Prediction device and prediction method

Abstract

To appropriately maintain the accuracy of prediction based on the data of at least one factor among a plurality of factors that affect the object to be predicted.SOLUTION: A prediction device according to the present invention builds a prediction model for each window width, on the basis of prediction target data for samples equal in quantity to window widths (time-series data of past values obtained regarding the prediction target) and factor data for samples equal in quantity to window widths (data having for each factor the time series of past values obtained regarding factors that can affect the prediction target), and calculates the degree of model suitability that is an index value relating to the prediction model. The prediction device inputs data having the time series of values for each factor that were used for building the prediction model, to the prediction model based on the predication target data for second samples equal in quantity to window widths that satisfy conditions and correspond to the model suitability, and factor data for second sample. The window widths represent the length of a past period reckoned from a prediction target date / time of the prediction target.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention generally relates to a technique for predicting future values for a prediction target.

Background Art

[0002] In energy business fields such as power and gas businesses, communication business fields, and transportation business fields such as taxis and deliveries, equipment operation and resource allocation are planned and executed in accordance with consumer demand based on predicted values of various indicator values such as demand up to a predetermined time in the future.

[0003] For example, it is known that power demand and power trading prices are affected by temperature. In this case, a prediction model for power demand and power trading prices is identified with outside air temperature as an explanatory variable (an example of a factor). A predicted value up to a predetermined time in the future is output using the identified prediction model. For example, when the prediction target is the power trading price, the predicted value is a value such as A yen or B yen.

[0004] There can be multiple types of factors. For example, in the case of power demand, temperature and humidity can be factors. Also, when predicting wide-area power demand, temperature and humidity at each of multiple locations are factors.

[0005] Also, there can be multiple factors of the same type. For example, when the type of factor is temperature, factors such as temperature can exist for each time zone or region (for example, multiple factors such as the temperature in the 1 o'clock hour, the temperature in the 2 o'clock hour, …, or multiple factors such as the temperature in the Kanto region and the temperature in the Tohoku region can exist).

[0006] When there are multiple factors as described above, the multiple factors include one or more types of factors.

[0007] Patent Document 1 discloses a method of generating an input-output model representing the relationship between each input candidate variable and an output variable, generating the sensitivity of each input variable to the output variable, and selecting and presenting the input variable based on the sensitivity.

[0008] Patent Document 2 discloses a method for generating a location model that outputs demand forecast values ​​for each location in a geographic area having multiple locations, and for generating an overall model based on coefficients set based on the contribution of weather values ​​at each location to the demand forecast values ​​and the location model. [Prior art documents] [Patent Documents]

[0009] [Patent Document 1] Japanese Patent Publication No. 2010-282547 [Patent Document 2] Japanese Patent Publication No. 2019-87027 [Overview of the project] [Problems that the invention aims to solve]

[0010] Among multiple factors, the major factors (factors that strongly influence the prediction target) are not always the same and change over time. For example, electricity demand is one of the major factors influencing electricity trading prices, and it is known that the degree of electricity demand congestion affects electricity trading prices. However, the areas where electricity demand is congested are not always the same, depending on weather conditions.

[0011] Thus, the degree to which each of the multiple factors influences the target of prediction changes over time. However, in Patent Documents 1 and 2, the degree to which the factors influence the target of prediction remains constant. Therefore, it is difficult to maintain appropriate prediction accuracy. [Means for solving the problem]

[0012] The prediction device outputs a predicted value for the target by inputting at least a portion of the first prediction factor data (data containing time series of future values ​​obtained for each factor that may influence the target of prediction) into a prediction model that takes the data of the values ​​of factors that may influence the target of prediction as input and outputs a value for the target of prediction. For each of several different window widths, the prediction device constructs a prediction model based on the second sample prediction target data, which is the data for that window width from the first sample prediction target data (time series data of past values ​​obtained for the target of prediction), and the second sample factor data, which is the data for that window width from the first sample factor data (data containing time series of past values ​​obtained for each factor that may influence the target of prediction), and calculates the model fit, which is an index value for the prediction model. The prediction device takes the second prediction factor data, which is at least a portion of the first prediction factor data, as input to the target prediction model. The target prediction model is a prediction model based on the second sample prediction target data and the second sample factor data for the window width corresponding to the model fit that satisfies predetermined conditions among the calculated model fit values. The second set of predictive factor data consists of time-series data representing the values ​​obtained for each factor used in constructing the target predictive model, from among the multiple factors represented by the first set of predictive factor data. The window width is the length of the past period starting from the target date and time of the prediction target. [Effects of the Invention]

[0013] According to the present invention, the accuracy of predictions based on data of the value of at least one of several factors that may influence the target of prediction can be appropriately maintained. [Brief explanation of the drawing]

[0014] [Figure 1] This figure shows the device configuration of the data prediction system according to the first embodiment. [Figure 2] This diagram shows the configuration of the prediction device. [Figure 3] This diagram shows the data flow of a data prediction system. [Figure 4] It is a diagram showing the processing flow of a data prediction system. [Figure 5] It is a diagram showing the data flow of a sample selection unit. [Figure 6] It is a diagram showing the data flow of a factor selection unit. [Figure 7] It is a diagram showing the data flow of an evaluation determination unit. [Figure 8] It is a diagram showing an example of processing data in an evaluation determination unit. [Figure 9] It is a diagram showing an example of processing data in an evaluation determination unit. [Figure 10] It is a diagram showing an example of processing data in an evaluation determination unit. [Figure 11] It is a diagram showing the data flow of a prediction unit. [Figure 12] It is a diagram showing an example of the effect of the first embodiment. [Figure 13] It is a diagram showing the data flow of factor selection processing according to the second embodiment. [Figure 14] It is a diagram showing the data flow of a data prediction system according to the fifth embodiment.

Embodiments for Carrying Out the Invention

[0015] In the following description, an "interface device" may be one or more interface devices. The one or more interface devices may be at least one of the following. · One or more I / O (Input / Output) interface devices. An I / O (Input / Output) interface device is an interface device for at least one of an I / O device and a remote display computer. The I / O interface device for the display computer may be a communication interface device. At least one I / O device may be either an input device such as a user interface device, for example, a keyboard and a pointing device, or an output device such as a display device. • One or more communication interface devices. One or more communication interface devices may be one or more identical communication interface devices (e.g., one or more NICs (Network Interface Cards)) or two or more different communication interface devices (e.g., a NIC and an HBA (Host Bus Adapter)).

[0016] Furthermore, in the following explanation, "memory" refers to one or more memory devices, which are typically main memory devices. At least one memory device in memory may be a volatile memory device or a non-volatile memory device.

[0017] Furthermore, in the following explanation, "persistent storage device" refers to one or more persistent storage devices. Persistent storage devices are typically non-volatile storage devices (e.g., auxiliary storage devices), specifically, for example, HDDs (Hard Disk Drives) or SSDs (Solid State Drives).

[0018] Furthermore, in the following explanation, "storage device" may refer to at least memory, including both memory and persistent storage.

[0019] Furthermore, in the following explanation, "processor" refers to one or more processor devices. At least one processor device is typically a microprocessor device such as a CPU (Central Processing Unit), but may be other types of processor devices such as a GPU (Graphics Processing Unit). At least one processor device may be single-core or multi-core. At least one processor device may be a processor core. At least one processor device may be a broader processor device such as a hardware circuit that performs some or all of the processing (e.g., an FPGA (Field-Programmable Gate Array) or ASIC (Application Specific Integrated Circuit)).

[0020] Furthermore, in the following explanation, functions may be described using the expression "yyy section," but a function may be implemented by the execution of one or more computer programs by a processor, by one or more hardware circuits (e.g., FPGA or ASIC), or by a combination thereof. When a function is implemented by the execution of a program by a processor, the defined processing is carried out using memory and / or interface devices as appropriate, so the function may be at least a part of the processor. Processing described with a function as the subject may be processing performed by the processor or a device having that processor. Programs may be installed from program source. Program source may be, for example, a program distribution computer or a computer-readable recording medium (e.g., a non-temporary recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

[0021] Several embodiments of the present invention will be described in detail below with reference to the drawings. (1) First Embodiment (1-1) Configuration of the data prediction system according to this embodiment

[0022] The data prediction system 1 according to this embodiment, when applied to the power business sector, for example, predicts values ​​such as the amount of electricity demand, the amount of electricity generated, and the market transaction price for a predetermined period in the future based on past electricity demand data. Based on the prediction results, it enables power supply and demand management, such as formulating and executing power generator operation plans and formulating and executing power procurement transaction plans from other electric power companies.

[0023] The data prediction system 1 consists of a prediction device 12 (in this embodiment, a device composed of a prediction calculation device 2 and a data management device 3), a planning management device 5, an information input / output terminal 4, a data observation device 6, a data distribution device 7, and a controlled device 9. The communication path 8 is a communication network such as a LAN (Local Area Network) or a WAN (Wide Area Network), and is a communication path that connects the various devices and terminals constituting the data prediction system 1 so that they can communicate with each other.

[0024] The data management device 3 stores sample data for the target of prediction and factors, as well as prediction data for factors, which are used to calculate the predicted values ​​of the target of prediction.

[0025] The sample data for prediction includes at least the sample prediction data, which is historical observational data of the prediction target observed over time. The sample data for factors includes at least the sample factor data, which is historical observational data of various factors that may influence the increase or decrease of the prediction value. The prediction data for factors includes at least the prediction data for each factor included in the sample factor data.

[0026] The targets of prediction may be, for example, at least some of the following, or other targets. • Consumption of energy such as electricity, gas, and water. • The amount of energy output from sources such as solar power and wind power (e.g., amount of electricity generated). • The volume and price of energy traded on wholesale exchanges. • The amount of data transmitted, as measured by communication base stations, etc. • Location history of moving objects such as automobiles.

[0027] The sample data to be predicted can be either data from a single measuring instrument or data from the sum of multiple measuring instruments.

[0028] Furthermore, the factors may be, for example, at least some of the following, or other subjects. Weather conditions such as temperature, humidity, solar radiation, wind speed, and atmospheric pressure. • Trading volume and price of fuels such as crude oil and natural gas. • Subject to transmission capacity and other aspects related to power transmission lines. • Generator operating status, such as the generator's operation or maintenance schedule. • Calendar date including year, month, and day, day of the week, and a flag value indicating the type of day (as arbitrarily set). • Whether or not sudden events such as typhoons or other unforeseen circumstances occur. • Economic conditions such as the number of energy consumers, industry trends, and business sentiment indices. • Information on the movement of people and vehicles, such as the occupancy rate of express trains, the number of passengers, the number of reserved seats, or road traffic conditions. • The number of communication terminals connected to a communication base station.

[0029] Factor data may include, in place of or in addition to the factor data described above, the historical observation data of the subject to prediction itself, or the predicted value itself for the subject to prediction at the specified date and time.

[0030] The data management device 3 stores sample data, including sample data from a pre-set past date and time to the most recent observation date and time, via the information input / output terminal 4. The data management device 3 also searches for and transmits sample data in response to data acquisition requests from other devices.

[0031] The prediction calculation unit 2 performs predictions using the data stored in the data management unit 3 and outputs prediction result data. Details of the prediction calculation unit 2 will be described later (the prediction calculation unit 2 and the data management unit 3 may be integrated).

[0032] The planning management device 5 creates an operation plan for physical equipment to achieve a predetermined target based on the prediction result data output by the prediction calculation device 2, and transmits control commands to the controlled device 9 based on the operation plan. In the energy sector, the operation plan for physical equipment is, for example, an operation plan for generators based on predicted future energy demand values, power output, and market transaction prices. Specifically, for example, the operation plan for physical equipment is a plan for the number of generators to be started and the distribution of their output, or a plan for the distribution of gas and water flow rates and pressures to be supplied to gas pipelines and water pipes. Alternatively, in the power demand adjustment control known as demand response, the operation plan for physical equipment may be a plan for the distribution of demand adjustment amounts to power consumers participating in demand response or to the demand equipment of power consumers (the adjustment control of this plan may be executed according to the control commands described above). In the telecommunications sector, the operation plan for physical equipment is, for example, a control plan for the number of communication terminals connected to each communication base station so as not to exceed the capacity of the communication base station. In the transportation sector, the operation plan for physical equipment is, for example, a plan for dispatching or driving vehicles (e.g., taxis) that can meet the predicted number of users. Therefore, the example of the controlled device 9 may be any of the following: a generator, a generator control device, a communication base station, and a vehicle.

[0033] Furthermore, the operation plan for physical equipment is not limited to plans used for direct execution by the entity utilizing the planning management device 5, but may also be a plan used for indirect execution. In the power sector, indirect execution of equipment operation refers, for example, to the operation of physical equipment by others based on direct bilateral trading contracts or trading contracts through exchanges. In this case, the execution plan of the trading contract corresponds to the equipment operation plan.

[0034] The information input / output terminal 4 inputs data to the prediction calculation device 2, the data management device 3, and the planning management device 5, and displays the data stored or output by these devices. The data observation device 6 periodically measures or collects sample prediction target data, sample factor data, and prediction factor data at predetermined time intervals and transmits them to the data distribution device 7 or the data management device 3. The data distribution device 7 stores the data received from the data observation device 6 and transmits the stored data to the data management device 3, the prediction calculation device 2, or both. (1-2) Device internal configuration

[0035] Figure 2 shows the configuration of the prediction device 12.

[0036] The data management device 3 consists of a CPU (Central Processing Unit) 31 that comprehensively controls the operation of the data management device 3, an input device 32, an output device 33, a communication device 34, and a storage device 35. The data management device 3 is an information processing device such as a personal computer, a server computer, or a handheld computer.

[0037] The input device 32 may consist of a keyboard or a mouse. The output device 33 may consist of a display or a printer. The communication device 34 may be equipped with a NIC (Network Interface Card) for connecting to a wireless LAN or a wired LAN. The storage device 35 may be a storage medium such as RAM (Random Access Memory) or ROM (Read Only Memory). The output results of each processing unit and intermediate results may be output as appropriate via the output device 33. The CPU 31 is connected to devices 32 to 35.

[0038] The storage device 35 may be provided with storage areas such as a storage area 351 for the first sample prediction target data 351A, a storage area 352 for the first sample factor data 352A, and a storage area 353 for the first prediction factor data 353A.

[0039] The first sample data for prediction 351A is time-series data of past values ​​obtained for the prediction target, for example, time-series data of past observed values ​​(an example of actual values) of the prediction target. The first sample factor data 352A is data that has time-series data of past values ​​obtained for each factor that affects the increase or decrease of the value of the prediction target, for example, data that includes time-series data of observed values ​​(an example of factor values ​​as actual values) for each factor. The prediction factor data 353A is time-series data of future values ​​obtained for various factors, for example, time-series data of the values ​​of each factor (factor values ​​as predicted values) used in calculating the predicted value of the prediction target (for example, temperature forecast data).

[0040] The predictive computing unit 2 consists of a CPU (Central Processing Unit) 21 that comprehensively controls the operation of the predictive computing unit 2, an input device 22, an output device 23, a communication device 24, and a storage device 25. The predictive computing unit 2 is an information processing device such as a personal computer, a server computer, or a handheld computer. The CPU 21 is connected to devices 22 to 25.

[0041] The storage device 25 has a storage area 255 where prediction result data 255A is stored. The storage device 25 also stores various computer programs for multiple functions, including a sample selection unit 251, a factor selection unit 252, an evaluation and judgment unit 253, and a prediction unit 254. When the computer programs are executed by the CPU 21, multiple functions, including the sample selection unit 251, the factor selection unit 252, the evaluation and judgment unit 253, and the prediction unit 254, are realized.

[0042] The sample selection unit 251 acquires the first sample prediction target data 351A and the first sample factor data 352A from the data management device 3, and outputs the second sample prediction target data, which is data from the first sample prediction target data 351A that is only within a specified date and time range, and the second sample factor data, which is data from the sample factor data 352A that is only within the same specified date and time range.

[0043] The factor selection unit 252 selects factors to be used in the prediction model using the second sample prediction target data and the second sample factor data, and outputs a third sample factor data which is the data of the selected factors from the second sample factor data. The factor selection unit 252 constructs a prediction model using the third sample factor data and outputs model fit data, which is an index value indicating the degree of fit of the constructed prediction model to the second sample prediction target data.

[0044] The evaluation and judgment unit 253 determines (evaluates) whether the model fit data values ​​meet predetermined criteria. If the judgment result is true, the evaluation and judgment unit 253 outputs data specifying the sample date and time and factor type to be used in the prediction model. If the judgment result is false, it changes the sample date and time specified in the sample selection unit 251 to a new date and time and outputs it.

[0045] The prediction unit 254 obtains sample data of the target and factors from the sample prediction target data 351A, sample factor data 352A, and prediction factor data 353A based on the sample date and time and factor specification data output by the evaluation judgment unit 253. It identifies a prediction model using the obtained sample prediction target data and sample factor data, inputs the prediction factor data into the identified prediction model, and outputs the predicted value of the target.

[0046] Communication devices 24 and 34 may be examples of interface devices for the prediction device 12. Storage devices 25 and 35 may be examples of storage devices for the prediction device 12. CPUs 21 and 31 may be examples of processors for the prediction device 12. (1-3) Overall processing and data flow of data prediction system 1

[0047] The processing and data flow of the data prediction system 1 in this embodiment will be described with reference to Figures 3 and 4.

[0048] The data management device 3 receives the first sample prediction target data 351A, the first sample factor data 352A, and the first prediction factor data 353A from the data observation device 6 and / or the data distribution device 7, and stores them in the memory areas 351, 352, and 353, respectively (S400).

[0049] In the prediction calculation device 2, the sample selection unit 251 selects sample data from the first sample prediction target data 351A and the first sample factor data 352A based on the date and time specified in the sample date and time specification data 253C2 (see Figure 5), which is output from the evaluation and judgment unit 253, or an identifier that identifies each sample. The selected sample data is then output as the second sample prediction target data 251B1 and the second sample factor data 251B2 (see Figure 5) (S401). The data specifying the sample date and time to be selected is input from the evaluation and judgment unit 253, but if no input is provided, pre-set initial data is used.

[0050] Next, the factor selection unit 252 receives the second sample prediction target data 251B1 and the second sample factor data 251B2 (see Figure 6), selects one or more factors from the second sample factor data 251B2 to apply to the prediction model, generates a third sample factor data from the second sample factor data 251B2 which is the factor data of the one or more selected factors, and constructs a prediction model using the third sample factor data and the second sample prediction target data 251B1 (S402). The factor selection unit 252 generates and outputs model fit data 252D (see Figure 6), which is an index value indicating the degree of fit of the constructed prediction model to the second sample prediction target data 251B1. The factor selection unit 252 also outputs the third sample factor data 252D1 (see Figure 6) used to construct the prediction model.

[0051] Next, the evaluation and determination unit 253 determines (evaluates) whether the model fit value of the model fit data 252D2 output by the factor selection unit 252 meets a predetermined standard (S404). If the evaluation and determination unit 253 determines that the determination result is true (S404:Yes), it outputs sample factor specification data 253C1; if the determination result is false (S404:No), it outputs new sample date and time specification data 253C2 (see Figure 7) (S405).

[0052] The prediction unit 254 then obtains data corresponding to the sample and factor specified in the sample factor designation data 253C1 from the sample prediction target data 351A, sample factor data 352A, and prediction factor data 353A. Using the obtained data, it constructs a prediction model and calculates predicted values, and outputs the calculated predicted values ​​as prediction data result data 255A (S406). The output prediction result data 255A is stored in the storage area 255 (for example, a non-volatile area), or, without being stored in the storage area 255, it is transmitted by the prediction unit 254 to the planning management device 5.

[0053] The planning management device 5 has functions such as a planning generation unit 501 and an instruction transmission unit 502. These functions may be realized in the planning management device 5 by having a computer program executed on the CPU. The planning generation unit 501 creates an operation plan for the equipment based on the prediction result data 255A from the prediction calculation device 2. The instruction transmission unit 502 generates control commands for the controlled device 9 based on the operation plan and transmits the control commands to the controlled device 9.

[0054] With the above processing, the data prediction process in this embodiment is completed.

[0055] Hereafter, detailed embodiments of each component will be described using Figures 5 to 7. (1-4) Details of each component (1-4-1) Sample selection section 251

[0056] Referring to Figure 5, the data flow and processing operation of the sample selection unit 251 will be explained.

[0057] The sample selection unit 251 includes a sample acquisition unit 251A. The sample selection unit 251 acquires sample factor data 352A and prediction factor data 353A from the data management device 3, and inputs these data 352A and 353A, along with the sample date and time specification data 253C2 from the evaluation and judgment unit 253, to the sample acquisition unit 251A. The sample acquisition unit 251A acquires sample data for the date and time set in the sample date and time specification data 253C2 from the sample factor data 352A and prediction factor data 353A, and outputs the acquired data as second sample prediction target data 251B1 and second sample factor data 251B2. Here, the sample date and time specification data 253C2 specifically includes data that specifies a date and time range, such as "the past 100 days starting from the last date and time of the sample," "the past 100 days from December 22, 2021," or "from December 22, 2021 to January 1, 2021," or data that directly specifies a date and time, such as "December 22, 2021, December 1, 2021," or combinations thereof. Furthermore, the sample date and time specification data 253C2 may include identifiers that identify each sample, not just date and time. In addition, if there is no input of sample date and time specification data 253C2 from the evaluation judgment unit 253, such as during the initial operation, the sample date and time specification data 253C2 input to the sample acquisition unit 251A may be the sample date and time specification data 253C2 set in advance as an initial value.

[0058] This concludes the operation of the sample selection unit 251. (1-4-2) Factor selection section 252

[0059] Referring to Figure 6, the data flow and processing operation of the factor selection unit 252 will be explained.

[0060] The factor selection unit 252 includes a factor fit calculation unit 252A, a factor data acquisition unit 252B, and a model fit calculation unit 252C.

[0061] The factor fit calculation unit 252A calculates the degree of fit (factor fit) for each factor represented by the second sample factor data 251B2 to the second sample prediction target data 251B1. Factor fit is an index value that indicates the strength of the influence of a factor on the second sample prediction target data 251B1 (for example, the strength of correlation, the strength of importance, or the strength of contribution), and known methods may be used to calculate factor fit. Known methods include, for example, correlation analysis methods that calculate the strength of linear or nonlinear correlations such as Pearson correlation coefficients and maximal information coefficients, methods that use regression coefficients in multiple regression models, ridge regression, and lasso regression, methods that calculate likelihood values ​​when a model is constructed for each factor using regression models that use likelihood such as Bayesian regression models and Gaussian process regression, and methods that calculate factor importance using ensemble trees such as random forests and gradient boosting trees.

[0062] The factor data acquisition unit 252B uses the factor fit of each factor calculated by the factor fit calculation unit 252A to acquire sample data of the Nth factors in descending order of factor fit from the second sample factor data 251B2, and outputs the acquired data as the third sample factor data to the model fit calculation unit 252C. Here, the value of N is set by the output from the model fit calculation unit 252C, but if there is no output from the fit calculation unit 252C or during the first operation, N=1 is acceptable (i.e., the sample data of the factor with the highest factor fit may be acquired from the second sample factor data 251B2).

[0063] The model fit calculation unit 252C then constructs a prediction model (e.g., a regression model) using the second sample prediction target data 251B1 and the third sample factor data output by the factor data acquisition unit 252B, and calculates and records the goodness of fit (model fit) of the constructed prediction model. If the model fit value does not reach a predetermined threshold, or if the model fit does not reach the maximum value, or if there are still factors that can be acquired by the factor data acquisition unit 252B, the model fit calculation unit 252C outputs setting information to the factor data acquisition unit 252B to increase the above value of N used by the factor data acquisition unit 252B. The factor data acquisition unit 252B then outputs the third sample factor data again based on that setting information, constructs a prediction model using the newly outputted third sample factor data, and calculates and records the goodness of fit (model fit) of the constructed prediction model. The model fit calculation unit 252 repeats the above process until the model fit value reaches a predetermined threshold, or the model fit reaches its maximum value, or until there are no more factors available for acquisition by the factor data acquisition unit 252B. After that, the model fit calculation unit 252 outputs the third sample factor data used for the M-th prediction models in descending order of model fit as the third sample factor data 252D1. At the same time, the model fit calculation unit 252 outputs the model fit of the M-th prediction models in descending order of model fit as the model fit data 252D2. The value of M above is a value of "1" or greater, which is set in advance.

[0064] The method used to construct the prediction model may be any known method. Examples of known methods include any of the following: Linear models such as ridge regression, lasso regression, and elastic nets. • Tree models such as regression trees, random forests, and boosting trees. • Kernel methods such as support vector regression, kernel ridge regression, and Gaussian process regression. Nonlinear models such as neural networks including recurrent networks and Long Short-Term Memory.

[0065] Furthermore, the calculation of the goodness of fit of the model can also be done using known methods. Known methods include, for example, the sum of squared residuals of a regression model when a regression model is constructed, the likelihood of the regression model, and information criteria such as AIC (Akaike's Information Criterion) and BIC (Bayesian Information Criterion) that can be calculated using the sum of squared residuals and likelihood.

[0066] This concludes the operation of the factor selection unit 252. (1-4-3) Evaluation and Judgment Unit 253

[0067] The data flow and processing operation of the evaluation and judgment unit 253 will be explained with reference to Figures 7 to 10.

[0068] The evaluation and determination unit 253 includes a model fit comparison unit 253A and a sample factor determination unit 253B.

[0069] The model fit comparison unit 253A records the second sample prediction target data 251B1, the third sample factor data 252D1, and the model fit data 252D2. An example of the recorded data is explained using Figure 8.

[0070] The "sample set" shown in the column direction of Figure 8 refers to an identifier indicating the content of the second sample prediction target data and second sample factor data selected by the sample selection unit 251. The content of the sample indicated by each identifier will be further explained using Figure 9. Figure 9 is data composed of information such as "sample set ID," which is an identifier indicating the content of the sample, and "sample content," which describes the content of the sample, and is data recorded in the evaluation selection unit 253. Here, "sample set 3" shown in column 801 of Figure 8 corresponds to the sample with "sample set ID" of "3" shown in row 901 of Figure 9, and the content of the sample means that it is a sample of the second sample prediction target data and second sample factor data selected by the sample selection unit 251 as a sample of "the past 90 days starting from the last observation data." "Starting from the last observation data" can be interpreted as "starting from the date and time represented by the last observation data (last observation date and time)."

[0071] Next, the "factor set" shown in the row direction of Figure 8 refers to an identifier indicating the content of the factors included in the third sample factor data 252D2 selected by the factor selection unit 252. The content of the factors indicated by each identifier will be further explained using Figure 10. Figure 10 is data that consists of information such as the "factor set ID," which is an identifier indicating the content of the factor, and the "factor content," which describes the content of the factor, and is stored in the evaluation selection unit 253. Here, "factor set 3" shown in row 802 of Figure 8 corresponds to the factor whose "factor set ID" shown in row 1001 of Figure 10 is "3," meaning that the third sample factor data in question is a sample that includes "factor 1" and "factor 2." The values ​​in the table shown in Figure 8 are the values ​​of the model fit data 252D2 calculated when the factor selection unit 252 constructed a model using each sample and factor, and in Figure 8, a larger value means a higher model fit. The model fit comparison unit 253A creates the data shown in Figure 8 and outputs that data to the sample factor determination unit 253B.

[0072] Next, the sample factor determination unit 253B refers to the data shown in Figure 8 output by the model fit comparison unit 253A and extracts the identifiers of the sample set and factor set with the highest model fit. In Figure 8, the model fit of the prediction model constructed using the sample set 3 shown in column 801 and the factor set 3 shown in row 802 is the highest at "324". Therefore, the sample factor determination unit 253B extracts "sample set 3" and "factor set 3". Then, by referring to the data recording the contents of the samples and factors shown in Figures 9 and 10, the sample factor determination unit 253B determines that the sample and factor with the highest model fit are the sample from the "past 90 days starting from the last observed data" and the factors "factor 1, factor 2", and outputs information representing the determined sample and factor as sample factor specification data 253C1. Furthermore, if the model fit stored in the data shown in Figure 8 does not reach a predetermined threshold, or does not reach the maximum value, or if there are still samples available for acquisition by the sample selection unit 251, the sample factor determination unit 253B does not output the sample factor specification data 253C1. Instead, it rewrites the sample acquisition specification set in the sample date and time specification data 253C2 used by the sample selection unit 251 and outputs it as new sample date and time specification data 253C2. The rewriting process updates the second sample prediction target data and sample factor data acquired by the sample selection unit 251, for example, to increase or decrease the amount.

[0073] With this, the operation of the evaluation selection unit 253 is completed. (1-4-4) Prediction unit 254

[0074] The data flow and processing operation of the prediction unit 254 will be explained with reference to Figure 11.

[0075] The prediction unit 254 includes a sample selection unit 254A, a factor selection unit 254B, a model identification unit 254C, and a predicted value calculation unit 254D.

[0076] The sample selection unit 254A extracts sample data (e.g., window width) from the sample prediction target data 351A and the sample factor data 352A that corresponds to the sample content specified in the sample factor specification data 253C1, and outputs this data as the second sample prediction target data and the second sample factor data.

[0077] Next, the factor selection unit 254B extracts the data of the factors specified in the sample factor designation data 253C1 from the second sample factor data and outputs this data as the third sample factor data. The factor selection unit 254B also extracts the data of the factors specified in the sample factor designation data 253C1 from the prediction factor data 353A and outputs this data as the second prediction factor data.

[0078] Next, the model identification unit 254C uses the second sample prediction target data output by the sample selection unit 254A and the third sample factor data output by the factor selection unit 254B to construct a prediction model for calculating the predicted value of the target, and outputs the prediction model.

[0079] The prediction value calculation unit 254D then calculates the predicted value of the target to be predicted by inputting the second prediction factor data output by the factor selection unit 254B to the prediction model output by the model identification unit 254C, and outputs it as prediction result data 255A. The outputted prediction result data 255A is stored in the memory area 255.

[0080] This concludes the operation of the prediction unit 254. (1-5) Description of the effects of this embodiment

[0081] Next, the effects of the data prediction system 12 in this embodiment will be explained with reference to Figure 12.

[0082] The table shown on the left side of Figure 12 shows the second set of sample factor data. Each of the labels 1201A, 1202A, and 1203A represents the second set of sample factor data that corresponds to the range of the sample selected by the sample selection unit 251 (in other words, the window width, which is the past time range starting from the target date and time for the prediction target) from the first set of sample factor data. For example, the second set of sample factor data 1201A is sample factor data that uses the most recent 10 days from "T-1" to "T-10" as the sample (data for each of Factor 1 to Factor 4, including the factor values ​​for each of the dates and times "T-1" to "T-10").

[0083] In Figure 12, for each factor, the black and white colors indicate whether or not it had a strong influence on the predicted value at each sample day ("T-1", "T-2", ..., "T-10") (i.e., whether or not the factor was a major factor at the time of the sample). Black indicates a strong influence. Therefore, for example, factor 1 was a major factor that had a strong influence on the predicted value at the sample days from T-10 to T-4.

[0084] For the second sample factor data 1201A, the factor fit is calculated for each factor. The set of factors selected by the factor selection unit 252 based on the factor fit values ​​is as shown in frame 1201B. In other words, in the second sample factor data 1201A, the factor fit of factors 1 and 2 is high, and therefore the data for factors 1 and 2 from the second sample factor data 1201A are selected as the third sample factor data.

[0085] The result of constructing a predictive model for the second sample target data using the selected third sample factor data is shown in the line graph on the right side of Figure 12. Specifically, the solid line graph 1200 shows the second sample target data (for example, the time series of past actual values ​​of the target) over the sample days from "T-10" to "T-1". The dotted line graph 1201C is the regression line of the target output from the constructed predictive model (the time series of predicted values ​​over the sample days from "T-10" to "T-1"). The factors used to construct this predictive model were "Factor 1" and "Factor 2", but the factors that influenced the period from "T-1" to "T-3" were "Factor 2", "Factor 3", and "Factor 4". Because these factors, "Factor 2," "Factor 3," and "Factor 4," could not be selected, the goodness of fit of the regression line 1201C from "T-1" to "T-3" to the target 1200 is relatively small (i.e., the difference between graph 1200 and 1201C is relatively large from "T-10" to "T-1"). Therefore, the goodness of fit value of the prediction model itself is a relatively small value of "120."

[0086] The sample range data 253C2, which specifies the sample date and time, is modified (the window width is shortened), and new second sample factor data 1202A is obtained. Factor fit is then recalculated for each factor. The sample range of the second sample factor data 1202A is from "T-6" to "T-1". In this sample, the factor with the highest factor fit is calculated to be "Factor 2" (see symbol 1202B), and therefore Factor 2 is selected. When a predictive model is constructed using the data for Factor 2 from the second sample factor data 1202A, the regression line will be as shown in the dotted line graph 1202C. In other words, for the period from "T-6" to "T-1", the difference between graphs 1200 and 1202C narrows (the degree of fit to the target of prediction increases), and as a result, the model fit value increases to "250".

[0087] The sample date and time specification data 253C2 is modified again (here the window width is further reduced), and the same process is performed. For example, there is a second sample factor data 1203A (sample factor data from "T-4" to "T-1"), and for this sample factor data 1203A, the factor fit is high for "Factor 2", "Factor 3", and "Factor 4". Therefore, "Factor 2", "Factor 3", and "Factor 4" are selected. For "T-4" to "T-1", the regression line of the prediction model constructed using the selected factors has a higher fit to the target of prediction, as shown in the dotted line graph 1203C, and therefore the model fit value is also a higher value of "420".

[0088] As described above, according to this embodiment, for each prediction target date and time, the following is performed. That is, the prediction device 12 decreases or increases the sample range (window width) starting from the prediction target date and time, and for each sample range, it selects factors with a high degree of fit from the sample factor data of that sample range (i.e., dynamically extracts the main factors), and constructs a prediction model using the data of the selected factors from the sample factor data. The prediction device 12 then compares the model fit of these prediction models and selects the prediction model with the highest degree of fit. This makes it possible to further reduce the error in the prediction of the prediction target, and therefore it is possible to improve the accuracy of operation planning and control of the prediction target using the predicted values. (2) Second embodiment (modified version of the factor selection unit)

[0089] In the above embodiment, the factor selection unit 252 directly uses the second sample factor data 251B2 output by the sample selection unit 251. However, the factor selection unit 252 may also extract representative factors in advance from the factors represented by the second sample factor data 251B2 and use the sample factor data of the extracted factors as new second sample factor data.

[0090] A modified example of the factor selection unit 252 will be explained with reference to Figure 13. The factor selection unit 252 has a representative factor extraction unit 252E. The representative factor extraction unit 252E receives input from the second sample factor data 251B2 output by the sample selection unit 251, and extracts one or more representative factors from among the multiple factors represented by the second sample factor data 251B2. Specifically, for example, the representative factor extraction unit 252E classifies similar factors as clusters based on the similarity of the sample data of the factors represented by the second sample factor data 251B2. The method for classifying into clusters can be a known method. Known methods include cluster analysis methods such as hierarchical clustering methods such as Ward's method, and non-hierarchical clustering methods such as k-means and spectral clustering. The number of clusters may be a predetermined number, or the number of classifications may be determined by calculating the value of an information criterion such as AIC while increasing or decreasing the number of clusters to be classified, and using the number of classifications when the value of the information criterion reaches an extreme value.

[0091] Next, the representative factor extraction unit 252E extracts a representative factor from each cluster. Specifically, for example, the representative factor extraction unit 252E calculates the factor fit of each factor classified into each cluster using the second sample prediction target data 251B1, and extracts the factor with the highest factor fit as the representative factor of that cluster. The representative factor extraction unit 252E performs the same process for all clusters, extracting a representative factor from each cluster. The representative factor extraction unit 252E then outputs the sample of the representative factor from each extracted cluster as new second sample factor data. The subsequent processing in the factor selection unit 252 is the same as in the first embodiment.

[0092] If the number of factors represented by the second sample factor data is enormous, the processing load on the factor selection unit 252 can become high. Furthermore, if the factors represented by the second sample factor data are strongly correlated with each other, the factor selection results in the factor selection unit 252 may output different results each time processing is performed, leading to instability. This embodiment reduces the processing load and stabilizes the factor selection results by reducing the number of factors in advance. Consequently, the modeling accuracy of the prediction model of the target of prediction becomes stable, and the prediction results and the operation and control of the equipment using the prediction results also become stable.

[0093] Furthermore, the factor selection unit 252 in the above embodiment uses, for example, the sum of squared residuals of the regression model, the likelihood of the regression model, or information criteria such as AIC and BIC that can be calculated using the sum of squared residuals or likelihood to calculate the goodness of model fit. However, it is not limited to these, and a coefficient or function that imposes a penalty effect for a small number of samples (a narrow window width) for the second sample prediction target data or the second sample factor data may be incorporated. The coefficient or function that imposes the penalty effect may, for example, be obtained by multiplying the reciprocal of the sample by a predetermined coefficient and dividing this value by the goodness of model fit calculated in the first embodiment. Alternatively, it may be an information criterion function that takes the number of samples into account, such as cAIC (conditional AIC). When the number of samples becomes extremely small, the regression model is prone to overfitting, and therefore the sum of squared residuals approaches zero, and the likelihood also increases. This embodiment makes it possible to calculate an appropriate goodness of model fit even when the number of samples of the second sample prediction target data or the second sample factor data is small.

[0094] Furthermore, although the factor selection unit 252 in the above embodiment was described as calculating the model fit and factor fit using the sum of squared residuals and likelihood of the regression model constructed using the second sample data, it is not limited to this, and the fit may be calculated by applying methods generally called cross-validation or cross-testing. Specifically, first, the factor selection unit 252 extracts a predetermined number of past N sample data from each of the second sample prediction target data and the second sample factor data, including the sample data for the most recent observation date and time (e.g., observed value). The extracted data becomes the validation prediction target data and the validation factor data. The factor selection unit 252 then resets the remaining data, after excluding the above validation prediction target data and validation factor data from the second sample prediction target data and the second sample factor data, as new second sample prediction target data and second sample factor data. The factor selection unit 252 then constructs a prediction model using the new second sample prediction target data and the second sample factor data, and calculates the validation prediction result data by inputting the aforementioned validation factor data into the constructed prediction model. Using the deviation between the calculated validation prediction result data and the aforementioned validation prediction target data, the reciprocal of the deviation is calculated as the model fit.

[0095] The smaller the number of samples in the second sample data input to the factor selection unit 252, the more likely overfitting is to occur, and therefore the model fit may not be calculated accurately. In contrast, in this embodiment, if overfitting occurs, the deviation from the validation data increases, making it possible to prevent overestimation of the model fit when overfitting occurs. Therefore, it is possible to prevent deterioration of the accuracy of the prediction result data and also prevent deterioration of the control accuracy of the equipment using the prediction result data.

[0096] Furthermore, the factor selection unit 252 in the above embodiment may perform the factor selection process described in the above embodiment in multiple stages. Specifically, for example, the factor selection unit 252 selects a group of factors (first group of factors) consisting of one or more factors by performing a factor selection process. The factor selection unit 252 calculates first factor selection result data (data representing one or more factors selected from the first group of factors) by performing a factor selection process on the first group of factors. Next, the factor selection unit 252 similarly performs a factor selection process on a second group of factors obtained by excluding the factors represented by the first factor selection result data from the first group of factors, and calculates second factor selection result data. Then, the factor selection unit 252 calculates final factor selection result data by performing a factor selection process on the group of factors represented by the first and second factor selection result data. Such factor selection processing is not limited to two stages (i.e., not limited to obtaining first and second factor selection result data), but may be performed in n stages (i.e., until n factor selection result data are obtained), or may be performed until there are no more factors to select.

[0097] When a group of factors contains a large number of factors, and many of them are interrelated, it is possible that some factors that should be selected may be missed during the first factor selection process. By configuring the selection process in multiple stages as in this embodiment, it is possible to prevent factors from being missed during selection.

[0098] Furthermore, while the factor selection unit 252 in the above embodiment processes using a general feature selection method, it is not limited to this and may perform selection processing based on predetermined rules. For example, the factor selection unit 252 may exclude factors whose values ​​do not change, or factors whose type of change is less than a predetermined amount, from the selection target in advance. The factor selection unit 252 may also exclude factors whose values ​​are missing at the target date and time for prediction in advance.

[0099] By excluding factor data to be selected in advance, the load on the factor selection process can be reduced. (3) Third embodiment (modified sample selection unit)

[0100] In the above embodiment, the sample selection unit 251 selects only one type of sample according to the specification of the sample date and time specification data 253C2, but it is not limited to this, and may select multiple types of samples in a single processing operation. In this case, similarly, the contents of the selected samples specified in the sample date and time specification data 253C2 may be multiple types. Also, the initial values ​​set in the sample selection unit 251 (initial values ​​of the data specifying the selected samples) may be multiple types.

[0101] This reduces the number of processing steps in the feedback loop from the evaluation selection unit 253 to the sample selection unit 251, and therefore reduces the processing load.

[0102] Furthermore, in the above embodiment, if there are missing values ​​in the sample values ​​(e.g., observed values) or factor values, the sample selection unit 251 may either exclude only the factors with missing values ​​in advance, or exclude the samples themselves that have missing values. Specifically, if the sample is time-series data and there are missing values ​​in the sample up to a predetermined past date, the sample selection unit 251 may exclude only the factors with missing values ​​without excluding the sample. By preserving the most recent past sample, it is possible to construct a prediction model that reflects the most recent behavioral characteristics of the target of prediction. Also, if there are missing values ​​in the past sample beyond a predetermined past date, the factor selection unit 252 may exclude the sample itself. By preserving many factors, it is possible to increase the likelihood of including important factors among the factors selected by the factor selection unit 252. (4) Fourth embodiment (modified version of the prediction unit)

[0103] In the above embodiment, the prediction unit 254 identifies the prediction model using a known method, but is not limited to that; it may also modify the model using an index value indicating the contribution of each factor to the prediction model, which is output by the factor selection unit 252.

[0104] Specifically, for example, if the factors used are two types, x1 and x2, and the prediction model is a linear regression model such as a multivariate regression model or an autoregressive model, the prediction model may be given by the following equation. Y = a × x1 × w1 + b × x2 × w2 + c

[0105] Here, Y is the value to be predicted. a, b, and c are parameters of the regression model. w1 and w2 are index values ​​indicating the contribution of each factor output by the factor selection unit 252. For example, if the value of w2 is larger than w1, it means that factor x2 has a higher contribution to the prediction model than factor x1. Also, in the case of a prediction model based on the similarity between data, such as the kernel method, where the similarity is calculated using the Euclidean distance, the similarity is given by the following formula. d_ij∝w1(x1_i-x1_j)^2+w2(x2_i-x2_j)^2

[0106] Here, d_ij is the Euclidean distance between the i-th and j-th samples, and w1 and w2 are index values ​​indicating the degree of contribution.

[0107] This allows the factor selection unit 252 to more strongly fit the prediction model to factors that have a higher contribution to the prediction model among the factors it has selected, thereby improving prediction accuracy. (5) Fifth embodiment (modified version of the prediction calculation device 2)

[0108] In the above embodiment, the prediction calculation device 2 starts processing when it receives an input operation from the device user or when a pre-set execution time is reached via the information input / output terminal 4. However, it is not limited to this, and may also monitor the sample prediction target data 351A and the sample factor data 352A and start processing when new sample data is added. A specific example will be explained with reference to Figure 14.

[0109] The prediction calculation device 2 has an activation determination unit 257. The activation determination unit 257 receives sample prediction target data 351A, sample factor data 352A, and sample factor specification data 253C1 as input, and outputs activation signal data 257A, which is a control signal to determine whether or not to activate the processing from the sample selection unit 251 onwards. More specifically, the activation determination unit 257 first generates second sample prediction target data and second sample factor data by obtaining a sample and factors from the sample prediction target data 351A and sample factor data 352A according to the contents of the sample and factors specified in the sample factor specification data 253C1. Then, the activation determination unit 257 extracts sample data for a predetermined number of past N samples from the second sample prediction target data and second sample factor data, starting with the sample from the most recent observation date and time, and outputs this data as third sample prediction target data and third sample factor data. The activation determination unit 257 also resets the remaining data of the second sample prediction target data and second sample factor data as new second sample prediction target data and second sample factor data. The activation determination unit 257 then constructs a prediction model using the new second sample prediction target data and second sample factor data, and calculates the prediction result data by inputting the aforementioned third sample factor data into the constructed prediction model. If the deviation between the calculated prediction result data and the aforementioned third sample prediction target data exceeds a predetermined threshold, the activation determination unit 257 generates control signal data to activate the processing of the sample selection unit 251 and outputs it as activation signal data 257A.

[0110] The embodiment described above is one of the methods generally known as cross-validation, but is not limited to it. The activation determination unit 257 may acquire second sample data according to the sample factor designation data 253C1 from each sample before and after new sample data for prediction and sample factor data are added (for example, before and after new date and time observation data is added to the sample data for prediction 351A, and / or before and after at least one of the new date and time observations for an existing factor and the observations for the new factor is added to the sample factor data 352A). The activation determination unit 257 may compare the model fit of the prediction model constructed using each second sample data (sample data for prediction and sample factor data). The activation determination unit 257 may determine whether the model fit of the prediction model constructed using the second sample data after the addition of the new sample has decreased by more than a predetermined threshold compared to the model fit of the prediction model constructed using the second sample data before the addition of the new sample. If the value drops below a threshold, the activation determination unit 257 may generate control signal data to activate the processing of the sample selection unit 251 and output it as activation signal data 257A.

[0111] In this embodiment, the prediction calculation device 2 detects when the prediction model, constructed using samples and factors selected according to the sample factor designation data 253C1 which contains the contents of samples and factors calculated by past operations of the prediction calculation device 2, no longer fits the recently recorded data in the sample prediction target data 351A and sample factor data 352A, and updates the contents of the samples and factors designated in the sample factor designation data 253C1. This allows the device to automatically track changes in the sample prediction target data and sample factor data over time, thereby maintaining the prediction accuracy of the prediction result data and stabilizing the control accuracy of the equipment using the prediction result data.

[0112] Although the prediction calculation device 2 in the above embodiment was described as a process for calculating prediction result data for a single type of prediction target, it is not limited to this and may be applied to a multi-stage process in which there are multiple prediction targets and the first prediction result data is used as one of the factors in calculating the second prediction result data. For example, first, the prediction calculation device 2 sets total power consumption data for a certain region as sample prediction target data 351A, temperature observation data at multiple points in the region as sample factor data 352A, and similarly sets predicted temperature data for those points as prediction factor data 353A. The prediction calculation device 2 then calculates the predicted value of the total power consumption data for the region that is the prediction target and outputs it as prediction result data 255A. Next, the prediction calculation device 2 adds the total power consumption data for the region set as sample prediction target data 351A as a new factor to sample factor data 352A, and adds the predicted result data of the total power consumption data for the region, which is the aforementioned prediction result data 255A, to prediction factor data 353A. The prediction calculation device 2 then sets the buy / sell bid volume in the electricity market as a new prediction target and sets sample data of buy / sell bid volume as new sample prediction target data 351A. The prediction calculation device 2 then calculates the predicted value of the buy / sell bid volume, which is the prediction target, and outputs it as prediction result data 255A. The prediction calculation device 2 then adds the buy / sell bid volume data set as sample prediction target data 351A as a new factor to sample factor data 352A, and adds the aforementioned prediction result data 255A, which is the predicted result data of buy / sell bid volume, to prediction factor data 353A. The prediction calculation device 2 then sets the electricity market price as a new prediction target and sets sample data of the electricity market price as new sample prediction target data 351A. The prediction calculation device 2 then calculates the predicted value of the electricity market price, which is the prediction target, and outputs it as prediction result data 255A.

[0113] In order to calculate prediction result data for a target to be predicted, it is sometimes necessary to first predict the data itself that will be used as factors. However, with this embodiment, even when calculating prediction result data for the factors themselves, it is possible to improve the prediction accuracy of each factor, and therefore the prediction accuracy of the final target to be predicted is also improved, as is the control accuracy of the equipment using the prediction result data.

[0114] In the above embodiment, the prediction calculation device 2 was described as a process that calculates the prediction result data of the target to be predicted only once. However, it is not limited to this, and may also be a multi-stage process in which the first prediction result data of the target to be predicted is incorporated as one of the new factors, and the prediction result data of the target to be predicted is calculated again. For example, if the target to be predicted is the electricity market price for 24 hours, calculated one prediction result data for each hour, and the first prediction result data of the target to be predicted is calculated as an array of 24 points. Next, the prediction calculation device 2 sets the first prediction result data as new factor data and similarly calculates one prediction result data for each hour, and calculates the second prediction result data of the target to be predicted as an array of 24 points. When calculating the nth point of the second prediction result data, the 23 points of the first prediction result data, obtained by subtracting the nth point from the first prediction result data, are set as a new factor. Alternatively, the calculations may be performed sequentially from the first time point to obtain the 24 points of the second prediction result data, but it is not limited to this, and a process may be adopted in which the calculation starts from the time point in the past where the average error of the first prediction result data over a predetermined period is smallest. Furthermore, while the above explanation has always described the prediction result data set as a new factor as the first prediction result data, this is not limited to this; when a second prediction result data is obtained, the new factor may be switched from the first prediction result data to the second prediction result data.

[0115] The values ​​to be predicted may be correlated across consecutive time points. Since the processing in this embodiment allows for the construction of a prediction model using values ​​from other time points, it is possible to achieve predictions that reflect the correlation between time points.

[0116] The prediction calculation device 2 in the above embodiment constructs a prediction model for each time period of the target to be predicted and calculates prediction result data for each time period, but is not limited to this. The prediction calculation device 2 may also perform a frequency transformation on the array of the target to be predicted, calculate prediction result data for each of the coefficients obtained by the transformation as a new target to be predicted, and then perform an inverse transformation on the array of prediction result data for each coefficient to calculate the prediction result data for the target to be predicted.

[0117] The values ​​to be predicted may have correlations over consecutive time points. The prediction calculation device 2 may perform frequency transforms such as Fourier transforms or wavelet transforms on the array of values ​​to be predicted at multiple time points to calculate coefficients that indicate the intensity of frequencies. The prediction calculation device 2 calculates the predicted value of each coefficient and performs an inverse transform on the array of predicted values ​​of the coefficients, thereby more precisely modeling the characteristics of the time-dependent transition of the array of values ​​to be predicted before the transformation. Furthermore, the prediction accuracy of each coefficient can be further improved by extracting factors and samples to be used for predicting each coefficient using the sample selection unit 251, the factor selection unit 252, and the evaluation and judgment unit 253, respectively.

[0118] Although several embodiments of the present invention have been described above, these are merely illustrative examples for the purpose of explaining the present invention and are not intended to limit the scope of the present invention to these embodiments. The present invention can be implemented in various other forms.

[0119] For example, it is possible to combine any two or more embodiments from the multiple embodiments described above.

[0120] Alternatively, instead of the prediction unit 254 identifying a prediction model, it may use a prediction model constructed (identified) in the preceding stage of the prediction unit 254 (for example, a prediction model determined to have the highest model fit). That is, the prediction unit 254 may obtain predicted values ​​for the target date and time of the prediction target by inputting second prediction factor data, which is data corresponding to the sample and factor specified in the sample factor designation data 253C1 from the first prediction factor data 353A, into the prediction model constructed in the preceding stage of the prediction unit 254.

[0121] Furthermore, the above explanation can be summarized as follows, for example.

[0122] The prediction device 12 comprises an interface device (24, 34), a storage device (25, 35), and processors (21, 31) connected thereto. The interface device receives input of first sample prediction target data (351A), first sample factor data (352A), and first prediction factor data (353A). The storage device stores the input first sample prediction target data, first sample factor data, and first prediction factor data. The processor outputs prediction result data including the predicted value of the prediction target by inputting at least a portion of the first prediction factor data into the prediction model. The first sample prediction target data is time series data of past values ​​obtained for the prediction target. The first sample factor data may be data having time series data of past values ​​obtained for each factor that may affect the prediction target. The first prediction factor data is data having time series data of future values ​​obtained for each factor that may affect the prediction target.

[0123] The processor constructs a prediction model for each of several different window widths, based on the second sample prediction data (251B1), which is the data portion of the first sample prediction data corresponding to that window width, and the second sample factor data (251B2), which is the data portion of the first sample factor data corresponding to that window width, and calculates the goodness of fit, which is an index value for the prediction model. The window width is the length of the past period starting from the prediction target date and time. The processor takes the second prediction factor data, which is at least a part of the first prediction factor data, as input to the target prediction model. The target prediction model is a prediction model based on the second sample prediction data and the second sample factor data corresponding to the window width that satisfies the model fit condition among the calculated model fit. The second prediction factor data is data that has a time series of values ​​obtained for each factor used in constructing the target prediction model from among the multiple factors represented by the first prediction factor data. In the embodiment described above, the "factors used to construct the target prediction model" are the factors specified in the sample factor specification data 253C1, but they may also be factors corresponding to factor values ​​specified by other methods and used to construct the target prediction model.

[0124] The processor may perform the following (X) and (Y) for each of the multiple window widths. (X) Based on the second sample target data and the second sample factor data, a factor fit is calculated for each of the multiple factors represented by the second sample factor data, which is an index value indicating the strength of the influence of that factor on at least one of the second sample target data and the predicted result data. One or more factors are selected for which one or more factor fits satisfying the factor fit conditions are obtained. (Y) A predictive model is constructed using the second sample data to be predicted and the third sample factor data (252D1), which consists of data from one or more selected factors from the second sample factor data.

[0125] The processor may perform the following (x1) through (x5) in (X) for each of the multiple window widths. (x1) From the multiple factors represented by the second sample factor data, select one or more factors for which one or more factor fits satisfying the factor fit condition is obtained. (x2) For each of the one or more factors remaining in the second sample factor data, after excluding the factors selected in (x1) and (x3), calculate the factor fit. From one or more factors in (x3)(x2), select one or more factors for which one or more factor fits satisfying the factor fit condition are obtained. (x4) If the condition is not met, perform (x2). (x5) If the condition is met, perform (Y).

[0126] The processor may calculate factor fit for each factor, applying a penalty value corresponding to the window width. The penalty value is a value that reduces factor fit, and it can be larger for shorter window widths.

[0127] The processor may perform (A) through (C) below. (A) Calculate the model fit of the constructed prediction model for a given or modified window width. (B) Determine whether a predetermined condition is met for at least one of the number of times (A) was performed and the calculated model fit. If the result of (C)(B) is false, change the window width and perform (A).

[0128] The processor may use the second predictive factor data as input to the predictive model if the result of (B) is true. The processor may also calculate the goodness of fit of the constructed predictive model for each of two or more window widths in a single (A).

[0129] For each of the multiple window widths, the processor may classify the multiple factors represented by the second sample factor data into one or more clusters, each of which is a set of one or more factors, based on the similarity of the data of the factor values, and determine a representative factor for each cluster. The "some factors" mentioned above may be the representative factor for each cluster.

[0130] The processor compares the information (P) below with the information (Q) below, and if the result of the comparison satisfies predetermined conditions, it may start building a predictive model and calculating the model fit for each window width. (P) Information on the model fit of the predictive model that satisfies the model fit condition before any changes were made to at least one of the first sample predictor data and the first sample factor data. (Q) Information regarding a predictive model based on a second set of predictable data and a second set of predictable data corresponding to a window width that satisfies the model fit condition after a change has been made to at least one of the first set of predictable data and the first set of predictable factor data.

[0131] The processor may include data containing the predicted values ​​output from the target prediction model for the target to be predicted, as factor data with the target to be predicted as a factor, in the factor data used as input to the prediction model.

[0132] The indicator values ​​for the prediction model may be either an indicator showing the degree of fit of the prediction model to the second sample data to be predicted, or an indicator showing the prediction accuracy of the prediction model, or both.

[0133] The prediction system 1 may include a prediction device (12) and a planning management device (5) that creates an operation plan based on the prediction result data of the prediction device and controls the controlled device based on the operation plan. [Explanation of symbols]

[0134] 1...Data prediction system, 2...Predictive calculation device, 3...Data management device

Claims

1. An interface device that accepts input of the first sample prediction target data, the first sample factor data, and the first prediction factor data, A storage device that stores the input first sample prediction target data, first sample factor data, and first prediction factor data, A processor connected to the interface device and the storage device, which inputs at least a portion of the first prediction factor data into the prediction model and outputs prediction result data including the predicted value of the target to be predicted. Equipped with, The first sample data for prediction is time-series data of past values ​​obtained for the prediction target. The first sample factor data mentioned above is data that has a time series of past values ​​obtained for each factor that may affect the target of prediction, The first predictive factor data is data that has a time series of future values ​​obtained for each factor that may affect the target of prediction, The processor constructs a prediction model for each of several different window widths based on a second set of prediction data, which is the data portion of the first set of prediction data corresponding to that window width, and a second set of prediction factor data, which is the data portion of the first set of prediction factor data corresponding to that window width, and calculates a model fit index value for the prediction model. The window width is the length of the past period starting from the date and time of the prediction target. The processor takes a second set of predictive factor data, which is at least a portion of the first set of predictive factor data, as input to the target predictive model. The target prediction model is a prediction model based on a second sample of prediction target data and a second sample of factor data corresponding to the window width of the model fit that satisfies the model fit condition among the calculated model fit values. The second predictive factor data is data that has a time series of values ​​obtained for each factor used in constructing the target predictive model, among the multiple factors represented by the first predictive factor data. Prediction device.

2. The processor, for each of the plurality of window widths, (X) Based on the second sample prediction target data and the second sample factor data, for each of the multiple factors represented by the second sample factor data, a factor fit is calculated, which is an index value indicating the strength of the influence of the factor on at least one of the second sample prediction target data and the prediction result data, and one or more factors are selected for which one or more factor fits that satisfy the factor fit conditions are obtained. (Y) Construct a predictive model using the second sample data to be predicted and the third sample factor data, which is the data of one or more selected factors from the second sample factor data. The prediction device according to claim 1.

3. The aforementioned processor, (A) Calculate the model fit of the constructed prediction model for a given or modified window width. (B) Determine whether a predetermined condition is met for at least one of the number of times (A) was performed and the calculated model fit. If the result of (C)(B) is false, change the window width and perform (A). The prediction device according to claim 1.

4. The processor, when the result of (B) is true, takes the second predictive factor data as input to the predictive model. The prediction device according to claim 3.

5. For each of the aforementioned multiple window widths, The processor classifies the multiple factors represented by the second sample factor data into one or more clusters, each of which is a set of one or more factors, based on the similarity of the data of the factor values. The aforementioned processor determines a representative factor for each cluster, Some of the aforementioned factors are representative factors for each cluster. The prediction device according to claim 2.

6. The aforementioned processor compares the following information (P) with the following information (Q), (P) Information regarding the model fit prediction model that satisfies the model fit condition before any changes are made to at least one of the first sample prediction target data and the first sample factor data. (Q) Information regarding a prediction model based on a second set of prediction data and a second set of prediction data corresponding to a window width that satisfies the model fit condition after a change has been made to at least one of the first set of prediction data and the first set of prediction factor data. The processor, when the results of the comparison meet predetermined conditions, starts building a predictive model and calculating the model fit for each window width. The prediction device according to claim 1.

7. The processor includes data containing the predicted values ​​output from the target prediction model for the target to be predicted, as data with the target to be predicted as a factor, in the factor data used as input to the prediction model. The prediction device according to claim 1.

8. The processor calculates a factor fit for each factor, to which a penalty value corresponding to the window width is applied. The penalty value is a value that reduces the factor fit, and the smaller the window width, the larger the value. The prediction device according to claim 2.

9. The processor calculates the model fit of the constructed predictive model for each of two or more window widths in a single (A) operation. The prediction device according to claim 3.

10. The indicator values ​​for the prediction model are either indicator values ​​that show the degree of fit of the prediction model to the second sample data to be predicted, or indicators that show the prediction accuracy of the prediction model, or both. The prediction device according to claim 1.

11. The processor, for each of the plurality of window widths, in (X), (x1) From the multiple factors represented by the second sample factor data, select one or more factors for which one or more factor fits satisfying the factor fit condition is obtained. (x2) For each of the one or more factors from the second sample factor data, excluding the factors selected in (x1) and (x3) from the multiple factors, calculate the factor fit. From one or more factors in (x3) and (x2), select one or more factors for which one or more factor fits satisfying the factor fit condition are obtained. (x4) If the condition is not met, perform (x2), (x4) If the condition is met, perform (Y). The prediction device according to claim 2.

12. A prediction method performed by a prediction device, For each of several different window widths, a prediction model is constructed based on the second set of prediction data, which is the data corresponding to that window width from the first set of prediction data, and the second set of prediction factor data, which is the data corresponding to that window width from the first set of prediction factor data. The model fit, which is an index value for this prediction model, is then calculated. The window width is the length of the past period starting from the date and time of the prediction target. The first sample data for prediction is time-series data of past values ​​obtained for the prediction target. The first sample factor data mentioned above is data that has a time series of past values ​​obtained for each factor that may affect the target of prediction, By inputting the second set of predictive factor data, which is at least a portion of the first set of predictive factor data, into the target prediction model, the model outputs prediction result data that includes the predicted value of the target. The target prediction model is a prediction model based on a second sample of prediction target data and a second sample of factor data corresponding to the window width of the model fit that satisfies the model fit condition among the calculated model fit values. The first predictive factor data is data that has a time series of future values ​​obtained for each factor that may affect the target of prediction, The second predictive factor data is data that has a time series of values ​​obtained for each factor used in the target predictive model, among the multiple factors represented by the first predictive factor data. Prediction method.

13. For each of several different window widths, a prediction model is constructed based on the second set of prediction data, which is the data corresponding to that window width from the first set of prediction data, and the second set of prediction factor data, which is the data corresponding to that window width from the first set of prediction factor data. The model fit, which is an index value for this prediction model, is then calculated. The window width is the length of the past period starting from the date and time of the prediction target. The first sample data for prediction is time-series data of past values ​​obtained for the prediction target. The first sample factor data mentioned above is data that has a time series of past values ​​obtained for each factor that may affect the target of prediction, By inputting a second set of predictive factor data, which is at least a portion of the first set of predictive factor data, into the target prediction model, the model outputs prediction result data that includes the predicted value of the target. Have the computer do it, The aforementioned prediction model is a prediction model based on a second sample of prediction target data and a second sample of factor data corresponding to a window width that satisfies the model fit condition among the calculated model fit values. The first predictive factor data is data that has a time series of future values ​​obtained for each factor that may affect the target of prediction, The second predictive factor data is data that has a time series of values ​​obtained for each factor used in the target predictive model, among the multiple factors represented by the first predictive factor data. Computer program.

14. The prediction device according to claim 1, A planning management device creates an operation plan based on the prediction result data of the aforementioned prediction device, and controls the controlled device based on the said operation plan. A prediction system equipped with the following features.

Images