Learned model generating method and learned model generating program

Abstract

To generate a learned model for predicting a state quantity of an object space with high accuracy.SOLUTION: Correcting a state quantity of the target space at an N-th time (N is a natural number) calculated by a simulation that outputs the state quantity of the target space, using measurement data of the state quantity at the N-th time acquired by a sensor provided in the target space, and outputting a state quantity of the target space at an (N + 1) - th time by performing a simulation using the corrected state quantity of the target space at the N-th time. Generating a learned model that outputs a state quantity of the target space after a specific time in response to an input of a state quantity of the target space at the specific time, based on learning data using a plurality of pairs of the corrected state quantity of the target space at the N-th time and the corrected state quantity of the target space at the (N + 1) - th time by repeatedly performing the arithmetic processing; SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a trained model generation method and a trained model generation program. [Background technology]

[0002] Patent Document 1 discloses a system for controlling the state quantities of space within a building. The system performs a simulation of the space using the state quantities of the space within the building at a first time and the operating conditions of equipment installed in the building, and outputs the state quantities of the space at a second time that is a time after the first time. The system then performs data assimilation using the state quantities of the space at the second time and detection information from sensors installed in the building to correct the state quantities of the space and output the corrected state quantities of the space. The system further performs a simulation of the space using the corrected state quantities of the space and the operating conditions of the equipment when a new simulation is performed, and outputs the state quantities of the space predicted by the simulation. The system controls the operation of the equipment based on the predicted state quantities of the space and a target state quantity of the space. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2024-2000 A Summary of the Invention [Problem to be solved by the invention]

[0004] For example, by having a trained model predict the state quantities of a space within a building, it is possible to obtain prediction results of the state quantities of the space more quickly than when the state quantities of the space are predicted using existing simulation technology. The training data for generating the trained model can be prepared by running a simulation.

[0005] However, when a trained model outputs a state quantity of a space predicted using the state quantity of the space after correction by data assimilation, as in the simulation performed by the system of Patent Document 1, the accuracy of the state quantity of the space output by the trained model may be low. This is because the state quantity of the space before correction by data assimilation is calculated in the simulation performed to prepare training data.

[0006] The present disclosure has been devised in view of the above-described conventional situation, and aims to generate a trained model that predicts state quantities of a target space with high accuracy. [Means for solving the problem]

[0007] The present disclosure provides a trained model generation method, which corrects a state quantity of a target space at an Nth time (N: natural number) calculated by a simulation that outputs a state quantity of a target space using measurement data of the state quantity at the Nth time acquired by a sensor installed in the target space, and performs the simulation using the corrected state quantity of the target space at the Nth time to output the state quantity of the target space at the (N+1)th time, by repeatedly performing a calculation process, and generates a trained model that outputs a state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time, based on training data using a plurality of pairs of the corrected state quantity of the target space at the Nth time and the state quantity of the target space at the (N+1)th time.

[0008] The present disclosure also provides a trained model generation program that causes a computing device to repeatedly perform a computation process in which a state quantity of a target space at an Nth time (N: natural number) calculated by a simulation that outputs a state quantity of a target space is corrected using measurement data of the state quantity at the Nth time acquired by a sensor provided in the target space, and the simulation is performed using the corrected state quantity of the target space at the Nth time, thereby outputting the state quantity of the target space at the (N+1)th time; and generates a trained model that outputs a state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time, based on training data using a plurality of pairs of the corrected state quantity of the target space at the Nth time and the state quantity of the target space at the (N+1)th time.

[0009] Any combination of the above components, and conversion of the expression of the present disclosure into a method, device, system, storage medium, computer program, etc., are also valid aspects of the present disclosure. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to generate a trained model that predicts the state quantities of a target space with high accuracy. [Brief explanation of the drawings]

[0011] [Figure 1] Block diagram showing a configuration example of a trained model generation system according to a first embodiment. [Figure 2] FIG. 1 is a schematic diagram illustrating a flow of generating a trained model according to a first embodiment. [Figure 3] Flowchart showing a process for generating a trained model according to the first embodiment [Figure 4] 1 is a flowchart showing an analysis and assimilation process according to the first embodiment. [Figure 5] FIG. 1 is a block diagram showing a configuration example of a space state control system according to a first embodiment. [Figure 6]FIG. 1 is a schematic diagram illustrating a flow of space state control according to the first embodiment. [Figure 7] 1 is a flowchart showing a process for predicting a state quantity of a target space according to the first embodiment; [Figure 8] 1 is a flowchart showing a control change determination process according to the first embodiment; DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, with appropriate reference to the drawings, embodiments specifically disclosing a trained model generation method and a trained model generation program according to the present disclosure will be described in detail. However, more detailed description than necessary may be omitted. For example, detailed descriptions of already well-known matters and redundant descriptions of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following description and to facilitate understanding by those skilled in the art. Note that the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure and are not intended to limit the subject matter of the claims.

[0013] (Embodiment 1) [Generate a trained model] First, with reference to FIGS. 1 to 4, a method for generating a trained model 50 (see FIG. 2) used for state control of a target space 40 (see FIG. 1) will be described.

[0014] FIG. 1 is a block diagram showing an example configuration of a trained model generation system 1 according to a first embodiment. In this embodiment, a trained artificial intelligence (hereinafter referred to as "AI") model is used to control a state quantity such as a temperature distribution in a target space 40. The trained model generation system 1 is a system for generating a trained AI model. The trained model generation system 1 includes a computing device 10 and a sensor 30.

[0015] The arithmetic device 10 may be configured using a general-purpose computer device such as a personal computer or a server computer. The arithmetic device 10 may be installed inside or outside the target space 40. The arithmetic device 10 includes a processor 11, a memory 15, an input device 16, a display device 17, a communication device 18, an external interface device 19, and an internal interface device 20. The components included in the arithmetic device 10 are configured to be able to communicate with each other via the internal interface device 20.

[0016] The processor 11 may be configured using, for example, a central processing unit (hereinafter referred to as "CPU"), a graphical processing unit (hereinafter referred to as "GPU"), a micro processing unit (hereinafter referred to as "MPU"), a digital signal processor (hereinafter referred to as "DSP"), or a field programmable gate array (hereinafter referred to as "FPGA"). The processor 11 realizes various functions of the arithmetic device 10, for example, by referencing various data stored in the memory 15 or reading out programs. For example, the processor 11 and the memory 15 work together to realize various functions of a data assimilation module 12, a computer-aided engineering (hereinafter referred to as "CAE") tool 13, and a model generation tool 14, which will be described later.

[0017] The memory 15 is a storage unit for storing various data, programs, etc. The memory 15 may be configured from a volatile / non-volatile storage device such as a Random Access Memory (hereinafter referred to as "RAM"), a Read Only Memory (hereinafter referred to as "ROM"), a Hard Disk Drive (hereinafter referred to as "HDD"), or a Solid State Drive (hereinafter referred to as "SSD").

[0018] The input device 16 receives, for example, operations and instructions from a user and may be configured with a mouse, a keyboard, a touch panel display, or the like.

[0019] The display device 17 displays various user interfaces to the user and may be configured as a liquid crystal display, a touch panel display, or the like.

[0020] The communication device 18 is an interface device for communicating with external devices such as the sensor 30 via a network (not shown). There are no particular limitations on the communication standard that the communication device 18 can support, and the communication device 18 may support either a wired or wireless communication standard. Furthermore, the communication device 18 may be capable of supporting multiple communication standards.

[0021] The external interface device 19 is an interface device for communicating with an external device such as a sensor 30. The arithmetic device 10 and the sensor 30 may be communicably connected via the communication device 18 and a network (not shown). The external device is not limited to the sensor 30 shown in FIG. 1 , but may be another device.

[0022] The sensor 30 acquires measurement data of a state quantity at a predetermined position in the target space 40. In this embodiment, it is assumed that the state quantity of the target space 40 is the temperature distribution of the target space 40, and the sensor 30 is a temperature sensor. However, the state quantity of the target space 40 is not limited to the temperature distribution and may include, for example, a humidity distribution or a wind speed distribution. Similarly, the sensor 30 is not limited to a temperature sensor and may include, for example, a humidity sensor or a wind speed sensor. Furthermore, the sensor 30 may be installed at a predetermined position inside the target space 40, or at a predetermined position outside the target space 40. For example, the sensor 30 is installed at a predetermined position inside the target space 40 and acquires measurement data of the temperature at the predetermined position. Furthermore, the sensor 30 may be configured to include multiple sensors.

[0023] The target space 40 may be, for example, a room or a hall in a building.

[0024] Next, a flow of generating the trained model 50 according to the first embodiment will be described with reference to Fig. 2. Fig. 2 is a schematic diagram for explaining a flow of generating the trained model 50 according to the first embodiment.

[0025] The CAE tool 13, which is executed by the processor 11 and the memory 15 working together, calculates and reproduces the state quantities of the target space 40 at a specific time through a simulation using an analytical model. The output of the CAE tool 13 may be referred to as an analytical output. The analytical output is used as training data for generating a trained model 50. In the following description, for data indicating state quantities at a specific time (e.g., training data), the specific time may be expressed in parentheses. For example, training data at time t may be expressed as training data (t). In the following description, time may be expressed as hours.

[0026] The data assimilation module 12 corrects the analysis output (t) using the measurement data (t) acquired by the sensor 30. That is, the data assimilation module 12 corrects the state quantities of the target space 40 at a certain time calculated by simulation using the measurement data acquired by the sensor 30 at the same time. The correction is a process also called assimilation (data assimilation). An outline of the assimilation process will be explained below. Here, it is assumed that the temperature distribution in the target space 40 at a specific time is reproduced by simulation, and that the temperature at a specific position in the target space 40 at the specific time is measured by a sensor 30 installed at the specific position. Through the assimilation process, the temperature at the specific position where the sensor 30 is installed in the temperature distribution is replaced with the temperature measured by the sensor 30. Furthermore, the temperature in the temperature distribution near the specific position where the sensor 30 is installed is corrected based on the temperature measured by the sensor 30. For example, it is assumed that the temperature at a specific position in the temperature distribution reproduced by simulation is lower than the temperature at that position measured by the sensor 30. At this time, the temperature at a specific position in the temperature distribution reproduced by the simulation is replaced with the temperature measured by the sensor 30. That is, the temperature at the specific position in the temperature distribution reproduced by the simulation is replaced with a higher temperature. Furthermore, the temperature near the specific position in the temperature distribution also becomes higher than the temperature reproduced by the simulation. At this time, positions closer to the specific position are more susceptible to the influence of the replaced temperature at the specific position than positions farther from the specific position. Furthermore, the influence of the temperature at the specific position after replacement decreases as the distance from the specific position increases.

[0027] If the sensor 30 is installed at a predetermined location outside the target space 40, the temperature distribution of the target space 40 reproduced by the simulation, for example, the temperature at a location close to the predetermined location, may be corrected by data assimilation based on the temperature measured by the sensor 30.

[0028] Hereinafter, the corrected learning data will be referred to as corrected learning data or assimilated output. The corrected learning data is fed back to the CAE tool 13. The fed back corrected learning data is used in a simulation by the CAE tool 13. The CAE tool 13 uses the fed back corrected learning data (t) in the simulation to reproduce the state quantities of the object space 40 at the time next to time t. The time next to time t may be, for example, time t+Δt.

[0029] By repeating the reproduction of state quantities by the CAE tool 13 and the assimilation of data by the data assimilation module 12, multiple pieces of training data (analysis output) and corrected training data (assimilation output) are generated. The model generation tool 14 pairs an assimilated output at a certain time with an analysis output at the time following that time. For example, the model generation tool 14 creates a pair of an assimilated output (t) and an analysis output (t+Δt), a pair of an assimilated output (t+Δt) and an analysis output (t+2Δt), etc. The model generation tool 14 then performs a model training process using multiple pairs as a training data set. This generates a trained model 50 that can output state quantities of the target space 40 after a specific time in response to input of state quantities of the target space 40 at the specific time. The trained model 50 may be generated using any learning algorithm so as to output state quantities of the target space 40 after the specific time in response to input of state quantities of the target space 40 at the specific time. The learning algorithm for generating the trained model 50 according to this embodiment is not particularly limited, but may be, for example, machine learning including neural network technology such as Multi Layer Perceptron (hereinafter referred to as "MLP") and deep learning technology. The trained model 50 may be stored in, for example, the memory 15 or an external storage device (not shown).

[0030] Next, a process for generating the trained model 50 according to the first embodiment will be described with reference to Fig. 3. Fig. 3 is a flowchart showing the process for generating the trained model 50 according to the first embodiment. Each process shown in Fig. 3 is executed by the processor 11 of the arithmetic device 10.

[0031] The processor 11 performs an initial setting of the number of times of processing of the analysis and assimilation process described below (step S100). Specifically, the processor 11 specifies the number of times of processing of the analysis and assimilation process, and sets the current number of times of processing to 0. The processor 11 may perform the initial setting of step S100 based on a user operation.

[0032] The processor 11 executes an analysis and assimilation process (step S101). Details of step S101 will be described later with reference to Fig. 4. Through the process of step S101, training data and corrected training data are generated.

[0033] The processor 11 increments the number of times the analysis and assimilation process has been performed (step S102). That is, the processor 11 increments the count of the number of times the analysis and assimilation process has been performed by one.

[0034] The processor 11 determines whether the number of times the analysis and assimilation process has been performed has reached the number of times designated in the process of step S100 (step S103).

[0035] If processor 11 determines that the number of times the analysis and assimilation process has been performed has not reached the number of times specified in step S100 (step S103: NO), processor 11 returns to step S101 and repeats the process. By repeating the process of step S101, processor 11 can obtain a large amount of training data and corrected training data.

[0036] If the processor 11 determines that the number of times the analysis and assimilation process has been performed has reached the number of times specified in the process of step S100 (step S103: YES), the processor 11 proceeds to the process of step S104.

[0037] The processor 11 creates a training data set including a plurality of pairs of training data and corrected training data generated by the analysis and assimilation process of step S101 (step S104).

[0038] The processor 11 generates the trained model 50 based on the training data set (step S105). Then, the processor 11 ends this processing flow. The trained model 50 is generated by a series of processes in this flowchart.

[0039] 4 is a flowchart showing the analysis and assimilation process according to Embodiment 1. Each process shown in FIG.

[0040] The processor 11 creates an analytical model to be used by the CAE tool 13 for the simulation (step S110). A known technique may be used to create the analytical model. Various conditions and parameters for the analytical model may be arbitrarily set by, for example, a user. The analytical model is created so that the state quantities of the target space 40 can be reproduced by the simulation executed by the CAE tool 13. Here, the description is given on the assumption that the analytical model is created so that the temperature distribution in the target space 40 can be reproduced.

[0041] The processor 11 performs an initial setting of the time of the target space 40 for which the state quantities are to be reproduced in the analytical model (step S111). For example, the processor 11 sets the time from which to which the state quantities of the target space 40 are to be reproduced. Alternatively, the processor 11 may set how long, for example, how many hours, the state quantities of the target space 40 are to be reproduced. The processor 11 also sets the initial time (initial time). Furthermore, the processor 11 sets how much the time (hours) is to be updated for each analysis. These settings may be performed based on, for example, a user operation.

[0042] The processor 11 performs data assimilation setting (step S112). Specifically, the processor 11 sets the position of the sensor 30 installed in the real target space 40 in the analytical model created in step S110. The position in the analytical model may be expressed using, for example, a three-dimensional coordinate system. Hereinafter, the position in the analytical model may be referred to as a coordinate in the analytical model. The processor 11 can use the temperature at a predetermined position measured by the sensor 30 to replace the temperature at the coordinate corresponding to the predetermined position in the analytical model. The processor 11 also sets parameters that specify the extent to which the temperature near the set coordinate should be corrected based on the temperature at the coordinate after replacement.

[0043] The processor 11 calculates the state quantities of the target space 40 at a certain time through a simulation using the analytical model created in step S110 (step S113). Here, the explanation will be continued assuming that the temperature distribution in the target space 40 at time t has been calculated. Also, the time t is assumed to be the time set as the initial time in step S111.

[0044] The processor 11 obtains the calculation result of the process in step S113 (step S114), thereby obtaining the analysis output, that is, the learning data.

[0045] The processor 11 performs data assimilation processing using the state quantities of the target space 40 at a certain time calculated by the simulation and the measurement data at that time (step S115). The measurement data acquired by the sensor 30 may be stored and held in the memory 15 of the computing device 10. For example, the processor 11 performs data assimilation processing on the temperature distribution in the target space 40 at time t calculated by the simulation using the temperature measured by the sensor 30 at time t. As a result, the temperature at a predetermined position where the sensor 30 is installed in the temperature distribution is replaced with the temperature measured by the sensor 30. Furthermore, the temperature in the temperature distribution near the predetermined position is also corrected based on the temperature measured by the sensor 30 and the parameters set in the processing of step S112.

[0046] The processor 11 obtains the assimilation result from the process of step S115 (step S116), thereby obtaining the assimilation output, that is, the corrected learning data.

[0047] The processor 11 updates the time based on the setting made in the process of step S111 (step S117). For example, if the initial time is set to time t, the time t is updated to time t+Δt.

[0048] Processor 11 determines whether the time updated in the process of step S117 has reached the time specified in the process of step S111 (step S118).

[0049] If processor 11 determines that the time has not reached the specified time (step S118: NO), it returns to step S113 and repeats the process. Note that in the processing of step S113 from the second time onwards, the assimilated output from the immediately previous time is used. For example, in the calculation by simulation using the analytical model in step S113 for the second time, the assimilated output acquired in step S116 for the first time is used. As a result, the result of the calculation in step S113 for the second time becomes closer to the temperature distribution in the actual target space 40.

[0050] If processor 11 determines that the time has reached the designated time (step S118: YES), it ends this processing flow, and then proceeds to step S102 of the processing flow shown in FIG.

[0051] Hereinafter, the series of processes from step S113 to step S117 may be collectively referred to as the “calculation process.” In the analysis and assimilation process, the calculation process is repeatedly executed until the updated time reaches the designated time.

[0052] In this way, the calculation device 10 corrects the state quantity of the object space 40 at the Nth time calculated by a simulation that outputs the state quantity of the object space 40, using the measurement data of the state quantity at the Nth time acquired by the sensor 30 installed in the object space 40. Here, N is a natural number. The Nth time may be, for example, time t. The calculation device 10 then performs a simulation using the state quantity of the object space 40 at the corrected Nth time, thereby outputting the state quantity of the object space 40 at the (N+1)th time. The (N+1)th time is a time that is a specified time after the Nth time and is the time after the Nth time is updated. The calculation device 10 repeats the above calculation process until the (N+1)th time reaches the specified time. The calculation device 10 also repeats the analysis and assimilation process a specified number of times. As a result, a large number of pairs of the state quantity of the object space 40 at the corrected Nth time and the state quantity of the object space 40 at the (N+1)th time are generated. That is, a large number of pairs of assimilation output and analysis output are generated. The computing device 10 generates a trained model 50 that outputs the state quantity of the target space 40 at a time specified time after a specific time in response to the input of the state quantity of the target space 40 at the specific time, based on training data using multiple pairs of assimilation output and analysis output.

[0053] The method for generating the trained model 50 used for state control of the target space 40 has been described above with reference to Figures 1 to 4. Next, a method for controlling the state quantity of the target space 40 using the generated trained model 50 will be described with reference to Figures 5 to 8.

[0054] [Control of spatial conditions] 5 is a block diagram showing an example configuration of a space state control system 2 according to embodiment 1. The space state control system 2 is a system that uses a trained model 50 to control the operation of a space control device 60 installed in a target space 40, and controls state quantities such as the temperature distribution in the target space 40. In the description of the space state control system 2, the same matters as those described for the trained model generation system 1 will be omitted or simplified.

[0055] The space state control system 2 includes a calculation device 10, a sensor 30, and a space control device 60. In this embodiment, the trained model 50 is described as being stored and held in a memory 15.

[0056] The space control device 60 is a device that controls the state quantities of the target space 40, and is, for example, an air conditioner, a ventilation fan, an air purifier, a circulator, etc. The space control device 60 may include not only one device but also multiple devices. The space control device 60 may be provided inside the target space 40, or may be provided at the boundary between the inside and outside of the target space 40. The space control device 60 operates based on the control of the calculation device 10. For example, the space control device 60 may include a control unit (not shown), and the control unit may realize the functions of the space control device 60 based on the control of the calculation device 10. Hereinafter, the space control device 60 may also be referred to as a device.

[0057] Next, the flow of space state control according to the first embodiment will be described with reference to Fig. 6. Fig. 6 is a schematic diagram for explaining the flow of space state control according to the first embodiment.

[0058] The CAE tool 13 can calculate and reproduce the state quantities of the target space 40 at a specific time through a simulation using an analytical model. In this embodiment, it is assumed that the state quantities of the target space 40 reproduced and output by the CAE tool 13 are stored in a database 70. The database 70 may be stored in, for example, the memory 15 or an external device (not shown). The state quantities of the target space 40 output by the CAE tool 13 may be stored and retained in the memory 15 without going through the database 70. The state quantities of the target space 40 stored in the database 70 are referred to as internal data. For example, internal data indicating the state quantities of the target space 40 at time t may be referred to as internal data (t). Hereinafter, the internal data indicating the state quantities of the target space 40 may be referred to as data indicating the state quantities of the target space 40.

[0059] The flow of controlling the temperature distribution in the target space 40 will be described. First, the calculation device 10 acquires measurement data (t) from the sensor 30. The measurement data (t) indicates the temperature at time t at a predetermined position inside or outside the target space 40 where the sensor 30 is installed. The data assimilation module 12 performs data assimilation processing using the acquired measurement data (t) and internal data at the same time. In other words, the data assimilation module 12 corrects the internal data (t) using the measurement data (t). Hereinafter, the corrected internal data may be referred to as corrected internal data.

[0060] The corrected internal data (t) is input to the trained model 50. The trained model 50 outputs the temperature distribution of the target space 40 at time t + Δt in response to the input of the corrected internal data (t). The temperature distribution of the target space 40 at time t + Δt output by the trained model 50 is predicted based on the temperature distribution of the target space 40 at time t. Therefore, the output of the trained model 50 may be referred to as predicted data. In other words, the trained model 50 outputs predicted data (t + Δt) in response to the input of the corrected internal data (t).

[0061] The predicted data output by the trained model 50 is fed back to the database 70, i.e., the internal data. As a result, the internal data is updated using the predicted data. A specific explanation will be given using an example in which predicted data (t+Δt) is output by the trained model 50. Specifically, the internal data (t) is replaced with the predicted data (t+Δt) to become new internal data (t+Δt). Thereafter, the new internal data (t+Δt) is corrected using the measurement data (t+Δt). Note that the database 70 may store internal data in time series. In this case, the internal data at the same time as the predicted data fed back to the database 70 may be replaced with the predicted data.

[0062] The predicted data output by the trained model 50 is also used to control the space control device 60. For example, the calculation device 10 determines whether or not it is necessary to control the operation of the space control device 60 based on the predicted data (t+Δt) output by the trained model 50 and a target value of the temperature distribution in the target space 40 at a predetermined time t+Δt. Then, for example, if the calculation device 10 determines that it is necessary to control the operation of the space control device 60, it executes control of the operation of the space control device 60. The calculation device 10 may control the operation of the space control device 60 based on, for example, a target value of the temperature distribution in the target space 40 at a certain time, the current temperature at a predetermined position in the target space 40, the current operating conditions of the space control device 60, etc.

[0063] Next, a control process of the state quantity of the target space 40 using the trained model 50 according to the first embodiment will be described with reference to Fig. 7 and Fig. 8. Here, the control of the temperature of the target space 40 will be described as an example. Fig. 7 is a flowchart showing a prediction process of the state quantity of the target space 40 according to the first embodiment. Each process shown in Fig. 7 is executed by the processor 11 of the arithmetic device 10.

[0064] In this embodiment, the calculation device 10 may determine at regular intervals whether or not to control the operation of the space control device 60. Here, the description will be given on the assumption that the calculation device 10 determines at regular intervals whether or not to control the operation of the space control device 60.

[0065] The processor 11 of the arithmetic device 10 performs an initial setting of time (step S120). Here, the processor 11 sets that the determination of whether or not to control the operation of the space control device 60 is to be performed at 100-second intervals. The processor 11 also sets the current time (in other words, the elapsed time of the processing) to 0 seconds. The processor 11 further sets how much the current time should be advanced by updating the time, which will be described later. Here, the processor 11 sets that the current time should be advanced by 1 second by updating the time. Note that these setting contents are examples for the purpose of explanation and are not limited to these. The setting in step S120 may be performed based on a user operation, for example.

[0066] The processor 11 performs initial setting of the internal data (step S121). As described with reference to FIG. 6, the processor 11 (data assimilation module 12) corrects the internal data indicating the temperature distribution in the target space 40 at a certain time using the measurement data at that time. In other words, initial internal data that is corrected by the measurement data is required. In step S121, the processor 11 sets the initial internal data, in other words, the initial temperature distribution in the target space 40. The setting in step S121 may be performed based on, for example, a user operation.

[0067] The processor 11 acquires measurement data from the sensor 30 at a predetermined position in the target space 40 where the sensor 30 is installed (step S122).

[0068] The processor 11 performs data assimilation processing using the internal data initially set in step S121 and the measurement data acquired in step S122 (step S123), thereby obtaining corrected internal data.

[0069] The processor 11 inputs the corrected internal data obtained in the processing of step S123 (step S124) to the trained model 50 (step S124). The trained model 50 outputs predicted data for a time a specified time after the time corresponding to the corrected internal data. Note that when there is a state quantity of the target space 40 at a certain time (hour) or various data indicating the state quantity, the time (hour) may be referred to as the time (hour) corresponding to the state quantity or the various data. Here, for convenience of explanation, it is assumed that the trained model 50 outputs predicted data one second after the time corresponding to the corrected internal data.

[0070] The processor 11 acquires the prediction data output from the trained model 50 and updates the internal data (step S125). When the data assimilation process of step S123 is performed again, the internal data to be corrected will be the new internal data updated in step S125, rather than the internal data set in step S121.

[0071] Processor 11 determines whether the current time satisfies the condition preset in step S120 (step S126). The preset condition refers to the interval set in step S120. Specifically, processor 11 determines whether the current time is a multiple of 100 seconds. If the current time is 0 seconds, processor 11 determines that the current time is a multiple of 100 seconds.

[0072] If processor 11 determines in step S120 that the current time does not satisfy the preset condition (step S126: NO), processor 11 proceeds to the process in step S128.

[0073] If processor 11 determines that the current time satisfies the preset condition in step S120 (step S126: YES), it executes a control change determination process (step S127). In the control change determination process in step S127, processor 11 determines whether or not control of the operation of space control device 60 is necessary. Details will be described later with reference to FIG. 8. After step S127, processor 11 proceeds to the process in step S128.

[0074] Processor 11 updates the time based on the setting content in step S120 (step S128). If the current time is 0 seconds, processor 11 updates the current time to 1 second through the processing in step S128. Then, processor 11 returns to step S122 and repeats the processing.

[0075] By processing this flowchart, a determination is made every 100 seconds as to whether or not control of the operation of the space control device 60 is necessary. Furthermore, by processing this flowchart, the temperature distribution in the target space 40 one second from now is repeatedly predicted every second with high accuracy. The state quantities of the target space 40 generated by the CAE tool 13 are assimilated (corrected) by measurement data. Furthermore, the trained model 50 is generated so that it can output state quantities after a specified time in response to the input of the assimilated state quantities. Therefore, for example, when a trained model is generated without data assimilation, in other words, when the simulation results are used directly as training data, the output of the trained model 50 is more accurate.

[0076] Although not shown in the figure, the processor 11 may output the prediction data output from the trained model 50 acquired in step S125 to the display device 17 or an external device (not shown). Similarly, although not shown in the figure, this processing flow may be terminated, for example, based on a user operation. Alternatively, this processing flow may be set in advance to terminate, for example, when a specified time has elapsed.

[0077] FIG. 8 is a flowchart showing a control change determination process according to the first embodiment. Each process shown in FIG. 8 is executed by processor 11 of arithmetic device 10. Processor 11 may process the flowchart shown in FIG. 8 and the flowchart shown in FIG. 7 in parallel. For example, processor 11 may perform a process of accurately predicting the temperature distribution in target space 40 one second from now until 100 seconds have elapsed, in other words, until the next interval arrives, while simultaneously performing a process of determining whether or not control of the operation of space control device 60 is necessary. Note that, as with FIG. 7, the description will be given on the assumption that trained model 50 predicts and outputs the state quantity of target space 40 one second from the time corresponding to the input state quantity of target space 40.

[0078] The processor 11 performs an initial setting of time (step S130). For example, the processor 11 sets a condition for determining whether or not to control the operation of the space control device 60 based on the predicted temperature distribution of the target space 40 100 seconds after the current time. That is, the processor 11 sets a condition for determining whether or not to control the operation of the space control device 60 that the time corresponding to the temperature distribution of the target space 40 predicted by the trained model 50 is 100 seconds after the current time. Here, the current time is the time corresponding to the temperature distribution of the target space 40 indicated by the internal data copied in step S131, which will be described later. Note that the above setting contents are an example for explanation purposes and are not limited thereto. The setting in step S130 may be performed based on a user operation, for example.

[0079] The processor 11 copies the current internal data (step S131). More specifically, in the process of step S126, the processor 11 copies the internal data at the timing at which it is determined that the current time satisfies the preset condition in step S120. This is because the internal data at that timing serves as a starting point for predicting the temperature distribution in the target space 40 100 seconds from now. The internal data is updated by repeatedly executing step S125 of the flowchart shown in FIG. 7. Therefore, the processor 11 copies the current internal data so that it is not updated. The copied internal data may be stored in the memory 15, for example.

[0080] The processor 11 inputs the state quantity of the target space 40, i.e., the temperature distribution, to the trained model 50 (step S132). The processing of step S132 may be executed repeatedly, but the temperature distribution of the target space 40 input to the trained model 50 in the first processing is the temperature distribution of the target space 40 indicated by the internal data copied in step S131. As a result, the trained model 50 outputs the temperature distribution of the target space 40 one second after the time corresponding to the temperature distribution of the target space 40 indicated by the internal data. In other words, the temperature distribution of the target space 40 one second after the current time is predicted.

[0081] The processor 11 acquires the prediction data output from the trained model 50 (step S133).

[0082] Processor 11 determines whether the time corresponding to the predicted data acquired in step S133 satisfies the condition set in step S130 (step S134). Here, the condition set in step S130 is that the time corresponding to the predicted temperature distribution in target space 40 is 100 seconds after the current time.

[0083] If the processor 11 determines that the time corresponding to the predicted data acquired in step S133 does not satisfy the condition set in step S130 (step S134: NO), it inputs the predicted data acquired in step S133 to the trained model 50 (step S135). As a result, if the time corresponding to the predicted data acquired in step S133, i.e., the temperature distribution of the target space 40 input to the trained model 50, is one second after the current time, the trained model 50 outputs the temperature distribution of the target space 40 two seconds after the current time.

[0084] In this way, the processor 11 can obtain the predicted result of the temperature distribution in the target space 40 100 seconds after the current time by repeatedly executing step S133, step S134, and step S135.

[0085] If the processor 11 determines that the time corresponding to the predicted data acquired in step S133 satisfies the condition set in step S130 (step S134: YES), the processor 11 proceeds to step S136.

[0086] The processor 11 determines whether operation control of the space control device 60 is necessary based on the state quantities of the target space 40 output from the trained model 50 and preset state quantities of the target space 40 (step S136). For example, a target value for the temperature distribution of the target space 40 at a certain time may be set in advance. The processor 11 may determine that operation control of the space control device 60 is necessary, for example, when the difference between the temperature distribution of the target space 40 100 seconds after the current time and the preset target value of the temperature distribution is outside a specified range. The processor 11 may also determine that operation control of the space control device 60 is not necessary, for example, when the difference between the temperature distribution of the target space 40 100 seconds after the current time and the preset target value of the temperature distribution is within a specified range.

[0087] If processor 11 determines that operation control of space control device 60 is not necessary (step S136: NO), processor 11 ends this processing flow. Then, processor 11 proceeds to step S128 of the flowchart shown in FIG.

[0088] When processor 11 determines that operation control of space control device 60 is necessary (step S136: YES), processor 11 executes operation control of space control device 60 (step S137). For example, when processor 11 determines that the temperature distribution in target space 40 100 seconds from the current time is below a preset target value of the temperature distribution, that is, when processor 11 determines that target space 40 will be colder in 100 seconds, processor 11 may control the operation of space control device 60 to increase the temperature of target space 40. For example, when processor 11 determines that the temperature distribution in target space 40 100 seconds from the current time is above a preset target value of the temperature distribution, that is, when processor 11 determines that target space 40 will be hotter in 100 seconds, processor 11 may control the operation of space control device 60 to decrease the temperature of target space 40.

[0089] After the process of step S137, the processor 11 ends this process flow, and then the processor 11 proceeds to the process of step S128 in the flowchart shown in FIG.

[0090] In this way, the calculation device 10 predicts the future temperature distribution of the target space 40 at predetermined intervals and controls the operation of the space control device 60 based on the prediction results. By using the trained model 50, the calculation device 10 can predict the future temperature distribution of the target space 40 more quickly than, for example, when using conventional simulation technology. Furthermore, the internal data copied in step S131 and input to the trained model 50 in step S132 is updated by the prediction data output from the trained model 50 in step S125, thereby reproducing the temperature distribution of the target space 40 with high accuracy. The internal data is then used as a starting point to predict the future temperature distribution of the target space 40. In other words, the calculation device 10 can predict the future temperature distribution of the target space 40 with high accuracy and high speed. This allows the calculation device 10 to control the space control device 60 in real time so that the temperature distribution of the target space 40 is appropriate.

[0091] 7 and 8, the calculation device 10 can, for example, predict the temperature distribution in the target space 40 one second from now with high accuracy every second, and determine whether operation control of the space control device 60 is necessary at 100-second intervals. When determining whether operation control of the space control device 60 is necessary, the calculation device 10 can quickly predict the temperature distribution in the target space 40 100 seconds from now and determine whether operation control of the space control device 60 is necessary based on the prediction result. Note that the number of seconds described above is an example, and may be arbitrarily set by the user, for example.

[0092] In this way, the arithmetic device 10 can control the space control device 60 provided in the target space 40 using internal data indicating the state quantities of the target space 40. Specifically, the arithmetic device 10 inputs the state quantities of the target space 40 indicated by the corrected internal data to a trained model 50 that can output the state quantities of the target space 40 after a specific time in response to the input of the state quantities of the target space 40 at the specific time. At this time, the arithmetic device 10 corrects the internal data using measurement data of the state quantities acquired by a sensor 30 provided in the target space 40, and inputs the state quantities of the target space indicated by the corrected internal data to the trained model 50. The arithmetic device 10 repeatedly inputs the state quantities of the target space 40 output from the trained model 50 to the trained model 50. Then, the arithmetic device 10 controls the space control device 60 based on the state quantities of the target space 40 output from the trained model 50 and the preset state quantities of the target space 40. Here, the arithmetic device 10 may determine whether or not operation control of the space control device 60 is necessary. When the calculation device 10 determines that operation control of the space control device 60 is necessary, the calculation device 10 may execute operation control of the space control device 60. Note that the internal data may be generated by a simulation and may indicate the state quantity of the target space at a specified time. Furthermore, the calculation device 10 may control the space control device 60 or determine whether control of the space control device 60 is necessary at predetermined intervals.

[0093] (Variation) In the above embodiment, the calculation device 10 is configured using a general-purpose computer device such as a personal computer or a server computer. However, this is not limited to this, and the calculation device 10 may be configured as a cloud-based system on a network. This allows for flexible response to changes in computational resources in response to changes in the system configuration, such as an increase or decrease in the number of sensors 30 or space control devices 60 installed in the target space 40. Furthermore, configuring the calculation device 10 as a cloud-based system on a network can facilitate, for example, the collection and analysis of various log data related to the target space 40 and the performance improvement of the trained model 50 or the data assimilation module 12 based on the analysis results. Examples of log data include measurement data acquired by the sensors 30, records of operation control of the space control device 60, and state quantities output from the trained model 50.

[0094] In the above embodiment, the calculation device 10 has been described as repeating the analytical assimilation process a specified number of times (see FIG. 3 ). However, this is not limited thereto, and the calculation device 10 may perform the analytical assimilation process for each type of state quantity of the target space 40. For example, the calculation device 10 may perform the analytical assimilation process for each of the temperature distribution, humidity distribution, and wind speed distribution of the target space 40. In this case, the calculation device 10 may generate a trained model 50 for each type of state quantity of the target space 40. In the above embodiment, an example has been described in which the space state control system 2 controls the temperature distribution of the target space 40. However, the space state control system 2 may control each of the various state quantities of the target space 40 using the trained model 50 generated for each type of state quantity of the target space 40.

[0095] Furthermore, in the above embodiment, the control change determination process (see FIG. 8 ) uses the trained model 50, which outputs the state quantities of the target space 40 one second after the time corresponding to the input state quantities of the target space 40. Furthermore, since the necessity of operation control of the space control device 60 is determined based on the prediction result of the temperature distribution in the target space 40 after 100 seconds, the input of the state quantities of the target space 40 to the trained model 50 and the output of the state quantities of the target space 40 by the trained model 50 are repeatedly performed. However, this is not limited to this. For example, the trained model 50 may be generated to output the temperature distribution of the target space 40 after 100 seconds after the time corresponding to the input state quantities of the target space 40. As a result, for example, when the necessity of operation control of the space control device 60 is determined based on the prediction result of the temperature distribution in the target space 40 after 100 seconds, as in the example described with reference to FIG. 8 , the necessity of operation control of the space control device 60 can be determined without repeating the input to the trained model 50.

[0096] In the above embodiment, the control change determination process (see FIG. 8 ) illustrates an example in which a determination is made as to whether operation control of the space control device 60 is necessary based on the prediction results of the state quantities of the target space 40 after a set time (e.g., 100 seconds) from the current time. However, this is not limiting, and a determination as to whether operation control of the space control device 60 is necessary may be made based on the prediction results of the state quantities of the target space 40 over a predetermined period of time, in other words, based on the prediction results of the time series of the state quantities of the target space 40. For example, in the example described with reference to FIG. 8 , the state quantities of the target space 40 1 second after the current time, the state quantities 2 seconds after the current time, ..., the state quantities 100 seconds after the current time are repeatedly predicted, thereby obtaining a time change in the temperature distribution of the target space 40 over 100 seconds. For example, the calculation device 10 may determine that operation control of the space control device 60 is necessary if, during the time change in the temperature distribution of the target space 40 over 100 seconds, there is a temperature distribution that deviates from the target temperature distribution value by a specified value or more at a certain timing (e.g., 20 seconds after the 100 seconds). Alternatively, the calculation device 10 may determine that operation control of the space control device 60 is necessary, for example, when there is a period of drastic fluctuations in the temperature distribution of the target space 40 during a 100-second time change in the temperature distribution. In this way, the calculation device 10 may determine whether operation control of the space control device 60 is necessary based on a plurality of state quantities obtained by repeatedly inputting the state quantities of the target space 40 to the trained model 50 and outputting the state quantities of the target space 40 from the trained model 50, and based on a preset state quantity. In this way, the calculation device 10 can control the state quantities of the target space 40 based on the predicted results of the transitions of the state quantities of the target space 40 over time.

[0097] In the above embodiment, an example was shown in which the state quantities generated by the CAE tool 13 were corrected by data assimilation using measurement data and used to generate the trained model 50. As described with reference to FIG. 7 , an example was shown in which the state quantities after data assimilation are input to the trained model 50 during operation of the space state control system 2. A simulation performed to prepare training data can obtain the state quantities of the target space 40 before data assimilation, but it is difficult to obtain the state quantities of the target space 40 after data assimilation. Therefore, by preparing training data using data assimilation and inputting the state quantities of the target space 40 after data assimilation to the trained model 50 as in the above embodiment, it is possible to obtain more accurate prediction results than when data assimilation is not used. For example, the state quantities obtained by simulation may be corrected by a method other than data assimilation, and the corrected state quantities may be used as training data. For example, random noise or white noise may be added to the state quantities obtained by simulation to form the training data. In this case, for example, compared to when data assimilation is used as in the above embodiment, there is an advantage in that there is no need to install a sensor 30 in the actual target space 40.

[0098] <Summary of the First Embodiment> The above description of the first embodiment discloses at least the following techniques. Note that the components corresponding to the first embodiment are shown in parentheses, but the present invention is not limited to these.

[0099] <Technology A1> The trained model generation method involves correcting the state quantity (e.g., analysis output (t)) of the target space at time N (N: natural number) calculated by a simulation that outputs the state quantity of the target space (e.g., target space 40) using measurement data (e.g., measurement data (t)) of the state quantity at time N acquired by a sensor (e.g., sensor 30) installed in the target space, and performing a simulation using the corrected state quantity of the target space at time N (e.g., assimilation output (t)) to output the state quantity of the target space at time (N+1) (e.g., analysis output (t+1)). Based on training data using multiple pairs of the corrected state quantity of the target space at time N and the state quantity of the target space at time (N+1), a trained model (e.g., trained model 50) is generated that outputs the state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time.

[0100] As a result, the trained model generation method can correct the state quantities of the target space calculated by simulation using measurement data acquired by a sensor installed in the target space. Furthermore, the trained model generation method can obtain the state quantities of the target space by simulation using the corrected state quantities. As a result, the trained model generation method can prepare a training dataset for generating a trained model. The trained model generation method can generate a trained model that predicts the state quantities of the target space with high accuracy based on the training dataset.

[0101] <Technology A2> In the trained model generation method described in Technology A1, the (N+1)th time is a time that is a specified time after the Nth time, and the trained model generation method may repeat an analysis and assimilation process that repeatedly performs calculation processing a specified number of times until the (N+1)th time reaches a specified time.

[0102] This allows the trained model generation method to generate a large amount of training data.

[0103] <Technology A3> In the trained model generation method described in Technology A1 or Technology A2, the state quantities of the target space include at least one of the temperature distribution, humidity distribution, and wind speed distribution of the target space, and the trained model generation method may perform calculation processing for each type of state quantity of the target space.

[0104] As a result, the trained model generation method can generate a trained model that predicts the state quantities of the target space with high accuracy for each type of state quantity of the target space.

[0105] <Technology A4> The trained model generation program causes the arithmetic device to repeatedly perform a calculation process in which the state quantity of the target space at the Nth time (N: natural number) calculated by a simulation that outputs the state quantity of the target space is corrected using measurement data of the state quantity at the Nth time acquired by a sensor installed in the target space, and the state quantity of the target space at the (N+1)th time is output by performing a simulation using the corrected state quantity of the target space at the Nth time, and generates a trained model that outputs the state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time, based on learning data using multiple pairs of the corrected state quantity of the target space at the Nth time and the state quantity of the target space at the (N+1)th time.

[0106] This allows the trained model generation program to achieve the same effect as technique A1.

[0107] <Technology B1> A space state control method for controlling equipment (e.g., space control device 60) installed in a target space using data indicating the state quantities of the target space corrects the data using measurement data of the state quantities acquired by a sensor installed in the target space, inputs the state quantities of the target space indicated by the corrected data into a trained model that can output the state quantities of the target space after a specific time in response to input of the state quantities of the target space at the specific time, repeatedly inputs the state quantities of the target space output from the trained model into the trained model, and controls the equipment based on the state quantities of the target space output from the trained model and the preset state quantities of the target space.

[0108] This allows the space state control method to control the state quantities of the target space with high precision at preset intervals. The space state control method can correct data indicating the state quantities of the target space using measurement data of the state quantities acquired by a sensor installed in the target space. The space state control method can also input the corrected data into a trained model. The space state control method can then predict, for example, future temperature distribution in the target space and control equipment based on the prediction results.

[0109] <Technology B2> In the space state control method described in Technique B1, the state quantity of the target space may include at least one of a temperature distribution, a humidity distribution, and a wind speed distribution of the target space.

[0110] This allows the space condition control method to control at least one of the temperature distribution, humidity distribution, and wind speed distribution in the target space.

[0111] <Technology B3> In the space state control method according to Technique B1 or Technique B2, the data may be generated by a simulation that outputs a state quantity of the target space at a specified time.

[0112] This allows the spatial state control method to input data indicating the state quantities of the target space generated by simulation into the trained model.

[0113] <Technology B4> The spatial state control method described in any one of Techniques B1 to B3 may control equipment based on a plurality of state quantities of a target space output from a trained model and a predetermined state quantity of a target space.

[0114] This allows the space state control method to control equipment based on the predicted results of the time-dependent transition of the state quantity of the target space.

[0115] <Technology B5> The space state control program is a space state control program that causes a calculation device to control equipment installed in a target space using data indicating the state quantities of the target space, corrects the data using measurement data of the state quantities acquired by a sensor installed in the target space, inputs the state quantities of the target space indicated by the corrected data into a trained model that can output the state quantities of the target space after a specific time in accordance with the input of the state quantities of the target space at the specific time, repeatedly inputs the state quantities of the target space output from the trained model into the trained model, and controls the equipment based on the state quantities of the target space output from the trained model and the predetermined state quantities of the target space.

[0116] This allows the space state control program to obtain the same effect as technique B1.

[0117] <Technology B6> The space state control system includes equipment installed in the target space and a computing device that controls the equipment using data indicating the state quantities of the target space. The computing device corrects the data using measurement data of the state quantities acquired by a sensor installed in the target space, inputs the state quantities of the target space indicated by the corrected data into a trained model that can output the state quantities of the target space after a specific time in response to the input of the state quantities of the target space at the specific time, repeatedly inputs the state quantities of the target space output from the trained model into the trained model, and controls the equipment based on the state quantities of the target space output from the trained model and the preset state quantities of the target space.

[0118] This allows the space state control system to obtain the same effect as technology B1.

[0119] Although the embodiments of the present disclosure have been described above with reference to the drawings, the present disclosure is not limited to such examples. It is clear that a person skilled in the art can conceive of various modifications, alterations, substitutions, additions, deletions, and equivalents within the scope of the claims, and it is understood that these also fall within the technical scope of the present disclosure. Furthermore, the components of the above-described embodiments may be combined in any manner without departing from the spirit of the invention. [Industrial Applicability]

[0120] The present disclosure is useful as a trained model generation method and a trained model generation program. [Explanation of symbols]

[0121] 1. Trained model generation system 2. Space State Control System 10 Arithmetic unit 11 processors 12 Data Assimilation Module 13 CAE Tools 14 Model Generation Tools 15 memory 16 Input Devices 17 Display device 18. Communications Equipment 19 External interface device 20 Internal interface device 30 sensors 40 Target Space 50 pre-trained models 60 Space Control Device 70 databases

Claims

1. correcting a state quantity of the target space at an N-th time (N: natural number) calculated by a simulation that outputs a state quantity of the target space using measurement data of the state quantity at the N-th time acquired by a sensor provided in the target space; performing the simulation using the corrected state quantity of the object space at the Nth time, thereby outputting the state quantity of the object space at the (N+1)th time; Repeated calculations are performed, generating a trained model that outputs a state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time, based on training data using a plurality of pairs of the corrected state quantity of the target space at the Nth time and the state quantity of the target space at the (N+1)th time; How to generate a trained model.

2. the (N+1) time is a time that is a specified time later than the N time, repeating the analysis and assimilation process, which repeatedly performs the calculation process, a specified number of times until the time (N+1) reaches a specified time; The trained model generation method according to claim 1 .

3. the state quantity of the target space includes at least one of a temperature distribution, a humidity distribution, and a wind speed distribution of the target space; performing the calculation process for each type of state quantity of the target space; The trained model generation method according to claim 1 .

4. The computing device correcting a state quantity of the target space at an Nth time (N: natural number) calculated by a simulation that outputs a state quantity of the target space using measurement data of the state quantity at the Nth time acquired by a sensor provided in the target space; performing the simulation using the corrected state quantity of the object space at the Nth time, thereby outputting the state quantity of the object space at the (N+1)th time; Repeatedly perform the calculation process, generating a trained model that outputs a state quantity of the target space after a specific time in response to an input of the state quantity of the target space at the specific time, based on training data using a plurality of pairs of the corrected state quantity of the target space at the Nth time and the state quantity of the target space at the (N+1)th time; A trained model generator.

Images