JP2001014295A - Data prediction method, data prediction device, and recording medium - Google Patents

Abstract

(57) [Summary] [Problem] To provide a data prediction method capable of coping with a temporal change in the property of time-series data, a data prediction device for implementing the method, and a computer program for causing a computer to function as the device Of a recording medium on which is recorded. A computer functioning as a data prediction device stores a plurality of prediction models in a prediction model database, and uses the plurality of prediction models to convert time-series data stored in an actual performance database. Calculate multiple future data. Then, referring to the first prediction history database 73 in which the prediction history is stored for each prediction model and the second prediction history database 74 in which the prediction history is stored for each evaluation period, the plurality of calculated future data are referred to. Select the best data from.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data prediction method for predicting future data based on time-series data, a data prediction device for performing the method, and a computer for causing a computer to function as the device. The present invention relates to a recording medium on which a program is recorded.

[0002]

2. Description of the Related Art As a prediction model for predicting future data from time-series data, a simple moving average model, an exponential smoothing model, an ARIMA (AutoRegressive Integrated
Moving Average) models, neural network models, and the like have been conventionally used. In addition, a prediction model in which elements such as seasonal variation or trend are incorporated in these models may be used. Further, apart from these models, there is a case where a prediction model focusing on the characteristics of the target time-series data is developed, focusing on the characteristics of the target time-series data.

[0003] Each of the above-mentioned prediction models has its own characteristic. In order to select a prediction model suitable for the target time-series data from these prediction models, these prediction models are actually used. It is necessary to calculate a predicted value using a prediction model, and to evaluate each prediction model based on an error between the predicted value and the actual value. As a result, the prediction model with the smallest error is determined to be the prediction model suitable for the target time-series data.

By the way, the properties of time-series data often change with the passage of time. In order to perform highly accurate prediction, it is necessary to change a prediction model applied in response to such a change. No. Therefore, in the data prediction method, it is important to determine when to change the prediction model.

In order to solve this problem, Japanese Patent Laid-Open No. 7-649 discloses
Japanese Patent Publication No. 65 (hereinafter referred to as “prior art 1”) discloses a sales volume prediction method capable of dynamically correcting a prediction model at an appropriate time by responding to a change in the properties of time-series data using a neural network. Has been proposed. Japanese Patent Application Laid-Open No. 9-179850 (hereinafter referred to as Conventional Technique 2)
Is to calculate more accurate errors by separating the errors that occur in prediction into steady errors inherent in the prediction model itself and disturbance errors due to factors that cannot be considered when building the prediction model. Therefore, a demand prediction model evaluation method capable of optimizing the change time of the prediction model has been proposed.

[0006]

However, in the prior art 1, the prediction model can be dynamically modified, but cannot be applied to prediction models other than the neural network, so that the applicable field is limited. There is a problem that is.

[0007] Further, in the prior art 2, it is possible to specify an optimal change time of the prediction model, but it is necessary to perform an operation of changing the prediction model separately from this. Forecasts of product sales volume etc. often target not only one product but also a plurality of different products.In this case, every time a forecast model is changed, the operator changes the forecast model to all products. This requires a large number of man-hours.

Further, in the prior art 2, when a prediction model is evaluated, a period to be evaluated (hereinafter referred to as an evaluation period) needs to be set in advance by an operator.
If the set evaluation periods are different, the prediction model may be changed at different times, and thus there is a problem that the evaluation can be highly arbitrary.

[0009] The present invention has been made in view of such circumstances, and a plurality of future data are predicted using a plurality of prediction models, and optimal data is selected from the plurality of predicted future data. Accordingly, an object of the present invention is to provide a data prediction method, a data prediction device, and a recording medium that can be used for predicting data in various fields.

Another object of the present invention is to automatically select an optimal prediction model by calculating an error between a plurality of prediction data already predicted and past time series data. It is an object of the present invention to provide a data prediction method, a device for implementing the method, and a recording medium on which a computer program for causing a computer to function as the device is recorded.

Another object of the present invention is to automatically calculate an optimum evaluation period by calculating an error between a plurality of prediction data already predicted for each of a plurality of evaluation periods and past time-series data. An object of the present invention is to provide a selectable data prediction method, a data prediction device, and a recording medium.

[0012]

A data prediction method according to a first aspect of the present invention predicts a plurality of future data from past time-series data using a plurality of prediction models, and evaluates the plurality of prediction models. A data prediction method for selecting specific data from the plurality of predicted future data based on the evaluation, wherein the plurality of evaluation data are selected for each of a plurality of evaluation periods that are target periods for evaluation of a prediction model. Selecting a specific prediction model from among the prediction models, selecting a specific evaluation period from the plurality of evaluation periods, and performing the prediction based on the selected specific prediction model and the specific evaluation period. Specific data is selected from the plurality of future data.

A data prediction method according to a second invention stores a plurality of prediction data already predicted by a plurality of prediction models, and uses the plurality of prediction models to convert future time series data from past time series data. After the plurality of predictions, the plurality of prediction models are evaluated based on the stored plurality of prediction data, and specific data is selected from the plurality of predicted future data based on the evaluation. Calculating a first error between the plurality of prediction data and the time-series data for each of a plurality of evaluation periods, which is a period for which a prediction model is evaluated. Selecting a specific prediction model for each of the evaluation periods based on the first error, and a second error between the prediction data predicted by the selected prediction model and the time-series data. Is calculated for each of the evaluation periods, based on the calculated second error, a step of selecting a specific evaluation period, and based on the selected specific prediction model and the specific evaluation period. Selecting specific data from the plurality of predicted future data.

[0014] A data prediction device according to a third aspect of the present invention includes a storage unit for storing a plurality of prediction data already predicted by a plurality of prediction models, and using the plurality of prediction models to convert past time series data. After predicting a plurality of future data, an evaluation is performed on the plurality of prediction models based on the plurality of prediction data stored in the storage unit, and based on the evaluation, the predicted plurality of future models is evaluated. In a data prediction device that selects specific data from data,
A first calculating unit that calculates a first error between the plurality of prediction data and the time-series data for each of a plurality of evaluation periods, which are periods for which a prediction model is evaluated, and a first calculating unit. Prediction model selection means for selecting a specific prediction model for each of the evaluation periods based on the calculated first error; prediction data predicted by the prediction model selected by the prediction model selection means; Second calculating means for calculating a second error from data for each of the evaluation periods, evaluation period selecting means for selecting a specific evaluation period based on the calculated second error, and the prediction model Based on the specific prediction model selected by the selection unit and the specific evaluation period selected by the evaluation period selection unit, specific data is selected from the plurality of predicted future data. Characterized in that it comprises a data selection means for-option.

A recording medium according to a fourth aspect of the present invention causes a computer to store a plurality of prediction data already predicted by a plurality of prediction models, and use the plurality of prediction models to convert past time-series data into future data. After predicting a plurality of data, an evaluation of the plurality of prediction models is performed based on the stored plurality of prediction data, and based on the evaluation, out of the plurality of predicted future data, In a computer-readable recording medium on which a program for selecting specific data is recorded, a computer is provided with a first error between the plurality of prediction data and the time-series data as an object of evaluation of a prediction model. Program code means for calculating for each of a plurality of evaluation periods, which is a period of time, and a computer for each of the evaluation periods based on the calculated first error. Program code means for selecting a constant prediction model; and program code means for causing a computer to calculate a second error between the prediction data predicted by the selected prediction model and the time-series data for each evaluation period. A program code means for causing the computer to select a specific evaluation period based on the calculated second error; and causing the computer to select the specific prediction model and the specific evaluation period,
A program having program code means for selecting specific data from the plurality of predicted future data is recorded.

According to the present invention, a plurality of prediction data already predicted are stored, and a plurality of future data are predicted from past time-series data using a plurality of prediction models. Is calculated between the predicted data of the above and the past time-series data. The plurality of prediction models are evaluated based on the calculated error, and an optimum prediction model is selected for each evaluation period according to the evaluation result.
Next, an error between the prediction data predicted by the selected prediction model and the time-series data is calculated. An optimal evaluation period is selected based on the calculated error. Then, data specified by the selected optimal prediction model and optimal evaluation period is selected, and the selected data is used as a final result.

As described above, not only a specific prediction model but also a plurality of prediction models can be applied, so that it can be used for prediction of data in various fields.

In addition, an optimal prediction model can be automatically selected and applied based on an error between a plurality of prediction data already predicted and past time-series data.
It is possible to specify an appropriate change time of the prediction model, and it is not necessary for the operator to change the prediction model.

Furthermore, the optimum evaluation period can be automatically selected based on the error between the plurality of prediction data already predicted for each of the plurality of evaluation periods and the past time-series data. This makes it possible to perform more objective data prediction in comparison with the present invention, and can cope with a cycle in which the property of the time-series data changes.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the drawings showing the embodiments. FIG. 1 is a block diagram showing a configuration of an embodiment of a data prediction device according to the present invention. In FIG. 1, reference numeral 1 denotes a computer used as a data prediction device. The computer 1 has a CPU 2, and is provided with a RAM 3 for storing data generated in the CPU 2, an input device such as a keyboard and a mouse. An input unit 4, a display 5 for displaying an input / output screen, a CD on which a program of the present invention is recorded
An external storage device 6 such as a CD-ROM drive or a flexible disk drive for reading a program of the data prediction device according to the present invention from a portable recording medium 10 such as a ROM or a flexible disk, and a book read by the external storage device 6 A hard disk 7 for storing the program of the invention and a communication interface 8 for communicating with the outside are connected. The hard disk 7
Results database 7 that stores past time-series data
1, a prediction model database 72 storing a plurality of prediction models, a first prediction history database 73 storing past prediction values for each prediction model, and storing past prediction values for each evaluation period. Second prediction history database 7
Four.

The computer 1 has a communication interface 8
Is connected to the communication line 20. Further, the communication line 20 is connected to an external network line 22 such as the Internet by a connection device 21 such as a router.

The program of the data prediction device according to the present invention is not only read from the portable recording medium 10 but also connected to the external server computer 23 via the external network line 22.
And stores the program in the hard disk 7 by downloading the program from the recording medium 24 in which the program built in the external server computer 23 is recorded to the computer 1, and stores the stored program in C
The computer 1 functions as the data prediction device of the present invention by the PU 2 loading the RAM 3. The computer 1 thus functioning as a data prediction device
Can execute the processing procedure of the data prediction method of the present invention described later.

FIGS. 2 to 4 show the results database 7 respectively.
FIG. 1 is a conceptual diagram of a first prediction history database 73 and a second prediction history database 74.

The result database 71 shown in FIG. 2 stores the actual value of the monthly demand amount of a certain product in the past. Note that the stored data may be any data that changes with time, such as, for example, the shipment amount or stock amount of the product and the temperature. The time-series data need not be monthly data, but may be data at fixed intervals such as weekly, daily, and hourly.

The first prediction history database 73 shown in FIG.
Stores the history of the predicted value of the demand calculated for each month for each prediction model m (m = 1 to M). This prediction value is calculated for the following month each time the demand for the month is determined. The prediction model m is stored in the prediction model database 72.
Supports all prediction models stored in.
Therefore, for example, when five prediction models are stored in the prediction model database 72, the first prediction history database 73 stores the prediction values calculated for each month for each prediction model m (m = 1 to 5). The history is stored.

The second prediction history database 74 shown in FIG.
Stores the history of the predicted value of the demand calculated for each month for each evaluation period n (n = 1 to N). This prediction value is a prediction value calculated by the prediction model having the highest evaluation in the evaluation period n, and is selected from prediction values stored in the first prediction history database 73 according to a procedure described later. It is. Here, the value of N is set in advance by the operator before executing the data prediction processing according to the present invention. The evaluation period n is a period that starts from the month of the latest time-series data stored in the performance database 71 and goes back in the past. Therefore, in the case of the present embodiment, as shown in FIG.
Since the month of the latest time-series data stored in the performance database 71 is December 1998, for example, the evaluation period n (n = 3) represents three months from October to December 1998. become.

The operation of the computer 1 functioning as a data prediction device will be described below. FIG. 5 is a flowchart showing a processing procedure in a data prediction process of the computer 1 functioning as a data prediction device.

First, an arbitrary prediction model m is obtained from a plurality of prediction models m (m = 1 to M) stored in the prediction model database 72 (S501). Next, using the acquired prediction model m, a predicted value of the demand amount is calculated from the actual value of the demand amount stored in the performance database 71 (S502), and the calculated predicted value is calculated for each prediction model m by the first prediction model m. It is stored in the prediction history database 73 (S50)
3). The above processing of S501 to S503 is performed on all the prediction models m stored in the prediction model database 72 (S504).

FIG. 6 is a conceptual diagram of the first prediction history database 73 after executing the processing of S501 to S504. As shown in FIG. 6, as a result of executing the processing of S501 to S504, the predicted value of the demand amount for January 1999, which is the month to be predicted, is added to the first prediction history database 73 for each prediction model m. Is stored.

Next, the actual value of the demand is acquired from the actual database 71, and the prediction history of the predicted value of the demand calculated by the arbitrary prediction model m is acquired from the first prediction history database 73 (S505). . An evaluation value representing an error between the obtained actual value of the demand amount and the predicted value of the demand amount calculated in the past is calculated for each evaluation period n (n = 1 to N) (S506), and the calculated evaluation value is calculated. The value is stored in the RAM 3 (S507). As the evaluation value, an average of the square errors, an average of the absolute values of the errors, and the like can be used. The processing from S505 to S507 is performed on all the prediction models m stored in the prediction model database 72 (S508).

Next, referring to all the evaluation values stored in the RAM 3, a prediction model having the best evaluation value is selected for each evaluation period n (S509). The predicted value of the demand calculated by the selected prediction model m is stored in the second prediction history database 74 for each evaluation period n (S510).

FIG. 7 is a conceptual diagram of the second prediction history database 74 after executing the processing of S505 to S510. As shown in FIG. 7, as a result of executing the processing of S505 to S510, the predicted value of the demand amount for January 1999, which is the month to be predicted, is added to the second prediction history database 74 for each evaluation period n. Is stored.

Next, the actual value of the demand amount from the actual performance database 71 and the prediction history of the predicted value of the demand amount for an arbitrary evaluation period n from the second prediction history database 73 are respectively stored for a predetermined period L. It is acquired (S511). An evaluation value representing an error between the acquired actual demand value and a predicted demand value calculated in the past is calculated (S512), and the calculated evaluation value is stored in the RAM 3 (S513). In this case, as described above, the average of the square errors and the average of the absolute values of the errors can be used as the evaluation values. The above processing of S511 to S513 is performed for all the evaluation periods n (S514).

Next, referring to all the evaluation values stored in the RAM 3, an evaluation period n having the best evaluation value is selected (S515). The predicted value of the demand amount for the selected evaluation period n is displayed on the display 5 as a final result (S
516).

The above processing is repeated every time the monthly demand is determined. Then, the predicted value of the demand calculated by the optimal prediction model in each case is used as the final result. Therefore, not only can the appropriate change time of the prediction model be specified, but at the same time, the work of changing the prediction model is being performed. Therefore, unlike the case of the prior art 1, it is not necessary for the operator to change the prediction model every time the prediction model is changed.

[0036]

As described above, not only a specific prediction model but also a plurality of prediction models can be applied.
It can be used to predict data in various fields.

Further, the most suitable prediction model can be automatically selected and applied based on the error between a plurality of prediction data already predicted and the past time series data.
It is possible to specify an appropriate change time of the prediction model, and it is not necessary for the operator to change the prediction model.

Further, the optimum evaluation period can be automatically selected based on the error between the plurality of prediction data already predicted for each of the plurality of evaluation periods and the past time-series data. The present invention has excellent effects, for example, it is possible to perform more objective data prediction compared to the present invention, and it is possible to cope with a cycle in which the property of the time-series data changes.

As a result of applying the present invention to the steel field,
The effects described above have been observed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a data prediction device according to the present invention.

FIG. 2 is a conceptual diagram of a performance database.

FIG. 3 is a conceptual diagram of a first prediction history database.

FIG. 4 is a conceptual diagram of a second prediction history database.

FIG. 5 is a flowchart showing a processing procedure in a data prediction process of a computer functioning as a data prediction device.

FIG. 6 is a conceptual diagram of a first prediction history database.

FIG. 7 is a conceptual diagram of a second prediction history database.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Computer 2 CPU 3 RAM 4 Input part 5 Display 6 External storage device 7 Hard disk 8 Communication interface 71 Result database 72 Prediction model database 73 First prediction history database 74 Second prediction history database

Continued on the front page (72) Inventor Megumi Nagai 4-5-33 Kitahama, Chuo-ku, Osaka City, Osaka Prefecture Inside Sumitomo Metal Industries, Ltd. (72) Inventor Taichi Imai 4-5-33 Kitahama, Chuo-ku, Osaka City, Osaka Prefecture Sumitomo Metal Industries Co., Ltd. (72) Inventor Matsu ▲ Saki ▼ Kenichi 4-53 Kitahama, Chuo-ku, Osaka City, Osaka Prefecture F-term (reference) 5B049 AA02 CC11 EE01 EE31

Claims (4)

[Claims] After predicting a plurality of future data from past time-series data using a plurality of prediction models, an evaluation is performed on the plurality of prediction models, and based on the evaluation, the predicted plurality of future models are evaluated. In a data prediction method of selecting specific data from among the data, a specific prediction model is selected from the plurality of prediction models for each of a plurality of evaluation periods, which are periods to be evaluated by the prediction model, A specific evaluation period is selected from a plurality of evaluation periods, and specific data is selected from the plurality of predicted future data based on the selected specific prediction model and the specific evaluation period. A data prediction method characterized by: 2. A method for storing a plurality of prediction data already predicted by a plurality of prediction models, predicting a plurality of future data from past time-series data using the plurality of prediction models, and storing the plurality of prediction data. A data prediction method for evaluating the plurality of prediction models based on a plurality of prediction data and selecting specific data from the plurality of predicted future data based on the evaluation; Calculating a first error between the predicted data and the time-series data for each of a plurality of evaluation periods, which are periods for which the prediction model is evaluated, based on the calculated first error, Selecting a specific prediction model for each evaluation period; calculating a second error between prediction data predicted by the selected prediction model and the time-series data for each evaluation period Selecting a specific evaluation period based on the calculated second error; and, based on the selected specific prediction model and the specific evaluation period, the predicted plurality of future periods. Selecting specific data from the data. 3. A storage unit for storing a plurality of prediction data already predicted by a plurality of prediction models, and after predicting a plurality of future data from past time-series data using the plurality of prediction models. Performing an evaluation on the plurality of prediction models based on the plurality of prediction data stored in the storage unit, and, based on the evaluation, identifying specific data from the plurality of predicted future data. In the data prediction device to be selected, a first calculation unit that calculates a first error between the plurality of prediction data and the time-series data for each of a plurality of evaluation periods that are target periods for evaluation of a prediction model; A prediction model selecting unit that selects a specific prediction model for each evaluation period based on the first error calculated by the first calculation unit; and a prediction model selected by the prediction model selecting unit. A second of the prediction data predicted by Dell and the time-series data
A second evaluation means for calculating an error of each of the evaluation periods, an evaluation period selection means for selecting a specific evaluation period based on the calculated second error, and a prediction model selection means. Data selecting means for selecting specific data from the plurality of predicted future data based on the specific prediction model and the specific evaluation period selected by the evaluation period selecting means. A data prediction device characterized by the above-mentioned. 4. A method in which a plurality of prediction data already predicted by a plurality of prediction models are stored in a computer, and a plurality of future data are predicted from past time-series data using the plurality of prediction models. A program for causing the plurality of prediction models to be evaluated based on the stored plurality of prediction data, and selecting specific data from the plurality of predicted future data based on the evaluation. A computer-readable recording medium on which a first error between the plurality of prediction data and the time-series data is stored in a computer. A program code means for calculating for each evaluation period, and a computer selecting a specific prediction model for each evaluation period based on the calculated first error. That the program code means, to a computer, a second error between said time-series data and the prediction data predicted by the selected prediction model,
Program code means for calculating for each evaluation period; program code means for causing a computer to select a specific evaluation period based on the calculated second error; and A program having program code means for selecting specific data from the plurality of predicted future data based on the specific evaluation period, wherein the program is readable by a computer. Recording medium.