JP7413843B2 - Information processing device, information processing method and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

Patent Document 1 describes predicting the amount of power demand on a prediction target day.

JP 2015-33203 Publication

When a predicted value of a prediction target (for example, electric power demand) is obtained based on a prediction model, an error occurs between the predicted value and the actual value of the prediction target. In addition, when the predicted value of a prediction target that changes over time is obtained based on a prediction model, the error between the predicted value and the actual value increases as time passes after the prediction model is generated, and the predicted value Accuracy may decrease.

An object of the present invention is to improve the prediction accuracy of a prediction target that changes over time.

A first aspect of the means for solving the above-mentioned problem includes a first acquisition unit that acquires first data indicating a predetermined factor that changes over time, and a first acquisition unit that acquires first data indicating a predetermined factor that changes over time, and a first acquisition unit that acquires first data that indicates a predetermined factor that changes over time, and a first acquisition unit that acquires first data that indicates a predetermined factor that changes over time, and a first acquisition unit that acquires first data that indicates a predetermined factor that changes over time. a second acquisition unit that acquires second data indicating an error between the first predicted value calculated by the method and the actual value of the prediction target; and a classification that determines the class according to the factor from among a plurality of classes. the classification model in which third data for correcting the first predicted value and calculating the second predicted value is associated with each of the classes; An information processing apparatus includes a classification model generation unit that generates a classification model based on.

Further, a second aspect of the means for solving the above-mentioned problem is a first acquisition process in which a computer acquires first data of a predetermined factor that changes over time, and a first acquisition process in which a computer acquires first data of a predetermined factor that changes over time. a second acquisition process for acquiring second data indicating an error between a first predicted value calculated based on a prediction model and the actual value of the prediction target; and a second acquisition process that selects the class from among a plurality of classes according to the factor. a classification model that determines a classification model in which third data for correcting the first predicted value and calculating a second predicted value is associated with each of the classes, and the first data and the third data for calculating the second predicted value. and a classification model generation process generated based on the second data.

Further, a third aspect of the means for solving the above-mentioned problem is to provide a computer with a first acquisition process of acquiring first data of a predetermined factor that changes over time, and a prediction target that changes over time. a second acquisition process for acquiring second data indicating an error between a first predicted value calculated based on a prediction model and the actual value of the prediction target; and a second acquisition process that selects the class from among a plurality of classes according to the factor. a classification model that determines a classification model in which third data for correcting the first predicted value and calculating a second predicted value is associated with each of the classes, and the first data and the third data for calculating the second predicted value. This program executes a classification model generation process that is generated based on the second data.

According to the present invention, it is possible to improve the prediction accuracy of a prediction target that changes over time.

FIG. 1 is a block diagram showing the hardware of the information processing device 1 of this embodiment. FIG. 2 is a block diagram showing the functions of the information processing device 1 of this embodiment. FIG. 3 is an explanatory diagram of data of actual values to be predicted. FIG. 4A is an explanatory diagram of weather data. FIG. 4B is an explanatory diagram of another weather data. FIG. 5A is an explanatory diagram of day of the week data. FIG. 5B is an explanatory diagram of another day of the week data. FIG. 6 is an explanatory diagram of the prediction model. FIG. 7 is an explanatory diagram of an example of a prediction model. FIG. 8 is a flow diagram of the classification model generation process. 9A to 9D are explanatory diagrams of the second data acquired by the second acquisition unit 212. FIG. 10 is an explanatory diagram of data used to generate a decision tree. FIG. 11 is an explanatory diagram of a decision tree for a time period from 0:00 to 1:00 on a certain prediction target day. FIG. 12 is an explanatory diagram of the third data generation method. FIG. 13 is an explanatory diagram of the decision tree for each time period. FIG. 14 is a flow diagram of the prediction method of this embodiment. FIG. 15 is an explanatory diagram of the prediction results of this embodiment. FIG. 16 is a graph showing the prediction results of this embodiment.

From the description of this specification and the attached drawings, at least the following matters will become clear. DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below based on one embodiment thereof with reference to the accompanying drawings.

=====This embodiment =====
<Configuration>
FIG. 1 is a block diagram showing the hardware of the information processing device 1 of this embodiment.

The information processing device 1 of this embodiment is a prediction device for obtaining a predicted value of a prediction target based on a prediction model. In this embodiment, the prediction target is the power demand amount, and the information processing device 1 calculates a predicted value of the power demand amount on the prediction target day, which is a future day at the time of prediction. However, the prediction target is not limited to the power demand amount, and may be, for example, the power market price. Further, the prediction target is not limited to those related to electricity, and may be, for example, sales of products (number of sales, sales amount, etc.). The information processing device 1 is configured by, for example, a computer such as a personal computer or a server. The information processing device 1 may be composed of one computer or may be composed of multiple computers.

The information processing device 1 includes a CPU 11, a storage device 12, and a communication device 13. The CPU 11 is an arithmetic processing unit that controls the prediction device. The storage device 12 is a main storage device such as a RAM, or an auxiliary storage device such as a hard disk drive or SSD (Solid State Drive). The storage device 12 stores programs, data, and the like. When the CPU 11 reads and executes a program stored in the storage device 12, various processes described below are executed. Note that the storage device 12 includes a non-transitory storage medium that records programs that cause the information processing device 1 (computer) to execute various processes to be described later. The communication device 13 is a communication module for connecting to the communication network 3. Note that the communication network 3 is, for example, a telephone line network (public telephone line network or mobile phone line network), a wireless communication network, the Internet, a LAN, etc., and the Internet is assumed here. The information processing device 1 may include a display device 14 (for example, a display) and an input device 15 (for example, a keyboard, a mouse).

FIG. 2 is a block diagram showing the functions of the information processing device 1 of this embodiment. The information processing device 1 includes a control section 20 and a data storage section 30.

The control unit 20 controls various processes that the information processing device 1 should perform. The control unit 20 is realized by the CPU 11 executing a control program stored in the storage device 12 to perform various controls. The control unit 20 has the functions of a data acquisition unit 21 , a predictive model generation unit 22 , a first prediction unit 23 , a classification model generation unit 24 , and a second prediction unit 25 . The data acquisition unit 21 has a function of acquiring various data. Note that the data acquisition section 21 includes an input data acquisition section 210, a first acquisition section 211, and a second acquisition section 212. The predictive model generation unit 22 has a function of generating a predictive model (described later). The first prediction unit 23 has a function of obtaining a first predicted value using a prediction model. The classification model generation unit 24 has a function of generating a classification model (described later) by analyzing the tendency of the error between the first predicted value and the actual value. The second prediction unit 25 has a function as a correction unit that corrects the first predicted value based on the correction value obtained using the classification model to obtain a second predicted value. Each of these functions will be described later.

The data storage section 30 is a storage section for storing predetermined data. The data storage unit 30 is realized by a part of the storage area of the storage device 12. Here, the data storage unit 30 stores various data such as actual values of prediction targets, factor data, predicted values, errors, and correction values.

FIG. 3 is an explanatory diagram of data of actual values to be predicted. In this embodiment, since the prediction target is the power demand amount, FIG. 3 shows the data of the actual value of the power demand amount. Note that the actual value indicates an actual value in the past. The prediction target of this embodiment changes over time, and the data of the actual value of the prediction target is data that associates a predetermined period with the actual value of the prediction target for the period. Here, the predetermined period is a one-hour time period obtained by dividing a day into 24, but it is not limited to this. For example, it may be a two-hour period (time period) or one day. It may be a period of The data acquisition unit 21 acquires the actual value of the prediction target for each predetermined period, and stores in the data storage unit 30 data that associates the period with the actual value. In the case of the present embodiment, the data acquisition unit 21 acquires the actual value of the power demand amount for each hour, associates the time period with the actual value of the power demand amount, and stores it in the data storage unit 30. By accumulating the data, the prediction target performance data shown in FIG. 3 is generated.

The factor data is data on factors that influence the prediction target. Data on factors that influence the prediction target also changes over time, and the factor data is data that associates a predetermined period with the actual value or predicted value of each factor for the period. Here, the factor data includes weather data and day of the week data. However, the factor data may include data other than the weather and the day of the week, or may include only one of the weather data and the day of the week data. The factor data may be actual values of factors in the past or predicted values of factors in the future.

FIG. 4A is an explanatory diagram of weather data. The weather data of this embodiment is data in which a predetermined period is associated with the temperature and humidity of the period. Here, the predetermined period is a one-hour time period obtained by dividing one day into 24, but is not limited to this. Furthermore, meteorological factors other than temperature and humidity (for example, discomfort index, sunshine hours, weather, etc.) may be associated. Furthermore, as shown in FIG. 4B, maximum temperature, minimum temperature, maximum humidity, and minimum humidity may be associated with one day's period. The data acquisition unit 21 acquires weather data for each predetermined period, stores the weather data in the data storage unit 30, and generates the weather data shown in FIG. 4A. Note that the data acquisition unit 21 may generate other weather data shown in FIG. 4B based on the weather data in FIG. 4A.

FIG. 5A is an explanatory diagram of day of the week data. Day of the week data is data in which a date is associated with a flag indicating an attribute of the date. Here, the value indicated by the flag distinguishes between Saturday, Sunday, Monday (the day after a holiday), and weekdays other than Monday. However, the flag is not limited to this; for example, as shown in FIG. It can be any type of flag. Additionally, there may be a flag indicating a holiday.

Note that other data (predicted values, errors, correction values, etc.) stored in the data storage unit 30 will become clear from the description below.

<Reference: About prediction models>
FIG. 6 is an explanatory diagram of the prediction model. A prediction model is a model for obtaining a prediction result (output data) according to input data, and has a function-like function. The first prediction unit 23 uses the prediction model to calculate a first predicted value (output data) from the factor data (input data). In this embodiment, the factor data (input data) input to the prediction model is, for example, weather data or day of the week data. Furthermore, in this embodiment, the predicted value (first predicted value) output from the prediction model is a predicted value of the amount of power demand at the request point (predetermined time on the prediction target day).

FIG. 7 is an explanatory diagram of an example of a prediction model. The prediction model in the figure is realized by the CPU 11 executing a control program stored in the storage device 12. The control unit 20 of this embodiment includes an input data acquisition unit 210, a first prediction unit 23, etc. for performing each process shown in the figure (see FIG. 2).

First, the input data acquisition unit 210 acquires factor data that becomes input data for a prediction model. Here, the input data acquisition unit 210 acquires predicted values of weather data at the request point, actual values of weather data immediately before the request point, day of the week data, etc. from the factor data stored in the data storage unit 30. do. Note that the input data acquisition unit 210 may acquire factor data (for example, predicted values of weather data) from an external database. The input data acquisition unit 210 passes the acquired factor data to the first prediction unit 23 as input data. Here, the input data is assumed to be a vector x.

The first prediction unit 23 performs various calculations on the input data x and calculates a distance calculation vector a. Here, the input data x is multiplied by the weight matrix W to calculate the distance calculation vector a (a=W×x). Note that the weight matrix W is generated in advance by machine learning by the predictive model generation unit 22.

Next, the first prediction unit 23 extracts similar days similar to the requested point from among the past days based on the distance calculation vector a (similar day extraction process). In the prediction model, a distance calculation vector b corresponding to each past day is prepared (generated) in advance, and the first prediction unit 23 specifies a distance calculation vector b similar to the distance calculation vector a. The past day corresponding to the specified distance calculation vector b is extracted as a similar day. Specifically, the first prediction unit 23 calculates the absolute value of the difference (distance between the vectors) between the distance calculation vector a and the distance calculation vector b of the past day for each past day, and calculates the distance between the vectors. The past day corresponding to the distance calculation vector b with the shortest distance is extracted as a similar day. In this embodiment, the first prediction unit 23 extracts a plurality of (for example, five) similar days. Note that the number of similar days extracted by the first predicted date may be one day. Further, the first prediction unit 23 may extract similar dates based on parameters other than the distance between vectors. Note that the database of distance calculation vectors b corresponding to past days is generated by the predictive model generation unit 22.

Next, the first prediction unit 23 calculates a first predicted value based on the similar date (first predicted value calculation process). Here, the first prediction unit 23 acquires the actual value of the power demand amount corresponding to each of the similar days for 5 days from the data storage unit 30, and calculates the average value of the actual value of the power demand amount for the 5 days. Calculate as the first predicted value. However, the first prediction unit 23 may calculate the first predicted value based on an intermediate value instead of the average value of the actual values of power demand on a plurality of similar days. Further, the first prediction unit 23 may calculate the first predicted value by multiplying each of the actual values of the power demand amount for a plurality of days by a weighting coefficient according to the distance (distance between the vectors described above). . The first prediction unit 23 outputs a first predicted value Vp1, which is a predicted value of power demand, as an output of the prediction model.

Note that the prediction model of this embodiment does not need to calculate a predicted value based on similar dates. For example, the prediction model generation unit 22 uses factor data (weather data and day of the week data) and actual values of electricity demand as learning data to generate a prediction model whose input is the factor data and whose output is the electricity demand. It may be generated.

By the way, in order to generate a predictive model, it is necessary to process a huge amount of data, so the cycle in which the predictive model generation unit 22 generates a predictive model may be relatively long. In the case of the present embodiment, the predictive model generation unit 22 generates the "weight matrix W" of the predictive model shown in FIG. ” database will be updated. On the other hand, when the predicted value of a prediction target that changes over time is obtained based on a prediction model, the error between the predicted value and the actual value increases as time passes after the prediction model is generated, and the accuracy of the predicted value increases. may decrease.

Therefore, in the present embodiment, the control unit 20 generates a classification model by analyzing the tendency of the error between the first predicted value and the actual value, and makes the first prediction based on the correction value obtained using the classification model. The second predicted value is obtained by correcting the value. In this embodiment, the classification model generation unit 24 generates a classification model. Further, in this embodiment, the second prediction unit 25 obtains a correction value using the classification model, corrects the first predicted value based on the correction value, and calculates the second predicted value. The classification model generation unit 24 and second prediction unit 25 of this embodiment will be described below.

<Classification model generation unit 24>
The classification model is a model for obtaining correction values (output data) according to factor data (input data). The classification model of this embodiment is constructed from a decision tree. The classification model generation unit 24 generates a decision tree for each hourly time period obtained by dividing one day into 24 (see FIG. 13).

FIG. 8 is a flow diagram of the classification model generation process. Each process in the figure is realized by the CPU 11 executing a control program stored in the storage device 12. The control unit 20 of this embodiment includes a first acquisition unit 211, a second acquisition unit 212, and a classification model generation unit 24 for performing each process shown in the figure (see FIG. 2).

The first acquisition unit 211 acquires first data of a predetermined factor from the data storage unit 30 (S101: first acquisition process). The first data is data on factors that are estimated to affect the error. Since the prediction target of this embodiment changes over time, the error also changes over time, so it is assumed that the cause of the error also changes over time. Furthermore, it is assumed that the factors that affect the error are also the factors that affect the predicted value of the prediction target. Therefore, the first data of this embodiment includes weather data and day of the week data, similar to the input data (factor data) of the prediction model. The first acquisition unit 211 of the present embodiment obtains, as first data, meteorological data of maximum temperature, minimum temperature, maximum humidity, and minimum humidity shown in FIG. 4B, and day-of-the-week data of flags corresponding to the days of the week shown in FIG. 5A. get. However, the first data may include other weather data, or may include data other than weather data or day of the week data.

The second acquisition unit 212 acquires second data indicating the error between the first predicted value and the actual value of the prediction target (S102: second acquisition process). 9A to 9D are explanatory diagrams of the second data acquired by the second acquisition unit 212.

As shown in FIG. 9A, the control unit 20 stores the first predicted value in the data storage unit 30. The first predicted value corresponds to the output of the aforementioned prediction model. Here, the first predicted value is a predicted value of the power demand amount at a predetermined time (request point) on the prediction target day, which is a future day at the prediction time. Therefore, as shown in FIG. 9A, at the time of prediction, the actual value (and error) of the prediction target corresponding to the first predicted value is unknown.

As time passes after the first predicted value is predicted, the actual value of the power demand amount, which was unknown at the time of the prediction, is determined. Therefore, the control unit 20 stores the actual value of the power demand amount in the data storage unit 30, as shown in FIG. 9B.

Furthermore, as time passes after the first predicted value is predicted, the error that was unknown at the time of prediction becomes finalized. Therefore, the control unit 20 stores the error between the first predicted value and the actual value in the data storage unit 30, as shown in FIG. 9C. Here, the error is defined as a ratio (error ratio) of the difference between the first predicted value and the actual value to the actual value. That is, when the first predicted value is Vp1 and the actual value is Vr, the error E is as shown in the following equation.
E = (Vp1-Vr)/Vr

In addition, the control unit 20 stores the error (and the first predicted value and the actual value) in the period for each predetermined period in the data storage unit 30, so that the period and the error can be adjusted as shown in FIG. 9D. Accumulate data that corresponds to Note that the error may be simply defined as the difference (error amount Ve) between the first predicted value and the actual value (Ve=Vp1-Vr). The second acquisition unit 212 of this embodiment acquires error data shown in FIG. 9D as second data.

Next, the classification model generation unit 24 generates a classification model (S103: classification model generation process). Some classification models determine a class from among multiple classes according to factors. In this embodiment, the classification model generation unit 24 generates a decision tree that uses first data (weather data, day of the week data) as an explanatory variable and second data as an objective variable. Furthermore, the classification model generation unit 24 generates a decision tree every predetermined period. In this embodiment, the classification model generation unit 24 generates a decision tree for each hourly time period obtained by dividing one day into 24. Here, a method for generating a decision tree in the time period from 0:00 to 1:00 on a certain prediction target day will be explained.

FIG. 10 is an explanatory diagram of data used to generate a decision tree. The classification model generation unit 24 receives first data and second data in the time period from 0:00 to 1:00 for the most recent past 30 days from the prediction target date from the first acquisition unit 211 and the second acquisition unit 212. As a result, the classification model generation unit 24 obtains the first data and second data for the time period from 0:00 to 1:00 for 30 days, as shown in FIG.

Next, the classification model generation unit 24 generates a decision tree using the first data (weather data, day of the week data) shown in FIG. 10 as explanatory variables and the second data (error) as an objective variable. For example, the classification model generation unit 24 generates a decision tree based on, for example, the CART algorithm. However, the decision tree is not limited to the CART algorithm, and the decision tree may be generated based on other algorithms. Furthermore, in order to suppress overfitting, the classification model generation unit 24 generates a decision tree (prunes the decision tree) after limiting the depth of the tree, the number of terminal nodes (number of classes), etc. . Note that the generation of a decision tree by the classification model generation unit 24 has a lighter computational load than the generation of a prediction model by the prediction model generation unit 22, so it can be executed faster than the generation of a prediction model, and it can be executed more frequently than the prediction model. can be updated to

FIG. 11 is an explanatory diagram of a decision tree for a time period from 0:00 to 1:00 on a certain prediction target day. Note that the decision tree in FIG. 11 is a decision tree generated using the first data (weather data, day of the week data) shown in FIG. 10 as explanatory variables and the second data (error E) as an objective variable.

As shown in FIG. 11, branching conditions based on factors (variables) included in the first data are set at branching nodes of the decision tree. At each branch node, a branch is set according to the branch condition. The next branch node (child node) or the class to be classified (terminal node) is set as the branch destination of each branch node. In the decision tree shown in FIG. 11, the first branch node ND1 (root node) has a branch condition set as "whether the minimum temperature is 14.58° C. or lower." At the first branch node ND1, "Class 1" is set as the branch destination when "the minimum temperature is 14.58 degrees Celsius or less", and the second branch is set as the branch destination when "the minimum temperature is greater than 14.58 degrees Celsius". Node ND2 (child node) is set. The second branch node ND2 has a branch condition set as "whether the day of the week flag is 1.5 or less." At the second branch node ND2, "Class 4" is set as the branch destination when "the day of the week flag is greater than 1.5", and the third branch node ND3 is set as the branch destination when the "day of the week flag is 1.5 or less". It is set. The third branch node ND3 has a branch condition set as “whether the maximum humidity is 86.52% or less”. At the third branch node ND3, "Class 2" is set as the branch destination when "the maximum humidity is 86.52% or less", and "Class 3" is set as the branch destination when "the maximum humidity is greater than 86.52%". is set.

Next, the classification model generation unit 24 associates third data with each class based on the second data (error) corresponding to each class of the decision tree (S104). The third data is data for correcting the first predicted value, and is the correction value C in this embodiment. Here, a method for generating the correction value C (third data) associated with "class 1" of the decision tree shown in FIG. 11 will be described.

FIG. 12 is an explanatory diagram of the third data generation method.

The classification model generation unit 24 determines each class according to 30 days of first data (weather data and day of the week data) based on the decision tree for the time period from 0:00 to 1:00. In the figure, the second data (error E) classified into class 1 is indicated by a thick line. As shown in the figure, class 1 is associated with 5 days of second data (error E). Then, the classification model generation unit 24 generates a correction value C for correcting the first predicted value based on the five pieces of second data (error E) associated with class 1. The correction value C in this embodiment is defined as a value obtained by adding 1 to the average value of a plurality of second data (error E) associated with each class. However, the correction value C associated with each class is not limited to this, and may be any data for correcting the first predicted value. Similarly, the classification model generation unit 24 also generates third data (correction value C) associated with other classes based on the plurality of second data (error E) associated with each class.

Through the above processing, a classification model for determining a class according to a factor from among a plurality of classes is generated based on the first data and the second data (S103). Furthermore, through the above processing, a classification model in which each class is associated with third data (correction value C for correcting the first correction value) is generated based on the first data and the second data. (S104).

FIG. 13 is an explanatory diagram of the decision tree for each time period. As already explained, in this embodiment, the classification model generation unit 24 generates a decision tree for each hourly time period obtained by dividing one day into 24. As shown in the figure, the decision trees for each time period are different. This means that the factors that influence the correction value differ depending on the time period. According to this embodiment, since the classification model generation unit 24 generates a decision tree for each time period, it is possible to generate third data (correction value C) suitable for each time period.

<Second prediction unit 25>
FIG. 14 is a flow diagram of the prediction method of this embodiment. Each process in the figure is realized by the CPU 11 executing a control program stored in the storage device 12. The control unit 20 of this embodiment includes a data acquisition unit 21, a first prediction unit 23, and a second prediction unit 25 for performing each process shown in the figure. Note that the second prediction unit 25 serves as a correction unit that corrects the first predicted value to the second predicted value.

The data acquisition unit 21 acquires factor data (S201). The data acquisition unit 21 of this embodiment acquires factor data for use in a prediction model and factor data for use in a classification model. Specifically, the data acquisition unit 21 uses the predicted value of the weather data at the request point (the date and time of the prediction target) and the request point as factor data to be used in the prediction model, as already explained (see FIG. 7). Obtain the actual value of the previous weather data, day of the week data, etc. The data acquisition unit 21 also acquires weather data and day of the week data at the request point (date and time of prediction target) as factor data for use in the classification model. Note that the factor data used in the classification model is data of factors corresponding to the first data used to generate the classification model, and here, the maximum temperature, minimum temperature, maximum humidity, minimum humidity, etc. at the request point, etc. 5A and flags corresponding to the days of the week shown in FIG. 5A. The data acquisition unit 21 passes the acquired factor data to the first prediction unit 23 and the second prediction unit 25 as input data.

Next, the first prediction unit 23 uses the prediction model to calculate a first predicted value (output data) from the factor data (input data) acquired in S201 (S202). As already explained, the first prediction unit 23 of this embodiment extracts similar days based on the weather data and day of the week data at the request point, and calculates the amount of power demand at the request point based on the extracted similar days. A first predicted value is calculated (see FIG. 7). However, the method of calculating the first predicted value (output data) from the factor data (input data) is not limited to this.

Next, the second prediction unit 25 uses the classification model to determine a class according to the factor data acquired in S201, and acquires third data associated with the determined class (S203). In this embodiment, since a decision tree is generated for each hourly time period obtained by dividing one day into 24 (see FIG. 13), the second prediction unit 25 selects the required point from among the 24 decision trees. A decision tree corresponding to (prediction target date and time) is selected (or the second prediction unit 25 causes the classification model generation unit 24 to generate a decision tree corresponding to the requested point). Next, the second prediction unit 25 determines a class according to the factor data based on the branching condition of the decision tree corresponding to the request point. Then, the second prediction unit 25 obtains third data (correction value C) associated with the determined class.

Next, the second prediction unit 25 corrects the first predicted value to the second predicted value based on the third data (correction value C) acquired in S202 (S204). Here, the second predicted value is a value obtained by dividing the first predicted value by the correction value C. That is, when the first predicted value is Vp1 and the correction value is C, the second predicted value Vp2 is as shown in the following equation.
Vp2 = Vp1/C

Note that in this embodiment, the second predicted value is obtained by dividing the first predicted value by the correction value, but the method for calculating the second predicted value (the method for correcting the first predicted value) is limited to this. It's not a thing. For example, the second predicted value Vp2 may be calculated by adding the correction amount Vc to the first predicted value Vp1 (Vp2=Vp1+C). The method for calculating the second predicted value (the method for correcting the first predicted value) is appropriately set according to the definition of the correction value (third data).

FIG. 15 is an explanatory diagram of the prediction results of this embodiment. In the present embodiment, the first prediction unit 23 calculates the first predicted value Vp1 of the power demand amount for each predetermined period (here, for each one-hour time period obtained by dividing one day into 24). Further, in the present embodiment, the second prediction unit 25 calculates the correction value C for each predetermined period based on a decision tree corresponding to the period. Furthermore, in the present embodiment, the second prediction unit 25 calculates the second predicted value by correcting the first predicted value based on the correction value C for each predetermined period. In addition, for reference, the actual value of the power demand amount, which was unknown at the time of prediction, is shown for each predetermined period in the figure.

FIG. 16 is a graph showing the prediction results of this embodiment. The horizontal axis of the graph indicates time. The vertical axis of the graph indicates the power demand (kWh). The solid line graph indicates the actual value Vr of the power demand amount. The dashed-dotted line graph indicates the first predicted value Vp1. The dotted line graph indicates the second predicted value Vp2. As can be understood from the graph, in this embodiment, the predicted value (second predicted value Vp2) can be determined with high accuracy.

===Summary===
The information processing device 1 of this embodiment includes a first acquisition section 211, a second acquisition section 212, and a classification model generation section 24. The first acquisition unit acquires first data (for example, weather data, day of the week data) indicating a predetermined factor that changes over time. The second acquisition unit includes second data (for example, an error Obtain E). The classification model generation unit generates a classification model based on the first data and the second data. Here, the classification model determines a class according to a factor from among multiple classes, and each class includes third data for correcting the first predicted value and calculating the second predicted value. (for example, correction value C) are associated with each other (see FIG. 13). According to this embodiment, as shown in the second predicted value of FIG. 16, it is possible to improve the prediction accuracy of a prediction target that changes over time.

By the way, when a prediction target is influenced by a plurality of factors, it is conceivable to associate a correction value with each individual factor and to correct the predicted value by superimposing the plurality of correction values. However, in such a correction method, if the degree of influence of a certain factor on the prediction target changes depending on other factors, even if multiple correction values associated with each factor are superimposed, It is difficult to correct predicted values with high accuracy. On the other hand, in this embodiment, each class of the classification model is classified in consideration of a combination of multiple factors. Therefore, in this embodiment, even in a situation where the degree of influence of a certain factor (e.g. temperature) on the prediction target changes depending on other factors (e.g. humidity), each class has a combination of multiple factors. Since the third data (correction value C) reflecting the above is associated, it is possible to correct the predicted value with high precision.

The classification model generation unit 24 of this embodiment generates a decision tree that uses first data (for example, weather data, day of the week data) as an explanatory variable and second data (for example, error E) as an objective variable (see FIG. (see FIG. 11), third data (correction value C) associated with each class is generated based on second data (for example, error E) associated with the class of the decision tree. Thereby, by using the generated decision tree, it is possible to obtain a correction value according to the factors that influence the error. However, the second data may be classified into a plurality of classes according to factors using a clustering analysis model different from a decision tree, and the third data may be calculated based on the second data classified into each class.

Furthermore, the classification model generation unit 24 of the present embodiment generates a classification model every predetermined period (for example, every hour) based on the first data and second data of the predetermined period. Thereby, suitable third data can be generated for each predetermined period, and predicted values can be determined with high accuracy.

Further, the classification model generation unit 24 of this embodiment generates the classification model in a period shorter than the period in which the prediction model is updated (for example, a four-month cycle or a one-year cycle). Furthermore, the classification model generation unit 24 of this embodiment generates a classification model based on the first data and second data after the prediction model has been updated. Thereby, even if time has passed since the prediction model was generated, by updating the classification model, it is possible to suppress a decrease in the accuracy of the predicted value.

Note that the classification model generation unit 24 may update the classification model every time the first predicted value is calculated based on the prediction model. Thereby, deterioration of the classification model due to the passage of time can be suppressed, and third data suitable for correcting the first predicted value can be obtained.

Furthermore, the classification model generation unit 24 of this embodiment generates a classification model based on the most recent first data and second data. This makes it possible to analyze the latest error trends and generate a classification model. Note that in this embodiment, the classification model is generated based on the first data and second data for the most recent 30 days, but the most recent first data and second data are limited to the most recent 30 days. It's not a thing.

Moreover, the prediction model of this embodiment calculates the first predicted value of the prediction target based on the predicted value of the factor (for example, the predicted value of weather data). However, the prediction model may calculate the first predicted value of the prediction target based on the actual value of the factor (for example, the actual value of weather data). In addition, in the above-described embodiment, the predicted value of the factor (weather data predicted value) was used. However, when calculating the predicted value (second predicted value) of the power demand amount, the first predicted value is calculated from the prediction model based on the predicted value of the factor (for example, the predicted value of weather data) (S202 in FIG. 14). ), when generating a classification model, the first predicted value may be calculated from the prediction model based on the actual value of the factor (for example, the actual value of weather data).

Further, the information processing device 1 of the present embodiment includes a second prediction unit 25 as a correction unit that corrects the first predicted value to a second predicted value. The second prediction unit 25 determines a class according to a factor using a classification model, and corrects the first predicted value to a second predicted value based on the third data associated with the determined class. Thereby, the information processing device 1 can correct the first predicted value to the second predicted value. However, the information processing device 1 does not need to include the second prediction unit 25. That is, the information processing device 1 does not need to have a function of only generating a classification model and not a function of correcting the predicted value of the prediction target (that is, using the classification model generated by the information processing device 1, (Another information processing device may correct the predicted value.)

In this embodiment, the prediction target is electric power demand, and the first data is data indicating temperature and humidity. However, the prediction target and the first data are not limited to these. For example, the prediction target may be electricity market prices, product sales, or the like. Further, in the present embodiment, the second data is the error E (the ratio of the difference between the first predicted value and the actual value to the actual value), but the second data is not limited to this. The second data may be any value that indicates the error between the first predicted value and the actual value of the prediction target, for example, it may be the amount of error (=Vp1-Vr), or it may be the amount of error (=Vp1-Vr), or the value obtained by adding 1 to the error E. It may be a value, or it may be another value. Similarly, in this embodiment, the third data is the correction value C, but the third data is not limited to this. The third data may be any data that allows calculation of the second predicted value from the first predicted value; for example, it may be the error E, which is the ratio of the difference between the first predicted value and the actual value to the actual value, or the It may be the average value of a plurality of errors E of the class, or it may be any other value.

===Others===
The above-described embodiments are provided to facilitate understanding of the present invention, and are not intended to be interpreted as limiting the present invention. Further, the present invention can be modified and improved without departing from the spirit thereof, and it goes without saying that the present invention includes equivalents thereof.

1 information processing device, 3 communication network,
11 CPU, 12 storage device,
13 communication device, 14 display device, 15 input device,
20 control unit, 21 data acquisition unit, 210 input data acquisition unit,
211 first acquisition unit, 212 second acquisition unit,
22 prediction model generation unit, 23 first prediction unit,
24 classification model generation unit, 25 second prediction unit,
30 Data storage section

Claims (12)

a first acquisition unit that acquires first data indicating a predetermined factor that changes over time;
a second acquisition unit that acquires second data indicating an error between a first predicted value calculated based on a prediction model for a prediction target that changes over time and an actual value of the prediction target;
A classification model that determines the class according to the factor from among a plurality of classes, wherein third data for correcting the first predicted value and calculating a second predicted value corresponds to each of the classes. a classification model generation unit that generates the attached classification model based on the first data and the second data;
Information processing equipment including. The information processing device according to claim 1,
The classification model generation unit includes:
Generating a decision tree using the first data as an explanatory variable and the second data as an objective variable,
The information processing apparatus is characterized in that the third data associated with each of the classes is generated based on the second data associated with the classes of the decision tree. The information processing device according to claim 1 or 2,
The information processing device is characterized in that the classification model generation unit generates the classification model for each predetermined period based on the first data and the second data of the predetermined period. The information processing device according to any one of claims 1 to 3,
The information processing device is characterized in that the classification model generation unit updates the classification model in a period shorter than a period in which the prediction model is updated. The information processing device according to claim 4,
The information processing device is characterized in that the classification model generation unit updates the classification model every time the first predicted value is calculated based on the prediction model. The information processing device according to any one of claims 1 to 5,
The information processing device is characterized in that the classification model generation unit generates the classification model based on the first data and the second data after the prediction model is updated. The information processing device according to any one of claims 1 to 6,
The information processing device is characterized in that the classification model generation unit generates the classification model based on the most recent first data and the second data. The information processing device according to any one of claims 1 to 7,
The information processing apparatus is characterized in that the prediction model calculates the first predicted value of the prediction target based on the predicted value of the factor. The information processing device according to any one of claims 1 to 8,
determining the class according to the factor using the classification model, and correcting the first predicted value to the second predicted value based on the third data associated with the determined class; An information processing device further comprising a section. The information processing device according to any one of claims 1 to 9,
The prediction target is electricity demand,
An information processing device, wherein the first data is data indicating temperature and humidity. The computer is
a first acquisition process of acquiring first data of a predetermined factor that changes over time;
a second acquisition process of acquiring second data indicating an error between a first predicted value calculated based on a prediction model for a prediction target that changes over time and an actual value of the prediction target;
A classification model that determines the class according to the factor from among a plurality of classes, wherein third data for correcting the first predicted value and calculating a second predicted value corresponds to each of the classes. a classification model generation process that generates the attached classification model based on the first data and the second data;
An information processing method that performs. to the computer,
a first acquisition process of acquiring first data of a predetermined factor that changes over time;
a second acquisition process of acquiring second data indicating an error between a first predicted value calculated based on a prediction model for a prediction target that changes over time and an actual value of the prediction target;
A classification model that determines the class according to the factor from among a plurality of classes, wherein third data for correcting the first predicted value and calculating a second predicted value corresponds to each of the classes. a classification model generation process that generates the attached classification model based on the first data and the second data;
A program to run.

Images