JP7511533B2 - Prediction device, prediction method, and program - Google Patents

Description

The present disclosure relates to a prediction device, a prediction method, and a program.

Patent Document 1 discloses the following data analysis system. That is, the data analysis system disclosed in Patent Document 1 holds learning data consisting of multiple records, each of which includes multiple explanatory variables and a target variable corresponding to the multiple explanatory variables. Then, based on the learning data, a prediction model is generated that predicts the target variable from the multiple explanatory variables. In addition, for each record, a reliability indicating the likelihood of a prediction based on the prediction model is calculated, and a subset is created from the multiple records that have a reliability within a predetermined range. Furthermore, based on the records belonging to the subset, a combination of explanatory variables that is highly correlated with the target variable is extracted from the multiple explanatory variables, and the extracted combination of explanatory variables is added to the learning data as a new explanatory variable, and a prediction model that predicts the target variable from the multiple explanatory variables is generated based on the learning data to which the new explanatory variables have been added. According to the data analysis system disclosed in Patent Document 1, it is said that the target variable can be predicted with high accuracy by adding an appropriate combination of explanatory variables to the data as a new explanatory variable.

JP 2016-4525 A

As described above, the data analysis system described in Patent Document 1 aims to improve the prediction accuracy of the objective variable by adding a new combination of explanatory variables to the training data and recreating the prediction model. However, there is a problem in that when a change in the characteristics occurs in the objective variable that cannot be adequately handled by adding a new combination of explanatory variables, it may not be possible to adequately improve the prediction accuracy of the objective variable.

The present disclosure has been made to solve the above problem, and aims to provide a prediction device, a prediction method, and a program that can improve the prediction accuracy of the dependent variable.

In order to solve the above problems, the prediction device according to the present disclosure includes one or more prediction units that predict each value of the objective variable using a plurality of trained prediction models, each of which defines a prediction algorithm, one or more explanatory variables, and a learning period, and in which the prediction algorithm, the one or more explanatory variables, and some or all of the learning periods are different from one another and each predicts the value of the same objective variable; an evaluation unit that evaluates the prediction accuracy of each value of the objective variable for each of the trained prediction models; and an output unit that outputs the value of the objective variable based on the evaluation result of the prediction accuracy.

The prediction method according to the present disclosure includes a step of predicting each value of the objective variable using a plurality of trained prediction models, each of which defines a prediction algorithm, one or more explanatory variables, and a learning period, and in which the prediction algorithm, the one or more explanatory variables, and some or all of the learning period are different from one another and each predicts the value of the same objective variable; a step of evaluating the prediction accuracy of each value of the objective variable for each of the trained prediction models; and a step of outputting the value of the objective variable based on the evaluation result of the prediction accuracy.

The program according to the present disclosure causes a computer to execute the steps of: predicting each value of the objective variable using a plurality of trained prediction models, each of which defines a prediction algorithm, one or more explanatory variables, and a learning period, and in which the prediction algorithm, the one or more explanatory variables, and some or all of the learning period are different from one another, and each of which predicts the value of the same objective variable; evaluating the prediction accuracy of each value of the objective variable for each of the trained prediction models; and outputting the value of the objective variable based on the evaluation result of the prediction accuracy.

The prediction device, prediction method, and program disclosed herein can improve the prediction accuracy of the objective variable.

1 is a block diagram illustrating an example configuration of a prediction device according to an embodiment of the present disclosure. 1 is a diagram showing an example of the configuration of a trained prediction model according to an embodiment of the present disclosure. FIG. 1 is a schematic diagram for explaining an example of the operation of a prediction device according to an embodiment of the present disclosure. FIG. 1 is a schematic diagram for explaining an example of the operation of a prediction device according to an embodiment of the present disclosure. 11 is a flowchart illustrating an example of the operation of the prediction device according to the embodiment of the present disclosure. FIG. 1 is a schematic block diagram illustrating a configuration of a computer according to at least one embodiment.

(Configuration of prediction device)
Hereinafter, a prediction device, a prediction method, and a program according to an embodiment of the present disclosure will be described with reference to FIGS. 1 to 6. FIG. 1 is a block diagram showing a configuration example of a prediction device according to an embodiment of the present disclosure. FIG. 2 is a diagram showing a configuration example of a trained prediction model according to an embodiment of the present disclosure. FIGS. 3 and 4 are schematic diagrams for explaining an operation example of a prediction device according to an embodiment of the present disclosure. FIG. 5 is a flowchart showing an operation example of a prediction device according to an embodiment of the present disclosure. FIG. 6 is a schematic block diagram showing a configuration of a computer according to at least one embodiment. Note that in each figure, the same or corresponding configurations are designated by the same reference numerals and the description is omitted as appropriate.

In the following, a case will be described in which the prediction device according to this embodiment uses the target variable to be predicted as the spot price in the electricity trading market (the system price, area price, etc. of spot market transactions (hereinafter referred to as "spot transactions"). However, the target variable is not limited to the spot price in the electricity trading market, and can be a variable representing any social or natural phenomenon that has a continuous response.

As shown in FIG. 1, the prediction device 1 has the following components as a functional configuration consisting of a combination of a computer such as a server or a personal computer, hardware such as peripheral devices of the computer, and software such as a program executed by the computer. That is, the prediction device 1 has a data collection unit 11, a collected data DB (database) 12, an evaluation unit 13, an output unit 14, multiple update units 15, a prediction model DB 16, multiple prediction units 17, and a prediction value DB 18.

The data collection unit 11 collects spot price performance data, predetermined open data, meteorological data such as temperature, humidity, and weather, market data, operation data of various facilities, environmental data, maintenance and field data of various facilities, etc. from one or more external servers 31-33, etc., via a communication network 2 such as the Internet, and stores the collected data in the collected data DB 12. Depending on the type of objective variable, the data collection unit 11 may be configured to collect various data within a certain premises using an on-premises communication network, etc., in addition to or instead of collecting data from outside. Alternatively, the data collection unit 11 may include a configuration for collecting data by accepting conditions such as equipment operation plans input by a person via an input unit such as a web screen (GUI (Graphical User Interface)).

The collected data DB 12 stores multiple input data time series 121, which are time series of multiple types of input data, and actual value time series 122, which are time series of actual values of the objective variable. Each input data time series 121 is a time series of predetermined open data, meteorological data such as temperature, humidity, and weather, market data, operation data of various facilities, environmental data, maintenance and field data of various facilities, etc., collected by the data collection unit 11. In this example, the actual value time series 122 is a time series of actual values of spot prices, and includes, for example, prices every 30 minutes.

The prediction model DB16 stores multiple trained prediction models 161. A prediction algorithm, one or more explanatory variables, and a learning period are defined for each trained prediction model 161. Furthermore, each trained prediction model 161 differs from the others in some or all of the prediction algorithm, one or more explanatory variables, and learning period. Each trained prediction model 161 is a trained prediction model (machine learning model) that each predicts the value of the same objective variable (in this example, the spot price).

FIG. 2 shows an example of the configuration of a trained prediction model 161 according to an embodiment of the present disclosure. FIG. 2 shows an example of the configuration of five trained prediction models with IDs (identification codes) ID1 to ID5. The five trained prediction models ID1 to ID5 all have the same target variable, which is the spot price. The three trained prediction models ID1 to ID3 have the same learning period. The three trained prediction models ID1 to ID3 have different prediction algorithms. The two trained prediction models ID4 and ID5 have the same explanatory variables (1) and (2) and the same prediction algorithm, but have different learning periods. The two trained prediction models ID4 and ID5 have only explanatory variables (1) and (2) as explanatory variables.

In the example shown in FIG. 2, the value of the objective variable is, for example, a time series of predicted spot prices in 30-minute intervals. The value of explanatory variable (1) is, for example, a time series of actual spot prices in 30-minute intervals or a time series or predicted spot prices. The explanatory variable (2) represents calendar information, and the value of the explanatory variable is, for example, a time series of one day indicating whether it is a weekday or a holiday. The value of explanatory variable (3) is an actual or predicted temperature value, for example, a time series of one hour. The value of explanatory variable (4) is an actual or predicted humidity value, for example, a time series of one hour. The explanatory variable (5) represents equipment operation information, and is information indicating whether a specific piece of equipment or multiple pieces of equipment is operating, or the actual or predicted power consumption or power generation value.

The learning period indicates the period of learning data when machine learning the learned prediction model 161. For example, if it is one year, the time series of the actual values of the objective variable and the time series of the actual values of one or more explanatory variables for the past year are used as learning data (teacher data) to machine learn the learned prediction model 161 (or the unlearned prediction model 161).

The prediction algorithm is a regression algorithm, and includes multiple types of algorithms, such as linear regression, nonlinear regression, random forest, and neural network. Method A, Method B, and Method C represent different types of prediction algorithms, or the same type but with different parameter settings that can be set in advance.

The multiple update units 15 are provided corresponding to the multiple trained prediction models 161 stored in the prediction model DB 16. Each update unit 15 re-learns and updates each trained prediction model 161. Each update unit 15 re-learns and updates each trained prediction model 161 for each update period that is preset based on each learning period, for example. The update period can be, for example, one day, three days, one week, one month, etc., regardless of the learning period. Alternatively, the learning period and the update period may be the same. Also, each update unit 15 may be omitted.

The multiple prediction units 17 are provided corresponding to the multiple learned prediction models 161 stored in the prediction model DB 16. Each prediction unit 17 predicts each value of the objective variable using each learned prediction model 161, and stores each predicted value as each predicted value time series 181 in the predicted value DB 18. Each prediction unit 17 inputs a time series of one or more explanatory variables set for each learned prediction model 161 to each learned prediction model 161 in parallel using the multiple learned prediction models 161, and predicts each value (time series or one value) of the objective variable as the output of each learned prediction model 161.

The evaluation unit 13 evaluates the prediction accuracy of each value of the objective variable for each trained prediction model 161. The output unit 14 outputs the value of the objective variable to a predetermined output medium based on the evaluation result of the prediction accuracy by the evaluation unit 13. Here, an overview of the processing of the evaluation unit 13 and the processing of the output unit 14 will be described with reference to FIG. 3. FIG. 3 shows an example of the actual line of the objective variable (an approximate curve of the time series of the actual measured value (solid curve)), the predicted values of the three types of trained prediction models 161 of ID1 to ID3 shown in FIG. 2 (plots represented by "black circles", "white squares" and "white triangles" in the figure), the values output by the output unit 14 ("white circles") and the prediction line (dashed curve), with the horizontal axis representing time and the vertical axis representing the value of the objective variable. In this case, the prediction algorithm for the predicted value "black circle" is Method A, the prediction algorithm for the predicted value "white square" is Method B, and the prediction algorithm for the predicted value "white triangle" is Method C.

The evaluation unit 13 compares the actual values (actual line) with each predicted value (the "black circle", "white square", and "white triangle" in the figure) during the evaluation period from the present to a predetermined time ago, and evaluates the prediction accuracy. In the example shown in FIG. 3, the evaluation unit 13 evaluates, for example, that the prediction accuracy of the predicted value of method A (the "black circle") was the highest. Meanwhile, the output unit 14 selects the trained prediction model 161 of ID1 (method A) which had the highest evaluation result for prediction accuracy, and outputs the value of the objective variable predicted by the selected trained prediction model 161 (the "white circle") and the prediction line (the dashed curve).

The evaluation unit 13 evaluates the prediction accuracy of each value (each predicted value) of the objective variable based on, for example, any one of the following evaluation indexes I1 to I6, or multiple values of multiple weighted evaluation indexes I1 to I6. However, the evaluation indexes are not limited to these.

Here, ya is the actual value, yp is the predicted value, and n is the number of predicted values in the evaluation period.

Index I1 is an index called RMSE (Root Mean Squared Error) and is used when evaluating the average error and wanting to increase the penalty as the error increases. Index I2 is an index called MAE (Mean Absolute Error) and is used when evaluating the average error and placing more importance on outliers (large errors) than on RMSE. Index I3 is an index called MAPE (Mean Absolute Percentage Error) and is used when evaluating the average error and is evaluated as an error rate, so it is less affected by the size of the value and can be evaluated in the same way even when the value itself is small and large. Index I4 is an index called ME (Max Error) and is used when evaluating the maximum error. Note that max(Y1, Y2, Y3, ...) is a function that outputs the maximum value of Y1, Y2, Y3, .... Index I5 is an index called the R2 Score, and is used to evaluate the average error trend, including the impact of overshooting and undershooting of the predicted value. Index I5 is used to evaluate the qualitative trend (magnitude relationship) against the threshold.

The output unit 14 may select and output the predicted value (output value) of the learned prediction model 161 that the evaluation unit 13 has evaluated as having the highest prediction accuracy as described with reference to FIG. 3, or may calculate and output the average value or weighted average value of the predicted values of the learned prediction models 161 with the highest prediction accuracy. For example, as shown in FIG. 4, the output unit 14 outputs the predicted values and prediction lines for the future (24 hours on the date of the month and year of the month ...

The predicted value DB 18 also stores a time series of each predicted value (or one predicted value) (predicted value time series) 181 predicted using each trained model 161, etc.

(Example of operation of prediction device)
An example of the operation of the prediction device 1 will be described with reference to Fig. 5. Fig. 5 shows a process flow when the prediction device 1 calculates and outputs a single predicted value (for one to multiple days in the future, for example). The variable k is a variable representing the prediction model ID of the trained prediction model 161. In addition, in the process shown in Fig. 5, N trained prediction models 161 with prediction model IDs = 1 to N are processed in parallel. The process shown in Fig. 5 is executed, for example, once a day.

When the process shown in FIG. 5 is started, the prediction device 1 first initializes the variable k to 0 (step S1). Next, the prediction device 1 increments the variable k by 1 (step S2). Next, in the prediction device 1, the update unit 15 corresponding to the trained prediction model 161 with the prediction model ID = k re-learns and updates the trained prediction model 161 with the prediction model ID = k using the learning data for the learning period set for the trained prediction model 161 for each update period (step S3). Next, the prediction unit 17 corresponding to the trained prediction model 161 with the prediction model ID = k calculates a predicted value using the trained prediction model 161 with the prediction model ID = k (step S4). Next, the evaluation unit 13 calculates an evaluation index for the trained prediction model 161 with the prediction model ID = k and stores it in a specified storage area (step S5). Next, the prediction device 1 determines whether the prediction model ID has reached N (step S6).

If the prediction model ID has not reached N (step S6: NO), the prediction device 1 executes the processing from step S2 onwards again.

When the prediction model ID reaches N (step S6: YES), the output unit 14 selects a trained prediction model 161 based on the evaluation index (step S7), outputs a predicted value based on the selected trained prediction model 161 (step S8), and ends the process shown in FIG. 5.

By the above processing, the prediction device 1 operates multiple trained prediction models 161 in parallel, each of which has a different prediction algorithm, one or more explanatory variables, and a part or all of a learning period. The prediction device 1 also re-learns and updates the multiple trained prediction models 161 in parallel, with the same or different learning periods. With this configuration, one or more optimal trained prediction models 161 can be selected from the multiple trained prediction models 161 processed in parallel, and a predicted value of the objective variable can be determined based on the value predicted by the selected trained prediction model 161. Therefore, even if a change in the characteristics occurs in the explanatory variables or objective variables that impairs continuity, for example, a predicted value can be obtained using a trained prediction model appropriate for the change, and the prediction accuracy of the objective variable can be appropriately improved.

(Action and Effect)
According to the prediction device, prediction method, and program configured as described above, it is possible to improve the prediction accuracy of the dependent variable.

Other Embodiments
Although the embodiment of the present disclosure has been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and includes design changes and the like within the scope of the gist of the present disclosure. For example, an embodiment is also conceivable in which a plurality of trained prediction models 161 are stored on a cloud, and data of each trained prediction model 161 on the cloud is called up one by one on a separate PC (Personal Computer) or the like, and the above-mentioned prediction, evaluation, and output are performed on the PC.

In the configuration example shown in FIG. 1, the prediction device 1 includes a plurality of update units 15 and a plurality of prediction units 17 corresponding to the plurality of learned prediction models 161 stored in the prediction model DB 16, but is not limited thereto. For example, one prediction unit 17 may predict the values of a plurality of objective variables by using a plurality of learned prediction models 161, for example, stepwise, or a plurality of prediction units 17 less than the number of learned prediction models 161 may predict the values of a plurality of objective variables by using a plurality of learned prediction models 161, for example, stepwise, in parallel. The same applies to the update unit 15. In this case, the prediction device 1 includes one prediction unit 17, a plurality of prediction units 17 less than the number of learned prediction models 161, or a plurality of prediction units 17 equal in number to the number of learned prediction models 161. Similarly, the prediction device 1 includes one update unit 15, a plurality of update units 15 less than the number of learned prediction models 161, or a plurality of update units 15 equal in number to the number of learned prediction models 161.

In the example of the five trained prediction models 161 shown in FIG. 2, the explanatory variable (1) "spot price" and the explanatory variable (2) "calendar information" are common, but there may be a trained prediction model 161 that does not have explanatory variables in common with other trained prediction models 161. That is, there may be a trained prediction model 161 that does not have any explanatory variables that are the same as other trained prediction models 161 (i.e., all explanatory variables are different from others). In the example shown in FIG. 2, for example, there may be a trained prediction model 161 that does not include "spot price", "calendar information", "temperature", "humidity", and "equipment operation information" as explanatory variables. In addition, there may be a trained prediction model 161 that is different from other trained prediction models 161 in all of the explanatory variables, learning period, and prediction algorithm, not limited to the explanatory variables.

(Computer Configuration)
FIG. 6, as described above, is a schematic block diagram illustrating a computer configuration according to at least one embodiment.
The computer 90 comprises a processor 91 , a main memory 92 , a storage 93 , and an interface 94 .
The prediction device 1 described above is implemented in a computer 90. The operations of each of the above-mentioned processing units are stored in the form of a program in a storage 93. The processor 91 reads the program from the storage 93, loads it in the main memory 92, and executes the above-mentioned processing in accordance with the program. The processor 91 also secures storage areas in the main memory 92 corresponding to each of the above-mentioned storage units in accordance with the program.

The program may be for realizing some of the functions to be performed by the computer 90. For example, the program may be for realizing the functions by combining with other programs already stored in the storage, or by combining with other programs implemented in other devices. In other embodiments, the computer may include a custom LSI (Large Scale Integrated Circuit) such as a PLD (Programmable Logic Device) in addition to or instead of the above configuration. Examples of PLDs include PAL (Programmable Array Logic), GAL (Generic Array Logic), CPLD (Complex Programmable Logic Device), and FPGA (Field Programmable Gate Array). In this case, some or all of the functions realized by the processor may be realized by the integrated circuit.

Examples of storage 93 include HDD (Hard Disk Drive), SSD (Solid State Drive), magnetic disk, magneto-optical disk, CD-ROM (Compact Disc Read Only Memory), DVD-ROM (Digital Versatile Disc Read Only Memory), semiconductor memory, etc. Storage 93 may be an internal medium directly connected to the bus of computer 90, or an external medium connected to computer 90 via interface 94 or a communication line. In addition, when this program is distributed to computer 90 via a communication line, computer 90 that receives the program may expand the program in main memory 92 and execute the above-mentioned processing. In at least one embodiment, storage 93 is a non-transitory tangible storage medium.

<Additional Notes>
The prediction device 1 described in the above embodiment can be understood, for example, as follows.

(1) A prediction device 1 according to a first aspect includes one or more prediction units 17 that predict each value of an objective variable using a plurality of trained prediction models 161, each of which defines a prediction algorithm, one or more explanatory variables, and a learning period, and in which the prediction algorithm, the one or more explanatory variables, and a part or all of the learning period are different from one another and each predicts a value of the same objective variable, an evaluation unit 13 that evaluates the prediction accuracy of each value of the objective variable for each of the trained prediction models 161, and an output unit 14 that outputs the value of the objective variable based on the evaluation result of the prediction accuracy. With this configuration, it is possible to improve the prediction accuracy of the objective variable.

(2) The prediction device 1 according to the second aspect is the prediction device 1 according to (1), further including one or more update units 15 that re-learn and update each of the learned prediction models in each of the learning periods. With this configuration, the multiple learned prediction models 161 can be maintained as prediction models based on, for example, the latest learning data.

(3) A prediction device 1 according to a third aspect is the prediction device 1 according to (1) or (2), in which the output unit selects one of the trained prediction models based on the evaluation result of the prediction accuracy, and outputs the value of the objective variable predicted by the selected trained prediction model. With this configuration, the value of the objective variable can be output by simple processing.

(4) The prediction device 1 according to the fourth aspect is the prediction device 1 according to any one of (1) to (3), in which the objective variable is the spot price in the electricity trading market.

1 Prediction device 11 Data collection unit 12 Collected data DB
13 Evaluation unit 14 Output unit 15 Update unit 16 Prediction model DB
17 Prediction unit 18 Prediction value DB
161 Trained predictive model

Claims (6)

one or more prediction units, each of which defines a prediction algorithm, a plurality of explanatory variables, and a learning period, and which predicts values of the same objective variable using a plurality of trained prediction models, each of which predicts values of the same objective variable, and in which the prediction algorithm, the plurality of explanatory variables, and the learning period are all or partly different from one another;
an evaluation unit that evaluates the prediction accuracy of each value of the objective variable for each of the trained prediction models;
an output unit that outputs the value of the response variable based on the evaluation result of the prediction accuracy;
Equipped with
The plurality of trained prediction models include at least those selected from a plurality of predetermined candidates of explanatory variables such that their combinations are different from each other, selected from candidates of prediction algorithms such that their combinations are different from each other, and selected from candidates of learning periods such that their combinations are different from each other .
Prediction device. The prediction device according to claim 1 , further comprising one or more update units configured to re-learn and update each of the trained prediction models in each of the learning periods. The prediction device according to claim 1 , wherein the output unit selects one of the trained prediction models based on a result of the evaluation of the prediction accuracy, and outputs a value of the dependent variable predicted by the selected trained prediction model. The prediction device according to claim 1 , wherein the objective variable is a spot price in an electricity trading market. a prediction unit of the prediction device predicts each value of the objective variable using a plurality of trained prediction models, each of which defines a prediction algorithm, a plurality of explanatory variables, and a learning period, the prediction algorithm, the plurality of explanatory variables, and a part or all of the learning period being different from one another, and each of which predicts a value of the same objective variable;
an evaluation unit of the prediction device evaluating a prediction accuracy of each value of the objective variable for each of the trained prediction models;
an output unit of the prediction device outputting a value of the dependent variable based on a result of the evaluation of the prediction accuracy;
Including,
The plurality of trained prediction models include at least those selected from a plurality of predetermined candidates of explanatory variables such that their combinations are different from each other, selected from candidates of prediction algorithms such that their combinations are different from each other, and selected from candidates of learning periods such that their combinations are different from each other .
Forecasting methods. a step of predicting each value of the objective variable using a plurality of trained prediction models, each of which defines a prediction algorithm, a plurality of explanatory variables, and a learning period, the prediction algorithm, the plurality of explanatory variables, and a part or all of the learning period being different from one another, and each of which predicts a value of the same objective variable;
evaluating the prediction accuracy of each value of the objective variable for each of the trained prediction models;
outputting a value of the response variable based on a result of the evaluation of the prediction accuracy;
Run the following on your computer:
The plurality of trained prediction models include at least those selected from a plurality of predetermined candidates of explanatory variables such that their combinations are different from each other, selected from candidates of prediction algorithms such that their combinations are different from each other, and selected from candidates of learning periods such that their combinations are different from each other .
program.

Images