Prediction model creation device and prediction model creation method

Abstract

To efficiently provide a prediction model with high prediction accuracy.SOLUTION: A prediction model creation device includes: a first processing unit that divides a plurality of prediction models for predicting a consumer load into a first group with higher priority orders and a second group with lower priority orders; a second processing unit that, when the priority orders are switched between the first group and the second group after reviewing the priority orders of the plurality of prediction models based on a difference between a prediction value and an actual value of the consumer load using the plurality of prediction models, reviews the priority orders of the plurality of prediction models until the priority orders of at least the plurality of prediction models are no longer switched between the first group and the second group; and a third processing unit that creates a prediction model for the consumer based on the prediction models belonging to the first group.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a prediction model creation device and a prediction model creation method for creating a prediction model for predicting the load of a customer. relates to.

Background Art

[0002] For example, when predicting how the load of a customer (the energy demand of the customer) changes under the influence of factors such as the temperature and day of the week at the location of the customer, it is known to predict using a prediction model employing techniques such as neural networks and multiple regression analysis (for example, Patent Document 1).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Since the load of customers varies from customer to customer, when creating the above prediction model, for example, an analyst analyzes the factors affecting the load of each customer, and based on this analysis result, determines the explanatory variables to be used in the prediction model so as to improve the prediction accuracy of the prediction model.

[0005] In recent years, with the spread of microgrids, which is a mechanism for collectively selling or lending the generated electricity of distributed power sources using renewable energy such as solar power and wind power to the customers in each region where the distributed power sources are installed, the number of customers requiring the above prediction model has been increasing.

[0006] When an analyst creates a prediction model, improving the prediction accuracy of the prediction model depends on the analyst's knowledge and experience, and it takes a lot of time to complete the prediction model. Therefore, when the number of customers increases, there is a risk that the creation of the prediction model cannot keep up.

[0007] The present invention has been made in view of the above problems, and an object thereof is to provide a prediction model creation device and a prediction model creation method that improve the prediction accuracy of a prediction model for predicting a customer's load and efficiently create a prediction model.

Means for Solving the Problems

[0008] One of the present inventions for achieving the above object is a prediction model creation device, which divides a plurality of prediction models for predicting a customer's load into a first group with a higher priority and a second group with a lower priority. A first processing unit for classifying; based on the difference between the predicted value and the actual value of the customer's load by the plurality of prediction models, after rechecking the priority order of the plurality of prediction models, if the priority order has been swapped between the first group and the second group, A second processing unit that rechecks the priority order of the plurality of prediction models and calculates a customer load prediction value until the priority order of at least the plurality of prediction models no longer swaps between the first group and the second group; and based on the prediction models belonging to the first group, A third processing unit for creating a prediction model for the customer, is included.

[0009] According to the prediction model creation device of the present invention, since a prediction model for a customer is created based on a prediction model in which the priority orders of a plurality of prediction models for predicting a customer's load continuously belong to the first group with a higher priority, it is possible to efficiently provide a prediction model with high prediction accuracy to the customer.

[0010] Another aspect of the present invention for achieving the above object is a prediction model creation device, wherein the first group is a group to which a plurality of the prediction models belong, and after the processing of the second processing unit is performed, based on the difference between the predicted value and the actual value of the load of the consumer by a plurality of the prediction models belonging to the first group, a fourth processing unit that rechecks the priority order of the plurality of the prediction models belonging to the first group and calculates a predicted value of the consumer load is included.

[0011] According to the prediction model creation device of the present invention, since the priority order of the plurality of prediction models belonging to the first group is further rechecked, it is possible to further improve the prediction accuracy of the prediction model provided to the consumer.

[0012] Another aspect of the present invention for achieving the above object is a prediction model creation device, wherein the second processing unit rechecks the priority order of the plurality of the prediction models until the number of times the priority order of the plurality of the prediction models does not continuously change between the first group and the second group reaches a predetermined number of times.

[0013] According to the prediction model creation device of the present invention, since the priority order of the prediction models belonging to the first group is stabilized, it is possible to further improve the prediction accuracy of the prediction model provided to the consumer.

[0014] Another aspect of the present invention for achieving the above object is a prediction model creation device, wherein in the processing of the second processing unit and the fourth processing unit, the predicted value of the load of the consumer by the prediction model belonging to the first group is weighted according to the prediction accuracy, evaluation rank average value, or priority order of the prediction model.

[0015] According to the prediction model creation device of the present invention, since the predicted value of the prediction model belonging to the first group is weighted, it is possible to further improve the prediction accuracy of the prediction model provided to the consumer.

[0016] Another aspect of the present invention for achieving the above object is a prediction model creation device, wherein the predicted value of the load of the customer by the prediction model belonging to the first group can be weighted using a linear or non-linear membership function.

[0017] In addition, the problems disclosed in the present application and the solutions thereto will be clarified by the column of the embodiments for carrying out the invention and the drawings.

Effects of the Invention

[0018] According to the present invention, it is possible to efficiently provide a prediction model with high prediction accuracy for customers.

Brief Description of the Drawings

[0019]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5A

Figure 5B

Figure 5C

Figure 5D

Figure 5E

Figure 5F

Figure 6

Figure 7

Figure 8

Figure 9

[0020] At least the following matters become clear from the description in this specification and the accompanying drawings. Hereinafter, the present invention will be described with reference to the accompanying drawings according to one embodiment thereof.

[0021] FIG. 1 is a block diagram showing a schematic configuration of the prediction model creation apparatus 1 according to the present embodiment.

[0022] The prediction model creation apparatus 1 is an apparatus that creates a prediction model for predicting the load of a customer, and is an apparatus that can automatically and efficiently create a prediction model with high prediction accuracy and provide it to the customer. Note that various models can be adopted as the prediction model, but in the present embodiment, for example, a prediction model using explanatory variables such as a neural network and multiple regression analysis will be adopted.

[0023] The prediction model creation device 1 includes an input unit 110, a control unit 120, a storage unit 130, and an output unit 140 as means for realizing the above functions, and is connected to an external computer 150 via a communication network 160.

[0024] In order to efficiently provide a prediction model with high prediction accuracy to the customers by the prediction model creation device 1, a plurality of prediction models with different explanatory variables capable of predicting the load of the customers are prepared in advance for each customer.

[0025] FIG. 2 is a flowchart showing an example of the procedure until a plurality of prediction models prepared for each customer are determined, and the procedure until a questionnaire is conducted for each customer and the prediction model ranked first in the initial priority is determined from among the plurality of prediction models. In the present embodiment, eight prediction models are prepared for each customer.

[0026] First, it is conceivable that there are cases where the daily change in the load of the customer varies depending on the temperature at the location where the customer exists and cases where it does not. Therefore, assuming that the daily change in the load of the customer is affected by the temperature at the location where the customer exists (S1000: YES), the "temperature" at the location where the customer exists is added as an explanatory variable of the prediction model (S1010).

[0027] Next, it is conceivable that there are cases where the daily change in the load of the customer varies depending on the day of the week and cases where it does not. Therefore, assuming that the daily change in the load of the customer is affected by the day of the week (S1020: YES), (a) the case where the daily change in the load of the customer is affected by the difference between weekdays (Monday to Friday) and holidays (Saturday and Sunday), (b) in addition to being affected by the difference between weekdays and holidays, also being affected by the difference in the day of the week within weekdays, (c) the case of being affected by the specified day of the week among the days of the week from Monday to Sunday, can be considered (S1030).

[0028] Therefore, assuming the case of (a) above, "day of the week" divided into weekdays and holidays is added as an explanatory variable to the prediction model (S1040). Then, a prediction model "Te·D1" is created, which includes "temperature" at the location where the customer exists and "day of the week" divided into weekdays and holidays as explanatory variables (S1050).

[0029] Also, assuming the case of (b) above, "day of the week" is divided into weekdays and holidays, and "day of the week" divided into each weekday from Monday to Friday is added as an explanatory variable to the prediction model (S1060). Then, a prediction model "Te·D2" is created, which includes "temperature" at the location where the customer exists and "day of the week" divided into weekdays and holidays and further divided into each weekday from Monday to Friday as explanatory variables (S1070).

[0030] Also, assuming the case of (c) above, "day of the week" specified among the days of the week from Monday to Sunday is added as an explanatory variable to the prediction model (S1080). Then, a prediction model "Te·D3" is created, which includes "temperature" at the location where the customer exists and "day of the week" specified among the days of the week from Monday to Sunday as explanatory variables (S1090).

[0031] Also, assuming that the daily change in the customer's load is affected by the temperature at the location where the customer exists but not by the day of the week (S1020: NO), a prediction model "Te" is created, which includes "temperature" at the location where the customer exists as an explanatory variable but does not include "day of the week" as an explanatory variable (S1100).

[0032] Next, although there is no difference in the daily change of the customer's load depending on the temperature at the location where the customer is located (S1000: NO), it is conceivable that there are cases where there is a difference in the daily change of the customer's load depending on the day of the week and cases where there is no difference. Therefore, assuming that the daily change of the customer's load is affected by the day of the week (S1110: YES), (d) the case where the daily change of the customer's load is affected by the difference between weekdays (Monday to Friday) and holidays (Saturday and Sunday), (e) in addition to being affected by the difference between weekdays and holidays, the case where it is also affected by the difference in the day of the week within weekdays, and (f) the case where it is affected by the specified day of the week among the days of the week from Monday to Sunday, can be considered (S1120).

[0033] Therefore, assuming the case of (d) above, "day of the week" divided into weekdays and holidays is added to the explanatory variables of the prediction model (S1130). Then, a prediction model "D1" including "day of the week" divided into weekdays and holidays as explanatory variables is created (S1140).

[0034] Also, assuming the case of (e) above, in the explanatory variables of the prediction model, while dividing the day of the week into weekdays and holidays, "day of the week" divided into each day of the week from Monday to Friday for weekdays is added (S1150). Then, a prediction model "D2" including "day of the week" that divides the day of the week into weekdays and holidays and further divides weekdays into each day of the week from Monday to Friday as explanatory variables is created (S1160).

[0035] Also, assuming the case of (f) above, "day of the week" specified among the days of the week from Monday to Sunday is added to the explanatory variables of the prediction model (S1170). Then, a prediction model "D3" including "day of the week" specified among the days of the week from Monday to Sunday as explanatory variables is created (S1180).

[0036] Also, assuming that the daily change of the customer's load is not affected by the day of the week (S1110: NO), a prediction model "N" that does not include "day of the week" as an explanatory variable is created (S1190).

[0037] Through the above procedures, eight prediction models, namely, "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", and "N", are prepared for each customer to predict the load based on the past performance of each customer. For example, when the eight prediction models, "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", and "N", are composed of neural networks, each prediction model is a learning model trained using learning data (supervised data) that associates data such as past temperature, day of the week, etc., which are feature quantities, and the date and time at the customer, with data such as past load changes at the customer, which are labels.

[0038] FIG. 3 is a diagram showing an example of the structure when the prediction model is a neural network.

[0039] The prediction model is composed of three layers: an input layer 210, an intermediate layer 220, and an output layer 230. The input layer 210 is a layer into which data of the same type as the feature quantities, such as the temperature and day of the week at the location where the customer exists, and the date and time are input. The intermediate layer 220 is a layer including one or more hidden layers composed of one or more nodes including parameters adjusted by learning using the above learning data. The intermediate layer 220 predicts the load of the customer based on the data input to the input layer 210. The output layer 230 is a layer that outputs the prediction result by the intermediate layer 220. The explanatory variables input to the input layer 210 are, for example, temperature, day of the week (weekday, holiday, etc.), month (January to December), time (hour, minute, etc.), etc., and the explanatory variables vary depending on each prediction model. For example, in the case of the prediction model "N", neither the temperature data nor the day-of-the-week data is included in the explanatory variables.

[0040] For each customer, a questionnaire such as the flowchart of FIG. 2 is conducted. Among the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N", the prediction model determined to be the most suitable explanatory variable for predicting the customer's load is set as the prediction model with the initial priority of 1st place. In this way, by determining the prediction model with the initial priority of 1st place based on the questionnaire results for each customer, it becomes possible to efficiently create a prediction model with high prediction accuracy by the processing of the present embodiment described later. Note that conducting a questionnaire for each customer is just an example and is not limited thereto. It may be possible to analyze the past load records of the customers and determine the prediction model with the initial priority of 1st place based on the analysis results.

[0041] Returning to FIG. 1, the storage unit 130 has a first storage area 131 and a second storage area 132. In the first storage area 131, a control program for operating the prediction model creation device 1 is stored. In the second storage area 132, work data and processed data in the process of the first processing unit 121 to the fourth processing unit 124 performing processing are stored. For example, in the second storage area 132, data indicating eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" associated with the initial priority, priority, and evaluation rank, prediction values, etc. are stored.

[0042] For each customer, the data indicating the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" and the data indicating the prediction model determined to have the initial priority of 1st place are stored in the second storage area 132 of the storage unit 130 after being input to the input unit 110.

[0043] The control unit 120 is configured to include a first processing unit 121, a second processing unit 122, a third processing unit 123, and a fourth processing unit 124.

[0044] FIG. 4 is a diagram showing an example of a prediction model with an initial priority of 1st place and prediction models with initial priorities from 2nd to 8th place following this prediction model.

[0045] The first processing unit 121 reads from the second storage area 132 of the storage unit 130, for each customer, data indicating eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N", and data indicating the prediction model determined to have the 1st place in the initial priority among the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N", and determines prediction models with initial priorities from 2nd to 8th place.

[0046] First, when the prediction model with the 1st place in the initial priority is, for example, "Te", the first processing unit 121 determines the prediction models with initial priorities from 2nd to 8th place to be "Te·D1", "Te·D2", "N", "D1", "D2", "Te·D3", "D3", respectively.

[0047] Second, when the prediction model with the 1st place in the initial priority is, for example, "Te·D1", the first processing unit 121 determines the prediction models with initial priorities from 2nd to 8th place to be "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3", respectively.

[0048] Third, when the prediction model with the 1st place in the initial priority is, for example, "Te·D2", the first processing unit 121 determines the prediction models with initial priorities from 2nd to 8th place to be "Te·D1", "Te", "D2", "D1", "N", "Te·D3", "D3", respectively.

[0049] Fourth, when the prediction model with the 1st place in the initial priority is, for example, "Te·D3", the first processing unit 121 determines the prediction models with initial priorities from 2nd to 8th place to be "D3", "Te", "N", "Te·D1", "Te·D2", "D1", "D2", respectively.

[0050] Also, as the fifth case, when the prediction model with the initial priority rank of 1st is, for example, "N", the first processing unit 121 determines the prediction models with the initial priorities from the 2nd to the 8th to be "D1", "D2", "Te", "Te·D1", "Te·D2", "D3", "Te·D3", respectively.

[0051] Also, as the sixth case, when the prediction model with the initial priority rank of 1st is, for example, "D1", the first processing unit 121 determines the prediction models with the initial priorities from the 2nd to the 8th to be "D2", "N", "Te·D1", "Te·D2", "Te", "D3", "Te·D3", respectively.

[0052] Also, as the seventh case, when the prediction model with the initial priority rank of 1st is, for example, "D2", the first processing unit 121 determines the prediction models with the initial priorities from the 2nd to the 8th to be "D1", "N", "Te·D2", "Te·D1", "Te", "D3", "Te·D3", respectively.

[0053] Finally, as the eighth case, when the prediction model with the initial priority rank of 1st is, for example, "D3", the first processing unit 121 determines the prediction models with the initial priorities from the 2nd to the 8th to be "Te·D3", "N", "Te", "D1", "D2", "Te·D1", "Te·D2", respectively.

[0054] The ranks of the prediction models with the initial priorities from the 2nd to the 8th are predetermined according to the prediction model with the initial priority rank of 1st, but may be appropriately changed by considering, for example, the actual performance values of past loads of various customers.

[0055] Further, among the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N", the first processing unit 121 groups three prediction models with initial priorities from, for example, the 1st to the 3rd (predetermined order) as main models into the first group, and groups five prediction models with initial priorities from, for example, the 4th to the 8th as support models into the second group, thus classifying the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" into two groups. Note that the number of prediction models belonging to the first group and the second group can be changed as appropriate. Note that it is also possible to randomly select three prediction models from the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" and group them into the first group, and group the remaining five prediction models into the second group.

[0056] The first processing unit 121 stores in the second storage area 132 of the storage unit 130 the data in which the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N", the initial priorities from the 1st to the 8th, and the first and second groups are associated with each other.

[0057] The second processing unit 122 calculates predicted values for eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" to which an initial priority or a priority after reviewing the initial priority is assigned. The second processing unit 122 performs weighting, which will be described later, on the predicted values before weighting of the three prediction models ranked from first to third belonging to the first group to calculate the predicted values after weighting, and uses them as the customer load predicted values. Then, for each evaluation period, the second processing unit 122 uses the predicted values before weighting and the actual values for the three prediction models in the first group, and for the five prediction models ranked from fourth to eighth belonging to the second group, uses the calculated predicted values and the actual values. The smaller the difference between the predicted value and the actual value, the higher the prediction accuracy. The second processing unit 122 assigns an evaluation rank to the prediction model. For each prediction model ranking review period, the second processing unit 122 calculates the average value of the evaluation ranks of each prediction model. The smaller the average value of the evaluation ranks, the higher the prediction accuracy. The second processing unit 122 performs a review process to review the priority so that the prediction model moves to the upper side. Then, for each evaluation period, the fourth processing unit 124 assigns an evaluation rank to the three prediction models in the first group. The smaller the difference between the predicted value before weighting and the actual value, the higher the prediction accuracy. The fourth processing unit 124 calculates the average value of the evaluation ranks of each prediction model for each prediction model ranking review period. The smaller the average value of the evaluation ranks, the higher the prediction accuracy. The fourth processing unit 124 performs a review process to review the priority so that the prediction model moves to the upper side. Note that the rank for each evaluation period is referred to as the evaluation rank, and the rank for each prediction model ranking review period is referred to as the priority rank.

[0058] Further, as a result of performing the above review process, if there is a swap in the ranking of the prediction models between the first group and the second group, such that any one of the eight prediction models "Te·D1", "Te·D2", "Te·D3", "Te", "D1", "D2", "D3", "N" moves from the first group to the second group, and accordingly, any other prediction model moves from the second group to the first group, the second processing unit 122 repeats the above review process until there is no longer a swap in the ranking of the prediction models between at least the first group and the second group.

[0059] After the second processing unit 122 repeats the above review process until there is no longer a swap in the ranking of the prediction models between at least the first group and the second group, the fourth processing unit 124 performs a review process of the priority ranking from the first to the third for only the three prediction models that have come to belong to the first group. Note that the predicted value of the prediction model within the first group when the second processing unit 122 and the fourth processing unit 124 perform the review process is the predicted value before weighting of the prediction model within the first group acquired via the input unit 110, and the predicted value calculated as the customer load predicted value is the predicted value after weighting obtained by performing weighting described later on the predicted value before weighting within the first group.

[0060] FIG. 5A is a diagram showing an example of numerical values in the first review process performed by the second processing unit 122. FIG. 5B is a diagram showing an example of numerical values in the second review process performed by the second processing unit 122 following FIG. 5A. FIG. 5C is a diagram showing an example of numerical values in the third review process performed by the second processing unit 122 following FIG. 5B. FIG. 5D is a diagram showing an example of numerical values in the fourth review process performed by the second processing unit 122 following FIG. 5C. FIG. 5E is a diagram showing an example of numerical values in the fifth review process performed by the fourth processing unit 124 following FIG. 5D. FIG. 5F is a diagram showing an example of numerical values in the sixth review process performed by the fourth processing unit 124 following FIG. 5E. In the present embodiment, as an example, the procedure of the review process when the initial priority of the prediction model "Te·D1" shown second in FIG. 4 is the first place will be described. Here, as an example of the review process, the evaluation period of the prediction model is a days, and the review period of the ranking of the prediction model is set to be b times when the a days of the evaluation period is counted once. In the present embodiment, for the sake of easy understanding of the explanation, the review period is set to be three times the evaluation period.

[0061] In FIG. 5A, the prediction models are arranged in the order of "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" with initial priorities from the 1st to the 8th. The top three prediction models "Te·D1", "Te·D2", "Te" belong to the first group, and the bottom five prediction models "D1", "D2", "N", "Te·D3", "D3" belong to the second group. First, when the evaluation was performed in the first evaluation cycle (Evaluation Cycle 1), the evaluation ranks of the prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" were 1st, 3rd, 6th, 2nd, 4th, 5th, 7th, and 8th, respectively. Next, when the evaluation was performed in the second evaluation cycle (Evaluation Cycle 2), the evaluation ranks of the prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" were 2nd, 4th, 6th, 1st, 3rd, 5th, 8th, and 7th, respectively. Next, when the evaluation was performed in the third evaluation cycle (Evaluation Cycle 3), the evaluation ranks of the prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" were 2nd, 4th, 5th, 1st, 3rd, 6th, 8th, and 7th, respectively. By performing evaluations in the evaluation cycles from the first to the third, the re-ranking cycle of the first prediction model was reached. Therefore, for the prediction models "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" with initial priorities from the 2nd to the 8th, the average value of the evaluation ranks from the first to the third is obtained, and this average value is used as the average evaluation rank. On the other hand, for the prediction model "Te·D1" with an initial priority of 1st, based on the questionnaire for customers, since it is considered the model with the highest priority, in order to prevent frequent changes in the priority, the average value of the evaluation ranks from the first to the third is obtained, and a value obtained by multiplying this average value by a priority rate p less than 1 (for example, p = 0.75) is used as the average evaluation rank.As a result, in the first review process shown in Fig. 5A, the average evaluation ranks of the prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" are "1.25", "3.67", "5.67", "1.33", "3.33", "5.33", "7.67", "7.33", respectively. Thus, the priority order of the prediction models after the first review process is "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" in that order. The prediction models "D1" and "D2" that belonged to the second group move to the first group, and the prediction models "Te·D2" and "Te" that belonged to the first group move to the second group, and an exchange of the prediction models occurs between the first group and the second group.

[0062] In FIG. 5B, after the first review process shown in FIG. 5A, the prediction models are arranged in the order of "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" with priorities from the first to the eighth. The top three prediction models "Te·D1", "D1", "D2" belong to the first group, and the bottom five prediction models "Te·D2", "N", "Te", "D3", "Te·D3" belong to the second group. First, when the evaluation was carried out in the fourth evaluation cycle (evaluation cycle 4), the evaluation ranks of the prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" were 2nd, 1st, 3rd, 5th, 4th, 6th, 8th, and 7th respectively. Next, when the evaluation was carried out in the fifth evaluation cycle (evaluation cycle 5), the evaluation ranks of the prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" were 2nd, 1st, 3rd, 5th, 6th, 4th, 7th, and 8th respectively. Next, when the evaluation was carried out in the sixth evaluation cycle (evaluation cycle 6), the evaluation ranks of the prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" were 2nd, 1st, 4th, 3rd, 6th, 5th, 7th, and 8th respectively. By performing the evaluation in the evaluation cycles from the fourth to the sixth, the second review cycle of the prediction model rankings was reached. Therefore, for the prediction models "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" with priorities from the second to the eighth among the prediction models after the first review process, the average value of the evaluation ranks from the fourth to the sixth is obtained, and this average value is taken as the average evaluation rank value. On the other hand, for the prediction model "Te·D1" with a priority of 1st among the prediction models after the first review process, similar to FIG. 5A, the average value of the evaluation ranks from the fourth to the sixth is obtained, and the value obtained by multiplying this average value by a priority rate p less than 1 is taken as the average evaluation rank value.As a result, in the second review process shown in FIG. 5B, the average values of the evaluation rankings of the prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" are "1.50", "1.00", "3.33", "4.33", "5.33", "5.00", "7.33", "7.67", respectively. Thus, the priority order of the prediction models after the second review process is "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3", and no replacement of the prediction models is performed between the first group and the second group.

[0063] In FIG. 5C, the prediction models after the second review process shown in FIG. 5B are arranged in the order of "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" with priorities from the first to the eighth. The top three prediction models "D1", "Te·D1", "D2" belong to the first group, and the bottom five prediction models "Te·D2", "Te", "N", "D3", "Te·D3" belong to the second group. First, when the evaluation was performed in the seventh evaluation cycle (evaluation cycle 7), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" were the second, first, third, fifth, sixth, fourth, seventh, and eighth, respectively. Next, when the evaluation was performed in the eighth evaluation cycle (evaluation cycle 8), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" were the first, second, third, fifth, sixth, fourth, eighth, and seventh, respectively. Next, when the evaluation was performed in the ninth evaluation cycle (evaluation cycle 9), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" were the second, first, fourth, third, fifth, sixth, eighth, and seventh, respectively. By performing the evaluation in the evaluation cycles from the seventh to the ninth, the third ranking review cycle of the prediction model was reached. Therefore, for the prediction models "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" with priorities from the second to the eighth among the prediction models after the second review process, the average value of the evaluation ranks from the seventh to the ninth is obtained, and this average value is used as the average evaluation rank value. On the other hand, for the prediction model "D1" with a priority of 1 among the prediction models after the second review process, similar to FIG. 5A, the average value of the evaluation ranks from the seventh to the ninth is obtained, and the value obtained by multiplying this average value by a priority rate p less than 1 is used as the average evaluation rank value.As a result, in the third review process shown in FIG. 5C, the average evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" are "1.25", "1.33", "3.33", "4.33", "5.67", "4.67", "7.67", "7.33", respectively. Thus, the priority order of the prediction models after the third review process is "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" in this order, and the replacement of the prediction models between the first group and the second group is not continuously performed twice in the second and third review processes.

[0064] In FIG. 5D, after the third review process shown in FIG. 5C, the prediction models are arranged in the order of "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" with priorities from the first to the eighth. The top three prediction models "D1", "Te·D1", "D2" belong to the first group, and the bottom five prediction models "Te·D2", "N", "Te", "Te·D3", "D3" belong to the second group. First, when the evaluation was carried out in the 10th evaluation cycle (evaluation cycle 10), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" were the 1st, 2nd, 4th, 5th, 3rd, 6th, 7th, and 8th respectively. Next, when the evaluation was carried out in the 11th evaluation cycle (evaluation cycle 11), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" were the 2nd, 1st, 3rd, 4th, 6th, 5th, 8th, and 7th respectively. Next, when the evaluation was carried out in the 12th evaluation cycle (evaluation cycle 12), the evaluation ranks of the prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" were the 1st, 2nd, 3rd, 5th, 4th, 6th, 8th, and 7th respectively. By performing evaluations in the evaluation cycles from the 10th to the 12th, the fourth ranking review cycle of the prediction models was reached. Therefore, for the prediction models "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" whose priorities after the third review process are from the 2nd to the 8th, the average value of the evaluation ranks from the 10th to the 12th is obtained, and this average value is taken as the average evaluation rank value. On the other hand, for the prediction model "D1" whose priority after the third review process is the 1st, similar to FIG. 5A, the average value of the evaluation ranks from the 10th to the 12th is obtained, and the value obtained by multiplying this average value by a priority rate p less than 1 is taken as the average evaluation rank value.As a result, in the fourth review process shown in FIG. 5D, the average values of the evaluation rankings of the prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" are "1.00", "1.67", "3.33", "4.67", "4.33", "5.67", "7.67", "7.33", respectively. Thus, the priority order of the prediction models after the fourth review process is "D1", "Te·D1", "D2", "N", "Te·D2", "Te", "D3", "Te·D3", and the swapping of the prediction models between the first group and the second group has not been continuously performed three times in the second, third, and fourth review processes.

[0065] As a result of reordering the prediction models, after the prediction model rankings were swapped between the first group and the second group, if the prediction model rankings were not swapped continuously, for example, M times (e.g., M = 3) between the first group and the second group, it is considered that the possibility of the three prediction models belonging to the first group moving to the second group becomes extremely low. After that, the review process is continued only for the prediction models "D1", "Te·D1", "D2" that belonged to the first group as a result of the fourth review process shown in FIG. 5D, and the review process is not performed for the prediction models "N", "Te·D2", "Te", "D3", "Te·D3" that belonged to the second group.

[0066] In addition, in this embodiment, for the sake of easy understanding, after the prediction model rankings were swapped between the first group and the second group, the number of times to confirm that the prediction model rankings do not continuously swap between the first group and the second group was set to 3 times. However, it may be set to any number of times of 0 or more. In this way, by performing the review processes of FIGS. 5A to 5D by the second processing unit 122, the three prediction models belonging to the first group are determined.

[0067] In FIG. 5E, after the fourth review process shown in FIG. 5D, the prediction models within the first group are arranged in the order of "D1", "Te·D1", "D2" with priorities assigned from the first to the third. As described above, the review process continues only for the top three prediction models "D1", "Te·D1", "D2". First, when the evaluation was carried out in the 13th evaluation cycle (evaluation cycle 13), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the first, third, and second respectively. Next, when the evaluation was carried out in the 14th evaluation cycle (evaluation cycle 14), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the second, first, and third respectively. Next, when the evaluation was carried out in the 15th evaluation cycle (evaluation cycle 15), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the second, first, and third respectively. By performing evaluations in the evaluation cycles from the 13th to the 15th, the fifth ranking review cycle of the prediction models was reached. Therefore, for the prediction models "Te·D1" and "D2" whose priorities were the second and third after the fourth review process, the average value of the evaluation rankings from the 13th to the 15th was obtained, and this average value was taken as the average evaluation ranking value. On the other hand, for the prediction model "D1" whose priority was the first after the fourth review process, similar to FIG. 5A, the average value of the evaluation rankings from the 13th to the 15th was obtained, and the value obtained by multiplying this average value by a priority rate p less than 1 was taken as the average evaluation ranking value. As a result, in the fifth review process shown in FIG. 5E, the average evaluation rankings of the prediction models "D1", "Te·D1", "D2" are "1.25", "1.67", and "2.67" respectively. Thereby, the priority order of the prediction models within the first group after the fifth review process is the same as the priority order after the fourth review process, in the order of "D1", "Te·D1", "D2".

[0068] In Fig. 5F, after the fifth review process shown in Fig. 5E, the prediction models within the first group are arranged in the order of "D1", "Te·D1", "D2" with priorities assigned from the first to the third. First, when evaluations were conducted in the 16th evaluation cycle (evaluation cycle 16), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the 2nd, 1st, and 3rd respectively. Next, when evaluations were conducted in the 17th evaluation cycle (evaluation cycle 17), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the 1st, 2nd, and 3rd respectively. Next, when evaluations were conducted in the 18th evaluation cycle (evaluation cycle 18), the evaluation rankings of the prediction models "D1", "Te·D1", "D2" were the 2nd, 1st, and 3rd respectively. By conducting evaluations in the evaluation cycles from the 16th to the 18th, the sixth review cycle of the prediction model rankings was reached. Therefore, for the prediction models "Te·D1" and "D2" whose priorities after the fifth review process were the 2nd and 3rd, the average value of the evaluation rankings from the 16th to the 18th was calculated and taken as the average evaluation ranking value. On the other hand, for the prediction model "D1" whose priority after the fifth review process was the 1st, similar to Fig. 5A, the average value of the evaluation rankings from the 16th to the 18th was calculated, and the value obtained by multiplying this average value by a priority rate p less than 1 was taken as the average evaluation ranking value. As a result, in the sixth review process shown in Fig. 5F, the average evaluation rankings of the prediction models "D1", "Te·D1", "D2" are "1.25", "1.33", and "3.00" respectively. Thereby, the priority order of the prediction models within the first group after the sixth review process is the same as the priority order after the fifth review process, namely in the order of "D1", "Te·D1", "D2". Thus, by conducting six review processes, prediction models with high prediction accuracy that can be provided to customers can be, for example, in the order of "D1", "Te·D1", "D2" according to the priority order.

[0069] The prediction model belonging to the first group is considered to be a model with higher prediction accuracy when predicting the load of the customer compared to the prediction model belonging to the second group. Therefore, the second processing unit 122 and the fourth processing unit 124 perform weighting such that the smaller the prediction error (for example, the root mean squared error RMSE (Root Mean Squared Error)), or the smaller the average value of the evaluation ranks, or the higher the priority, the greater the weight for the unweighted predicted value predicted by the prediction model belonging to the first group, and calculate the weighted predicted value. Note that the predicted value of the prediction model belonging to the first group when the second processing unit 122 performs the review process shown in FIGS. 5A to 5D and the predicted value of the prediction model belonging to the first group when the fourth processing unit 124 performs the review process shown in FIGS. 5E to 5F are the unweighted predicted values, and the predicted value calculated as the customer load predicted value is the weighted predicted value obtained by performing this weighting on the unweighted predicted value of the prediction model belonging to the first group.

[0070] FIG. 6 is a diagram showing an algorithm for performing weighting on the unweighted predicted value of the prediction model belonging to the first group using a linear membership function. Note that in FIG. 6, t: Time when load prediction is performed L fore (t): Weighted predicted value weighted by any of prediction error, average value of evaluation ranks, and priority i: Any one of the prediction models belonging to the first group (i = 1, 2, 3) L main (t) i : Unweighted predicted value of the prediction model belonging to the first group Edist i : Any one of prediction error, average value of evaluation ranks, and priority of the prediction model belonging to the first group Edist max : Edist i : Maximum value of E(Edist i ): Grade N main : Number of prediction models belonging to the first group to be used It is as follows.

[0071] In FIG. 6, using a linear membership function, a grade corresponding to Edist i is determined, and the weighted predicted value L fore (t) is calculated from Equation (1). Incidentally, the grade in the case of calculating the predicted value L fore (t) using fuzzy inference is as shown in Equation (2). Also, Edist i is updated by any one of the prediction error, the average value of the evaluation ranks, and the priority of the prediction models belonging to the first group.

[0072] FIG. 7 is a diagram showing an example of a non-linear membership function for weighting the predicted value before weighting of the prediction models belonging to the first group. By using the non-linear membership function, for the prediction models belonging to the first group, the smaller the prediction error and the higher the prediction accuracy, or the smaller the average value of the evaluation ranks, or the higher the priority, the grade can be determined so that the weight becomes larger compared to the case of FIG. 6. Using this non-linear membership function, L fore (t), which is the predicted value after weighting, may be calculated.

[0073] Returning to FIG. 1, the third processing unit 123 creates a prediction model to be provided to the customer based on the three prediction models "D1", "Te·D1", and "D2" that have come to belong to the first group by the review processing shown in FIGS. 5A to 5F, and stores it in the second storage area 132 of the storage unit 130. For example, the third processing unit 123 may select the prediction models "D1", "Te·D1", and "D2" as creation models in the order of priority, or may create a model having the average value of the predicted values of any two or all three of the prediction models "D1", "Te·D1", and "D2" as the predicted value as the creation model.

[0074] The output unit 140 reads out the data indicating the prediction model created by the third processing unit 123 to be provided to the customer, which is stored in the second storage area 132 of the storage unit 130, and outputs it externally. Also, the output unit 140 reads out the customer load prediction value calculated by the second processing unit 122 and the fourth processing unit 124, which is stored in the second storage area 132 of the storage unit 130, and outputs it externally. The data indicating the prediction model and the customer load prediction value output from the output unit 140 are stored, for example, in a portable storage medium (not shown) that can be carried around or in the storage unit 151 of the external computer 150.

[0075] The external computer 150 has a storage unit 151, stores the data indicating the prediction model created by the third processing unit 123 in the storage unit 151, and reads out the data indicating the prediction model from the storage unit 151 as needed and outputs it to the prediction model creation device 1.

[0076] The characteristics of the customer's load may change due to changes in the customer's holidays or equipment. Therefore, the prediction model creation device 1 is configured such that in the control unit 120, the review processes of FIGS. 5A to 5F by the first processing unit 121 to the fourth processing unit 124 are repeated regularly (for example, every three months, six months, one year, etc.), and the prediction model most suitable for the characteristics of the customer's load can be automatically and efficiently provided.

[0077] FIG. 8 is a block diagram showing an example of the hardware of the information processing apparatus 200 used to realize the prediction model creation device 1.

[0078] The information processing apparatus 200 includes a processor 2010, a main memory device 2020, an auxiliary storage device 2030, an input device 2040, an output device 2050, and a communication device 2060. The information processing apparatus 200 is, for example, a personal computer, an office computer, various server devices, a general-purpose machine, or the like. The information processing apparatus 200 may be realized using virtual information processing resources provided using virtualization technology, such as a virtual server provided by a cloud system, for all or part of it. The prediction model creation apparatus 1 may be realized using a plurality of information processing apparatuses 200 connected communicably.

[0079] The processor 2010 is configured using, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), an AI (Artificial Intelligence) chip, or the like.

[0080] The main memory device 2020 is a device that stores programs and data, and is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), a non-volatile memory (NVRAM (Non Volatile RAM)), or the like.

[0081] The auxiliary storage device 2030 is, for example, an SSD (Solid State Drive), a hard disk drive, an optical storage device (such as a CD (Compact Disc), a DVD (Digital Versatile Disc), etc.), a storage system, an IC card, an SD card, a reading / writing device for a recording medium such as an optical recording medium, a storage area of a cloud server, etc. Programs and data can be read into the auxiliary storage device 2030 via a reading device for a recording medium or a communication device 2060. Programs and data stored in the auxiliary storage device 2030 are read into the main storage device 2020 at any time.

[0082] The input device 2040 is an interface that receives external inputs, and is, for example, a keyboard, a mouse, a touch panel, a card reader, a pen-input type tablet, a voice input device, etc.

[0083] The output device 2050 is an interface that outputs various information such as the progress of processing and the results of processing. The output device 2050 is, for example, a display device (such as an LCD (Liquid Crystal Display), a graphics card, etc.) that visualizes the above various information, a device (such as a voice output device (a speaker, etc.)) that converts the above various information into voice, a device (such as a printing device, etc.) that converts the above various information into characters. Note that the information processing device 200 may be configured to input and output information to and from other devices via the communication device 2060.

[0084] The input device 2040 and the output device 2050 constitute a user interface that receives information from and presents information to the user.

[0085] The communication device 2060 is a device that realizes communication (wired communication or wireless communication) with other devices via a communication infrastructure such as a communication network 400, and is configured using, for example, a NIC (Network Interface Card), a wireless communication module, a USB module, etc.

[0086] The information processing apparatus 200 may also introduce, for example, an operating system, a file system, a DBMS (Data Base Management System) (relational database, NoSQL, etc.), a KVS (Key-Value Store), and the like.

[0087] The functions of the prediction model creation apparatus 1 are realized by the processor 2010 of the information processing apparatus 200 executing a program (control program) read into the main storage device 2020, or are realized by the functions of the hardware (FPGA, ASIC, AI chip, etc.) that constitutes the prediction model creation apparatus 1 itself. For example, the storage unit 130 in FIG. 1 is realized by the main storage device 2020 and the auxiliary storage device 2030, the input unit 110 in FIG. 1 is realized by the input device 2040, the control unit 120 in FIG. 1 is realized by the processor 2010, and the output unit 140 in FIG. 1 is realized by the output device 2050 and the communication device 2060.

[0088] FIG. 9 is a flowchart showing an example of a process for creating a prediction model to be provided to a customer by the control unit 120 of the prediction model creation apparatus 1. In FIG. 4, the second eight prediction models with the initial priority of the prediction model "Te·D1" being the first will be taken as an example. Also, the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" with the initial priorities from the first to the eighth are, as shown in FIG. 5A, divided by the first processing unit 121 into a first group to which three prediction models "Te·D1", "Te·D2", "Te" with the initial priorities from the first to the third belong as main models, and a second group to which five prediction models "D1", "D2", "N", "Te·D3", "D3" with the initial priorities from the fourth to the eighth belong as support models.

[0089] First, in evaluation period 1 within the first review cycle shown in FIG. 5A, the second processing unit 122 calculates predicted values for all of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3". Also, the second processing unit 122 performs weighted grading determined by the linear membership function shown in FIG. 6 or the non-linear membership function shown in FIG. 7 on the predicted values of the three prediction models "Te·D1", "Te·D2", "Te" ranked from first to third in the first group, calculates the weighted predicted values, and uses them as the customer load predicted values. Further, the second processing unit 122 acquires the actual values at the prediction times of the predicted values of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" via the input unit 110. Then, the second processing unit 122 calculates the difference between the predicted value before weighting and the actual value for each of the three prediction models "Te·D1", "Te·D2", "Te" ranked from first to third, and calculates the difference between the predicted value and the actual value for each of the five prediction models "D1", "D2", "N", "Te·D3", "D3" ranked from fourth to eighth (S2000). The process of step S2000 is for one day's processing.

[0090] If the process of step S2000 is continued for a days, which is the evaluation period (S2010: YES), the process returns to step S2000 and repeats. On the other hand, if the process of step S2000 is not continued for a days, which is the evaluation period (S2010: NO), the series of processes ends.

[0091] When the process of step S2000 for a days, which is the evaluation period, is completed, the second processing unit 122 calculates the evaluation rankings in evaluation period 1 of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" based on the differences between the predicted values and the actual values (S2020). The evaluation rankings in evaluation period 1 of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", "D3" are first, third, sixth, second, fourth, fifth, seventh, and eighth.

[0092] For the evaluation cycles 2 and 3 within the first review cycle, the second processing unit 122 repeats the processes of steps S2000, S2010, and S2020 in the same manner as in evaluation cycle 1. The evaluation rankings of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", and "D3" in evaluation cycle 2 are the 2nd, 4th, 6th, 1st, 3rd, 5th, 8th, and 7th respectively. Also, the evaluation rankings of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", and "D3" in evaluation cycle 3 are the 2nd, 4th, 5th, 1st, 3rd, 6th, 8th, and 7th respectively.

[0093] After the second processing unit 122 calculates the evaluation rankings of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", and "D3" in evaluation cycles 1 to 3, next, it calculates the average value of the evaluation rankings for evaluation cycles 1 to 3 (S2030). For the prediction model "Te·D1" whose initial priority ranking was 1st, the value obtained by multiplying the average value of the evaluation rankings by an additional priority rate (p%: for example, p = 0.75) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the eight prediction models "Te·D1", "Te·D2", "Te", "D1", "D2", "N", "Te·D3", and "D3" in evaluation cycles 1 to 3 are "1.25", "3.67", "5.67", "1.33", "3.33", "5.33", "7.67", and "7.33".

[0094] Next, the second processing unit 122 rearranges the rankings of the eight prediction models in ascending order of the average value of the evaluation rankings in evaluation cycles 1 to 3 (S2040). At this time, the priority rankings of the eight prediction models after rearrangement are in the order of "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3".

[0095] Next, in the second processing unit 122, depending on any one of the prediction errors, average values of the evaluation rankings, and priority rankings of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", and "Te·D3", Edist iUpdate it (S2050).

[0096] Next, as a result of the rearrangement of the eight prediction models in step S2040, the second processing unit 122 determines whether there has been an exchange of any prediction models between the first group and the second group (S2060).

[0097] In the first review cycle, since the prediction models "Te·D2" and "Te" moved from the first group to the second group, and the prediction models "D1" and "D2" moved from the second group to the first group, the second processing unit 122 determines that there has been an exchange between the prediction models "Te·D2", "Te" and the prediction models "D1", "D2" between the first group and the second group (S2060: YES).

[0098] In the second processing unit 122, for each review cycle of the ranking of the prediction models, after the exchange of the prediction models between the first group and the second group, the number of times (m) that the exchange of the prediction models has not been continuously performed is incremented by 1. In the determination of step S2060, since there has been an exchange of the prediction models between the first group and the second group, the second processing unit 122 resets m and does not count with m = 0 remaining (S2070), ends the first review process for the ranking of the eight prediction models, and proceeds to the second review process.

[0099] In the evaluation period 4 within the second review cycle shown in FIG. 5B, the second processing unit 122 calculates predicted values for all of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3". Further, the second processing unit 122 performs weighted grading determined by the linear membership function shown in FIG. 6 or the non-linear membership function shown in FIG. 7 on the predicted values of the three prediction models "Te·D1", "D1", "D2" ranked first to third in the first group, calculates the weighted predicted values, and uses them as the customer load predicted values. Also, the second processing unit 122 acquires the actual values at the prediction times of the predicted values of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" via the input unit 110. Then, the second processing unit 122 calculates the difference between the predicted value before weighting and the actual value for each of the three prediction models "Te·D1", "D1", "D2" ranked first to third, and calculates the difference between the predicted value and the actual value for each of the five prediction models "Te·D2", "N", "Te", "D3", "Te·D3" ranked fourth to eighth (S2000).

[0100] If the process of step S2000 is continued for a days, which is the evaluation period (S2010: YES), the process returns to step S2000 and repeats. On the other hand, if the process of step S2000 is not continued for a days, which is the evaluation period (S2010: NO), the series of processes ends.

[0101] When the process of step S2000 in the evaluation period of a days is completed, the second processing unit 122 calculates the evaluation ranks in the evaluation period 4 of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" based on the difference between the predicted value and the actual value (S2020). The evaluation ranks in the evaluation period 4 of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" are 2nd, 1st, 3rd, 5th, 4th, 6th, 8th, 7th.

[0102] For the evaluation cycles 5 and 6 within the second review cycle, the second processing unit 122 repeats the processes of steps S2000, S2010, and S2020 in the same manner as in evaluation cycle 4. The evaluation rankings of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" in evaluation cycle 5 are the 2nd, 1st, 3rd, 5th, 6th, 4th, 7th, and 8th. Also, the evaluation rankings of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" in evaluation cycle 6 are the 2nd, 1st, 4th, 3rd, 6th, 5th, 7th, and 8th.

[0103] After the second processing unit 122 calculates the evaluation rankings of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" in evaluation cycles 4 to 6, next, it calculates the average value of the evaluation rankings for evaluation cycles 4 to 6 (S2030). For the prediction model "Te·D1" which had the first priority, the value obtained by multiplying the average value of the evaluation rankings by the priority rate (p%) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the eight prediction models "Te·D1", "D1", "D2", "Te·D2", "N", "Te", "D3", "Te·D3" in evaluation cycles 4 to 6 are "1.50", "1.00", "3.33", "4.33", "5.33", "5.00", "7.33", "7.67".

[0104] Next, the second processing unit 122 rearranges the rankings of the eight prediction models in ascending order of the average value of the evaluation rankings in evaluation cycles 4 to 6 (S2040). At this time, the rankings of the eight prediction models after rearrangement are in the order of "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3".

[0105] Next, in the second processing unit 122, Edist i is updated based on any one of the prediction errors, average values of the evaluation rankings, and priority rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" (S2050).

[0106] Next, as a result of rearranging the eight prediction models in step S2040, the second processing unit 122 determines whether there has been an exchange of any prediction models between the first group and the second group (S2060).

[0107] In the second review cycle, since there has been no exchange of prediction models between the first group and the second group, the second processing unit 122 determines that there is no exchange of prediction models between the first group and the second group (S2060: YES).

[0108] The second processing unit 122 increments m by 1 and sets m = 1 (S2080).

[0109] The second processing unit 122 determines whether the count value of m has reached M (for example, 3) (S2090). At this time, since m = 1 (S2090: NO), the second review process for the ranking of the eight prediction models is terminated, and the process proceeds to the third review process.

[0110] In the evaluation period 7 within the third review cycle shown in FIG. 5C, the second processing unit 122 calculates predicted values for all of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3". Also, the second processing unit 122 performs weighted grading determined by the linear membership function shown in FIG. 6 or the non-linear membership function shown in FIG. 7 on the predicted values of the three prediction models "D1", "Te·D1", "D2" ranked first to third in the first group, calculates the weighted predicted values, and uses them as the customer load predicted values. Further, the second processing unit 122 acquires the actual values at the prediction times of the predicted values of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" via the input unit 110. Then, the second processing unit 122 calculates the difference between the predicted value before weighting and the actual value for each of the three prediction models "D1", "Te·D1", "D2" ranked first to third, and calculates the difference between the predicted value and the actual value for each of the five prediction models "Te·D2", "Te", "N", "D3", "Te·D3" ranked fourth to eighth (S2000).

[0111] If the process of step S2000 is continued for a days which is the evaluation period (S2010: YES), the process returns to step S2000 and repeats. On the other hand, if the process of step S2000 is not continued for a days which is the evaluation period (S2010: NO), the series of processes ends.

[0112] When the process of step S2000 in the evaluation period of a days is completed, the second processing unit 122 calculates the evaluation rankings in the evaluation period 7 of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" based on the differences between the predicted values and the actual values (S2020). The evaluation rankings in the evaluation period 7 of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", "Te·D3" are 2nd, 1st, 3rd, 5th, 6th, 4th, 7th, 8th.

[0113] For the evaluation cycles 8 and 9 within the third review cycle, the second processing unit 122 repeats the processes of steps S2000, S2010, and S2020 in the same manner as in evaluation cycle 7. The evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", and "Te·D3" in evaluation cycle 8 are the 1st, 2nd, 3rd, 5th, 6th, 4th, 8th, and 7th places. Also, the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", and "Te·D3" in evaluation cycle 9 are the 2nd, 1st, 4th, 3rd, 5th, 6th, 8th, and 7th places.

[0114] After calculating the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", and "Te·D3" in evaluation cycles 7 to 9, the second processing unit 122 then calculates the average value of the evaluation rankings for evaluation cycles 7 to 9 (S2030). For the prediction model "D1" which had the first priority, the value obtained by multiplying the average value of the evaluation rankings by the priority rate (p%) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "Te", "N", "D3", and "Te·D3" in evaluation cycles 7 to 9 are "1.25", "1.33", "3.33", "4.33", "5.67", "4.67", "7.67", and "7.33".

[0115] Next, the second processing unit 122 rearranges the rankings of the eight prediction models in ascending order of the average value of the evaluation rankings in evaluation cycles 7 to 9 (S2040). At this time, the rankings of the eight prediction models after rearrangement are in the order of "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3".

[0116] Next, in the second processing unit 122, Edist i is updated based on any one of the prediction errors, average values of the evaluation rankings, and priority rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3" (S2050).

[0117] Next, as a result of rearranging the eight prediction models in step S2040, the second processing unit 122 determines whether there has been an exchange of any prediction models between the first group and the second group (S2060).

[0118] In the third review cycle, since there was no exchange of prediction models between the first group and the second group, the second processing unit 122 determines that there is no exchange of prediction models between the first group and the second group (S2060: YES).

[0119] The second processing unit 122 increments m by 1 to make m = 2 (S2080).

[0120] The second processing unit 122 determines whether the count value of m has reached M (for example, 3) (S2090). At this time, since m = 2 (S2090: NO), the third review process for the ranking of the eight prediction models is terminated, and the process proceeds to the fourth review process.

[0121] In evaluation period 10 within the fourth review cycle shown in FIG. 5D, the second processing unit 122 calculates predicted values for all of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3". Also, the second processing unit 122 performs weighted grading on the predicted values of the three prediction models "D1", "Te·D1", "D2" ranked first to third belonging to the first group using the linear membership function shown in FIG. 6 or the non-linear membership function shown in FIG. 7, calculates the weighted predicted values, and sets them as the consumer load predicted values. Further, the second processing unit 122 acquires the actual values at the prediction times of the predicted values of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" via the input unit 110. Then, the second processing unit 122 calculates the difference between the predicted value before weighting and the actual value for each of the three prediction models "D1", "Te·D1", "D2" ranked first to third, and calculates the difference between the predicted value and the actual value for each of the five prediction models "Te·D2", "N", "Te", "Te·D3", "D3" ranked fourth to eighth (S2000).

[0122] If the process of step S2000 is continued for a days which is the evaluation period (S2010: YES), the process returns to step S2000 and repeats. On the other hand, if the process of step S2000 is not continued for a days which is the evaluation period (S2010: NO), the series of processes ends.

[0123] When the process of step S2000 in the evaluation period of a days is completed, the second processing unit 122 calculates the evaluation rankings in evaluation period 10 of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" based on the difference between the predicted value and the actual value (S2020). The evaluation rankings in evaluation period 10 of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", "D3" are first, second, fourth, fifth, third, sixth, seventh, and eighth.

[0124] For the evaluation cycles 11 and 12 within the fourth review cycle, the second processing unit 122 repeats the processes of steps S2000, S2010, and S2020 in the same manner as the evaluation cycle 10. The evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3" in the evaluation cycle 11 are the 2nd, 1st, 3rd, 4th, 6th, 5th, 8th, and 7th. Also, the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3" in the evaluation cycle 12 are the 1st, 2nd, 3rd, 5th, 4th, 6th, 8th, and 7th.

[0125] After the second processing unit 122 calculates the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3" in the evaluation cycles 10 to 12, next, it calculates the average value of the evaluation rankings in the evaluation cycles 10 to 12 (S2030). Note that for the prediction model "D1" which had the first priority, the value obtained by multiplying the average value of the evaluation rankings by the priority rate (p%) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the eight prediction models "D1", "Te·D1", "D2", "Te·D2", "N", "Te", "Te·D3", and "D3" in the evaluation cycles 10 to 12 are "1.00", "1.67", "3.33", "4.67", "4.33", "5.67", "7.67", and "7.33".

[0126] Next, the second processing unit 122 rearranges the rankings of the eight prediction models in ascending order of the average values of the evaluation rankings in the evaluation cycles 10 to 12 (S2040). At this time, the rankings of the eight prediction models after rearrangement are in the order of "D1", "Te·D1", "D2", "N", "Te·D2", "Te", "D3", "Te·D3".

[0127] Next, in the second processing unit 122, Edist i is updated based on any one of the prediction errors, average values of the evaluation rankings, and priority rankings of the eight prediction models "D1", "Te·D1", "D2", "N", "Te·D2", "Te", "D3", and "Te·D3" (S2050).

[0128] Next, as a result of the rearrangement of the eight prediction models in step S2040, the second processing unit 122 determines whether there has been an exchange of any prediction models between the first group and the second group (S2060).

[0129] In the fourth review cycle, since there was no exchange of prediction models between the first group and the second group, the second processing unit 122 determines that there is no exchange of prediction models between the first group and the second group (S2060: YES).

[0130] The second processing unit 122 determines whether the count value of m has reached M (for example, 3) (S2090). Since m = 3 (S2090: YES), the second processing unit 122 stops the re - evaluation process of the priority order for the five prediction models "N", "Te·D2", "Te", "D3", "Te·D3" belonging to the second group at present (S2100), ends the fourth review process, and proceeds to the fifth review process only for the three prediction models "D1", "Te·D1", "D2" belonging to the first group at present. Note that the fourth processing unit 124 performs the fifth review process.

[0131] In the evaluation cycle 13 in the fifth review cycle shown in FIG. 5E, the fourth processing unit 124 calculates prediction values for the three prediction models "D1", "Te·D1", "D2". Further, the fourth processing unit 124 performs weighted grading on the prediction values of the three prediction models "D1", "Te·D1", "D2" using the linear membership function shown in FIG. 6 or the non - linear membership function shown in FIG. 7, calculates the weighted prediction values, and uses them as the customer load prediction values. Also, the fourth processing unit 124 acquires the actual values at the prediction times of the prediction values of the three prediction models "D1", "Te·D1", "D2" via the input unit 110. Then, the fourth processing unit 124 calculates the difference between the prediction value before weighting and the actual value for each of the three prediction models "D1", "Te·D1", "D2" (S2110).

[0132] When the process of step S2110 is continued for a days which is the evaluation period (S2120: YES), the process returns to step S2110 and repeats. On the other hand, when the process of step S2110 is not continued for a days which is the evaluation period (S2120: NO), a series of processes ends.

[0133] When the process of step S2110 in the evaluation period of a days is completed, the fourth processing unit 124 calculates the evaluation rankings in the evaluation period 13 of the three prediction models "D1", "Te·D1", and "D2" based on the difference between the predicted value before weighting and the actual value (S2130). The evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 13 are the 1st, 3rd, and 2nd.

[0134] The fourth processing unit 124 also repeats the processes of steps S2110, S2120, and S2130 for the evaluation periods 14 and 15 in the fifth review period in the same manner as the evaluation period 13. The evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 14 are the 2nd, 1st, and 3rd. Also, the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 15 are the 2nd, 1st, and 3rd.

[0135] When the fourth processing unit 124 calculates the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation periods 13 to 15, next, it calculates the average value of the evaluation rankings in the evaluation periods 13 to 15 (S2140). For the prediction model "D1" whose priority rank was the 1st, the value obtained by multiplying the average value of the evaluation rankings by the priority rate (p%) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation periods 13 to 15 are "1.25", "1.67", and "2.67".

[0136] Next, the fourth processing unit 124 rearranges the ranks of the three prediction models "D1", "Te·D1", and "D2" in ascending order of the average rank value in the evaluation periods 13 to 15 (S2150). The ranks of the three prediction models "D1", "Te·D1", and "D2" remain unchanged as they were when the fourth review process was performed, and they are still in the order of "D1", "Te·D1", "D2".

[0137] Next, in the fourth processing unit 124, Edist i is updated based on any one of the prediction errors, average evaluation rank values, and priorities of the three prediction models "D1", "Te·D1", and "D2" (S2160).

[0138] Next, in step S2100, the fourth processing unit 124 determines whether a certain period (for example, 3 months, 6 months, 1 year, etc.) has elapsed since the review process of the priorities for the five prediction models "N", "Te·D2", "Te", "D3", and "Te·D3" belonging to the second group was stopped (S2170). If the certain period has not elapsed (S2170: NO), the fourth processing unit 124 ends the fifth review process and proceeds to the sixth review process.

[0139] In the evaluation period 16 within the sixth review cycle shown in FIG. 5F, the fourth processing unit 124 calculates predicted values for the three prediction models "D1", "Te·D1", and "D2". Further, the fourth processing unit 124 performs weighted grading on the predicted values of the three prediction models "D1", "Te·D1", and "D2" using the linear membership function shown in FIG. 6 or the non-linear membership function shown in FIG. 7, calculates the weighted predicted values, and uses them as the customer load predicted values. Also, the fourth processing unit 124 acquires the actual values at the prediction times of the predicted values of the three prediction models "D1", "Te·D1", and "D2" via the input unit 110. Then, the fourth processing unit 124 calculates the difference between the predicted value before weighting and the actual value for each of the three prediction models "D1", "Te·D1", and "D2" (S2110).

[0140] When the process of step S2110 is continued for a days which is the evaluation period (S2120: YES), the process returns to step S2110 and repeats. On the other hand, when the process of step S2110 is not continued for a days which is the evaluation period (S2120: NO), the series of processes ends.

[0141] When the process of step S2110 in the evaluation period of a days is completed, the fourth processing unit 124 calculates the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 16 based on the difference between the predicted value before weighting and the actual value (S2130). The evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 16 are the second, first, and third.

[0142] For the evaluation periods 17 and 18 in the sixth review period, the fourth processing unit 124 also repeats the processes of steps S2110, S2120, and S2130 in the same manner as in the evaluation period 16. The evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 17 are the first, second, and third. Also, the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation period 18 are the second, first, and third.

[0143] After the fourth processing unit 124 calculates the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation periods 16 to 18, next, it calculates the average value of the evaluation rankings of the evaluation rankings in the evaluation periods 16 to 18 (S2140). For the prediction model "D1" whose priority rank was the first, the value obtained by multiplying the average value of the evaluation rankings by the priority rate (p%) is used as the average value of the evaluation rankings. As a result, the average values of the evaluation rankings of the three prediction models "D1", "Te·D1", and "D2" in the evaluation periods 16 to 18 are "1.25", "1.33", and "3.00".

[0144] Next, the fourth processing unit 124 rearranges the ranks of the three prediction models "D1", "Te·D1", and "D2" in ascending order of the average rank value in the evaluation periods 16 to 18 (S2150). The ranks of the three prediction models "D1", "Te·D1", and "D2" remain unchanged as they were when the fifth review process was performed, and they remain in the order of "D1", "Te·D1", and "D2".

[0145] Next, in the fourth processing unit 124, depending on any one of the prediction errors, average evaluation rank values, and priorities of the three prediction models "D1", "Te·D1", and "D2", Edist i is updated (S2160).

[0146] Next, the fourth processing unit 124 determines whether a certain period has elapsed since the processing in step S2100 (S2170). If the certain period has not elapsed (S2170: NO), the fourth processing unit 124 ends the sixth review process and proceeds to the seventh and subsequent review processes (not shown). On the other hand, if the certain period has elapsed (S2170: YES), the processing from step S2000 and subsequent steps will be performed again by the second processing unit 122 instead of the fourth processing unit 124.

[0147] In this way, the third processing unit 123 can create a prediction model considering the latest characteristics of the load on the customer.

[0148] As described above, the prediction model creation device 1 includes a first processing unit 121 that divides eight prediction models for predicting the load of a customer into a first group with a higher priority and a second group with a lower priority, and based on the difference between the predicted value and the actual value of the load of the customer by the eight prediction models, when the priorities after reevaluating the priorities of the eight prediction models are swapped between the first group and the second group, until the priorities of at least the eight prediction models no longer swap between the first group and the second group, a second processing unit 122 that reevaluates the priorities of the eight prediction models, and a third processing unit 123 that creates a prediction model for the customer based on three prediction models belonging to the first group.

[0149] According to the prediction model creation device 1, a prediction model for the consumer is created based on three prediction models in which the priorities of eight prediction models for predicting the consumer's load continuously belong to the first group. Therefore, it is possible to efficiently provide a prediction model with high prediction accuracy to the consumer.

[0150] Also, in the prediction model creation device 1, the first group is a group to which three prediction models belong. After the processing of the second processing unit 122 is performed, based on the difference between the predicted value and the actual value of the consumer's load by the three prediction models belonging to the first group, a fourth processing unit 124 that rechecks the priorities of the three prediction models belonging to the first group is included.

[0151] According to the prediction model creation device 1, since the priorities of the three prediction models that come to belong to the first group are further rechecked, it is possible to further improve the prediction accuracy of the prediction model provided to the consumer.

[0152] Also, in the prediction model creation device 1, the second processing unit 122 rechecks the priorities of the eight prediction models until the number of times the priorities of the eight prediction models do not continuously change between the first group and the second group reaches M times.

[0153] According to the prediction model creation device 1, since the priorities of the prediction models belonging to the first group are stabilized, it is possible to further improve the prediction accuracy of the prediction model provided to the consumer.

[0154] Also, in the processing of the second processing unit 122 and the fourth processing unit 124 in the prediction model creation device 1, the predicted value of the consumer's load by the prediction models belonging to the first group is weighted according to the prediction accuracy of the prediction model, the average value of the evaluation ranks, or the priorities.

[0155] According to the prediction model creation device 1, since weighting is performed on the predicted values of the prediction models belonging to the first group, it is possible to further improve the prediction accuracy of the predicted values of the load of the customer by the prediction model provided to the customer.

[0156] Also, in the prediction model creation device 1, the predicted value of the load of the customer by the prediction model belonging to the first group can be weighted using a linear or non-linear membership function.

[0157] Note that the above embodiments are for facilitating the understanding of the present invention and are not for limiting and interpreting the present invention. The present invention can be changed and improved without departing from its gist, and equivalents thereof are also included in the present invention.

Explanation of Reference Numerals

[0158] 1 Prediction model creation device 110 Input unit 120 Control unit 121 First processing unit 122 Second processing unit 123 Third processing unit 124 Fourth processing unit 130, 151 Storage unit 131 First storage area 132 Second storage area 140 Output unit 150 External computer 160 Communication network

Claims

1. A first processing unit that divides a plurality of prediction models for predicting the load of a customer into a first group with a higher priority and a second group with a lower priority; Based on the difference between the predicted value and the actual value of the customer's load by the plurality of prediction models, when the priority order of the plurality of prediction models after re-evaluating the priority order is swapped between the first group and the second group, at least until the priority order of the plurality of prediction models no longer swaps between the first group and the second group, a second processing unit that re-evaluates the priority order of the plurality of prediction models; A third processing unit that creates a prediction model for a customer based on the prediction models belonging to the first group; A prediction model creation device including the above.

2. The prediction model creation device according to Claim 1, wherein the first group is a group to which a plurality of the prediction models belong, after the processing of the second processing unit is performed, based on the difference between the predicted value and the actual value of the customer's load by the plurality of prediction models belonging to the first group, a fourth processing unit that re-evaluates the priority order of the plurality of prediction models belonging to the first group A prediction model creation device including the above.

3. The prediction model creation device according to Claim 1, wherein the second processing unit re-evaluates the priority order of the plurality of prediction models until the number of times the priority order of the plurality of prediction models does not continuously swap between the first group and the second group reaches a predetermined number of times A prediction model creation device.

4. The prediction model creation device according to Claim 2, in the processing of the second processing unit and the fourth processing unit, the predicted value of the customer's load by the prediction models belonging to the first group is weighted according to the prediction accuracy or the average value of the evaluation ranks or the priority order of the prediction models A prediction model creation device.

5. The prediction model creation device according to Claim 4, wherein the predicted value of the customer's load by the prediction models belonging to the first group is weighted using a linear or non-linear membership function A prediction model creation device.

6. A first step of dividing a plurality of prediction models for predicting the load of a customer into a first group with a higher priority and a second group with a lower priority; After re - evaluating the priorities of the plurality of prediction models based on the difference between the predicted value and the actual value of the customer's load by the plurality of prediction models, if the priorities between the first group and the second group are swapped, a second step of re - evaluating the priorities of the plurality of prediction models until at least the priorities of the plurality of prediction models no longer swap between the first group and the second group; A third step of creating a prediction model for a customer based on the prediction models belonging to the first group; A prediction model creation method including the above.

7. The prediction model creation method according to claim 6, wherein the first group is a group to which the plurality of prediction models belong; After the process of the second step is performed, a fourth step of re - evaluating the priorities of the plurality of prediction models belonging to the first group based on the difference between the predicted value and the actual value of the customer's load by the plurality of prediction models belonging to the first group A prediction model creation method including the above.

8. The prediction model creation method according to claim 6, wherein in the second step, the priorities of the plurality of prediction models are re - evaluated until the number of times the priorities of the plurality of prediction models do not continuously swap between the first group and the second group reaches a predetermined number of times A prediction model creation method.

9. The prediction model creation method according to claim 7, wherein in the processes of the second step and the fourth step, the predicted value of the customer's load by the prediction models belonging to the first group is weighted according to the prediction accuracy or the average value of the evaluation ranks or the priorities of the prediction models A prediction model creation method.

10. The prediction model creation method according to claim 9, wherein the predicted value of the customer's load by the prediction models belonging to the first group is weighted using a linear or non - linear membership function A prediction model creation method.

Images