control plan for a control target device based on a predicted value of a physical quantity predicted by inference using an ML model; a device control unit that controls the control target device using the first control plan; a re-learning necessity determination unit that determines whether or not re-learning of the ML model is necessary based on the predicted value and an actual measurement value of the physical quantity when the controlled device is controlled by the device control unit; Equipped with the re-learning necessity determination unit evaluates accuracy of the inference based on the predicted value and the actual measured value; the control plan generation unit generates a second control plan for the controlled device based on the actual measurement value when the re-learning necessity determination unit evaluates that accuracy of the inference has deteriorated; The relearning necessity determination unit determines whether a negative difference occurs in a first control result when the controlled device is controlled using the first control plan compared to a second control result when the controlled device is controlled using the second control plan, and determines that relearning of the ML model is necessary if it is determined that a negative difference occurs.

2. 2. The ML model management device according to claim 1, the physical quantity includes at least one of an amount of power generated by renewable energy capable of supplying the power to a power consumption device that operates by consuming the power, and an amount of power consumed by the power consumption device; The negative difference includes at least one of a decrease in the amount of stored electricity in a power storage device capable of storing surplus electricity for the power consumption device and a decrease in the utilization rate of the renewable energy in the amount of power consumption.

3. 3. The ML model management device according to claim 2, The control target device is a charging / discharging device that charges and discharges the power storage device, The device control unit is an ML model management device that controls the charging / discharging device.

4. 3. The ML model management device according to claim 2, The control target device is the power consuming device, The device control unit is an ML model management device that controls the workload of the power consuming device.

5. 3. The ML model management device according to claim 2, The relearning necessity determination unit determines that relearning of the ML model is necessary when the amount of stored power in the storage device becomes saturated, even if a positive difference occurs in the first control result with respect to the second control result.

6. 3. The ML model management device according to claim 2, The negative difference is a decrease in the utilization rate of the renewable energy in the amount of power consumption, The relearning necessity determination unit calculates the surplus increase in the stored energy amount due to a decrease in the utilization rate of the renewable energy, and determines that the ML model needs to be relearned when the accumulated value of the surplus increase becomes equal to or greater than the power consumption required to relearn the ML model.

7. 3. The ML model management device according to claim 2, The negative difference is a decrease in the utilization rate of the renewable energy in the amount of power consumption, The relearning necessity determination unit calculates additional costs due to a decrease in the utilization rate of the renewable energy, and determines that relearning of the ML model is necessary when the accumulated value of the additional costs becomes equal to or greater than the cost required to relearn the ML model.

8. 5. The ML model management device according to claim 4, The control plan generation unit generates a plan for shifting the workload as the first control plan; The relearning necessity determination unit evaluates accuracy of the inference based on the predicted value and the actual measured value when the power consuming device is controlled using the first control plan.

9. A method for managing an ML model using a computer, comprising: generating a first control plan for a controlled device based on a predicted value of a physical quantity predicted by inference using an ML model, controlling the control target device using the first control plan by the computer; evaluating accuracy of the inference based on the predicted value and an actual measurement value of the physical quantity when the controlled device is controlled by the computer; generating a second control plan for the control target device based on the actual measurement value when the computer determines that the accuracy of the inference has deteriorated; determining, by the computer, whether or not a negative difference occurs in a first control result obtained when the control-target device is controlled using the first control plan compared to a second control result obtained when the control-target device is controlled using the second control plan; The ML model management method, wherein the computer determines that re-learning of the ML model is necessary when it determines that the negative difference occurs.

10. 10. The ML model management method according to claim 9, the physical quantity includes at least one of an amount of power generated by renewable energy capable of supplying the power to a power consumption device that operates by consuming the power, and an amount of power consumed by the power consumption device; The ML model management method, wherein the negative difference includes at least one of a decrease in the amount of stored electricity in a power storage device capable of storing surplus electricity for the power consumption device and a decrease in a utilization rate of the renewable energy in the amount of power consumption.

11. 11. The ML model management method according to claim 10, The ML model management method includes determining, by the computer, that re-learning of the ML model is necessary if the amount of stored electricity in the storage device becomes saturated even if a positive difference occurs in the first control result with respect to the second control result.

12. 11. The ML model management method according to claim 10, The negative difference is a decrease in the utilization rate of the renewable energy in the amount of power consumption, The computer calculates an excess increase in the amount of stored electricity due to a decrease in the utilization rate of the renewable energy, The ML model management method, wherein the computer determines that re-learning of the ML model is necessary when the accumulated value of the surplus increment becomes equal to or greater than the power consumption required for re-learning the ML model.

13. 11. The ML model management method according to claim 10, The negative difference is a decrease in the utilization rate of the renewable energy in the amount of power consumption, The computer calculates an additional cost due to a decrease in the utilization rate of the renewable energy, the computer determines that re-learning of the ML model is necessary when an accumulated value of the additional cost becomes equal to or greater than a cost required for re-learning the ML model.

14. 11. The ML model management method according to claim 10, The control target device is the power consuming device, generating, by the computer, a plan for shifting the workload of the power consuming device as the first control plan; The ML model management method further comprises: evaluating, by the computer, accuracy of the inference based on the predicted value and the actual measured value when the power consuming device is controlled using the first control plan.

Images