Computer system and information processing method

Abstract

This computer system holds information that pertains to a prediction algorithm and a data set that includes pieces of observation data which contain observed values used for controlling a system being controlled. The computer system generates predicted observation data by replacing an observed value of a piece of target observation data in the data set with an observed value that is calculated on the basis of the prediction algorithm, and uses the data set and the predicted observation data to generate a model for correcting an observed value. In the model generation process, an observed value contained in the predicted observation data is replaced with a correction result that is output from the model to which the predicted observation data and observation data other than the target observation data have been input, a risk evaluation value that represents a risk regarding the control of the system is calculated using the data set and the predicted observation data, and the model is updated using the risk evaluation value so as to lower the risk.

Description

Computer System and Information Processing Method Incorporation by Reference

[0001] This application claims the priority of Japanese Patent Application No. 2024-196872 filed on November 11, 2024, and incorporates the content thereof by reference into this application.

[0002] The present invention relates to a technique for complementing data deficiencies.

[0003] Techniques for controlling a system based on observed values obtained from a system to be controlled composed of elements such as devices and software are known. For example, an EMS (Energy Management System) obtains the power generation amount or charging amount of power generation / electric power facilities as observed values and performs efficient operation of the facilities based on the observed values.

[0004] There are cases where observed values cannot be obtained due to various factors such as communication errors. As the number of connected devices and the communication frequency increase, many deficiencies in observed values occur. In real-time control, the deficiencies in observed values as described above become a major problem. As a technique for complementing missing values, for example, the technique described in Patent Document 1 is known.

[0005] Patent Document 1 describes that "a data complementing server includes a receiving unit that receives power generation amount data indicating the power generation amount of a distributed power source that generates power using natural energy from the distributed power source, and a control unit that complements the deficiency of the power generation amount data during a deficiency period of the power generation amount data. When the model of the distributed power source is a target model in which the transmission of the power generation amount data is performed by the generated electric power of the distributed power source, the control unit performs zero complementation to complement the power generation amount during the deficiency period as zero."

[0006] Japanese Patent Application Laid-Open No. 2019-208341

[0007] In the prior art, observed values are complemented from the perspective of minimizing the error between the actual observed values and the complemented observed values. Even when the error is small, there is a possibility that a large difference may occur in the control of the system. Therefore, it is necessary to complement the observed values in consideration of the risk related to the control of the system.

[0008] The present invention aims to realize a method for supplementing observed values ​​that takes into account the risks related to system control.

[0009] A typical example of the invention disclosed in this application is as follows: A computer system comprising a processor and a storage device connected to the processor, wherein the storage device stores a dataset consisting of a plurality of observation data, including observation values ​​used to control a system to be controlled and the time when the observation values ​​were acquired, and information of a prediction algorithm for predicting the observation values, the processor selects one target observation data from the dataset, replaces the observation values ​​included in the target observation data with observation values ​​calculated based on the prediction algorithm to generate predicted observation data, and uses the dataset and the predicted observation data to execute a model generation process for generating a model that corrects the observation values ​​predicted based on the prediction algorithm, the model is the predicted observation data The model accepts a data set and at least one of the aforementioned observation data, excluding the predicted observation data, as input and outputs a correction result of the predicted observation value. The processor repeatedly performs the following steps in the model generation process: a first process of setting an initial model and replacing the observation value included in the predicted observation data with the correction result output from the model that has been input with at least one of the aforementioned observation data included in the dataset other than the predicted observation data and the target observation data; a second process of calculating a risk assessment value representing the risk related to the control of the system using the dataset and the predicted observation data; and a third process of updating the model to reduce the risk based on an update algorithm using the risk assessment value.

[0010] According to the present invention, observed values ​​can be supplemented while taking into account risks related to system control. Other issues, configurations, and effects will be clarified by the following description of the embodiments.

[0011] This figure shows an example of the configuration of the correction model generation device in Example 1. This figure shows an example of the data structure of the observation data DB in Example 1. This figure shows an example of the data structure of the prediction algorithm DB in Example 1. This figure shows an example of the data structure of the risk assessment algorithm DB in Example 1. This figure shows an example of the data structure of the correction model DB in Example 1. This figure shows an example of the screen displayed by the correction model generation device in Example 1. This is a flowchart illustrating an example of the processing performed by the correction model generation device in Example 1. This figure shows a specific example of the processing performed by the correction model generation device in Example 1. This figure shows an example of the screen displayed by the correction model generation device in Example 1. This is a flowchart illustrating an example of the correction model generation process performed by the correction model generation device in Example 1. This figure illustrates an example of the method for calculating the risk assessment value in Example 1. This figure shows an example of the system configuration in Example 2. This figure shows an example of the system configuration in Example 3.

[0012] The embodiments of the present invention will be described below with reference to the drawings. However, the present invention is not to be construed as being limited to the embodiments described below. It will be readily apparent to those skilled in the art that the specific configuration can be modified without departing from the spirit or intent of the present invention.

[0013] In the configuration of the invention described below, identical or similar components or functions are denoted by the same reference numerals, and redundant descriptions are omitted.

[0014] The designations "First," "Second," "Third," etc., used in this specification are for the purpose of identifying constituent elements and do not necessarily limit their number or order.

[0015] Figure 1 shows an example of the configuration of the correction model generation device of Example 1.

[0016] The correction model generation device 100 includes a processor 101, a storage device 102, an input device 103, an output device 104, and a network interface 105. Each hardware element is connected via a bus 106.

[0017] The processor 101 executes a program stored in the memory device 102. By executing processing according to the program, the processor 101 operates as a functional unit (module) that realizes a specific function. In the following description, when the processing is described with a functional unit as the subject, it indicates that the processor 101 is executing a program that realizes that functional unit.

[0018] The processor 101 in Embodiment 1 functions as a learning data generation unit 110, a completion unit 111, a risk assessment unit 112, and a correction model generation unit 113. Note that for each functional unit of the correction model generation device 100, multiple functional units may be combined into a single functional unit, or a single functional unit may be divided into multiple functional units according to its function.

[0019] The storage device 102 is a memory or the like, and stores the program executed by the processor 101 and various data. The storage device 102 also includes a work area used by the program. The correction model generation device 100 may have a large-capacity storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive).

[0020] The input device 103 is a device that accepts input such as data and commands, and is, for example, a keyboard, mouse, or touch panel. The output device 104 is a device that outputs various types of data, and is, for example, a display. The network interface 105 communicates with external devices via a network.

[0021] The storage device 102 stores the observation data DB 120, the prediction algorithm DB 121, the risk assessment algorithm DB 122, and the correction model DB 123.

[0022] The observation data DB120 is a database for managing observation data acquired from the controlled system. The observation data includes the observed values ​​used to control the system and the date and time when the observed values ​​were acquired.

[0023] The system to be controlled is a power system that includes elements such as power generation equipment, charging equipment, power grids, and demand generation. The observed values ​​are, for example, the amount of power generated by a power generation device and the amount of energy stored by an energy storage device.

[0024] The prediction algorithm DB121 is a database for managing prediction algorithms used to predict missing observations.

[0025] The Risk Assessment Algorithm DB122 is a database for managing algorithms for evaluating the risks to system control based on observational data supplemented with predicted observations.

[0026] Risks include, for example, financial losses in the operation of the system, the stability of the system, and the security of the system.

[0027] The correction model DB123 is a database for managing correction models that correct interpolated observations. A correction model is a mathematical formula or a machine learning model such as a neural network. The correction model takes observation data containing the interpolated observation data and at least one observation data from the training data used for learning, excluding the interpolated observation data, as input, and outputs a corrected value for the interpolated observations.

[0028] The training data generation unit 110 generates training data for generating a correction model. The interpolation unit 111 predicts missing observations based on a prediction algorithm and interpolates them using the predicted observations. The risk assessment unit 112 assesses the risk based on a risk assessment algorithm. The correction model generation unit 113 generates a correction model.

[0029] The functions of the correction model generation device 100 may be implemented using a computer system composed of multiple computers.

[0030] The database managed by the correction model generation device 100 will be explained using Figures 2 to 5.

[0031] Figure 2 shows an example of the data structure of the observation data DB120 in Example 1.

[0032] The observation data DB 120 stores, for example, tables 200, as shown in Figure 2, for each type of element constituting the system being controlled. Table 200 stores records containing the date and time 201 and the observed value 202. Each record corresponds to one observation data.

[0033] Observation value 202 is a field that stores the observed value. Date and time 201 is a field that stores the date and time when the observed value was obtained.

[0034] Figure 3 shows an example of the data structure of the prediction algorithm DB121 in Example 1.

[0035] The prediction algorithm DB 121 stores, for example, a table 300 as shown in Figure 3. Table 300 stores records containing an algorithm ID 301, a type 302, and a prediction algorithm 303. Each record corresponds to one prediction algorithm.

[0036] Algorithm ID 301 is a field that stores the ID of the prediction algorithm. Type 302 is a field that stores the type of prediction algorithm. Possible types of prediction algorithms include rule-based algorithms, statistical methods, and machine learning models. Prediction algorithm 303 is a field that stores the program (tool) and configuration information for implementing the prediction algorithm.

[0037] If type 302 is "machine learning model", the prediction algorithm 303 stores hyperparameters and neural network weights, etc.

[0038] Figure 4 shows an example of the data structure of the risk assessment algorithm DB122 in Example 1.

[0039] The risk assessment algorithm DB122 stores, for example, a table 400 as shown in Figure 4. Table 400 stores records containing algorithm ID 401 and risk assessment algorithm 402. Each record corresponds to one prediction algorithm.

[0040] The algorithm ID 401 is a field for storing the ID of the risk assessment algorithm. The risk assessment algorithm 402 is a field for storing a program (tool) for implementing the risk assessment algorithm, setting information, and the like.

[0041] In the risk assessment algorithm 402, for example, a simulator that simulates system control using observation data including complemented observation values and a risk assessment formula based on the simulation result of system control are stored.

[0042] FIG. 5 is a diagram showing an example of the data structure of the correction model DB 123 in the first embodiment.

[0043] The correction model DB 123 stores, for example, a table 500 as shown in FIG. 5. The table 500 stores records including a model ID 501, a characteristic 502, a type 503, and a correction model 504. One record corresponds to one correction model.

[0044] The model ID 501 is a field for storing the ID of the correction model. The characteristic 502 is a field for storing the characteristics of the correction model. For example, the seasonality of the input observation data and the type of the observation value are stored. The type 503 is a field for storing the type of the correction model. As the type of the correction model, a machine learning model, a mathematical formula, and the like can be considered.

[0045] The correction model 504 is a field for storing information for defining the correction model. When the correction model is a machine learning model, hyperparameters, weights of the neural network, and the like are stored in the correction model 504. When the correction model is a mathematical formula, the mathematical formula and the definition of the parameters of the mathematical formula are stored in the correction model 504.

[0046] The process executed by the correction model generation device 100 will be described using FIGS. 6 to 9.

[0047] FIG. 6 is a diagram showing an example of a screen presented by the correction model generation device 100 of Example 1. FIG. 7 is a flowchart for explaining an example of the process executed by the correction model generation device 100 of Example 1. FIG. 8 is a diagram showing a specific example of the process of the correction model generation device 100 of Example 1. FIG. 9 is a diagram showing an example of a screen presented by the correction model generation device 100 of Example 1.

[0048] When the correction model generation device 100 receives an access from a user, it presents a screen 600. The screen 600 includes input fields 601, 602, 603, 604, 605 and a button 606.

[0049] The input field 601 is a field for inputting learning data. Here, the learning data is composed of observation data arranged in chronological order. The observation data included in the learning data is collected in advance from the system. In the input field 601, the path of a file storing the learning data or the like is input.

[0050] The input field 602 is a field for inputting a prediction algorithm. When a prediction algorithm is set in advance, the prediction algorithm is displayed in the input field 602 in a pull-down format. When a prediction algorithm is not set in advance, the path of a file storing the prediction algorithm or the like is input in the input field 602.

[0051] The input field 603 is a field for inputting a risk assessment algorithm. When a risk assessment algorithm is set in advance, the risk assessment algorithm is displayed in the input field 603 in a pull-down format. When a risk assessment algorithm is not set in advance, the path of a file storing the risk assessment algorithm or the like is input in the input field 603.

[0052] The input field 604 is a field for inputting parameters used in the correction model generation process.

[0053] The input field 605 is a field for inputting the output destination of information regarding the correction model. In the input field 605, the path of a file for saving the correction model or the like is input.

[0054] Button 606 is a button used to instruct the correction model generation device 100 to execute a process.

[0055] The user enters various information into input fields 601, 602, 603, 604, and 605, and then presses button 606. The correction model generation device 100 starts processing when button 606 is pressed.

[0056] The correction model generation device 100 performs initial setup according to the inputs in input fields 601, 602, 603, 604, and 605 (step S101).

[0057] Specifically, the correction model generation device 100 stores the learning data in the observation data DB 120 and sets the prediction algorithm and the risk assessment algorithm.

[0058] As shown in Figure 8, the training data generation unit 110 generates missing training data by deleting the observed value of one of the observed data included in the training data (step S102). The observed data from which the observed value is deleted may be selected randomly or based on an arbitrary rule.

[0059] In the following explanation, observation data from which observation values ​​have been removed will be referred to as target observation data, observation data from which observation values ​​have been removed will be referred to as missing observation data, observation data from which observation values ​​have been filled in will be referred to as filled observation data, and observation data from which observation values ​​have been corrected will be referred to as corrected observation data. Furthermore, training data in which target observation data has been replaced with missing observation data will be referred to as missing training data, training data in which target observation data has been replaced with filled observation data will be referred to as filled training data, and training data in which target observation data has been replaced with corrected observation data will be referred to as corrected training data.

[0060] The interpolation unit 111 generates interpolated learning data by interpolating the observed values ​​of missing observation data included in the missing learning data, as shown in Figure 8, based on the prediction algorithm set in the input field 602 (step S103).

[0061] The correction model generation unit 113 executes a correction model generation process using the training data and complementary training data (step S104). Details of the correction model generation process will be described later.

[0062] The correction model generation unit 113 generates output information including information about the correction model and outputs the output information to the output destination according to the settings in the input field 605 (step S105).

[0063] For example, the correction model generation unit 113 calculates the prediction error and risk assessment value when using only the prediction algorithm, and also calculates the prediction error and risk assessment value when using both the prediction algorithm and the correction model, and includes the calculation results in the output information.

[0064] Based on the output information, a screen 900 as shown in Figure 9 is presented. The screen 900 includes tables 901 and 902 and a display field 903.

[0065] Table 901 is a table that displays training data, supplementary training data, and corrective training data.

[0066] Table 902 is a table that displays prediction accuracy and risk assessment values. Specifically, Table 902 displays the error (prediction accuracy) of the observed values ​​of the target observation data and the supplementary observation data, and the risk assessment value when the system is controlled based on the supplementary learning data. In addition, Table 902 also displays the error (prediction accuracy) of the observed values ​​of the target observation data and the corrected observation data, and the risk assessment value when the system is controlled based on the corrected learning data.

[0067] Display field 903 is a field that displays the difference between the observed values ​​when using only the prediction algorithm and the observed values ​​when using both the prediction algorithm and the correction model. If the prediction algorithm is an algorithm that determines the interpolated values ​​based on the probability distribution of the interpolated values, the probability distribution of the interpolated values ​​will be displayed in display field 903, as shown in Figure 9.

[0068] Figure 10 is a flowchart illustrating an example of the correction model generation process performed by the correction model generation device 100 of Example 1.

[0069] Here, the correction model is described as a machine learning model. In this case, the correction model accepts complementary training data as input and outputs corrected values ​​of the observed values ​​of the complementary observation data included in the complementary training data.

[0070] The correction model generation unit 113 initializes the parameters θ of the correction model (step S201). The parameters θ are, for example, weights. Note that any method can be used to initialize the parameters θ.

[0071] The correction model generation unit 113 initializes the variable j to 1 (step S202). The variable j represents the number of parameter candidate values ​​θ_j to be generated, and is an integer from 1 to J. J is an integer greater than 1.

[0072] The correction model generation unit 113 calculates the parameter candidate θ_j (step S203).

[0073] For example, the correction model generation unit 113 calculates a candidate parameter θ_j by adding a random perturbation Δθ to the parameter θ.

[0074] The correction model generation unit 113 generates correction learning data using a correction model in which the parameter candidate θ_j has been set (step S204).

[0075] Specifically, the correction model generation unit 113 calculates correction values ​​for observed values ​​of the complementary observation data included in the complementary learning data by inputting the complementary learning data into a correction model in which the parameter candidate θ_j has been set. The correction model generation unit 113 generates correction learning data by replacing the observed values ​​of the complementary observation data included in the complementary learning data with the correction values.

[0076] The correction model generation unit 113 calculates the difference between the observed value of the target observation data included in the training data and the observed value of the correction observation data included in the correction training data as the prediction error e_j (step S205).

[0077] The correction model generation unit 113 calls the risk assessment unit 112. The risk assessment unit 112 calculates the risk assessment value r_j based on the risk assessment algorithm set in the input field 603 (step S206). The risk assessment unit 112 outputs the risk assessment value r_j to the correction model generation unit 113.

[0078] For example, the risk assessment unit 112 calculates a risk assessment value r_j based on the simulation results of system control using complementary learning data and the simulation results of system control using corrected learning data. A specific example of how to calculate the risk assessment value r_j will be described later. Alternatively, the risk assessment unit 112 may calculate the risk assessment value r_j using a machine learning model that takes learning data and complementary learning data as input and outputs a risk assessment value r_j.

[0079] In this embodiment, it is assumed that the magnitude of the risk and the magnitude of the risk assessment value are proportional. That is, a large risk assessment value indicates a large risk, and a small risk assessment value indicates a small risk.

[0080] The correction model generation unit 113 calculates the loss l_j of the parameter candidate θ_j based on the prediction error e_j and the risk assessment value r_j (step S207).

[0081] For example, the loss l_j can be defined as shown in equation (1). Here, w is the weight set in input field 604, and is a value greater than 0 and less than 1. w is a parameter for adjusting the balance between prediction accuracy and risk assessment value. When generating a correction model that prioritizes prediction accuracy, the user sets w to a small value, and when generating a correction model that prioritizes risk assessment value, the user sets w to a large value.

[0082]

[0083] The correction model generation unit 113 determines whether or not the variable j is J (step S208).

[0084] If variable j is not J, the correction model generation unit 113 sets the value obtained by adding 1 to variable j as the new variable j (step S209), and then returns to step S203.

[0085] If the variable j is J, the correction model generation unit 113 selects N parameter candidates θ_j based on the loss l_j of the parameter candidate θ_j (step S210). Here, N is an integer less than J.

[0086] Specifically, the correction model generation unit 113 selects N parameter candidates θ_j in order of increasing loss l_j.

[0087] The correction model generation unit 113 updates the parameter θ using N parameter candidates θ_j (step S211).

[0088] For example, the correction model generation unit 113 updates the parameter θ using equation (2). Here, U is the set of variables j of the selected parameter candidate θ_j.

[0089]

[0090] By updating the parameter θ using a parameter candidate θ_j with a small loss l_j, i.e., a parameter with a small prediction error and risk assessment value, a corrected model with a small prediction error and risk can be generated.

[0091] The correction model generation unit 113 determines whether or not the termination condition is met (step S212).

[0092] Various termination conditions are possible. For example, if the number of parameter updates θ is greater than a threshold, the correction model generation unit 113 determines that the termination condition is met. Also, if the average value of the loss l_j of the N parameter candidates θ_j is less than a threshold, the correction model generation unit 113 determines that the termination condition is met.

[0093] If the termination conditions are not met, the correction model generation unit 113 returns to step S202. If the termination conditions are met, the correction model generation unit 113 terminates the series of processes.

[0094] Note that the method for generating the correction model shown in Figure 10 is just one example and is not limited thereto. For example, the correction model may be generated using reinforcement learning techniques.

[0095] Here, we will explain an example of a method for calculating risk assessment values. Figure 11 is a diagram illustrating an example of a method for calculating risk assessment values ​​in Example 1.

[0096] Figure 11 illustrates the control of charging and discharging of the battery by the Energy Management System (EMS). The solar panels supply power to the battery. When power supply to the battery is not needed, the solar panels supply power to electricity consumers. The battery stores power supplied from the solar panels or the power grid and supplies power to electricity consumers.

[0097] The EMS acquires observational data, including the power generation amount of the solar panels, and controls the charging and discharging of the battery. Let x_t be the actual power generation amount of the solar panels at time t. The observational data includes the power generation amount x_t at time t. If the power generation amount in the observational data for time t is missing, let x^_t be the power generation amount calculated using the prediction algorithm and correction model. Let B_t be the remaining charge of the battery at time t. Let d_t be the power demand of the electricity consumer at time t.

[0098] The EMS controls the charging and discharging of the battery at time t based on the difference between x_t or x^_t and d_t. Here, the difference between x_t or x^_t and d_t is denoted as b_t. Also, the absolute value of the difference between x_t and x^_t is denoted as v_t.

[0099] (Case 1) Battery charging control (b_t > 0) When x_t is less than x^_t, it is necessary to purchase v_t of electricity from the power grid. In this case, the value obtained by multiplying the purchase price C(-) by v_t is the expenditure. When x_t is greater than x^_t, it is possible to sell v_t of electricity to electricity consumers. In this case, the value obtained by multiplying the selling price C(+) by v_t is the revenue.

[0100] (Case 2) Battery discharge control (b_t < 0) When the remaining charge of the battery is insufficient, that is, when B_t is less than (-b_t), the cost is the sum of the purchased electricity unit price C(-) plus B_t and b_t.

[0101] When the expenditure or income at time t is defined as cost c_t, the risk assessment unit 112 calculates the sum of costs c_t over the time range in the training data as a risk index, as shown in equation (3).

[0102]

[0103] According to Example 1, the correction model generation device 100 generates a correction model that corrects the interpolated values ​​to reduce the risk using a risk assessment value representing the risk related to system control. By using this correction model, it is possible to suppress the risk related to system control and achieve highly accurate interpolation of defect values.

[0104] Example 2 describes a system that actually corrects missing values ​​using a correction model. Below, Example 2 will be described focusing on the differences from Example 1.

[0105] Figure 12 shows an example of the system configuration of Example 2.

[0106] The system consists of a correction model generation device 100 and a complementary device 1200. The correction model generation device 100 and the complementary device 1200 are connected via a network 1201 such as a LAN (Local Area Network).

[0107] The hardware and software configurations of the correction model generation device 100 in Example 2 are the same as those in Example 1. The method for generating the correction model in Example 2 is the same as in Example 1.

[0108] The supplementary device 1200 acquires correction model information from the correction model generation device 100 in advance. The correction model to be acquired can be selected according to the type of observation data included in the observation data and the timing of the observation data acquisition.

[0109] The interpolation device 1200 accepts the input of observation data. If there is observation data with missing values, the interpolation device 1200 predicts the observed values ​​based on the prediction algorithm, corrects the predicted observed values ​​using the correction model generated by the correction model generation device 100, and interpolates the observed values ​​of the observation data.

[0110] The hardware configuration of the interpolation device 1200 is the same as that of the correction model generation device 100, so its description is omitted. Functionally, the interpolation device 1200 has an interpolation unit 1210 and a correction unit 1211, and also holds a correction model DB 1220. The correction model DB 1220 stores information on the correction model acquired from the correction model generation device 100.

[0111] The interpolation unit 1210 interpolates the observed values ​​of the observation data based on a pre-set prediction algorithm. The correction unit 1211 corrects the interpolated observed values ​​using a correction model.

[0112] The functional configuration of the correction model generation device 100 and the interpolation device 1200 may be implemented using a single computer or using a computer system composed of multiple computers.

[0113] According to Example 2, the prediction error is small, and the observed values ​​can be supplemented to reduce the risk related to the control of the controlled system.

[0114] Example 3 describes a correction method for controlling a system using prediction results from a prediction algorithm. The system to be controlled is, for example, a logistics system and a manufacturing system.

[0115] Figure 13 shows an example of the system configuration of Example 3.

[0116] The system consists of a correction model generation device 100 and a prediction device 1300. The correction model generation device 100 and the prediction device 1300 are connected via a network 1301 such as a LAN (Local Area Network).

[0117] The hardware configuration of the correction model generation device 100 in Example 3 is the same as that of Example 1. The software configuration of the correction model generation device 100 in Example 3 differs in part from that of Example 1. Specifically, the correction model generation device 100 in Example 3 has a prediction unit 1311 instead of a interpolation unit 111. The prediction unit 1311 predicts future observed values ​​based on a prediction algorithm. Other functional units are the same as those in Example 1. Also, the method for generating the correction model in Example 3 is the same as that in Example 1.

[0118] The prediction algorithm DB121 in Example 3 manages prediction algorithms for predicting observed values ​​from past observation data.

[0119] The correction model of Example 3 takes observational data, including observational values ​​predicted based on the prediction algorithm, and at least one past observational data as input, and outputs a corrected value of the predicted observational value.

[0120] The prediction device 1300 obtains correction model information from the correction model generation device 100 in advance. The correction model to be obtained can be selected according to the type of observation data included in the observation data and the timing of the observation data acquisition.

[0121] The prediction device 1300 accepts the input of observation data. The prediction device 1300 predicts the observed values ​​based on the prediction algorithm and corrects the predicted observed values ​​using the correction model generated by the correction model generation device 100.

[0122] The hardware configuration of the prediction device 1300 is the same as that of the correction model generation device 100, so a description will be omitted. The prediction device 1300 has a prediction unit 1320 and a correction unit 1321 as its functional configuration, and also holds a correction model DB 1330. The correction model DB 1330 stores information on the correction model acquired from the correction model generation device 100.

[0123] The prediction unit 1320 predicts observed values ​​based on a pre-set prediction algorithm. The correction unit 1321 corrects the predicted observed values ​​using a correction model.

[0124] The functional configuration of the correction model generation device 100 and the prediction device 1300 may be implemented using a single computer or using a computer system composed of multiple computers.

[0125] According to Example 3, it is possible to predict observed values ​​that have a small prediction error and reduce the risk related to the control of the controlled system.

[0126] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are included. Furthermore, for example, the embodiments described above are detailed explanations of the configuration in order to clearly illustrate the present invention, and are not necessarily limited to those having all the configurations described. In addition, some of the configurations in each embodiment can be added to, deleted from, or replaced with other configurations.

[0127] Furthermore, each of the above-mentioned configurations, functions, processing units, processing means, etc., may be implemented in hardware, either partially or entirely, by designing them as integrated circuits, for example. The present invention can also be implemented by software program code that realizes the functions of the embodiment. In this case, a storage medium on which the program code is recorded is provided to a computer, and the processor of that computer reads the program code stored in the storage medium. In this case, the program code read from the storage medium itself realizes the functions of the embodiment described above, and the program code itself and the storage medium on which it is stored constitute the present invention. Examples of storage media used to supply such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, SSDs (Solid State Drives), optical disks, magneto-optical disks, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, and the like.

[0128] Furthermore, the program code that implements the functions described in this embodiment can be implemented in a wide range of programming or scripting languages, such as assembler, C / C++, Perl, Shell, PHP, Python, and Java.

[0129] Furthermore, the program code for the software that implements the functions of the embodiment may be distributed via a network and stored in a storage means such as a computer's hard disk or memory, or in a storage medium such as a CD-RW or CD-R, and the computer's processor may read and execute the program code stored in the storage means or storage medium.

[0130] In the above-described embodiment, the control lines and information lines shown are those deemed necessary for explanation and do not necessarily represent all control lines and information lines in the actual product. All components may be interconnected.

Claims

1. A computer system comprising a processor and a storage device connected to the processor, wherein the storage device stores a dataset consisting of multiple observation data, including observation values ​​used to control a system to be controlled and the time when the observation values ​​were acquired, and information on a prediction algorithm for predicting the observation values, the processor selects one target observation data from the dataset, replaces the observation values ​​included in the target observation data with observation values ​​calculated based on the prediction algorithm to generate predicted observation data, executes a model generation process to generate a model for correcting the predicted observation values ​​based on the prediction algorithm using the dataset and the predicted observation data, the model is a model that accepts the predicted observation data and at least one of the observation data excluding the predicted observation data as input and outputs the correction result of the predicted observation values, the processor sets an initial model in the model generation process, A computer system characterized by repeatedly performing the following: a first process of replacing the observed values ​​included in the predictive observation data with the correction result output from the model into which at least one of the observed data included in the dataset other than the predictive observation data and the target observation data is input; a second process of calculating a risk assessment value representing the risk related to the control of the system using the dataset and the predictive observation data; and a third process of updating the model to reduce the risk based on an update algorithm using the risk assessment value.

2. A computer system according to claim 1, wherein the processor, in the second process, simulates the control of the system to be controlled using the dataset and the predictive observation data, and calculates the risk assessment value based on the results of the simulation.

3. A computer system according to claim 2, wherein the model is a machine learning model, and the update algorithm is a machine learning algorithm that uses a loss calculated using the risk assessment value.

4. A computer system according to claim 3, wherein the processor calculates, in the second process, the error between the observed value included in the target observation data and the observed value included in the prediction observation data, and the update algorithm is a machine learning algorithm that uses a loss calculated using the error and the risk assessment value.

5. A computer system according to claim 3, wherein the processor acquires new observation data, predicts the observed values ​​based on the new observation data and the prediction algorithm, and outputs the correction result using the model.

6. A computer system according to claim 3, wherein the prediction algorithm is an algorithm for predicting observed values ​​to supplement the observed data in which the observed values ​​are missing.

7. A computer system according to claim 3, characterized in that the risk is at least one of the following: financial loss in the operation of the system, stability of the system, and safety of the system.

8. An information processing method to be performed by a computer system, wherein the computer system has a processor and a storage device connected to the processor, the storage device stores a dataset consisting of a plurality of observation data, including observation values ​​used to control a system to be controlled and the time when the observation values ​​were acquired, and information of a prediction algorithm for predicting the observation values, the information processing method includes: a first step in which the processor selects one target observation data from the dataset and generates predicted observation data by replacing the observation values ​​included in the target observation data with observation values ​​calculated based on the prediction algorithm; a second step in which the processor uses the dataset and the predicted observation data to perform a model generation process for generating a model that corrects the observation values ​​predicted based on the prediction algorithm, the model is a model that accepts the predicted observation data and at least one of the observation data excluding the predicted observation data as input and outputs a correction result of the predicted observation values, the second step includes: the processor setting an initial model, An information processing method characterized by the processor repeatedly performing the following steps: replacing the observed values ​​included in the predictive observation data with the correction results output from the model that has been input to at least one of the observed data included in the dataset other than the predictive observation data and the target observation data; calculating a risk assessment value representing the risk related to the control of the system using the dataset and the predictive observation data; and updating the model based on an update algorithm using the risk assessment value so that the risk is reduced.

Images