Model generation device, prediction device, model generation method, and prediction method

Abstract

Improve the accuracy of forecasting the amount of electricity demand at facilities. [Solution] A model generation device (10B) that generates a prediction model for predicting the amount of energy demand in a facility selects a special lag from among multiple lags other than the minimum lag based on the autocorrelation coefficients of actual values ​​of energy demand corresponding to each of multiple lags and working status data indicating daily working status at the facility, derives a new feature by setting the special lag to the actual value, generates a first model based on first time series data in which the new feature is added to time series data indicating the time progression of the actual value, derives a further new feature by setting the minimum lag to the actual value, and further generates a second model based on second time series data in which the further new feature is added to the first time series data.

Description

[Technical field]

[0001] One aspect of the present invention relates to a model generating device that generates a prediction model for predicting an amount of power demand in a facility. [Background technology]

[0002] In recent years, machine learning has been used to predict various types of energy demand in facilities. For example, Patent Document 1 below discloses a technology for predicting the heat load in a thermal storage utilization system by using a neural network. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 9-89348 Summary of the Invention [Problem to be solved by the invention]

[0004] An aspect of the present invention aims to improve the accuracy of predicting the amount of power demand in a facility. [Means for solving the problem]

[0005] A model generation device according to one embodiment of the present invention is a model generation device that generates a prediction model for predicting an amount of energy demand in a facility, wherein the model generation device sets multiple lags for time series data indicating a time progression of an actual value of the amount of energy demand, and derives autocorrelation coefficients of the actual value corresponding to each of the multiple lags, selects a special lag from among the multiple lags other than a minimum lag that is the minimum value of the multiple lags based on the autocorrelation coefficient and work status data indicating daily work status at the facility, and derives a new feature by setting the special lag for the actual value, generates a first model as the prediction model by machine learning based on first time series data in which the new feature is added to the time series data, derives a further new feature by setting the minimum lag for the actual value, and further generates a second model different from the first model by machine learning based on second time series data in which the further new feature is added to the first time series data.

[0006] A prediction device according to one aspect of the present invention is a prediction device that derives a predicted value of an amount of energy demand in a facility by using a prediction model for predicting an amount of energy demand in a facility, the prediction model having been generated in advance in a learning phase, in a prediction phase, wherein in the learning phase, a plurality of lags are set for time series data showing a time transition of an actual value of the amount of energy demand, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived, and a special lag is selected from among the plurality of lags other than a minimum lag, which is the minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and the ... a first model is generated by machine learning as the prediction model based on first time series data in which the new feature is added to the time series data, a further new feature is derived by setting the minimum lag to the actual value, and a second model different from the first model is further generated by machine learning based on second time series data in which the further new feature is added to the first time series data, and the prediction device predicts the demand energy by using the first model and the second model in the prediction phase.

[0007] A model generation method according to one embodiment of the present invention is a model generation method for generating a prediction model for predicting an amount of energy demand in a facility, the model generation method including the steps of: setting a plurality of lags for time series data indicating a time transition of an actual value of the amount of energy demand, thereby deriving an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags; selecting a special lag from among the plurality of lags other than a minimum lag, which is a minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data indicating a daily working status in the facility; deriving a new feature by setting the special lag for the actual value; generating a first model as the prediction model by machine learning based on first time series data in which the new feature is added to the time series data; deriving a further new feature by setting the minimum lag for the actual value; and further generating a second model different from the first model by machine learning based on second time series data in which the further new feature is added to the first time series data.

[0008] A prediction method according to one aspect of the present invention is a prediction method for deriving a predicted value of an amount of energy demand in a facility by using, in a prediction phase, a prediction model for predicting an amount of energy demand in a facility, the prediction model having been generated in advance in a learning phase. In the learning phase, a plurality of lags are set for time series data showing a time transition of an actual value of the amount of energy demand, and autocorrelation coefficients of the actual value corresponding to each of the plurality of lags are derived. A special lag is selected from among the plurality of lags other than a minimum lag, which is the minimum value of the plurality of lags, based on the autocorrelation coefficients and working status data showing daily working status in the facility, and the special lag is set as a prediction model for predicting an amount of energy demand in a facility. a first model is generated by machine learning as the prediction model based on first time series data in which the new feature has been added to the time series data, a further new feature is derived by setting the minimum lag to the actual value, and a second model different from the first model is further generated by machine learning based on second time series data in which the further new feature has been added to the first time series data, and the prediction method includes a step of predicting the demanded energy by using the first model and the second model in the prediction phase. Effect of the Invention

[0009] According to one aspect of the present invention, it is possible to improve the accuracy of predicting the amount of power demand in a facility. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram illustrating a configuration of a main part of an information processing device according to a reference embodiment. [Diagram 2] 13 is an example of performance data as training data in a reference form. [Diagram 3] 13 is an example of employment status data as training data in a reference form. [Figure 4]4 is an example of data generated by performing prevalue interpolation on the working status data of FIG. 3. [Diagram 5] 5 is an example of data generated by combining the performance data of FIG. 2 with the data of FIG. 4. [Figure 6] 13 is an example of a table showing a correspondence relationship between lags and ACs. [Figure 7] 1 is an example of a graph showing the relationship between lag and AC. [Figure 8] FIG. 6 is an example of data generated by adding lag features to the data in FIG. 5. [Figure 9] 8 is an example of data generated by adding a rolling feature component and rolling features to the data in FIG. [Figure 10] FIG. 10 is an example of data generated by removing the rolling feature component from the data in FIG. [Figure 11] FIG. 11 is an example of data generated by adding a time point equivalent to the data of FIG. 10. [Figure 12] 12 is an example of data generated by adding days of the week to the data in FIG. 11. [Figure 13] 13 is an example of data generated by standardizing each feature amount in the data of FIG. 12. [Figure 14] 13 is a flowchart illustrating a process of selecting an arithmetic progression element. [Figure 15] 13 is an example of pre-processed input data in a prediction phase in the reference embodiment. [Figure 16] 1 is a block diagram illustrating a configuration of a main part of an information processing device according to a first embodiment. [Figure 17] An example of a process for deriving a workday non-rolling lag with a focus on workdays will be shown below. [Figure 18] An example of a process for deriving a holiday non-rolling lag with a focus on holidays will be described below. [Figure 19] 17 and 18 show a list of non-rolling lags determined by the processes of FIG. [Figure 20]An example of a process for deriving a workday rolling lag with a focus on workdays will be shown below. [Figure 21] An example of a process for deriving a holiday non-rolling lag with a focus on holidays will be described below. [Figure 22] The table below shows the rolling lags determined by the processing in Figures 20 and 21. [Figure 23] 4 is an example of new time series data generated in the first embodiment. [Figure 24] 13 is another example of new time series data generated in the first embodiment. [Diagram 25] 13 is another example of new time series data generated in the first embodiment. [Figure 26] 13 is another example of new time series data generated in the first embodiment. [Figure 27] FIG. 11 is a block diagram illustrating the configuration of a main part of an information processing device according to a second embodiment. [Figure 28] 13 is an example of a graph showing an error of a first model at each time point in a prediction period. [Figure 29] 13 is an example of a graph showing the error of the second model at each time point in the prediction period. [Diagram 30] 13 is an example of a graph showing an error of a first model and an error of a second model at each time point in a prediction period. [Diagram 31] 13 shows an example of the total error corresponding to each of a plurality of switching time point candidates. [Diagram 32] 1 shows an example of input data for a first model in the prediction phase. [Diagram 33] 13 shows an example of input data for the second model in the prediction phase. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] [Reference form] Prior to the description of the information processing device 1A of the first embodiment, the information processing device 1 as a reference embodiment will be described. For convenience of description, in each of the following embodiments, components having the same functions as those described in the reference embodiment will be denoted with the same reference numerals, and the description will not be repeated. For simplicity, the description of matters similar to those in the known art will be omitted as appropriate. Each component and each numerical value described in this specification is merely an example unless there is a particular contradiction. Therefore, for example, unless there is a particular contradiction, the positional relationship and connection relationship of each component are not limited to the examples in each figure. Also, please note that the correspondence between the date and the day of the week in the examples in each figure does not necessarily match that of an actual calendar. In this specification, unless there is a particular contradiction, the notation "A to B" for two numbers A and B means "greater than or equal to A and less than or equal to B."

[0012] (Overview of information processing device 1) 1 is a block diagram illustrating a configuration of a main part of an information processing device 1 according to a reference embodiment. The information processing device 1 may include a control device 9 and a storage unit 90. The control device 9 comprehensively controls each part of the information processing device 1. The control device 9 may include a model generating device 10 and a prediction device 20.

[0013] As described below, the model generation device 10 generates a prediction model for predicting the amount of power demand in a facility by machine learning. For this reason, the model generation device 10 may be called a learning device. The prediction device 20 derives a prediction value of the amount of power demand in the facility by using the prediction model generated in advance by the model generation device 10.

[0014] The facility according to one embodiment of the present invention may be a facility in which the demand for electric power may fluctuate depending on at least one of the time of day, the day of the week, and the working status of the employees in the facility. Examples of the facility according to one embodiment of the present invention include industrial facilities such as factories, commercial facilities such as shopping malls, and medical facilities such as hospitals.

[0015] The storage unit 90 stores various data and programs used for processing by the control device 9. In the example of Fig. 1, it is assumed that actual data and working status data are stored in the storage unit 90. In this specification, actual data collectively refers to data (more specifically, data structure) indicating actual values ​​of the amount of power demand in a facility, recorded with a predetermined time resolution. In this specification, unless otherwise contradictory, the term "amount of power demand" refers to the actual value of the amount of power demand.

[0016] The actual data is data in which a time point and an amount of power demand are associated with each other, and is therefore an example of time-series data that indicates the time transition of the amount of power demand. In the examples of this specification, the time resolution of the actual data is one minute. In the examples of this specification, the time point in data with a time resolution of one minute is expressed in a "year / month / day hour / minute" format ("yyyy / mm / dd / hh / mm" format). The time point may be used as a key in the data structure.

[0017] In this specification, working status data collectively refers to data indicating the working status of a facility on a daily basis. As an example, working status data may be data indicating whether a certain day is a working day for the facility. Since working status data is data in which dates and working statuses are associated with each other, it is an example of time series data indicating daily changes in working status. In the example of this specification, the time resolution of working status data before the preceding value interpolation described below is one day. In the example of this specification, the date in data with a time resolution of one day is expressed in the "year / month / day" format ("yyyy / mm / dd" format).

[0018] (Learning phase in reference form) The following describes an example of the operation of the information processing device 1. The processing of the information processing device 1 is roughly divided into processing in a learning phase (a series of processing by the model generating device 10) and processing in a prediction phase (a series of processing by the prediction device 20). First, an example of the learning phase will be described.

[0019] 1, the model generating device 10 may include a data acquiring unit 11, a preprocessing unit 12, a learning unit 13, and an evaluation unit 14. The data acquiring unit 11 may be referred to as a learning-time data acquiring unit or a first data acquiring unit. The preprocessing unit 12 may be referred to as a learning-time preprocessing unit or a first preprocessing unit.

[0020] As shown in Fig. 1, the storage unit 90 may store performance data DZ and working status data DS used in the learning phase. Hereinafter, for example, the performance data DZ is simply abbreviated as DZ. The data acquisition unit 11 acquires DZ and DS as input data sets in the learning phase. Then, the data acquisition unit 11 divides the input data set into training data and validation data.

[0021] DZ0 in FIG. 2 is an example of actual data as training data. DZ0 is a part of DZ. In the example of this specification, the unit of the demand energy recorded in the actual data is kWh. DS0 in FIG. 3 is an example of working status data as training data. DS0 is a part of DS. In the example of this specification, the working status recorded in the working status data is converted in advance into a qualitative variable (more specifically, a nominal scale). In the example of FIG. 3, a working status value of "1" indicates that the day (the day corresponding to the working status) is a working day. On the other hand, a working status value of "0" indicates that the day is a holiday.

[0022] However, the type of working status recorded in the working status data is not limited to the above example. For example, a working status type of "long vacation day" may be recorded in the working status data. A long vacation day means a holiday that belongs to a specified long vacation (e.g., summer vacation period). In this case, for example, a value of "-1" in the working status may indicate that the day is a long vacation day.

[0023] The preprocessing unit 12 generates preprocessed training data by performing a series of preprocessing processes on the training data prior to execution of the machine learning algorithm by the learning unit 13. Next, the learning unit 13 executes the machine learning algorithm to generate a prediction model based on the preprocessed training data.

[0024] First, the pre-processing unit 12 generates new data by performing pre-value interpolation on the working status recorded in DS0. DS1 in FIG. 4 is an example of the new data. DS1 may be referred to as working status data after pre-value interpolation. Specifically, the pre-processing unit 12 converts DS0 into new data having a time resolution (e.g., 1 minute) equal to that of DZ0 by pre-value interpolation. As shown in FIG. 4, as a result of the pre-value interpolation, in DS1, working statuses belonging to the same date have the same value at all time points.

[0025] Next, the preprocessing unit 12 generates new data by combining DZ0 and DS1 using the time point as a key. DZ1 in FIG. 4 is an example of the new data. For example, the preprocessing unit 12 generates DZ1 by adding a column of "working status" in DS1 to DZ0. DZ1 may be referred to as actual data with working status data.

[0026] Next, the preprocessing unit 12 calculates an autocorrelation coefficient of the demand energy in the DZ1. In this specification, the autocorrelation coefficient is abbreviated as AC. Specifically, the preprocessing unit 12 changes the lag in various ways and calculates the AC for each lag. The lag represents a time delay amount set for the feature amount as time-series data.

[0027] The pre-processing unit 12 may set the lag as a product of the lag unit amount and the lag order. Then, the pre-processing unit 12 may calculate AC for each determined lag. As an example, the lag unit amount may be a fixed value previously set by a user of the information processing device 1. In this specification, a case where the lag unit amount is 30 minutes is illustrated. However, the lag unit amount may be set by the pre-processing unit 12. For example, the pre-processing unit 12 may set the time resolution of DZ1 (e.g., 1 minute) as the lag unit amount.

[0028] The preprocessing unit 12 may generate a table showing the correspondence between lags and ACs. LT in FIG. 6 shows an example of the table. In the example of FIG. 6, the lag order may be used as a key in LT. In the example of FIG. 6, the preprocessing unit 12 sets 1008 lags by changing the lag order from 1 to 1008. Thus, the preprocessing unit 12 calculates 1008 ACs. The lag in the examples of this specification is a quantity representing a time length in days.

[0029] As mentioned above, the lag unit amount is 30 minutes, so a lag order of 1 corresponds to 30 minutes. Therefore, a lag order of 48 corresponds to 1 day (=30 minutes x 48). And a lag order of 336 (=48 x 7) corresponds to 7 days, or 1 week. And a lag order of 672 (=336 x 2) corresponds to 14 days, or 2 weeks, and a lag order of 1008 (=336 x 3) corresponds to 21 days, or 3 weeks.

[0030] In this specification, the element number of the data structure is generally represented by an integer i. In the example of FIG. 6, i may be equal to the lag order. Therefore, in the example of FIG. 6, the lag corresponding to i is represented as lag[i]. Also, the AC corresponding to i is represented as AC[i]. As is clear from the above description, lag[i] for i>1 represents the i-th shortest lag. In this specification, unless otherwise contradicted, the minimum value of the element number of the data structure is 0 according to the notation of a general programming language. Although not shown in FIG. 6, lag[0]=0 and AC[0]=1, as will be clear to those skilled in the art.

[0031] The preprocessing unit 12 selects (extracts) a special lag from among the multiple lags that have been set based on the AC. As an example, the preprocessing unit 12 may select, as the special lag, a lag that has a value equal to or greater than a time length threshold and corresponds to an AC that has a value equal to or greater than an autocorrelation threshold among the multiple lags.

[0032] Generally, in the field of machine learning, when the correlation coefficient between two pieces of data is 0.7 or more, it can be considered that there is a strong positive correlation between the pieces of data. Therefore, as an example, the autocorrelation threshold may be 0.7 to 1. The autocorrelation threshold may be a value that can be arbitrarily set by the user. In this specification, a case where the autocorrelation threshold is 0.7 is exemplified.

[0033] Fluctuations in the amount of power demand in a facility are generally considered to have a cycle of about one day in the short term. Therefore, the time length threshold may be set to a value that is considered appropriate for grasping the relatively short-term fluctuation trend of the amount of power demand. The time length threshold may also be a value that can be arbitrarily set by the user. In this specification, a case where the time length threshold is one day is illustrated as an example.

[0034] As will be clear to those skilled in the art, when the lag is short, an AC close to 1 is likely to occur. Therefore, by setting the time length threshold as described above, it is possible to prevent many short lags from being selected as special lags. In other words, it is possible to reduce the number of special lags selected (see the rows of lag orders 1 to 3 in the example of FIG. 6). As a result, as will be understood from the explanation below, it is possible to improve the quality of the prediction model.

[0035] As mentioned above, the time length threshold of 1 day corresponds to i=48. In the example of FIG. 6, in the range of i≧48, AC

[48] =0.78 AC

[0336] =0.73 AC

[0672] =0.70 AC

[1008] =0.72 has a value equal to or greater than the autocorrelation threshold. Therefore, the preprocessing unit 12 lag

[48] =1 lag

[0336] =7 lag

[0672] =14 lag

[1008] =21 is chosen as the special lag.

[0036] The inventors used performance data obtained from a certain factory to derive the relationship between lag and AC in that factory. FIG. 7 is a graph showing the relationship between lag and AC in that factory derived by the inventors. In the graph of FIG. 7, the horizontal axis indicates the lag degree, and the vertical axis indicates AC. The horizontal axis in the graph of FIG. 7 can be read as lag. The trend of the graph of FIG. 7 roughly coincides with the example of FIG. 6.

[0037] Next, the pre-processing unit 12 may sort the extracted multiple special lugs in ascending order. Then, the pre-processing unit 12 may generate an array Plag including the sorted special lugs as elements. In the example of the reference embodiment, the pre-processing unit 12 generates Plag=[1, 7, 14, 21]. Then, the pre-processing unit 12 selects elements that form an arithmetic progression (arithmetic progression elements) from among the elements of Plag. The method of selecting the arithmetic progression elements will be described later.

[0038] In the example of the reference embodiment, the preprocessing unit 12 selects the first and subsequent elements of Plag, Plag[1:]=[7,14,21], as arithmetic progression elements. These arithmetic progression elements form an arithmetic progression whose first term has a value of 7 and whose common difference is 7.

[0039] Then, the preprocessing unit 12 selects elements of Plag excluding the arithmetic progression elements as the remaining elements. The remaining elements may be called non-arithmetic progression elements (elements that do not form an arithmetic progression). In the example of the reference embodiment, the preprocessing unit 12 selects the 0th element, Plag[0]=[1], as the remaining element of Plag.

[0040] The pre-processing unit 12 derives a new feature by setting the special lag selected as described above to the power demand. For example, the pre-processing unit 12 may derive at least one of a rolling feature and a lag feature as the new feature.

[0041] Considering the above-mentioned FIG. 7, it is considered that the special lag corresponding to the arithmetic progression element is suitable for expressing the relatively long-term periodicity of the fluctuation in the demand energy. Therefore, the pre-processing unit 12 may select the special lag corresponding to the arithmetic progression element as the rolling lag, which is the special lag for deriving the rolling feature. In the example of the reference embodiment, the pre-processing unit 12 selects the lags "7, 14, and 21" as the rolling lags. The pre-processing unit 12 derives the rolling feature based on the rolling lags.

[0042] Then, the pre-processing unit 12 may select the special lag corresponding to the remaining element as a non-rolling lag, which is a special lag for deriving a feature other than the rolling feature. In the example of the reference embodiment, the pre-processing unit 12 selects the lag "1" as the non-rolling lag. The non-rolling lag is considered to be suitable for expressing the relatively short-term periodicity of the fluctuation in the power demand. Therefore, the pre-processing unit 12 derives the lag feature based on the non-rolling lag.

[0043] In the reference embodiment, a case where a lag feature is derived prior to the derivation of a rolling feature is illustrated. However, as will be apparent to those skilled in the art, a rolling feature may be derived prior to the derivation of a lag feature. The processes in the examples of this specification may be performed in any order unless otherwise specified.

[0044] The preprocessing unit 12 derives a lag feature as a new feature corresponding to the demand power amount by setting a non-rolling lag for the demand power amount. DZ2 in FIG. 8 shows an example of data generated by adding a lag feature to DZ1. In the example of the reference form, since the non-rolling lag is 1, the preprocessing unit 12 derives the "demand power amount one day ago" as the lag feature. Therefore, in the example of FIG. 8, the "demand power amount one day ago" belonging to the row of the time point "2021 / 4 / 1 0:00" represents the "demand power amount at 2021 / 3 / 31 0:00". In addition, the "demand power amount one day ago" belonging to the row of the time point "2022 / 3 / 31 23:59" represents the "demand power amount at 2022 / 3 / 30 23:59".

[0045] Next, the preprocessing unit 12 derives a rolling feature as a new feature corresponding to the demand power amount by setting a non-rolling lag for the demand power amount. First, the preprocessing unit 12 acquires the demand power amount at a time point that is the rolling lag before a certain time point as a constituent element (rolling feature component) of the rolling feature corresponding to the time point. In the example of the reference embodiment, since the rolling lags are 7, 14, and 21, the preprocessing unit 12 acquires the "demand power amount 7 days ago", the "demand power amount 14 days ago", and the "demand power amount 21 days ago" as the rolling feature components. The preprocessing unit 12 adds the rolling feature components to the DZ2.

[0046] Next, the preprocessing unit 12 derives a statistical value of the rolling feature component as a rolling feature. In this specification, a case where the statistical value is an average value is illustrated. Therefore, the preprocessing unit 12 derives an average value of "demand energy amount 7 days ago", "demand energy amount 14 days ago", and "demand energy amount 21 days ago". The preprocessing unit 12 further adds the average value to the DZ2 as a rolling feature.

[0047] DZ3 in FIG. 9 shows an example of data generated by adding a rolling feature component and a rolling feature to DZ2. In the example of FIG. 9, the "demand energy amount 7 days ago", "demand energy amount 14 days ago", and "demand energy amount 21 days ago" belonging to the row of the time point "2021 / 4 / 1 0:00" represent "demand energy amount at 2021 / 3 / 25 0:00", "demand energy amount at 2021 / 3 / 18 0:00", and "demand energy amount at 2021 / 3 / 11 0:00", respectively. And the "average (7, 14, 21 days ago) demand energy amount" belonging to the row of the time point "2021 / 4 / 1 0:00" represents the average value of "demand energy amount 7 days ago", "demand energy amount 14 days ago", and "demand energy amount 21 days ago" belonging to the row of the same time point.

[0048] After the rolling features are acquired, the rolling feature components are considered to be no longer necessary. Therefore, the preprocessing unit 12 may delete the rolling feature components from DZ3 after the rolling features are acquired. DZ4 in FIG. 10 shows an example of data generated by deleting the rolling feature components from DZ3.

[0049] Next, the preprocessing unit 12 derives a time equivalent, which is a feature corresponding to the time, by converting the time recorded in DZ4 into a predetermined format. Then, the preprocessing unit 12 adds the time equivalent to DZ4. As an example, the preprocessing unit 12 may set the time equivalent as an amount representing the elapsed time from the reference time "0:00" to the target time in hours (h). DZ5 in FIG. 11 shows an example of data generated by adding a time equivalent to DZ4.

[0050] In the example of Fig. 11, one minute of elapsed time from "0:00" corresponds to a time equivalent of 1 / 60 (≈0.01667). In the example of the reference embodiment, the time equivalent is derived using only the hours (h) and minutes (m) without considering the year (y), month (m), and day (d) of the time. However, as will be clear to those skilled in the art, an amount derived in a form that takes into account at least one of the year, month, and day may be derived as the time equivalent.

[0051] However, as will be apparent to those skilled in the art, the time equivalent does not necessarily have to be derived, because the prediction model according to one aspect of the present invention only needs to be configured to output a predicted value of the energy demand using a new feature derived by setting a special lag for the energy demand as an explanatory variable.

[0052] Next, the preprocessing unit 12 determines the day of the week corresponding to the date recorded in DZ5, for example by referring to calendar data. Then, the preprocessing unit 12 adds the day of the week to DZ5. DZ6 in FIG. 12 is an example of data generated by adding a day of the week to DZ5. In the example of FIG. 12, the preprocessing unit 12 generates DZ6 by adding the day of the week converted into a qualitative variable (nominal scale) to DZ5. In the example of this specification, days of the week 0 to 6 represent Sunday to Saturday in that order. Therefore, day of the week 4 in the example of FIG. 12 represents Thursday.

[0053] As shown in Fig. 12, in DZ6, the demand energy and the new feature derived based on the demand energy may have a larger value than, for example, the feature indicating the day of the week and the working state (calendar feature). Therefore, the preprocessing unit 12 may perform scaling on each feature included in DZ6. This makes it possible to align the numerical range of each feature to the same extent, thereby making it possible to generate a prediction model with higher performance.

[0054] In this specification, a case where the pre-processing unit 12 performs standardization as scaling is exemplified. The pre-processing unit 12 derives the average value (μ) and standard deviation (σ) of a certain feature included in DZ6. Then, the pre-processing unit 12 standardizes the feature using the derived μ and σ. In this specification, the μ and σ used for standardization before generating the learning model are referred to as standardization parameters. DZ6S in FIG. 13 is an example of data generated by applying standardization to each feature of DZ6. In the reference example, DZ6 is used as pre-processed training data.

[0055] The learning unit 13 generates a prediction model by machine learning based on the time-series data of the demanded energy after the new feature amount derived based on the demanded energy is added. Therefore, as an example, the learning unit 13 may generate a prediction model based on the preprocessed training data (e.g., DZ6S) by executing a predetermined machine learning algorithm.

[0056] The prediction model according to one embodiment of the present invention may be a trained model capable of outputting a predicted value of the power demand at the time of prediction as a target variable. Therefore, the type of machine learning algorithm is not particularly limited as long as it is capable of solving a regression task. Examples of the machine learning algorithm include a neural network (NN), a support vector machine (SVM), and a decision tree (DT).

[0057] In an example of the reference form, the learning unit 13 uses (i) the demand energy amount at any prediction time included in DZ6S as the true value (correct data) of the objective variable, and (ii) each feature other than the demand energy amount prior to the prediction time as an explanatory variable, thereby generating a prediction model that derives the objective variable from the explanatory variables.

[0058] After generating the prediction model, the learning unit 13 may incorporate standardized parameters of each feature into the prediction model. The learning unit 13 may further incorporate a rolling lag and a non-rolling lag into the prediction model.

[0059] The evaluation unit 14 evaluates the prediction model generated by the learning unit 13. Specifically, the evaluation unit 14 evaluates the prediction performance of the prediction model. Prior to the evaluation by the evaluation unit 14, the preprocessing unit 12 generates preprocessed verification data by performing a series of preprocessing in the same format as the above-mentioned training data on the verification data. Therefore, in the example of the reference form, the preprocessed verification data has a data structure in the same format as DZ6S. Therefore, in the preprocessed verification data in the reference form, each feature is standardized using the standardization parameters incorporated in the prediction model.

[0060] The evaluation unit 14 evaluates the prediction model using the preprocessed verification data. Specifically, the evaluation unit 14 inputs the preprocessed verification data to the prediction model, causing the prediction model to output a predicted value. Then, the evaluation unit 14 derives an index value indicating the prediction accuracy of the prediction model based on the predicted value and the true value of the energy demand indicated in the preprocessed verification data.

[0061] The index value is not particularly limited as long as it is used in the machine learning field to evaluate the prediction accuracy in a regression task. Examples of the index value include Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The evaluation unit 14 may store the prediction model in the storage unit 90 after completing the derivation of the index value.

[0062] In addition, when the index value falls below a predetermined threshold, the model generation device 10 may cause the learning unit 13 to generate the prediction model again. In this case, prior to regeneration of the prediction model, for example, hyperparameters of the prediction model may be reset. In this way, in order to guarantee the performance of the prediction model used in the prediction phase, the generation of the prediction model may be repeated until an index value equal to or greater than the threshold is obtained. In this case, the evaluation unit 14 may store in the storage unit 90 the prediction model for which an index value equal to or greater than the threshold has been obtained for the first time.

[0063] (Example of a method for selecting elements of an arithmetic progression) 14 is a flowchart showing an example of a process flow for selecting arithmetic progression elements. First, the preprocessing unit 12 acquires the number of elements NPlag of the above-mentioned array Plag prior to S1. In the example of the reference embodiment, Plag=[1, 7, 14, 21], so NPlag=4.

[0064] In S1, the preprocessing unit 12 determines whether NPlag is 2 or more. If NPlag is 2 or more (YES in S1), the process proceeds to S2. On the other hand, if NPlag is 1 or less (NO in S1), the preprocessing unit 12 ends the process in Fig. 14. In this case, the preprocessing unit 12 determines that Plag does not include an arithmetic progression element.

[0065] In S2, the preprocessing unit 12 initializes an element number i for indicating an element of Plag to 0. In S3, the preprocessing unit 12 initializes a flag flg indicating that Plag includes an arithmetic progression element to True.

[0066] In S4, the preprocessing unit 12 judges whether i is smaller than NPlag-2. If i is smaller than NPlag-2 (YES in S4), the process proceeds to S5. On the other hand, if i is equal to or larger than NPlag-2 (NO in S4), the process proceeds to S8.

[0067] In S5, the preprocessing unit 12 assigns Plag[i] to a0. a0 is the first term of the assumed arithmetic progression. Thus, in S5, the preprocessing unit 12 sets a0 to a value equal to Plag[i]. In S6, the preprocessing unit 12 assigns Plag[i+1]-a0 to d. d is the common difference of the assumed arithmetic progression. Thus, in S6, the preprocessing unit 12 sets d to a value equal to Plag[i+1]-a0. Next, in S7, the preprocessing unit 12 assigns i+2 to element number j for indicating an element of Plag. Then, the process proceeds to S11.

[0068] S8 to S10 are processes paired with S5 to S7, respectively. In S8, the pre-processing unit 12 assigns 0 to a0. That is, in S8, the pre-processing unit 12 sets a0 to 0. In S9, the pre-processing unit 12 assigns Plag[i]-a0 to d. Thus, in S9, the pre-processing unit 12 sets d to a value equal to Plag[i]-a0. Next, in S10, the pre-processing unit 12 assigns i+1 to j. Then, the process proceeds to S11.

[0069] In S11, the preprocessing unit 12 assigns Plag[j]-Plag[j-1] to next_d. Next_d is the next common difference of d in the assumed arithmetic progression. In this way, the preprocessing unit 12 sets next_d to a value equal to Plag[j]-Plag[j-1].

[0070] In S12, the preprocessing unit 12 determines whether next_d is equal to d. If next_d is equal to d (YES in S12), the preprocessing unit 12 counts up j by 1 in S13. Then, the process proceeds to S15. On the other hand, if next_d is not equal to d (NO in S12), the preprocessing unit 12 updates flg to False in S14. Then, the process proceeds to S16.

[0071] In S15, the preprocessing unit 12 judges whether or not j is equal to or greater than NPlag. If j is equal to or greater than NPlag (YES in S15), the process proceeds to S16. On the other hand, if j is smaller than NPlag (NO in S15), the process returns to S11. Thus, the processes of S11 to S13 and S15 are repeated until j becomes equal to NPlag.

[0072] In S16, the preprocessing unit 12 determines whether or not flg is False. If flg is False (YES in S16), in S17 the preprocessing unit 12 counts up i by 1. Then, the process proceeds to S19. On the other hand, if flg is not False (NO in S16), that is, if flg is True, the process proceeds to S18.

[0073] In S18, the preprocessing unit 12 assigns Plag[i:], which is the i-th and subsequent elements of Plag, to an array ap. The array ap is an array for storing arithmetic progression elements. After completing S18, the preprocessing unit 12 ends the processing of FIG. 14. If flg is True at the end of the processing (in other words, if the array ap is not an empty array at the end of the processing), the preprocessing unit 12 determines that Pla contains arithmetic progression elements. In this case, the preprocessing unit 12 may output ap as a selection result of arithmetic progression elements.

[0074] In S19, the preprocessing unit 12 judges whether i is equal to or greater than NPlag. If i is equal to or greater than NPlag (YES in S19), the process proceeds to S20. On the other hand, if i is smaller than NPlag (NO in S19), the process returns to S3. Thus, the processes of S3 to S16, S17, and S19 are repeated until i becomes equal to NPlag.

[0075] In S20, the preprocessing unit 12 assigns an empty array to the array ap. That is, the preprocessing unit 12 sets the array ap to an empty array. After completing S20, the preprocessing unit 12 ends the processing of FIG. 14. If flg is False at the end of the processing (in other words, if the array ap is an empty array at the end of the processing), the preprocessing unit 12 determines that Plag does not include an arithmetic progression element. In this case, the preprocessing unit 12 may output ap as an empty array as a selection result of an arithmetic progression element.

[0076] In the example of the reference embodiment, after i is set to 0 in S2, a0 is set to Plag[0]=1 in S5. Therefore, d is set to Plag[1]-a0=7-1=6 in S6. Then, j is set to 2 in S7. Therefore, next_d is set to Plag[2]-Plag[1]=14-7=7 in S11. In this case, since d is not equal to next_d, flg is set to False in S14. Then, i is counted up to 1 in S17.

[0077] After flg is updated to True in S3, a0=Plag[1]=7 is set in S5. Therefore, d=Plag[2]-a0=14-7=7 is set in S6. Then, j=3 is set in S7. Therefore, next_d=Plag[3]-Plag[2]=21-14=7 is set in S11. Therefore, since d is equal to next_d and flg is True, Plag[1:]=[7,14,21] is set as ap in S18.

[0078] (Forecasting phase in reference form) Next, referring back to Fig. 1, an example of a prediction phase following the learning phase will be described. The prediction device 20 may include a data acquisition unit 21, a preprocessing unit 22, a prediction calculation unit 23, and a prediction result output unit 24. The data acquisition unit 21 may be referred to as a prediction-time data acquisition unit or a second data acquisition unit. The preprocessing unit 22 may be referred to as a prediction-time preprocessing unit or a second preprocessing unit.

[0079] As shown in FIG. 1, the memory unit 90 may store actual data DZP and working status data DSP used in the prediction phase. In the prediction phase, a demand energy amount prediction may be performed in real time. Therefore, the actual data DZP may be updated in real time. The data acquisition unit 21 acquires the DZP and the DSP as input data in the prediction phase. The DZP has a data structure in the same format as the DZ. The DSP has a data structure in the same format as the DS. In addition, the data acquisition unit 21 acquires the prediction model stored in the memory unit 90 by the model generation device 10 in the learning phase.

[0080] The pre-processing unit 22 performs a series of pre-processing steps on the input data set in the prediction phase in the same manner as in the learning phase described above, thereby generating pre-processed input data. An example of the processing performed by the pre-processing unit 22 is described below.

[0081] First, the preprocessing unit 22 performs pre-value interpolation on the working state in the DSP to generate working state data after pre-value interpolation in the prediction phase (for convenience, referred to as DSP1). As a result of the pre-value interpolation, DSP1 has a time resolution (e.g., 1 minute) equal to that of the DZP.

[0082] Next, the pre-processing unit 22 generates actual data with working status data in the prediction phase (for convenience, referred to as DZP1) by combining DZP and DSP1 using the time point as a key.

[0083] Next, the pre-processing unit 22 obtains the special lag determined in the learning phase. For example, the pre-processing unit 22 may read the rolling lag and the non-rolling lag determined in the learning phase from the prediction model. The pre-processing unit 22 derives new features in the prediction phase by setting the special lag determined in the learning phase to the power demand in the prediction phase.

[0084] The pre-processing unit 22 may derive a lag feature as a new feature in the prediction phase by setting a non-rolling lag for the demand energy in the prediction phase. Then, the pre-processing unit 22 may generate new data (for convenience, referred to as DZP2) by adding the lag feature to DZP1. In the example of the reference embodiment, since the non-rolling lag is 1, the pre-processing unit 22 acquires the "demand energy one day ago" as the lag feature in the prediction phase.

[0085] Next, the preprocessing unit 22 may derive a rolling feature as a new feature in the prediction phase by setting a rolling lag for the demand energy in the prediction phase. First, the preprocessing unit 22 acquires a rolling feature component in the prediction phase according to the rolling lag. In the example of the reference embodiment, since the rolling lag is 7, 14, and 21, the preprocessing unit 22 acquires "demand energy 7 days ago", "demand energy 14 days ago", and "demand energy 21 days ago" in the prediction phase as rolling feature components in the prediction phase. Then, the preprocessing unit 22 adds the rolling feature component to the DZP2.

[0086] Next, the preprocessing unit 22 derives an average value of the "energy demand 7 days ago", the "energy demand 14 days ago", and the "energy demand 21 days ago" in the prediction phase, and acquires the average value as a rolling feature in the prediction phase. Then, the preprocessing unit 22 further adds the rolling feature to the DZP2.

[0087] As described above, the preprocessing unit 22 may generate new data (for convenience, referred to as DZP3) by adding the rolling feature component and the rolling feature in the prediction phase to DZP2. After acquiring the rolling feature, the preprocessing unit 22 may generate new data (for convenience, referred to as DZP4) by deleting the rolling feature component from DZP3.

[0088] Next, the preprocessing unit 22 converts the time points recorded in the DZP4 into a time point equivalent amount in the prediction phase. Then, the preprocessing unit 22 generates new data (for convenience, referred to as DZP5) by adding the time point equivalent amount to the DZ4.

[0089] Next, the pre-processing unit 22 generates new data (for convenience, referred to as DZP6) by adding the day of the week (the day of the week converted into a qualitative variable) corresponding to the date recorded in DZP5 to DZ5.

[0090] Next, the preprocessing unit 22 reads out the standardized parameters determined in the learning phase from the prediction model. Then, the preprocessing unit 22 standardizes each feature of DZP6 using the standardized parameters. DZP6S in FIG. 15 is an example of data generated by standardizing each feature of DZP6. DZP6S is used as preprocessed input data in the prediction phase. In FIG. 15, the key time points are omitted. That is, in FIG. 15, only explanatory variables used to predict the demand energy in the prediction phase are shown.

[0091] As an example, in the forecast phase, a demand energy forecast may be performed every 30 minutes. In addition, in the forecast phase, a forecast value of the demand energy over a future period of, for example, more than 24 hours (e.g., from the present time to 36 hours later) may be derived. For this reason, in the forecast phase, the values ​​of some cells, for example, of the "demand energy one day ago" may be unknown. In the example of FIG. 15, a cell having an unknown value (i.e., a blank value) is indicated by the notation NULL. If a blank value exists in the DZP6S, the preprocessing unit 22 may interpolate the blank value using a predetermined interpolation method.

[0092] As an example, the preprocessing unit 22 may use the prediction model to interpolate a blank value based on a known value of the demand energy amount one day before. For example, the preprocessing unit 22 may input a known value of the demand energy amount one day before corresponding to a blank value into the prediction model, and cause the prediction model to output a predicted value according to the input. Then, the preprocessing unit 22 may set the predicted value as an interpolated value for the blank value.

[0093] The prediction calculation unit 23 causes the prediction model to output a predicted value of the amount of power demand based on the time-series data of the amount of power demand after a new feature value derived based on the amount of power demand in the prediction phase is added. Therefore, the prediction calculation unit 23 may cause the prediction model to output a predicted value by inputting preprocessed input data (e.g., DZP6S) to the prediction model. Specifically, the prediction calculation unit 23 supplies each feature value in the preprocessed input data to the prediction model as an explanatory variable. The prediction model outputs a predicted value according to the explanatory variable as a response variable.

[0094] The prediction result output unit 24 generates prediction result data based on the predicted value of the power demand output by the prediction model. For example, the prediction result output unit 24 may generate data in which each predicted value is associated with each time point as the prediction result data. The prediction result output unit 24 outputs the generated prediction result data.

[0095] Furthermore, the prediction result output unit 24 may convert the predicted value of the energy demand (unit: kWh) into another unit. As an example, the prediction result output unit 24 may convert the predicted value of the energy demand into a predicted value of the power demand (unit: kW). In this case, the prediction result output unit 24 may generate data in which each converted predicted value is associated with each time point as prediction result data.

[0096] (Effects of information processing device 1) According to the model generating device 10, in the learning phase, a special lag can be selected based on AC (autocorrelation coefficient of the power demand). Then, a new feature can be derived by setting the special lag for the power demand. In this way, according to the model generating device 10, the new feature can be derived based on AC, which is a feature expected to represent the periodicity of the fluctuation of the power demand (e.g., the above-mentioned rolling feature and lag feature). Therefore, it is possible to obtain the new feature as an explanatory variable expected to be useful for predicting the power demand.

[0097] The model generation device 10 can generate a prediction model based on time-series data to which new features are added to the power demand (for example, see DZ4 in FIG. 9 above). Therefore, a prediction model can be generated using the new features as explanatory variables, so that the prediction accuracy of the prediction model can be improved. In other words, a high-quality prediction model can be obtained.

[0098] The fluctuation trend of the power demand in a facility may differ depending on, for example, the type of business of the facility. For this reason, it is not necessarily easy for a user to manually set an appropriate lag according to the facility prior to generating a prediction model. Furthermore, if the lag set by the user is inappropriate, the quality of the prediction model may be degraded.

[0099] However, as described above, the model generation device 10 can select a special lag that is expected to be suitable for grasping the periodicity of fluctuations in the amount of power demand based on the AC. Therefore, a prediction model can be generated without the need for the user to manually set the lag. This makes it possible to obtain a high-quality prediction model suited to the type of business of the facility while improving user convenience.

[0100] Furthermore, the model generation device 10 can generate a prediction model using the employment status of the facility as an additional explanatory variable. By using the employment status as an additional explanatory variable, a prediction model that reflects the business condition of the facility in more detail can be generated. Therefore, it is possible to further improve the prediction accuracy of the prediction model.

[0101] Incidentally, Patent Document 1 shows the idea of ​​using multiple NNs in order to improve the accuracy of heat load prediction. Specifically, Patent Document 1 discloses that three separate NNs, an NN for weekdays, an NN for Saturdays, and an NN for Sundays, are generated in advance, and each NN is used depending on the day of the week of the prediction date.

[0102] However, depending on the type of business of the facility, at least one of Saturdays and Sundays may be set as working days. Alternatively, depending on the type of business of the facility, a specific weekday (e.g. Monday) may be set as a holiday. Alternatively, depending on the type of business of the facility, all of weekdays, Saturdays, and Sundays may be working days.

[0103] Therefore, the use of each NN shown in Patent Document 1 cannot flexibly deal with the various types of facilities described above. For this reason, when the use of each NN shown in Patent Document 1 is applied to power demand prediction, it is not necessarily possible to achieve high prediction accuracy.

[0104] On the other hand, as described above, the model generation device 10 can obtain a high-quality prediction model that is suited to the business condition of the facility. Therefore, unlike the technology of Patent Document 1, the model generation device 10 can generate a single prediction model that can achieve high prediction accuracy. In this way, the model generation device 10 can obtain a prediction model that is more versatile than the conventional ones.

[0105] Then, the prediction device 20 can derive a predicted value of the energy demand in the prediction phase by using the prediction model generated by the model generation device 10 in the learning phase. As described above, according to one aspect of the present invention, it is possible to improve the prediction accuracy of the energy demand in a facility compared to the conventional method.

[0106] [Modifications of the Reference Embodiment] (1) In the above example of the reference embodiment, a method is illustrated in which a lag having a value equal to or greater than a time length threshold and corresponding to an AC having a value equal to or greater than an autocorrelation threshold is selected as a special lag from among a plurality of lags set in the learning phase. However, as will be apparent to those skilled in the art, the method of selecting a special lag in the reference embodiment is not limited to the above example.

[0107] For example, the preprocessing unit 12 may select, from among a plurality of lags, a lag corresponding to a maximum value of AC in a data sequence indicating the relationship between the lag and AC as a special lag candidate. Thus, for example, the preprocessing unit 12 may select, as a special lag candidate, a lag corresponding to a maximum value of AC in the graph of Fig. 7. Next, the preprocessing unit 12 may select, from among the special lag candidates, a lag having a value equal to or greater than the time length threshold and corresponding to an AC having a value equal to or greater than the autocorrelation threshold as a special lag.

[0108] Depending on the type of business of the facility, there may be cases where there are many lags corresponding to ACs having values ​​equal to or greater than the autocorrelation threshold value in the vicinity of the lag corresponding to the maximum value of AC. In such cases, if the selection method in the above example in the reference embodiment is adopted, many special lags may be selected. In this case, many new explanatory variables (e.g., many rolling features and lag features) are derived. As is clear to those skilled in the art, if the number of explanatory variables is excessive, the quality of the prediction model may decrease. In addition, an excessive number of explanatory variables increases the time required to execute the machine learning algorithm.

[0109] On the other hand, in the above case, when the selection method of this modified example is adopted, a special lag is selected from among the special lag candidates, unlike the selection method of the above example of the reference form. Therefore, according to the selection method of this modified example, the number of selected special lags can be reduced compared to the selection method of the above example of the reference form.

[0110] Therefore, the selection method of this modification can reduce the number of new explanatory variables derived compared to the selection method of the above example in the reference embodiment. As a result, the selection method of this modification can generate a high-quality prediction model even in the above case. In addition, the time required to execute the machine learning algorithm can be reduced.

[0111] (2) As will be apparent to those skilled in the art, explanatory variables in the prediction model are not limited to the above examples. Any feature that is considered to affect the amount of power demand in a facility may be used as an explanatory variable. For example, depending on the type of business of the facility, meteorological conditions such as temperature, weather, and solar radiation affect the amount of power demand. Therefore, when there is a correlation between the meteorological conditions and the amount of power demand, a feature indicating the meteorological conditions (weather feature) can also be used as an additional explanatory variable.

[0112] The meteorological feature may be an actual value obtained as a result of measurement at a facility, or a predicted value provided by a weather forecast service provider. Thus, for example, in the learning phase, a prediction model may be generated using the actual value of the meteorological feature. Then, in the prediction phase, the predicted value of the meteorological feature may be used to predict the power demand by the prediction model. In this case, in the prediction phase, the predicted value of the meteorological feature may be corrected using the actual value of the meteorological feature in the learning phase.

[0113] (3) The preprocessing unit 12 calculates the partial autocorrelation coefficient (PCC) of the demand energy in DZ1. In this specification, the autocorrelation coefficient is abbreviated as PAC. Specifically, the preprocessing unit 12 may calculate the PAC for each lag by varying the lag.

[0114] In many cases, it is expected that there is some relationship between the PAC and the AC. Therefore, the preprocessing unit 12 may select a special lag based on the PAC. In this manner, the preprocessing unit 12 may derive new features (e.g., lag features and rolling features) based on the PAC.

[0115] In addition, it is expected that the use of both AC and PAC can take into account the periodicity of the time fluctuation of the demanded energy in more detail than the use of only AC or PAC. Therefore, the preprocessing unit 12 may select a special lag based on AC and PAC. In this way, the preprocessing unit 12 can also derive new features based on AC and PAC.

[0116] [Summary of reference format] Each item described in the reference embodiment can be expressed as follows.

[0117] A model generation device according to a reference embodiment is a model generation device that generates a prediction model for predicting an amount of energy demand in a facility, in which the model generation device sets multiple lags for time series data indicating a time progression of an actual value of the amount of energy demand, derives autocorrelation coefficients of the actual value corresponding to each of the multiple lags, selects a special lag from among the multiple lags based on the autocorrelation coefficients, derives a new feature by setting the special lag for the actual value, and generates the prediction model by machine learning based on the time series data to which the new feature has been added.

[0118] In the model generating device of the reference embodiment, in the above aspect 1, a lag having a value equal to or greater than a time length threshold and corresponding to the autocorrelation coefficient having a value equal to or greater than an autocorrelation threshold may be selected as the special lag from among the plurality of lags.

[0119] A model generating device according to a reference embodiment may select, from among the plurality of lags, a lag corresponding to a maximum value of the autocorrelation coefficient in a data series indicating the relationship between the lags and the autocorrelation coefficient as a special lag candidate, and may select, from among the special lag candidates, a lag having a value equal to or greater than a time length threshold and corresponding to the autocorrelation coefficient having a value equal to or greater than an autocorrelation threshold as a special lag.

[0120] In the model generating device according to the reference embodiment, the autocorrelation threshold may be a value not less than 0.7 and not more than 1.

[0121] A model generation device according to a reference embodiment may select, as a rolling lag, a plurality of special lags forming an arithmetic progression in an array in which the plurality of special lags are sorted in ascending order, and may derive a rolling feature as the new feature by setting the rolling lag to the actual value.

[0122] The model generation device according to the reference embodiment may select the special lags in the array excluding the rolling lag as non-rolling lags, and may derive a lag feature as the new feature by setting the non-rolling lag to the actual value.

[0123] In the model generating device according to the reference embodiment, the time-series data to which the new feature has been added may further include data indicating daily working conditions in the facility.

[0124] A prediction device according to a reference embodiment is a prediction device that derives a predicted value of an amount of demand energy by using a prediction model for predicting an amount of demand energy in a facility, the prediction model having been generated in advance in a learning phase, in a prediction phase. In the learning phase, for time series data in the learning phase showing a time progression of an actual value of the amount of demand energy, a plurality of lags are set for the actual value, thereby deriving an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags, a special lag is selected from among the plurality of lags based on the autocorrelation coefficients, a new feature is derived by setting the special lag for the actual value, and the prediction model is generated by machine learning based on the time series data to which the new feature has been added. The prediction device derives a new feature for the prediction phase by setting the special lag selected in the learning phase for the time series data in the prediction phase showing a time progression of the actual value, and outputs the predicted value to the prediction model based on the time series data in the prediction phase to which the new feature has been added.

[0125] A model generation method according to a reference embodiment is a model generation method for generating a predictive model for predicting the amount of electricity demand in a facility, and includes the steps of: setting multiple lags for actual values ​​of time series data indicating the time progression of actual values ​​of the amount of electricity demand in the facility, and deriving an autocorrelation coefficient of the actual values ​​for each of the multiple lags; selecting a special lag from among the multiple lags based on the autocorrelation coefficients; deriving a new feature by setting the special lag for the actual value; and generating the predictive model by machine learning based on the time series data to which the new feature has been added.

[0126] A prediction method according to a reference embodiment is a prediction method for deriving a predicted value of an amount of energy demand in a facility by using, in a prediction phase, a prediction model for predicting an amount of energy demand in the facility, which has been generated in advance in a learning phase. In the learning phase, a plurality of lags are set for time-series data in the learning phase that indicates a time transition of an actual value of the amount of energy demand in the facility, and an autocorrelation coefficient of the actual value is derived for each of the plurality of lags. A special lag is selected from among the plurality of lags based on the autocorrelation coefficient. A new characteristic is derived by setting the special lag for the actual value in the learning phase. a feature quantity is derived, and the prediction model is generated by machine learning based on the time series data in the learning phase to which the new feature quantity in the learning phase has been added, and the prediction method includes: a step of deriving a new feature quantity in the prediction phase by setting the special lag selected in the learning phase to the actual value in the prediction phase for the time series data in the prediction phase that indicates a time progression of an actual value of the energy demand in the facility; and a step of causing the prediction model to output the predicted value based on the time series data in the prediction phase to which the new feature quantity in the prediction phase has been added.

[0127] [Embodiment 1] FIG. 16 is a block diagram illustrating the configuration of the main parts of an information processing device 1A in the first embodiment. In the first embodiment, differences from the reference embodiment will be described. The information processing device 1A in the example of FIG. 16 includes a control device 9A instead of the control device 9. The control device 9A includes a model generation device 10A and a prediction device 20A instead of the model generation device 10 and the prediction device 20. The model generation device 10A includes a pre-processing unit 12A instead of the pre-processing unit 12. The prediction device 20A includes a pre-processing unit 22A instead of the pre-processing unit 22.

[0128] (Learning Phase in the First Embodiment) In the reference embodiment, the case where the model generation device 10 selects a special lag based on AC has been exemplified. The model generation device 10A of the first embodiment selects a special lag based on AC and the working status data DS. That is, the model generation device 10A selects a special lag while also taking into account the working status of the facility.

[0129] In the first embodiment, a case where the special lag includes a rolling lag and a non-rolling lag is also illustrated. In the first embodiment, the "relatively long-term periodicity" of the time fluctuation of the energy demand described in the reference embodiment is referred to as a "first periodicity." As is clear from the description of the reference embodiment, the rolling lag corresponds to the first periodicity.

[0130] On the other hand, in the first embodiment, the "relatively short-term periodicity" of the time fluctuation of the energy demand described in the reference embodiment is referred to as the "second periodicity." Therefore, the second periodicity is a periodicity that is shorter than the first periodicity. As is clear from the description of the reference embodiment, the non-rolling lag corresponds to the second periodicity.

[0131] In the reference embodiment, an example is shown in which the pre-processing unit 12 of the model generating device 10 selects, as a special lag, a lag among a plurality of lags that has a value equal to or greater than a time length threshold and corresponds to an AC that has a value equal to or greater than an autocorrelation threshold.

[0132] In the first embodiment, the preprocessing unit 12A may select, as a special lag candidate, a lag having a value equal to or greater than a time length threshold and corresponding to the autocorrelation coefficient having a value equal to or greater than an autocorrelation threshold, among a plurality of lags. In this manner, the preprocessing unit 12A may select, as a special lag candidate, an element that was selected as a special lag in the reference embodiment.

[0133] The pre-processing unit 12A may select rolling lag candidates and non-rolling lag candidates from among the special lag candidates by performing the same processing as in the reference embodiment. For example, the pre-processing unit 12A may sort the extracted multiple special lag candidates in ascending order. Then, the pre-processing unit 12A may generate an array Plag_A that includes the sorted special lag candidates as elements. In the first embodiment, for convenience of explanation, a case where Plag_A=[1, 7, 14] is illustrated.

[0134] The preprocessing unit 12A may select elements (arithmetic progression elements) that form an arithmetic progression among the elements of Plag_A as rolling lag candidates. That is, the preprocessing unit 12A may select multiple special lag candidates that form an arithmetic progression in Plag_A as rolling lag candidates. In the example of the first embodiment, the preprocessing unit 12A selects lags "7" and "14" as rolling lag candidates.

[0135] In this specification, a rolling lag corresponding to a facility's workdays is referred to as a workday rolling lag, and a rolling lag corresponding to a facility's holidays is referred to as a holiday rolling lag. Furthermore, a rolling feature corresponding to a workday rolling lag is referred to as a workday rolling feature, and a rolling feature corresponding to a holiday rolling lag is referred to as a holiday rolling feature.

[0136] The preprocessing unit 12A may derive the workday rolling lag and the holiday rolling lag from the rolling lag candidates based on the working status data DS. An example of the process of deriving the workday rolling lag and the holiday rolling lag from the rolling lag candidates will be described later.

[0137] The preprocessing unit 12A may derive a workday rolling feature by setting a workday rolling lag to the actual value of the demand energy. The preprocessing unit 12A may derive a holiday rolling feature by setting a holiday rolling lag to the actual value of the demand energy. The workday rolling feature and the holiday rolling feature are examples of new features in the first embodiment.

[0138] The pre-processing unit 12A may select the elements of Plag_A excluding the arithmetic progression elements as the remaining elements. Then, the pre-processing unit 12A may select the special lag candidates corresponding to the remaining elements as the non-rolling lag candidates. In this manner, the pre-processing unit 12A may select the special lag candidates excluding the rolling lag candidates in Plag_A as the non-rolling lag candidates. In the example of embodiment 1, the pre-processing unit 12A selects lag "1" as the non-rolling lag candidate.

[0139] In this specification, a non-rolling lag corresponding to a facility's workdays is referred to as a workday non-rolling lag, and a non-rolling lag corresponding to a facility's holidays is referred to as a holiday non-rolling lag. Furthermore, a lag feature corresponding to a workday non-rolling lag is referred to as a workday lag feature, and a lag feature corresponding to a holiday non-rolling lag is referred to as a holiday lag feature.

[0140] The preprocessing unit 12A may derive the workday non-rolling lag and the holiday non-rolling lag from the non-rolling lag candidates based on the working status data DS. An example of the process of deriving the workday non-rolling lag and the holiday non-rolling lag from the non-rolling lag candidates will be described later.

[0141] Then, the preprocessing unit 12A may derive the workday lag feature by setting the workday non-rolling lag to the actual value of the demand energy. The preprocessing unit 12A may derive the holiday lag feature by setting the holiday non-rolling lag to the actual value of the demand energy. The workday lag feature and the holiday lag feature are also examples of new features in the first embodiment.

[0142] (Example of process for deriving workday non-rolling lag and holiday non-rolling lag) An example of a process for deriving a workday non-rolling lag and a holiday non-rolling lag will be described with reference to Figs. 17 to 19. In this example, a rolling lag is set for each day from the 17th to the 31st of a certain month based on the working status data DS. As shown in item No. 0 of Fig. 17, in this example, the "17th", "19th", "24th", "30th", and "31st" are holidays, and the other days are working days. The notation "org" in Fig. 17 is an abbreviation of "orignal". No. 0 may be derived from the working status data DS.

[0143] An example of a process for deriving a workday non-rolling lag with a focus on workdays is shown in Fig. 17. The process flow in Fig. 17 is as follows: Process 1A to Process 6A.

[0144] Process 1A: The preprocessing unit 12A converts each day in item No. 0 into a qualitative variable and derives item No. 1. In the example of Fig. 17, since the focus is on workdays, a logical value of 1 is assigned to "workdays" and a logical value of 0 is assigned to "holidays."

[0145] Process 2A: The preprocessing unit 12A derives item No. 2 by shifting each value in item No. 1 back by one day. The notation "Shift1" in FIG. 17 represents a process of shifting each value in a certain item back by one day. This shift value corresponds to the lag. For this reason, the shift value for deriving item No. 2 is set as a value equal to the non-rolling lag candidate. Note that the logical value of "17th" in item No. 2 is equal to the logical value of "16th" in item No. 1 (not shown in FIG. 17).

[0146] Process 3A: The preprocessing unit 12A judges whether or not there is a workday to which a logical value of 0 is assigned among the current items. If there is no workday to which a logical value of 0 is assigned, the process proceeds to process 6A described below. On the other hand, if there is a workday to which a logical value of 0 is assigned, the process proceeds to process 4A described below.

[0147] Process 4A: If there is a workday to which a logical value of 0 is assigned, the preprocessing unit 12A extracts that workday. In the example of FIG. 17, the preprocessing unit 12A extracts "the 18th, 20th, and 25th" in No. 2. As shown in FIG. 17, a workday to which a logical value of 0 is assigned may be the day after a holiday. It may not be appropriate to apply a lag of "1" to a workday that is the day after a holiday.

[0148] Next, the preprocessing unit 12A increases the current shift value by 1. Then, the preprocessing unit 12A derives a new item by shifting each value in the current item using the increased shift value. In the example of FIG. 17, the preprocessing unit 12A increases the shift value from 1 to 2. Then, the preprocessing unit 12A derives item No. 3 by shifting each value in item No. 1 back by two days. Shifting each value in item No. 1 back by two days is equivalent to shifting each value in item No. 2 back by one day.

[0149] Process 5A: Next, the preprocessing unit 12A derives a new item by calculating the logical sum (OR) of each value in the latest item in which workdays to which a logical value 0 is assigned have been extracted and each value in the latest post-shift item for each day. For example, the preprocessing unit 12A derives item No. 4 by calculating the logical sum of each value in item No. 2 and each value in item No. 3 for each day. The logical value of a certain day in item No. 4 is set as the logical sum of the logical value of that day in item No. 2 and the logical value of that day in item No. 3. By calculating the logical sum, the number of workdays to which a logical value 0 is assigned can be reduced.

[0150] After the process 5A is completed, the process returns to the process 3A. According to the above flow, the processes 3A to 5A are repeated until there are no more work days to which the logical value 0 is assigned.

[0151] Process 6A: The preprocessing unit 12A sets the current item as the final item. In the example of FIG. 17, the preprocessing unit 12A sets item No. 6 as the final item. Then, for each workday in the final item, the preprocessing unit 12A determines the immediately preceding workday to be referenced. The determination of the immediately preceding workday to be referenced may be performed by referring to the working status data DS.

[0152] In the example of FIG. 17, the pre-processing unit 12A determines that the immediately preceding workday to be referenced for the workday "18th" is the "15th." That is, the pre-processing unit 12A sets the workday non-rolling lag corresponding to the workday "18th" to "3" (a value obtained by subtracting 15 from 18). On the other hand, for example, the pre-processing unit 12A determines that the immediately preceding workday to be referenced for the workday "29th" is the "28th." That is, the pre-processing unit 12A sets the workday non-rolling lag corresponding to the workday "29th" to "1" (a value obtained by subtracting 28 from 29).

[0153] In this way, the preprocessing unit 12A can set different workday non-rolling lags for the workday that is the day after a holiday (e.g., the 18th) and the workday that is the day after a workday (e.g., the 29th). As described above, by taking the working status data DS into consideration, it is possible to flexibly set the workday non-rolling lag.

[0154] Next, reference will be made to Fig. 18. Fig. 18 shows an example of a process for deriving a holiday non-rolling lag with a focus on holidays. The flow of the process in Fig. 18 is as follows: Process 1B to Process 6B.

[0155] Process 1B: The preprocessing unit 12A converts each day in item No. 0 into a qualitative variable and derives item No. 1. The example in FIG. 18, unlike FIG. 17, focuses on holidays. Therefore, in the example in FIG. 18, unlike FIG. 17, a logical value of 1 is assigned to "holidays" and a logical value of 0 is assigned to "workdays."

[0156] Process 2B: The preprocessing unit 12A shifts each value in the item No. 1 backward by one day to derive the item No. 2. This process is equivalent to the above-mentioned process 2A.

[0157] Process 3B: Preprocessing unit 12A determines whether or not there is a holiday to which a logical value of 0 is assigned among the current items. If there is no holiday to which a logical value of 0 is assigned, the process proceeds to process 6B described below. On the other hand, if there is a holiday to which a logical value of 0 is assigned, the process proceeds to process 4B described below.

[0158] Process 4B: If there is a holiday to which a logical value of 0 is assigned, the preprocessing unit 12A extracts that holiday. In the example of FIG. 18, the preprocessing unit 12A extracts "the 19th, 24th, and 30th" in No. 2. As shown in FIG. 18, holidays to which a logical value of 0 is assigned may be the day after a workday. It may not be appropriate to apply a lag of "1" to holidays that are located the day after a workday.

[0159] Next, the preprocessing unit 12A increments the current shift value by 1. Then, the preprocessing unit 12A derives a new item by shifting each value in the current item using the incremented shift value. This process is equivalent to the above-mentioned process 4A.

[0160] Process 5B: Next, the preprocessing unit 12A derives a new item by calculating the logical sum of each value in the latest item from which holidays assigned a logical value of 0 have been extracted for each day and each value in the latest post-shift item.

[0161] After the process 5B is completed, the process returns to the process 3B. According to the above flow, the processes 3B to 5B are repeated until there are no more holidays to which the logical value 0 is assigned.

[0162] Process 6B: The preprocessing unit 12A sets the current item as the last item. In the example of FIG. 18, the preprocessing unit 12A sets item No. 12 as the last item. Then, the preprocessing unit 12A determines the immediately preceding holiday to be referenced for each workday in the last item. The determination of the immediately preceding holiday to be referenced may be performed by referring to the working status data DS.

[0163] In the example of FIG. 18, the pre-processing unit 12A determines that for the holiday "30th," the immediately preceding holiday to be referenced is "24th." That is, the pre-processing unit 12A sets the holiday non-rolling lag corresponding to the holiday "30th" to "6" (a value obtained by subtracting 24 from 30). On the other hand, for example, the pre-processing unit 12A determines that for the holiday "31st," the immediately preceding holiday to be referenced is "30th." That is, the pre-processing unit 12A sets the holiday non-rolling lag corresponding to the holiday "31st" to "1" (a value obtained by subtracting 30 from 31).

[0164] According to the pre-processing unit 12A, for example, different holiday non-rolling lags can be set for the holiday that falls on the day after a workday (e.g. the 30th) and the holiday that falls on the day after a holiday (e.g. the 31st). As described above, by taking into account the working status data DS, it is also possible to flexibly set the holiday non-rolling lag. As shown in Fig. 18, the arrangement of holidays can be more irregular than the arrangement of workdays, so being able to flexibly set the holiday non-rolling lag is particularly beneficial.

[0165] Fig. 19 shows a list of non-rolling lags determined by the processes of Fig. 17 and Fig. 18. As is clear from Fig. 19, the pre-processing unit 12A allows the non-rolling lag to be flexibly set for both workdays and holidays.

[0166] (Example of process for deriving workday rolling lag and holiday rolling lag) An example of the process for deriving the workday rolling lag and the holiday rolling lag will be described with reference to Figures 20 to 22. Figures 20 to 22 are pairs of Figures 17 to 19, respectively.

[0167] An example of a process for deriving a workday rolling lag with a focus on workdays is shown in Fig. 20. The process flow in Fig. 20 is as follows: Process 1C to Process 6C.

[0168] Process 1C: The preprocessing unit 12A converts each day in item No. 0 into a qualitative variable to derive item No. 1. This process is equivalent to the above-mentioned process 1A.

[0169] Process 2C: The preprocessing unit 12A derives item No. 2 by shifting each value in item No. 1 back by seven days. The shift value "7" for deriving item No. 2 is set to a value equal to the minimum value of the rolling lag candidates. In other words, the shift value is set to a value equal to the first term of the arithmetic progression formed by the rolling lag candidates. The logical value of "17th" in item No. 2 is equal to the logical value of "16th" in item No. 1 (not shown in FIG. 17).

[0170] Process 3C: The preprocessing unit 12A determines whether or not there is a workday to which the logical value 0 is assigned among the current items. If there is no workday to which the logical value 0 is assigned, the process proceeds to process 6C described below. On the other hand, if there is a workday to which the logical value 0 is assigned, the process proceeds to process 4C described below.

[0171] Process 4C: If a workday to which a logical value of 0 is assigned exists, the preprocessing unit 12A extracts that workday.

[0172] Next, the preprocessing unit 12A increases the current shift value to a value equal to the next largest rolling lag candidate after the previously referenced rolling lag candidate term. For example, the preprocessing unit 12A sets the current shift value to "14". Then, the preprocessing unit 12A derives a new item by shifting each value in the current item using the increased shift value. The preprocessing unit 12A derives item No. 3 by shifting each value in item No. 1 back by 14 days. The process of shifting each value in item No. 1 back by 14 days is equivalent to the process of shifting each value in item No. 2 back by 7 days.

[0173] However, the shift value in the example of FIG. 20 does not exceed the maximum value (e.g., 14) of the workday rolling lag candidate. The maximum value in the example of FIG. 20 corresponds to two weeks. In the example of FIG. 20, when the shift value reaches the maximum value, the pre-processing unit 12A may set the next shift value to 1. Thereafter, the pre-processing unit 12A may increase the current shift value by 1. This also applies to the example of FIG. 21.

[0174] Process 5C: Next, the preprocessing unit 12A derives new items by calculating the logical sum (OR) of each value in the latest item in which the workdays to which the logical value 0 is assigned are extracted and each value in the latest post-shift item for each day. This process is equivalent to the above-mentioned process 5A.

[0175] After the process 5C is completed, the process returns to the process 3C. According to the above flow, the processes 3C to 5C are repeated until there are no more work days to which the logical value 0 is assigned.

[0176] Process 6C: The preprocessing unit 12A sets the current item as the final item. In the example of FIG. 20, the preprocessing unit 12A sets item No. 6 as the final item. Then, for each workday in the final item, the preprocessing unit 12A determines the workdays up to two weeks prior to be referenced. The determination of the workdays to be referenced may be performed by referring to the working status data DS.

[0177] In the example of Fig. 20, the logical value 1 is assigned to the workday "18th" in both item No. 2 corresponding to shift value 7 and item No. 3 corresponding to shift value 14. This means that both 7 and 14 may be set as the workday rolling lag for the workday "18th". Therefore, the pre-processing unit 12A sets the workday rolling lag corresponding to the workday "18th" to "7" and "14".

[0178] Next, according to the determined workday non-rolling lag, the preprocessing unit 12A determines the workdays to be referenced that correspond to a certain workday. In the example of Fig. 20, the preprocessing unit 12A determines that the workdays to be referenced that correspond to the workday "18th" are "4th" (the date 14 days before the 18th) and "11th" (the date 7 days before the 18th).

[0179] On the other hand, in the example of FIG. 20, the logical value 1 is assigned to the workday "26th" only in item No. 3 corresponding to shift value 14. This means that only "14" may be set as the workday rolling lag for the workday "26th". Therefore, the pre-processing unit 12A sets the workday rolling lag corresponding to the workday "26th" to "14". Then, the pre-processing unit 12A determines that the workday to be referred to corresponding to the workday "26th" is the "12th" (the date 14 days before the 26th). As described above, by taking into account the working status data DS, it is also possible to flexibly set the workday rolling lag.

[0180] Next, reference will be made to Fig. 21. Fig. 21 shows an example of a process for deriving a holiday non-rolling lag with a focus on holidays. The flow of the process in Fig. 21 is as follows: Process 1D to Process 6D.

[0181] Process 1D: The preprocessing unit 12A converts each day in item No. 0 into a qualitative variable to derive item No. 1. This process is equivalent to the above-mentioned process 1B.

[0182] Process 2D: The preprocessing unit 12A shifts each value in item No. 1 back by 7 days to derive item No. 2. This process is equivalent to process 2C described above.

[0183] Process 3D: The preprocessing unit 12A determines whether or not there is a holiday to which a logical value of 0 is assigned among the current items. If there is no holiday to which a logical value of 0 is assigned, the process proceeds to process 6D described below. On the other hand, if there is a holiday to which a logical value of 0 is assigned, the process proceeds to process 4D described below.

[0184] Process 4D: If there is a holiday to which a logical value of 0 is assigned, the preprocessing unit 12A extracts that holiday. The other processes are the same as those in process 4C described above.

[0185] Process 5D: Next, the preprocessing unit 12A calculates the logical sum of each value in the latest item in which holidays assigned with a logical value 0 are extracted for each day and each value in the latest post-shift item, thereby deriving a new item. This process is equivalent to the above-mentioned process 5B.

[0186] After the process 5D is completed, the process returns to the process 3D. According to the above flow, the processes 3D to 5D are repeated until there are no more holidays to which a logical value 0 is assigned.

[0187] Process 6D: The preprocessing unit 12A sets the current item as the final item. In the example of FIG. 21, the preprocessing unit 12A sets item No. 8 as the final item. Then, for each workday in the final item, the preprocessing unit 12A determines holidays up to two weeks prior to be referenced. The determination of holidays to be referenced may be performed by referring to the working status data DS.

[0188] In the example of FIG. 21, the logical value 1 is assigned to the holiday "30th" only in item No. 3, which corresponds to a shift value of 14. This means that only 14 may be set as the holiday rolling lag for the holiday "30th". Therefore, pre-processing unit 12A sets the holiday rolling lag corresponding to the holiday "30th" to 14. Then, pre-processing unit 12A determines that the holiday to be referenced that corresponds to the holiday "30th" is "16th" (the date 14 days before the 30th).

[0189] On the other hand, in the example of FIG. 21, the holiday "31st" is assigned a logical value of 1 in both item No. 2 corresponding to shift value 7 and item No. 3 corresponding to shift value 14. This means that both 7 and 14 may be set as the holiday rolling lag for the holiday "31st". Therefore, the pre-processing unit 12A sets the holiday rolling lag corresponding to the holiday "31st" to 7 and 14. The pre-processing unit 12A then determines that the holidays to be referenced corresponding to the holiday "31st" are the "17th" (the date 14 days before the 31st) and the "24th" (the date 7 days before the 31st). As described above, by taking into account the working status data DS, it is possible to flexibly set the holiday rolling lag.

[0190] Fig. 22 shows a list of rolling lags determined by the processes of Fig. 20 and Fig. 21. As is clear from Fig. 22, the pre-processing unit 12A allows the rolling lag to be flexibly set for both working days and holidays.

[0191] (Example of deriving new features) As is clear from the above explanations, the pre-processing unit 12A can derive new features corresponding to the working status of the facility based on special lags (e.g., non-rolling lag and rolling lag) derived taking into account the working status of the facility.

[0192] As an example, let us focus on the workday "18th" in the examples of Figs. 17 to 22 described above. As shown in Fig. 19, the workday non-rolling lag corresponding to the workday "18th" is set to "3". Therefore, the pre-processing unit 12A selects "energy demand at the same time on the 15th" as the workday lag feature corresponding to a certain time on the workday "18th".

[0193] 22, the workday rolling lags corresponding to the workday "18th" are set to "7" and "14". Therefore, the preprocessing unit 12A selects "demand energy at the same time on the 4th" and "demand energy at the same time on the 11th" as workday rolling feature components corresponding to a certain time on the workday "18th". Therefore, the preprocessing unit 12A calculates the average value of "demand energy at the same time on the 4th" and "demand energy at the same time on the 11th" as a workday rolling feature corresponding to a certain time on the workday "18th".

[0194] As another example, let us focus on the holiday "30th" in the examples of Figs. 17 to 22 described above. As shown in Fig. 19, the holiday non-rolling lag corresponding to the holiday "30th" is set to "6". Therefore, the pre-processing unit 12A selects "energy demand at the same time on the 24th" as the holiday lag feature corresponding to a certain time on the holiday "30th".

[0195] 22, the holiday rolling lag corresponding to the holiday "30th" is set to "14". Therefore, the pre-processing unit 12A selects "energy demand at the same time on the 16th" as a holiday rolling feature component corresponding to a certain time on the holiday "30th". Therefore, the pre-processing unit 12A selects "energy demand at the same time on the 16th" as a holiday rolling feature component corresponding to a certain time on the holiday "30th".

[0196] (Generation of Prediction Model in First Embodiment) As is clear from the above explanations, the learning unit 13 in the first embodiment can generate a prediction model based on time-series data to which new features derived according to the working status of the facility are added. That is, the learning unit 13 in the first embodiment can generate a prediction model that reflects the working status of the facility. In this way, according to the first embodiment, it is possible to generate a prediction model of higher quality than the reference embodiment.

[0197] The learning unit 13 may incorporate the rolling lag and non-rolling lag derived in the learning phase into the prediction model, so that the rolling lag and non-rolling lag derived in the learning phase can be easily used in the subsequent prediction phase.

[0198] (Prediction Phase in the First Embodiment) In the first embodiment, the preprocessing unit 22A of the prediction device 20A may acquire the special lag in the learning phase set by the preprocessing unit 12A of the model generation device 10A. As an example, the preprocessing unit 22A may read out the rolling lag and the non-rolling lag set in the learning phase from the prediction model.

[0199] Next, the preprocessing unit 22A may derive the specific lag in the prediction phase from the specific lag in the learning phase based on the working state data DSP in the prediction phase.

[0200] As an example, the preprocessing unit 22A may convert the workday rolling lags and holiday rolling lags of each day set in the learning phase into workday rolling lags and holiday rolling lags of each day in the prediction phase. As an example, the conversion of these rolling lags may be performed according to the correspondence between the daily working status indicated in the working status data DS in the learning phase and the daily working status indicated in the working status data DSP in the prediction phase.

[0201] In addition, the preprocessing unit 22A may convert the workday non-rolling lag and the holiday non-rolling lag of each day set in the learning phase into the workday non-rolling lag and the holiday non-rolling lag of each day in the prediction phase. The conversion of these non-rolling lags may be performed according to the above-mentioned correspondence relationship.

[0202] Next, the preprocessing unit 22A calculates a new feature amount in the prediction phase based on the special lag derived in the prediction phase. Specifically, the preprocessing unit 22A calculates a new feature amount in the prediction phase by setting the special lag to the demand power amount in the prediction phase (i.e., the demand power amount indicated in the actual data DZP).

[0203] In the example of the first embodiment, the preprocessing unit 22A derives a workday rolling feature value and a holiday rolling feature value in the prediction phase by setting a workday rolling lag and a holiday rolling lag in the prediction phase for the demand energy in the prediction phase. The preprocessing unit 22A derives a workday lag feature value and a holiday lag feature value in the prediction phase by setting a workday non-rolling lag and a holiday non-rolling lag in the prediction phase for the demand energy.

[0204] The preprocessing unit 22A generates new time series data by adding new feature quantities in the prediction phase to the time series data of the power demand in the prediction phase. In the example of the first embodiment, the new time series data includes the work status data after pre-value interpolation. Then, the prediction calculation unit 23 outputs a prediction value of the power demand in the prediction phase to the prediction model based on the new time series data.

[0205] As described above, according to the first embodiment, even in the prediction phase, new feature values ​​according to the working status of the facility can be derived. Then, the prediction model can be made to perform prediction based on the new feature values. Thus, according to the first embodiment, the prediction model can be made to perform prediction according to the working status of the facility. Therefore, the prediction accuracy can be further improved compared to the reference embodiment.

[0206] FIG. 23 is an example of new time series data generated by the pre-processing unit 22A. FIG. 23 illustrates a case where the working state is "1" throughout the entire time range to be predicted. In the example of FIG. 23, the workday non-rolling lag is assumed to be "1". Therefore, the lag feature in the example of FIG. 23 represents the energy demand one day before. The lag feature in the example of FIG. 23 is a workday lag feature.

[0207] In the example of Figure 23, no change in working status occurs over the entire time range to be predicted, so the blank values ​​(NULL) in the example of Figure 23 may be interpolated with predicted values ​​obtained from the prediction model.

[0208] Fig. 24 is another example of new time series data generated by the preprocessing unit 22A. Fig. 24 is a diagram paired with Fig. 23. Unlike the example of Fig. 23, Fig. 24 illustrates a case where a change in working status occurs in the prediction target time range. In the example of Fig. 24, the working status on 7 / 11, which is the initial date of the prediction target time range, is "0", and the working status on subsequent dates is "1".

[0209] In the example of Fig. 24, both the holiday non-rolling lag and the workday non-rolling lag are assumed to be "1". Therefore, the lag feature in Fig. 24 also represents the energy demand one day before. The lag feature in the example of Fig. 24 represents the workday lag feature when the working state is "1", and represents the holiday lag feature when the working state is "0".

[0210] In the example of Fig. 24, the working status "1" on 7 / 12, which is the date on which a blank value exists, does not match the working status "0" of the previous day. For this reason, it is desirable to interpolate the blank values ​​in the example of Fig. 24 using an interpolation method other than that of the example of Fig. 23. For example, the blank values ​​in Fig. 24 may be interpolated based on the actual value of the known power demand that corresponds to the working status "1".

[0211] FIG. 25 is another example of new time series data generated by the preprocessing unit 22A. As described in the reference embodiment, the new time series data may include additional features other than the lag feature, the rolling feature, and the working state. The new time series data in the example of FIG. 25 includes a time equivalent and a day of the week as additional features. In the example of FIG. 25, for convenience of explanation, the actual value of the demanded energy is shown alongside the lag feature and the rolling feature. The actual value of the demanded energy itself may be used as a feature.

[0212] (Additional information on employment status data) As described in the embodiment, the working status data can be interpolated to generate working status data that indicates the working status at each time in the facility. By resetting the working status data so as to represent the actual fluctuation of the working status in a day, the working status of the facility can be taken into consideration in more detail.

[0213] Therefore, for example, the working status data may include at least one of information indicating the start time of work at a facility (work start time information) and information indicating the end time of work at the facility (work end time information). This makes it possible to express the fluctuation of the working status of the facility within a day, taking into account at least one of the start time and the end time of work at the facility.

[0214] Fig. 26 is yet another example of new time series data generated by the pre-processing unit 22A. Fig. 26 is a diagram paired with Fig. 25. Fig. 26 illustrates a case where the working status data includes work start time information. The work start time in the example of Fig. 26 is 8:00 a.m.

[0215] The preprocessing unit according to an embodiment of the present invention may pre-interpolate the working state data and then reset the working state at each time in the working state data based on the work start time information. As an example, when the working state of a certain day is different from the working state of the day before the day, and the certain time of the day is before the working time, the preprocessing unit may replace the logical value representing the working state at the certain time with the logical value at the last time (23:59) of the previous day (see FIG. 26).

[0216] By considering the fluctuations in the facility's working status within a day, it is possible to derive a special lag that corresponds to the fluctuations. As a result, it is possible to derive new features that correspond to the fluctuations. In the example of Fig. 26, lag features and rolling features that are closer to the actual value of the demand energy amount than in the example of Fig. 25 are obtained at each time point from "2023 / 5 / 8 0:00" to "2023 / 5 / 8 7:59".

[0217] As an example, the lag feature in FIG. 26 will be described. In the example of FIG. 26, unlike the example of FIG. 25, the working status at each time point from "2023 / 5 / 8 0:00" to "2023 / 5 / 8 7:59" is equivalent to the working status at the same time on the previous day, "2023 / 5 / 7". Therefore, in the example of FIG. 26, unlike the example of FIG. 25, the lag feature at each time point from "2023 / 5 / 8 0:00" to "2023 / 5 / 8 7:59" is obtained by citing the actual value of the demand energy at each time point from "2023 / 5 / 7 0:00" to "2023 / 5 / 7 7:59".

[0218] For example, the lag feature of "2023 / 5 / 8 0:00" in FIG. 26 is equal to the actual value of the demand energy amount of "2023 / 5 / 7 0:00". The lag feature of "2023 / 5 / 8 0:01" is equal to the actual value of the demand energy amount of "2023 / 5 / 7 0:01". The lag feature of "2023 / 5 / 8 7:59" is equal to the actual value of the demand energy amount of "2023 / 5 / 7 7:59".

[0219] Summary of the First Embodiment The items described in the first embodiment can be expressed as follows.

[0220] The model generation device of embodiment 1 is a model generation device that generates a prediction model for predicting the amount of energy demand in a facility, and the model generation device derives autocorrelation coefficients of the actual values ​​corresponding to each of the multiple lags by setting multiple lags for time series data indicating the time progression of the actual values ​​of the amount of energy demand, selects a special lag from among the multiple lags based on the autocorrelation coefficients and work status data indicating daily work status in the facility, derives a new feature by setting the special lag for the actual values, and generates the prediction model by machine learning based on the time series data to which the new feature has been added.

[0221] In the model generation device of embodiment 1, the special lag may include a rolling lag corresponding to a first periodicity of the time fluctuation of the demand energy, and the model generation device may select, from the multiple lags, a lag having a value equal to or greater than a time length threshold and corresponding to the autocorrelation coefficient having a value equal to or greater than an autocorrelation threshold as a special lag candidate, and in an array in which the multiple special lag candidates are sorted in ascending order, may select, as rolling lag candidates, multiple special lag candidates that form an arithmetic progression, and may derive the rolling lag including a workday rolling lag corresponding to workdays of the facility and a holiday rolling lag corresponding to holidays of the facility from the rolling lag candidates based on the working status data.

[0222] In the model generation device according to embodiment 1, the new features may include rolling features corresponding to the rolling lag, and the rolling features may include a workday rolling feature and a holiday rolling feature, and the model generation device may derive the workday rolling feature by setting the workday rolling lag to the actual value, and may derive the holiday rolling feature by setting the holiday rolling lag to the actual value.

[0223] In the model generation device of embodiment 1, the special lag may include a non-rolling lag corresponding to a second periodicity of the time fluctuation of the demanded energy, and the second periodicity may be a shorter periodicity than the first periodicity. The model generation device may select the special lag candidates in the array excluding the rolling lag candidates as non-rolling lag candidates, and may derive the non-rolling lag including a workday non-rolling lag corresponding to the workday and a holiday non-rolling lag corresponding to the holiday from the non-rolling lag candidates based on the working status data.

[0224] In the model generation device of embodiment 1, the new features may include lag features corresponding to the non-rolling lag, and the lag features may include workday lag features and holiday lag features, and the model generation device may derive the workday lag features by setting the workday non-rolling lag to the actual value, and may derive the holiday lag features by setting the holiday non-rolling lag to the actual value.

[0225] In the model generating device according to the first embodiment, the working status data may further indicate working status at each time in the facility.

[0226] In the model generating device according to the first embodiment, the working status data may include at least one of information indicating a work start time at the facility and information indicating a work end time at the facility.

[0227] A prediction device according to a first embodiment is a prediction device that derives a predicted value of an amount of energy demand in a facility by using, in a prediction phase, a prediction model for predicting an amount of energy demand in a facility, which has been generated in advance in a learning phase. In the learning phase, for time-series data showing a time transition of an actual value of the amount of energy demand, a plurality of lags are set for the actual value, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived. A special lag is selected from the plurality of lags based on the autocorrelation coefficient and working status data showing a daily working status in the facility. A new feature amount is derived by setting the special lag for the actual value. and the prediction model is generated by machine learning based on the time series data to which the new feature has been added, and the prediction device derives a special lag for the prediction phase from the special lag in the learning phase based on the working status data in the prediction phase, and derives a new feature for the prediction phase by setting the special lag in the prediction phase to the actual value in the prediction phase for the time series data in the prediction phase that shows the time progression of the actual value, and outputs the predicted value to the prediction model based on the time series data in the prediction phase to which the new feature has been added.

[0228] The model generation method of embodiment 1 is a model generation method for generating a predictive model for predicting the amount of electricity demand in a facility, and includes the steps of: setting multiple lags for time series data indicating the time progression of actual values ​​of the amount of electricity demand, thereby deriving autocorrelation coefficients of the actual values ​​corresponding to each of the multiple lags; selecting a special lag from among the multiple lags based on the autocorrelation coefficients and work status data indicating daily work status in the facility; deriving a new feature by setting the special lag for the actual value; and generating the predictive model by machine learning based on the time series data to which the new feature has been added.

[0229] A prediction method according to a first embodiment is a prediction method for deriving a predicted value of an amount of power demand in a facility by using, in a prediction phase, a prediction model for predicting an amount of power demand in a facility, which has been generated in advance in a learning phase. In the learning phase, for time-series data showing a time transition of an actual value of the amount of power demand, a plurality of lags are set for the actual value, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived. A special lag is selected from the plurality of lags based on the autocorrelation coefficient and working status data showing a daily working status in the facility. A new feature amount is derived by setting the special lag for the actual value. The prediction model is generated by machine learning based on the time series data to which a feature has been added, and the prediction method includes: a step of deriving a special lag for the prediction phase from the special lag in the learning phase based on working status data in the prediction phase; a step of deriving a new feature for the prediction phase by setting the special lag in the prediction phase to the actual value in the prediction phase for the time series data in the prediction phase that indicates a time progression of the actual value; and a step of causing the prediction model to output the predicted value based on the time series data in the prediction phase to which the new feature in the prediction phase has been added.

[0230] [Embodiment 2] FIG. 27 is a block diagram illustrating the configuration of the main part of an information processing device 1B in the second embodiment. In the second embodiment, differences from the first embodiment will be described. The information processing device 1B in the example of FIG. 27 includes a control device 9B instead of the control device 9A. The control device 9B includes a model generation device 10B and a prediction device 20B instead of the model generation device 10A and the prediction device 20A. The model generation device 10B includes a pre-processing unit 12B instead of the pre-processing unit 12A. And, the model generation device 10B includes a learning unit 13B and an evaluation unit 14B instead of the learning unit 13 and the evaluation unit 14. The prediction device 20B includes a pre-processing unit 22B instead of the pre-processing unit 22A. And, the prediction device 20B includes a prediction calculation unit 23B instead of the prediction calculation unit 23.

[0231] (Learning Phase in the Second Embodiment) In the second embodiment, the minimum value of the multiple lags is referred to as the "minimum lag." The minimum lag is an amount equivalent to the lag unit amount described in the reference embodiment. Therefore, in the second embodiment, a case where the minimum lag is 30 minutes is exemplified. The minimum lag value of 30 minutes corresponds to "1 / 48 day" when converted to a day unit (see also FIG. 6 in the reference embodiment).

[0232] As described in the embodiment, the lag is set as a product of the lag unit amount and the lag order, so that each of the multiple lags greater than the minimum lag is equal to an integer multiple of the minimum lag.

[0233] In the second embodiment, the minimum lag is set to a value less than the time length threshold. That is, in the second embodiment, the minimum lag is not selected as a special lag. Therefore, the preprocessing unit 12B selects a special lag from among a plurality of lags other than the minimum lag based on the AC and the working state data DS.

[0234] For example, the preprocessing unit 12B selects a rolling lag and a non-rolling lag as special lags from among a plurality of lags other than the minimum lag based on the AC and the working status data DS. In the second embodiment as well, the preprocessing unit 12B selects a workday rolling lag and a holiday rolling lag as rolling lags, and selects a workday non-rolling lag and a holiday non-rolling lag as non-rolling lags.

[0235] In the second embodiment as well, the preprocessing unit 12B derives a new feature based on the special lag. That is, the preprocessing unit 12B derives a new feature by setting the special lag to the actual value of the demand energy. For example, the preprocessing unit 12B derives a rolling feature based on the rolling lag, and derives a lag feature based on the non-rolling lag. In the second embodiment as well, the preprocessing unit 12B derives a workday rolling feature and a holiday rolling lag as the rolling feature, and derives a workday lag feature and a holiday lag feature as the lag feature.

[0236] In this specification, the time series data in which the new feature is added to the time series data indicating the time transition of the actual value of the energy demand is referred to as the first time series data. The first time series data corresponds to the time series data described in the learning phase of the first embodiment. The pre-processing unit 12B may generate the first time series data by adding the new feature derived as described above to the training data.

[0237] In this specification, the prediction model generated by machine learning based on the first time series data is referred to as the first model. In the second embodiment, the learning unit 13B generates the first model based on the first time series data. The first model corresponds to the prediction model described in the first embodiment. The learning unit 13B may incorporate the special lag derived by the pre-processing unit 12B into the first model.

[0238] Unlike the pre-processing unit 12A, the pre-processing unit 12B derives a further new feature based on the minimum lag. Specifically, the pre-processing unit 12B derives a further new feature by setting the minimum lag to the actual value of the demand energy. In this specification, the further new feature is also referred to as the "immediately preceding demand energy."

[0239] In the second embodiment, the period for which the actual value of the demand energy is predicted is called the prediction period. The time step (time interval) of the prediction period is assumed to be equal to the minimum lag. The start point of the prediction period is called the prediction start time, and the end point of the prediction period is called the prediction end time.

[0240] The previous demand energy represents the actual value of the demand energy at a time point immediately prior to a certain time point within the forecast period (e.g., the actual value of the demand energy 30 minutes prior to a certain time point). As described above, the minimum lag does not fall under the special lag. Therefore, the previous demand energy does not fall under the feature derived based on the special lag. In this way, the previous demand energy is a feature different from both the rolling feature and the lag feature. By introducing the previous demand energy as a further new explanatory variable, the actual value of the demand energy immediately prior to the current time point can be taken into account. Therefore, further improvement in the prediction accuracy of the prediction model is expected.

[0241] In this specification, the time series data in which the immediately preceding demand energy amount is added to the first time series data is referred to as second time series data. The pre-processing unit 12B may generate the second time series data by adding the immediately preceding demand energy amount to the first time series data generated as described above.

[0242] In this specification, the prediction model generated by machine learning based on the second time series data is referred to as the second model. In the second embodiment, the learning unit 13B generates the second model based on the second time series data. As described above, unlike the learning unit 13 in the first embodiment, the learning unit 13B further generates the second model. As a result, in the second embodiment, it is possible to predict the demanded power amount by using the first model and the second model in combination. The learning unit 13B may incorporate the minimum lag into the second model.

[0243] The evaluation unit 14B evaluates each of the first model and the second model generated by the learning unit 13B. Prior to the evaluation by the evaluation unit 14B, the preprocessing unit 12B performs a series of preprocessing processes on the validation data in the same manner as the above-mentioned training data, thereby generating preprocessed validation data.

[0244] In the example of the second embodiment, the preprocessing unit 12B generates preprocessed verification data for the first model and preprocessed verification data for the second model. The preprocessed verification data for the first model includes the above-mentioned new features (e.g., rolling features and lag features). The preprocessed verification data for the second model further includes the immediately preceding demand energy in addition to the new features.

[0245] In this specification, the predicted value of the energy demand output from the first model is referred to as the first predicted value. The evaluation unit 14B may cause the first model to output the first predicted value at each time point in the prediction target period by supplying the preprocessed verification data for the first model to the first model. Also, in this specification, the error between the actual value of the energy demand and the first predicted value is referred to as the first error. The evaluation unit 14B may derive the first error.

[0246] In the example of the second embodiment, the error between the actual value and the predicted value is displayed in percentage. The index value of the prediction model described in the reference embodiment may be used as the error. Therefore, the error in the second embodiment may be, for example, MAE or RMSE.

[0247] FIG. 28 is an example of a graph showing the error of the first model (i.e., the first error) at each point in the prediction period. In the example of FIG. 28, the length of the prediction period is set to 36 hours. The horizontal axis of the graph represents the elapsed time from the prediction start point. Therefore, the value "0" on the horizontal axis corresponds to the prediction start point, and the value "36" corresponds to the prediction end point. Each point in time described below can be read as the elapsed time from the prediction start point "0" unless there is a contradiction in the context.

[0248] In addition, in this specification, the predicted value of the energy demand output from the second model is referred to as the second predicted value. The evaluation unit 14B may cause the second model to output the second predicted value at each time point in the prediction target period by supplying the preprocessed verification data for the second model to the second model. In addition, in this specification, the error between the actual value of the energy demand and the second predicted value is referred to as the first error. The evaluation unit 14B may derive the second error.

[0249] FIG. 29 is an example of a graph showing the error (i.e., the second error) of the second model at each point in the forecast period. FIG. 29 is a diagram paired with FIG. 28. Comparing FIG. 28 and FIG. 29, when the time elapsed from the forecast start point is not so large, the second error is smaller than the first error. This shows that the second model is superior to the first model in forecasting short-term time ranges. This characteristic of the second model is due to the fact that the second model uses the immediately preceding demand energy amount as an explanatory variable.

[0250] However, in the example of FIG. 29, the second error tends to increase as the time elapsed from the prediction start point increases. This characteristic of the second model is also due to the fact that the second model uses the immediately preceding demand energy amount as an explanatory variable. In the second model, it may be necessary to use the predicted value of the immediately preceding demand energy amount to derive the second predicted value at each point in the prediction period (see the explanation of the prediction phase described later). As a result, as time passes, the error between the predicted value of the immediately preceding demand energy amount and the actual value accumulates, and the second error may increase.

[0251] On the other hand, in the example of FIG. 28, even if the time elapsed from the prediction start point becomes large, the first error does not increase so much. Unlike the second model, the first model does not use the immediately preceding demand energy amount as an explanatory variable. Therefore, the first model does not accumulate the errors described above with respect to the second model. Therefore, when the time elapsed from the prediction start point is relatively large, the first error becomes smaller than the second error. This shows that the first model is superior to the second model in prediction of a long-term time range.

[0252] As described above, the second model is suitable for short-term predictions, and the first model is suitable for long-term predictions. Therefore, by appropriately combining the first and second models, better predictions can be achieved than when using only one of the models.

[0253] Therefore, the evaluation unit 14B may determine a time point at which to switch between using the first model and using the second model (hereinafter referred to as a "switching time point") based on the first predicted value and the second predicted value. As an example, the evaluation unit 14B may select a switching time point from among a plurality of switching time point candidates based on the first predicted value and the second predicted value.

[0254] The evaluation unit 14B may set at least a part of each time point belonging to the prediction target period as a switching time point candidate. In the second embodiment, the evaluation unit 14B sets all the time points belonging to the prediction target period as switching time point candidates. Therefore, the evaluation unit 14B sets 73 switching time point candidates from 0h to 36h (see also FIG. 31 described later).

[0255] Next, the evaluation unit 14B causes the second model to output the second predicted value for each time point from the prediction start time point to the switching time point candidate. Then, the evaluation unit 14B causes the first model to output the first predicted value for each time point from the time point next to the switching time point candidate to the prediction end time point. As an example, consider a case where the switching time point candidate is set to 3.5h. In this case, the evaluation unit 14B causes the second model to output the second predicted value for 8 time points from 0h to 3.5h among the 73 time points belonging to the prediction target period. Then, the evaluation unit 14B causes the first model to output the first predicted value for 65 time points from 4h to 36h among the 73 time points.

[0256] In the example of the second embodiment, the i-th switching time candidate is referred to as the i-th switching time candidate. In the example of the second embodiment, i is any integer greater than or equal to 0 and less than or equal to (Nc-1). Nc represents the total number of time points belonging to the prediction target period. In the above example, Nc=73. In the example of the second embodiment, the i-th switching time candidate is defined as the time point that is the product of i and the minimum lag after the prediction start time point. Therefore, the 0th switching time candidate corresponds to the prediction start time point, and the (Nc-1)th switching time candidate corresponds to the prediction end time point.

[0257] As described above, for each of all i greater than or equal to 0 and less than or equal to (Nc-1), the evaluation unit 14B causes the second model to output the second predicted value for (i+1) time points from the prediction start time point to the i-th switching time point candidate. Then, the evaluation unit 14B causes the first model to output the first predicted value for (Nc-i-1) time points from the (i+1)-th switching time point candidate to the prediction end time point.

[0258] The evaluation unit 14B may select a switching time from among the multiple switching time candidates based on (i) the second predicted value at each time point from the prediction start time point to the switching time candidate, and (ii) the first predicted value at each time point from the time point next to the switching time candidate to the prediction end time point. That is, the evaluation unit 14B may select a switching time from among the Nc switching time candidates based on the second predicted value for the above-mentioned (i+1) time points and the first predicted value for the above-mentioned (Nc-i-1) time points.

[0259] Next, for each of the multiple candidate switching time points, the evaluation unit 14B may calculate an overall error between the first predicted value and the second predicted value for the prediction period based on (i) the actual values ​​of power demand at each time point from the prediction start time to the prediction end time, (ii) the second predicted values ​​at each time point from the prediction start time to the candidate switching time point, and (iii) the first predicted values ​​at each time point from the time point next to the candidate switching time point to the prediction end time.

[0260] The total error in the second embodiment may be any statistical quantity calculated based on the actual value, the first predicted value, and the second predicted value. The evaluation unit 14B may select a switching time point from among a plurality of switching time point candidates based on the total error.

[0261] As an example, the evaluation unit 14B may calculate the second error at each time point from the prediction start time point to the switching time point candidate based on the actual value of the power demand and the second predicted value at each time point. Then, the evaluation unit 14B may calculate the cumulative value of the second error at each time point from the prediction start time point to the switching time point candidate. That is, the evaluation unit 14B may calculate the cumulative value of the second error for the above-mentioned (i+1) time points.

[0262] Furthermore, the evaluation unit 14B may calculate a first error at each time point from the next time point of the switching time point candidate to the prediction end time point based on the actual value and the first predicted value of the power demand at each time point. Then, the evaluation unit 14B may calculate a cumulative value of the first error at each time point from the next time point of the switching time point candidate to the prediction end time point. That is, the evaluation unit 14B may calculate a cumulative value of the first error for the above-mentioned (Nc-i-1) time points.

[0263] As is clear from the above descriptions, the total error may be any statistical amount calculated based on the first error and the second error. As an example, the evaluation unit 14B may calculate the sum of the accumulated value of the first error and the accumulated value of the second error. Then, the evaluation unit 14B may calculate the total error by dividing the sum by the above-mentioned Nc.

[0264] Fig. 30 is an example of a graph showing the first error and the second error at each time point of the prediction target period calculated by the evaluation unit 14B. Fig. 30 illustrates a case where the switching time candidate is set to 3.5h. This switching time candidate corresponds to the seventh switching time candidate. The values ​​on the vertical axis of the graph in the example of Fig. 30 represent the second error in the time period from 0h to 3.5h, and represent the first error in the time period from 4h to 36h.

[0265] The example in Fig. 30 shows that by appropriately using the first model and the second model in combination, it is possible to realize a better prediction than both the case where only the first model is used (Fig. 28) and the case where only the second model is used (Fig. 29). Therefore, according to the second embodiment, it is possible to further improve the prediction accuracy compared to the first embodiment.

[0266] Fig. 31 shows an example of the total error corresponding to each of a plurality of switching time candidate. As is clear from the above explanations, the total error when the switching time candidate is set to the minimum value "0h" (the total error corresponding to the 0th switching time candidate) corresponds to the average value of the graph plots in Fig. 28. On the other hand, the total error when the switching time candidate is set to the maximum value "36h" (the total error corresponding to the 72nd switching time candidate) corresponds to the average value of the graph plots in Fig. 29.

[0267] As an example, the evaluation unit 14B may select, as the switching time point, the switching time point candidate corresponding to the minimum value of the total error from among a plurality of switching time point candidates. In the example of FIG. 31, the minimum value of the total error is obtained at the switching time point candidate "3.5h". Therefore, the evaluation unit 14B selects 3.5h as the switching time point. In the example of FIG. 31, as the switching time point candidate increases from 0h to 3.5h, the total error decreases. Next, as the switching time point candidate increases from 4h to 36h, the total error increases.

[0268] The evaluation unit 14B may incorporate the prediction start time and prediction end time in the learning phase into at least one of the first model and the second model. In addition, the evaluation unit 14B may incorporate the switching time determined as described above into at least one of the first model and the second model. This makes it possible to easily use these time points in the prediction phase described below. As an example, the evaluation unit 14B may incorporate these time points into the second model.

[0269] (Prediction Phase in the Second Embodiment) The prediction device 20B predicts the power demand in the prediction phase by using the first model and the second model generated by the model generation device 10B in the learning phase. The process in which the prediction device 20B predicts the power demand using the first model is generally equivalent to the example of the first embodiment. Therefore, in the second embodiment, the process related to the second model will be mainly described.

[0270] The preprocessing unit 22B of the prediction device 20B may read out the prediction start time in the learning phase and the switching time determined in the learning phase from the second model. Then, the preprocessing unit 22B may calculate the time length from the prediction start time in the learning phase to the switching time as the switching time length. In the example of the second embodiment, the switching time length is 3.5 hours.

[0271] Next, the preprocessing unit 22B may obtain a prediction start time and a prediction end time of the prediction phase from the prediction target period of the prediction phase. Then, the preprocessing unit 12B sets a time obtained by adding the switching time length to the prediction start time of the prediction phase as a switching time of the prediction phase. In the second embodiment, the preprocessing unit 12B sets the switching time of the prediction phase as "a time 3.5 hours after the prediction start time of the prediction phase."

[0272] Then, the prediction calculation unit 23B causes one of the first model and the second model to output a prediction value of the power demand for each time point belonging to the prediction target period in the prediction phase. In the prediction target period in the prediction phase, the time step is also set to a value equal to the minimum lag.

[0273] Specifically, the prediction calculation unit 23B causes the second model to output the second predicted value for each time point from the prediction start time point to the switching time point in the prediction phase, and causes the first model to output the first predicted value for each time point from the time point next to the switching time point to the prediction end time point.

[0274] More specifically, in the example of embodiment 2, the pre-processing unit 22B generates input data for the first model and input data for the second model in the prediction phase by performing a series of pre-processing equivalent to the pre-processing described in the learning phase.

[0275] FIG. 32 shows an example of input data for the first model in the prediction phase. As shown in FIG. 32, the input data for the first model in the prediction phase includes the new features (e.g., rolling features and lag features) described above. FIG. 33 shows an example of input data for the second model in the prediction phase. As shown in FIG. 33, the input data for the second model in the prediction phase further includes the immediately preceding demand energy amount in addition to the new features.

[0276] The prediction calculation unit 23B supplies the input data for the second model to the second model, thereby causing the second model to output a second predicted value for each time point from the prediction start time point to the switching time point in the prediction phase.The prediction calculation unit 23B supplies the input data for the first model to the first model, thereby causing the first model to output a first predicted value for each time point from the time point next to the switching time point to the prediction end time.

[0277] In the example of Figures 32 to 33, the prediction end start time is "2021 / 7 / 11 12:30", and the prediction end time is "2021 / 7 / 13 0:00". In the example of the second embodiment, the switching time in the prediction phase is set as "2021 / 7 / 11 16:00".

[0278] Therefore, the prediction calculation unit 23B causes the second model to output the second predicted value for each time point from "2021 / 7 / 11 12:30" to "2021 / 7 / 11 16:00" by supplying the input data for the second model to the second model. Then, the prediction calculation unit 23B causes the first model to output the first predicted value for each time point from "2021 / 7 / 11 16:30" to "2021 / 7 / 13 0:00" by supplying the input data for the first model to the first model.

[0279] In Fig. 33, the immediately preceding demand energy amount at the prediction end start time point "2021 / 7 / 11 12:30" is the actual value of the demand energy amount at the time point the minimum lag before the prediction end start time point. In other words, the immediately preceding demand energy amount is the actual value of the demand energy amount at the time point "2021 / 7 / 11 12:00".

[0280] In the example of FIG. 33, the actual value of the demand energy at a time point after the time point "2021 / 7 / 11 12:00" has not yet been stored in the storage unit 90. Therefore, in the example of FIG. 33, the previous demand energy at each time point after "2021 / 7 / 11 13:00" is a blank value. These blank values ​​may be interpolated by any interpolation method.

[0281] (Supplementary Note on the Second Embodiment) (1) In the learning phase, the method by which the preprocessing unit 12B determines the switching time point based on the overall error is not limited to the above-mentioned example method. For example, the preprocessing unit 22B may select a switching time point from among a plurality of switching time point candidates based on the increasing trend of the overall error with an increase in the counter variable i indicating the number of the switching time point candidate.

[0282] In this specification, the total error corresponding to the i-th switching time candidate is denoted as TE(i). The difference between TE(i) and TE(i-1) is denoted as Δ. Δ represents the difference in the total error when i increases by 1. The preprocessing unit 12B may calculate Δ.

[0283] If the tendency of Δ to increase with an increase in i continues, it is considered that the possibility of finding a global minimum value of the total error within the range of a larger i is not very high. Therefore, when Δ increases K times consecutively with an increase in i, the preprocessing unit 22B may select the minimum value of the total error obtained within the range of i searched up to that point as the switching point. K may be set as any integer not less than 0 and not more than (Nc-1).

[0284] By selecting the switching time point based on the increasing trend of the total error (more specifically, based on the increasing trend of Δ), it is not necessary to calculate TE(i) for all i. As a result, the time required for the series of calculations to set the switching time point can be reduced.

[0285] (2) In the second embodiment, the learning unit 13B generates the first model and the second model. However, the learning unit 13B may further generate a prediction model other than the first model and the second model. For example, the learning unit 13B may further generate another prediction model based on a feature other than the lag feature, the rolling feature, and the immediately preceding demand energy amount.

[0286] As described above, the learning unit 13B may generate three or more prediction models. In this case, in the prediction phase, the prediction device 20B can predict the power demand by using the three or more prediction models in combination.

[0287] [Software implementation example] The functions of the information processing devices 1 to 1B (hereinafter referred to as "devices" for convenience) can be realized by a program for causing a computer to function as the devices, and by a program for causing a computer to function as each control block of the devices (in particular, each part included in the model generating devices 10 to 10B and the prediction devices 20 to 20B).

[0288] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program to realize each function described in each of the above embodiments.

[0289] The program may be non-transitory and may be recorded in one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be provided to the device via any wired or wireless transmission medium.

[0290] In addition, some or all of the functions of each of the control blocks can be realized by a logic circuit. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of each of the control blocks can be realized by, for example, a quantum computer.

[0291] As is clear from the description of each of the above embodiments, each process in each of the above embodiments can be executed by AI (Artificial Intelligence). The AI ​​may be one that runs on the control device, or may be one that runs on another device (for example, an edge computer or a cloud server).

[0292] 〔summary〕 A model generation device according to a first aspect of the present invention is a model generation device that generates a prediction model for predicting an amount of energy demand in a facility, wherein the model generation device sets multiple lags for time series data indicating a time progression of an actual value of the amount of energy demand, and derives autocorrelation coefficients of the actual value corresponding to each of the multiple lags, selects a special lag from among the multiple lags other than a minimum lag that is the minimum value of the multiple lags based on the autocorrelation coefficient and work status data indicating daily work status at the facility, and derives a new feature by setting the special lag for the actual value, generates a first model as the prediction model by machine learning based on first time series data in which the new feature is added to the time series data, derives a further new feature by setting the minimum lag for the actual value, and further generates a second model different from the first model by machine learning based on second time series data in which the further new feature is added to the first time series data.

[0293] In the model generation device according to aspect 2 of the present invention, in aspect 1, the predicted value of the demand energy output from the first model is referred to as a first predicted value, and the predicted value of the demand energy output from the second model is referred to as a second predicted value, and the model generation device may set at least a portion of the points in time belonging to a prediction period of the actual value as a plurality of switching time candidates, and may select a switching time from among the plurality of switching time candidates based on (i) the second predicted value at each point from a prediction start point which is the start point of the prediction period to the switching time candidate, and (ii) the first predicted value at each point from a point next to the switching time candidate to a prediction end point which is the end point of the prediction period.

[0294] The model generation device according to aspect 3 of the present invention, in aspect 2, may calculate, for each of the multiple switching time candidate points, a total error between the first predicted value and the second predicted value in the prediction period based on (i) the actual value at each time point from the prediction start time to the prediction end time, (ii) the second predicted value at each time point from the prediction start time to the switching time candidate point, and (iii) the first predicted value at each time point from the time point next to the switching time candidate point to the prediction end time, and may select the switching time from among the multiple switching time candidate points based on the total error.

[0295] In a model generation device according to a fourth aspect of the present invention, in the third aspect, the error between the actual value and the first predicted value is referred to as a first error, and the error between the actual value and the second predicted value is referred to as a second error, and the model generation device may calculate the sum of (i) a cumulative value of the second error at each time point from the prediction start time to the candidate switching time point and (ii) a cumulative value of the first error at each time point from the time point next to the candidate switching time point to the prediction end time, and may calculate the total error as the value obtained by dividing the sum by the total number of time points belonging to the prediction target period.

[0296] The model generating device according to a fifth aspect of the present invention, in the third or fourth aspect, may select, as the switching time point, from among the multiple switching time point candidates, a switching time point candidate corresponding to a minimum value of the overall error.

[0297] A prediction device according to a sixth aspect of the present invention is a prediction device that derives a predicted value of an amount of energy demand in a facility by using a prediction model for predicting an amount of energy demand in a facility, the prediction model having been generated in advance in a learning phase, in a prediction phase, wherein in the learning phase, a plurality of lags are set for time series data showing a time progression of an actual value of the amount of energy demand, thereby deriving an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags, and a special lag is selected from among the plurality of lags other than a minimum lag, which is the minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is selected from among the plurality of lags, based on the autocorrelation coefficient and the ... a first model is generated by machine learning as the prediction model based on first time series data in which the new feature is added to the time series data, a further new feature is derived by setting the minimum lag to the actual value, and a second model different from the first model is further generated by machine learning based on second time series data in which the further new feature is added to the first time series data, and the prediction device predicts the demand energy by using the first model and the second model in the prediction phase.

[0298] In the prediction device according to a seventh aspect of the present invention, in the sixth aspect, a predicted value of the energy demand output from the first model is referred to as a first predicted value, and a predicted value of the energy demand output from the second model is referred to as a second predicted value, and in the learning phase, at least a part of each time point from a prediction start time point to a prediction end time point may be set as a plurality of switching time point candidates, and a switching time point is selected from among the plurality of switching time point candidates based on (i) the second predicted value at each time point from the prediction start time point to the switching time point candidate, and (ii) the first predicted value at each time point from a time point next to the switching time point candidate to the prediction end time point. Preferably, the prediction device may calculate the length of time from the prediction start point in the learning phase to the switching point as a switching time length, may set the point obtained by adding the switching time length to the prediction start point in the prediction phase as the switching point in the prediction phase, may cause the second model to output the second predicted value for each point in time from the prediction start point in the prediction phase to the switching point in the prediction phase, and may cause the first model to output the first predicted value for each point in time from the point next to the switching point in the prediction phase to the prediction end point in the prediction phase.

[0299] A model generation method according to an eighth aspect of the present invention is a model generation method for generating a prediction model for predicting an amount of energy demand in a facility, the model generation method including the steps of: setting a plurality of lags for time series data indicating a time progression of an actual value of the amount of energy demand, thereby deriving an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags; selecting a special lag from among the plurality of lags other than a minimum lag, which is the minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data indicating a daily working status in the facility; deriving a new feature by setting the special lag to the actual value; generating a first model as the prediction model by machine learning based on first time series data in which the new feature is added to the time series data; deriving a further new feature by setting the minimum lag to the actual value; and further generating a second model different from the first model by machine learning based on second time series data in which the further new feature is added to the first time series data.

[0300] A prediction method according to a ninth aspect of the present invention is a prediction method for deriving a predicted value of an amount of energy demand in a facility by using a prediction model for predicting an amount of energy demand in a facility, the prediction model having been generated in advance in a learning phase, in a prediction phase. In the learning phase, a plurality of lags are set for time series data showing a time progression of an actual value of the amount of energy demand, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived. A special lag is selected from among the plurality of lags other than a minimum lag, which is the minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data showing a daily working status in the facility, and the special lag is a first model is generated by machine learning as the prediction model based on first time series data in which the new feature has been added to the time series data, a further new feature is derived by setting the minimum lag to the actual value, and a second model different from the first model is further generated by machine learning based on second time series data in which the further new feature has been added to the first time series data, and the prediction method includes a step of predicting the demanded energy by using the first model and the second model in the prediction phase.

[0301] [Additional Notes] One aspect of the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of one aspect of the present invention. [Explanation of symbols]

[0302] 1B Information processing equipment 10B Model generation device 20B Prediction Device DZ performance data during the learning phase DZP Actual data during forecast phase DS Employment status data during the learning phase Employment status data during the DSP prediction phase

Claims

1. A model generation device that generates a prediction model for predicting an amount of power demand in a facility, The model generating device comprises: A plurality of lags are set for time series data indicating a time transition of an actual value of the energy demand, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived; selecting a special lag from among the plurality of lags, other than a minimum lag that is a minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data indicating daily working status in the facility; deriving a new feature quantity by setting the special lag to the performance value; generating a first model as the prediction model by machine learning based on first time-series data in which the new feature amount is added to the time-series data; deriving a further new feature by setting the minimum lag to the actual value; a model generating device that further generates, by the machine learning, a second model different from the first model, based on second time series data in which the further new feature is added to the first time series data.

2. The predicted value of the amount of power demand output from the first model is referred to as a first predicted value, The predicted value of the amount of power demand output from the second model is referred to as a second predicted value, The model generating device comprises: setting at least some of the time points belonging to the prediction period of the actual value as a plurality of switching time point candidates; 2. The model generating device according to claim 1, wherein a switching time point is selected from among the multiple switching time point candidates based on: (i) the second predicted value at each time point from a prediction start time point that is a start point of the prediction target period to the switching time point candidate; and (ii) the first predicted value at each time point from a time point next to the switching time point candidate to a prediction end time point that is an end point of the prediction target period.

3. The model generating device comprises: for each of the plurality of switching time candidates, (i) calculating an overall error between the first predicted value and the second predicted value in the prediction target period based on the actual value at each time point from the prediction start time point to the prediction end time point, (ii) the second predicted value at each time point from the prediction start time point to the switching time candidate, and (iii) the first predicted value at each time point from the time point next to the switching time candidate to the prediction end time point; The model generating device according to claim 2 , further comprising: a switch time point selected from among a plurality of switch time point candidates based on the overall error.

4. The error between the actual value and the first predicted value is referred to as a first error, The error between the actual value and the second predicted value is referred to as a second error, The model generating device comprises: (i) calculating a sum of an accumulated value of the second error at each time point from the prediction start time point to the switching time point candidate, and (ii) calculating a sum of an accumulated value of the first error at each time point from a time point next to the switching time point candidate to the prediction end time point; The model generating device according to claim 3 , wherein the total error is calculated by dividing the sum by a total number of time points belonging to the prediction target period.

5. 5. The model generating device according to claim 3, wherein the model generating device selects, as the switching time point, a switching time point candidate corresponding to a minimum value of the overall error from among the plurality of switching time point candidates.

6. A prediction device that derives a predicted value of an amount of power demand in a facility by using a prediction model for predicting an amount of power demand in a facility, the prediction model having been generated in advance in a learning phase, the prediction device comprising: In the learning phase, a plurality of lags are set for time-series data indicating a time transition of an actual value of the energy demand, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived; a special lag is selected from among the plurality of lags other than a minimum lag that is a minimum value of the plurality of lags based on the autocorrelation coefficient and employment status data indicating daily employment status in the facility; A new feature is derived by setting the special lag to the actual performance value, a first model is generated as the prediction model by machine learning based on first time-series data in which the new feature is added to the time-series data; A further new feature is derived by setting the minimum lag to the actual value, a second model different from the first model is further generated by the machine learning based on second time series data in which the further new feature is added to the first time series data; The prediction device predicts the amount of power demand by using the first model and the second model in the prediction phase.

7. The predicted value of the amount of power demand output from the first model is referred to as a first predicted value, The predicted value of the amount of power demand output from the second model is referred to as a second predicted value, In the learning phase, At least some of the time points from the prediction start time point to the prediction end time point are set as a plurality of switching time point candidates, (i) a switching time point is selected from among the switching time point candidates based on the second predicted value at each time point from the prediction start time point to the switching time point candidate, and (ii) the first predicted value at each time point from a time point next to the switching time point candidate to the prediction end time point; The prediction device comprises: A time length from the prediction start time point to the switching time point in the learning phase is calculated as a switching time length; A time point obtained by adding the switching time length to a prediction start time point in the prediction phase is set as a switching time point in the prediction phase; causing the second model to output the second predicted value for each time point from the prediction start time point in the prediction phase to the switching time point in the prediction phase; The prediction device according to claim 6 , further comprising: causing the first model to output the first predicted value for each time point from a time point next to the switching time point in the prediction phase to a prediction end time point in the prediction phase.

8. A model generation method for generating a prediction model for predicting an amount of power demand in a facility, comprising: The model generation method includes: setting a plurality of lags for time-series data indicating a time transition of an actual value of the energy demand, and deriving an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags; selecting a special lag from among the plurality of lags, other than a minimum lag that is the minimum value of the plurality of lags, based on the autocorrelation coefficient and working status data indicating daily working status in the facility; deriving a new feature quantity by setting the special lag to the performance value; generating a first model as the prediction model by machine learning based on first time-series data in which the new feature amount is added to the time-series data; deriving a further new feature quantity by setting the minimum lag to the actual performance value; and further generating, by the machine learning, a second model different from the first model based on second time series data in which the further new feature is added to the first time series data.

9. A prediction method for deriving a predicted value of an amount of power demand in a facility by using a prediction model for predicting an amount of power demand in a facility, the prediction model having been generated in advance in a learning phase, the prediction method comprising: In the learning phase, a plurality of lags are set for time-series data indicating a time transition of an actual value of the energy demand, and an autocorrelation coefficient of the actual value corresponding to each of the plurality of lags is derived; a special lag is selected from among the plurality of lags other than a minimum lag that is a minimum value of the plurality of lags based on the autocorrelation coefficient and employment status data indicating daily employment status in the facility; A new feature is derived by setting the special lag to the actual performance value, a first model is generated as the prediction model by machine learning based on first time-series data in which the new feature is added to the time-series data; A further new feature is derived by setting the minimum lag to the actual value, a second model different from the first model is further generated by the machine learning based on second time series data in which the further new feature is added to the first time series data; The prediction method includes, in the prediction phase, predicting the amount of demanded energy by using the first model and the second model.

Images