Learning device, power demand inference device, grid control system, supply and demand control system, facility formation support system and program

Abstract

To obtain a learning device which can realize an accurate inference of a power demand without using the recorded value of the amount of solar radiation.SOLUTION: The present invention includes: an acquisition unit 11 for acquiring a recorded value of a power demand and datetime information showing the date and time corresponding to the recorded value; a feature amount generation unit 12 for generating a feature amount including an index regarding a solar altitude and information showing whether a day is a holiday, by using the datetime information; and a learning model generation unit 13 for generating a learned model to infer a power demand from the feature amount by using the recorded value as correct data and the feature amount corresponding to the recorded value, by machine learning.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a learning device that performs learning to infer power demand, a power demand inference device, a grid control system, a supply and demand control system, a facility formation support system, and a program. [Background technology]

[0002] In power systems, predictions and estimations of power demand are performed for various purposes, including supply and demand control, facility formation, and power system voltage control. For example, Patent Document 1 discloses a technology for performing demand prediction at time t by using, as input data, "average power demand at time t for a predetermined period (e.g., 49 days) from the time the prediction is made," "the difference between the average demand at the time the prediction is made and the actual demand," "heating and cooling effect variable (temperature)," "Saturday dummy variable (a flag for identifying Saturday)," "Sunday / holiday dummy variable (a flag for identifying Sunday and public holiday)," "solar radiation effect variable (solar radiation amount)," and the like, and performing demand prediction using multiple regression analysis. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2017-220980 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology described in Patent Document 1 requires acquisition of actual values ​​of solar radiation to predict solar radiation by regression analysis. However, there are many areas where actual values ​​of solar radiation cannot be acquired, and thus installation of a pyranometer is required to acquire the solar radiation amount for a desired area, which increases costs. Therefore, a technology for accurately estimating power demand without using actual values ​​of solar radiation is desired.

[0005] The present disclosure has been made in consideration of the above, and aims to provide a learning device that can achieve accurate inference of power demand without using actual values ​​of solar radiation. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems and achieve the objectives, the learning device according to the present disclosure includes an acquisition unit that acquires actual values ​​of power demand and date and time information indicating the date and time corresponding to the actual values; a feature generation unit that uses the date and time information to generate features including an index related to the altitude of the sun and information indicating whether or not it is a holiday; and a learning model generation unit that uses the actual values, which are ground truth data, and the features corresponding to the actual values ​​to generate a trained model by machine learning for inferring power demand from the features. [Effects of the Invention]

[0007] The learning device according to the present disclosure has the effect of being able to achieve accurate inference of power demand without using actual values ​​of solar radiation. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating a configuration example of a power system control system including a learning device according to a first embodiment. [Figure 2] FIG. 1 is a diagram showing an example of customer information according to the first embodiment; [Figure 3] 1 is a flowchart showing an example of a processing procedure during learning in the learning device of the first embodiment. [Figure 4] An example of the theoretical solar radiation intensity at noon [Figure 5] 1 is a flowchart showing an example of a processing procedure for inferring power demand in the power demand inference device according to the first embodiment. [Figure 6] FIG. 10 is a diagram showing the verification results of inference using the trained model of the first embodiment. [Figure 7] FIG. 10 is a diagram showing the verification results of inference using the trained model of the first embodiment. [Figure 8] FIG. 1 is a diagram showing an example of the configuration of a computer system that realizes the learning device and the power demand inference device according to the first embodiment. [Figure 9]FIG. 10 is a diagram illustrating a configuration example of a power system control system including a learning device according to a second embodiment. [Figure 10] FIG. 10 is a diagram showing an example of actual values ​​of power demand used as feature quantities in the second embodiment. [Figure 11] 10 is a flowchart showing an example of a processing procedure for inferring power demand in the power demand inference device according to the second embodiment. [Figure 12] FIG. 10 shows the verification results of inference using the trained model of the second embodiment. [Figure 13] FIG. 10 shows the verification results of inference using the trained model of the second embodiment. [Figure 14] FIG. 13 is a diagram showing an example of actual values ​​of power demand used as feature quantities in the third embodiment. [Figure 15] FIG. 10 shows the verification results of inference using the trained model of the third embodiment. [Figure 16] FIG. 10 shows the verification results of inference using the trained model of the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] A learning device, a power demand inference device, a grid control system, a supply and demand control system, a facility formation support system, and a program according to embodiments will be described in detail below with reference to the accompanying drawings.

[0010] Embodiment 1 FIG. 1 is a diagram illustrating an example of the configuration of a power grid control system including a learning device according to a first embodiment. A power grid control system 4 of this embodiment includes a learning device 1, a power demand inference device 2, and a control device 3. The learning device 1 generates a trained model for predicting power demand through machine learning using the "theoretical solar radiation intensity at meridian time," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday" as feature quantities and the power demand as correct answer data. The power demand inference device 2 infers power demand by inputting the feature quantities of the day to be inferred into the trained model generated by the learning device 1. The control device 3 uses the power demand inferred by the power demand inferred device 2 to control current and voltage, for example, using power grid control equipment or distributed power sources.

[0011] The power demand given as the correct data is, for example, a measurement value by a power metering device called a smart meter, or a value calculated from the measurement value of the metering device. Specifically, for example, for a consumer without a power generation facility, the power consumed by the load is measured by the metering device, and therefore the measurement value of the metering device can be used as the correct data. Furthermore, for a consumer under a contract for a full-volume purchase system (hereinafter referred to as a full-volume purchase contract), power generation output and power demand are measured separately, and therefore the measured power demand can be used as the correct data. On the other hand, for a consumer under a contract for a surplus power purchase system (hereinafter referred to as a surplus power purchase contract), the metering device measures the residual demand, which is the power demand minus the power generation output, and therefore the power demand obtained by adding the measured residual demand to the power generation output can be used as the correct data. The power generation output used here may be a value measured separately from the metering device, or a value estimated by any method. The power demand given as the correct data is not limited to the above-mentioned example, and may be an actual value measured by other means or calculated by another method.

[0012] The date to be inferred when the power demand inference device 2 performs inference may be in the future or in the past. That is, the power demand inference device 2 may predict power demand using a trained model, or may use a trained model to estimate current or past power demand. In the example shown in FIG. 1 , the date to be inferred is, for example, in the future because the model is used to control at least one of the voltage and current of the power grid. Alternatively, when the control device 3 calculates power demand using measured values ​​of voltage, current, or power energy in the power grid and performs control using the calculated power demand, the power demand inferred by the power demand inference device 2 may be used as a substitute for the current value of the measurement result when a measurement is missing.

[0013] FIG. 1 illustrates an example in which the learning device 1 and the power demand inference device 2 are included in a grid control system 4, and the power demand inferred by the power demand inference device 2 is used for grid control. However, the present invention is not limited to this example. The inferred power demand may also be used, for example, for supply and demand control, facility formation, etc. Furthermore, the power demand inferred by the power demand inference device 2 may be used for controlling a virtual power plant (VPP) or a demand response (DR), or for energy management at a consumer's facility. When used for controlling a VPP or DR, the power demand to be inferred and the correct answer data may be the total value of the power demands of multiple consumers. Furthermore, when the power demand inferred by the power demand inference device 2 is used for energy management at a consumer's facility, for example, an energy management device at the consumer's facility may display the estimated power demand, or may use the estimated power demand when controlling a distributed power source (such as a consumer's storage battery).

[0014] When the power demand inferred by the power demand inference device 2 is used for supply and demand control, for example, a supply and demand control system includes the learning device 1 and the power demand inference device 2, as well as a supply and demand control device that controls supply and demand, and the supply and demand control device controls the power supply and demand using the power demand predicted by the power demand inference device 2. When the power demand inferred by the power demand inference device 2 is used to support facility planning, a facility planning support system includes the learning device 1 and the power demand inference device 2, as well as a state management device that manages the state of the power system, and the state management device manages the state of the power system using the power demand inferred by the power demand inference device 2 and presents the state of the power system to an operator, thereby supporting facility planning. For example, the state management device may determine the future state of the power system using the power demand predicted by the power demand inference device 2, or may estimate the past state of the power system using the power demand inferred by the power demand inference device 2 in place of missing values ​​from a metering device. That is, the inference results by the power demand inference device 2 may be used to fill in missing values.

[0015] Furthermore, a trained model generated by the learning device 1 may be used by multiple power demand inference devices 2. For example, one learning device 1 may be provided, and a power demand inference device 2 that stores the trained model generated by the learning device 1 may be provided in each of the supply and demand control system and the facility formation support system.

[0016] Next, we will explain configuration examples of the learning device 1 and the power demand inference device 2. As shown in Figure 1, the learning device 1 includes an acquisition unit 11, a feature generation unit 12, a learning model generation unit 13, a model storage unit 14, and a data storage unit 15.

[0017] The acquisition unit 11 acquires an actual value of power demand and date and time information indicating the date and time corresponding to the actual value. Specifically, the acquisition unit 11 acquires power demand data, which is power demand with a time stamp, and stores the acquired power demand data in the data storage unit 15. The time stamp is date and time information indicating the date and time corresponding to the power demand data, which is the actual value of power demand. In the example described below, the actual value of power demand is assumed to be the measured value of the metering device or the power demand calculated using the measured value of the metering device. The time stamp is assigned, for example, at the time of measurement by the metering device. However, this is not limited to this, and the time stamp may be assigned by another device, such as a server that collects measured values, at the time the other device receives the measured value. The acquisition unit 11 acquires power demand data, for example, from a device not shown. In this example, it is assumed that the metering device transmits measured values ​​at 30-minute intervals, and the time-stamped power demand value is also obtained every 30 minutes. In addition, when the total value of the electricity demand of multiple consumers is the subject of inference, the electricity demand data acquired by the acquisition unit 11 may itself be the total value of the electricity demand of multiple consumers, or the electricity demand data acquired by the acquisition unit 11 may be on a consumer-by-consumer basis, and the feature generation unit 12 may calculate the total value of the electricity demand of multiple consumers using the consumer data stored in the data storage unit 15.

[0018] The feature generation unit 12 generates feature amounts using timestamps assigned to the power demand data stored in the data storage unit 15, and stores the generated feature amounts in association with the power demand data in the data storage unit 15. Specifically, the feature generation unit 12 calculates the "theoretical solar radiation intensity at noon," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday" from the timestamps. The method for calculating feature amounts will be described later.

[0019] The learning model generation unit 13 generates a trained model for inferring power demand from the feature quantities by machine learning using actual power demand values, which are correct data, and feature quantities corresponding to the actual values. Specifically, the learning model generation unit 13 generates a trained model by machine learning using the feature quantities stored in the data storage unit 15 and power demand data, which is correct data corresponding to the feature quantities, and stores the generated trained model in the model storage unit 14. A trained model is created for each time slot in a day. For example, a trained model is generated every 30 minutes, which is the same as the measurement value acquisition cycle. That is, 48 ​​trained models are generated, such as a trained model corresponding to midnight, a trained model corresponding to 12:30, a trained model corresponding to 12:00, and so on. For example, the trained model corresponding to 12:30 is generated using power demand data whose timestamp is 12:30. Furthermore, each measurement value is, for example, an integrated value for 30 minutes from the time indicated by the timestamp, but is not limited to this, and may be an integrated value for 30 minutes centered on the time indicated by the timestamp, or an integrated value for 30 minutes from 30 minutes before the time indicated by the timestamp to the time indicated by the timestamp, and the relationship between the timestamp and the corresponding 30-minute integration period is not limited to this example. Note that, although an example in which the time unit for creating a trained model is 30 minutes will be described here, the time unit for creating a trained model is not limited to 30 minutes.

[0020] Examples of machine learning algorithms that can be used to generate a trained model include random forests, regression trees, multiple regression, support vector regression, and neural networks, but are not limited to these and any regression-type algorithm may be used.

[0021] The model storage unit 14 stores trained models. The model storage unit 14 may also store trained models that are currently being trained. The data storage unit 15 stores consumer information including the latitude of the consumer, power demand data, and feature quantities. The latitude of the consumer is used to calculate the "theoretical solar irradiance at the time of noon," as described below.

[0022] FIG. 2 is a diagram showing an example of consumer information according to the present embodiment. The consumer information is created, for example, based on contract information indicating the contract details and address of each consumer. In the example shown in FIG. 2, the consumer information includes the contract power, address, and latitude for each consumer. The latitude is calculated in advance based on the address and stored as consumer information. Here, an example in which the latitude is calculated in advance is described, but the latitude may also be calculated based on the address when generating the feature. Furthermore, the example shown in FIG. 2 is merely an example, and the consumer information may include the latitude of the consumer or information necessary to calculate the latitude of the consumer, and the configuration of the consumer information is not limited to the example shown in FIG. 2.

[0023] Returning to the explanation of Fig. 1, the power demand inference device 2 includes a model storage unit 21, an acquisition unit 22, a feature generation unit 23, an inference unit 24, an output unit 25, and a data storage unit 26.

[0024] The model storage unit 21 stores a trained model generated by the learning device 1. This trained model may be stored in advance in the model storage unit 21, or a communication unit (not shown) of the power demand inference device 2 may receive this trained model from the learning device 1 and store it in the model storage unit 21.

[0025] The acquisition unit 22 acquires inference target date and time information indicating the target date and time for inferring power demand, and outputs the acquired inference target date and time information to the feature generation unit 23 and the inference unit 24. The acquisition unit 22 may acquire the inference target date and time information by receiving input from an operator or the like, or may acquire the inference target date and time information by receiving the inference target date and time information from another device. The feature generation unit 23 generates features using the inference target date and time information in the same way as the feature generation unit 12, and stores them in the data storage unit 26.

[0026] The inference unit 24 infers the power demand by inputting the feature quantities stored in the data storage unit 26 into the trained model stored in the model storage unit 21, and stores the inference result in the data storage unit 26. When predicting power demand, the inference result is a predicted value of power demand, and when used to complement missing values, the inference result is an estimated value of power demand corresponding to the date and time when the missing value occurred. The output unit 25 reads out the inference result stored in the data storage unit 26 and transmits the read inference result to the control device 3. The output unit 25 may have a display function and display the inference result.

[0027] The data storage unit 26 stores the consumer information, the feature amount, and the inference result. The consumer information is the same as the consumer information stored in the data storage unit 15 of the learning device 1.

[0028] 1, the learning device 1 and the power demand inference device 2 are provided separately, but this is not limiting, and the learning device 1 and the power demand inference device 2 may be integrated by the power demand inference device 2 having the functions of the learning device 1. In this case, a learning model generation unit 13 is added to the power demand inference device 2, and the learning model generation unit 13 stores a learned model in a model storage unit 21. Also, in this case, the acquisition unit 11, the feature generation unit 12, and the data storage unit 15 may be added to the power demand inference device 2, or the acquisition unit 22, the feature generation unit 23, and the data storage unit 26 may also have the functions of the acquisition unit 11, the feature generation unit 12, and the data storage unit 15, respectively, so that the acquisition unit 22, the feature generation unit 23, and the data storage unit 26 may be shared between learning and inference.

[0029] Next, the operation of this embodiment will be described. Fig. 3 is a flowchart showing an example of a processing procedure during learning in the learning device 1 of this embodiment. As shown in Fig. 3, the learning device 1 acquires power demand data (step S1). In detail, the acquisition unit 11 acquires the power demand data with a timestamp and stores it in the data storage unit 15.

[0030] Next, the learning device 1 generates features (step S2). In detail, the feature generator 12 calculates the following features: "theoretical solar radiation intensity at noon," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday" as follows:

[0031] For example, the feature generator 12 stores calendar information including, for each date, the day of the week and information indicating whether the date is a public holiday, and calculates the "day of the week (holidays are treated separately)" and "weekday / weekend / holiday type" for the day corresponding to the timestamp using the calendar information. The calendar information may be information indicating a calendar showing general days of the week and public holidays, information indicating a calendar reflecting local holidays, or information indicating a calendar reflecting working days and holidays specific to a consumer, such as a company or business. For the "day of the week (holidays are treated separately)," the feature generator 12, for example, calculates the day of the week corresponding to each timestamp, extracts public holidays, and sets the "day of the week (holidays are treated separately)" corresponding to the timestamp extracted as a public holiday as the "holiday," and for times other than public holidays, uses the day of the week as the "day of the week (holidays are treated separately)." For the "month," the feature generator 12 extracts the portion corresponding to the "month" from the timestamp.

[0032] Furthermore, the feature generating unit 12 calculates the "theoretical solar radiation intensity at the time of noon" corresponding to the power demand data based on the time stamp and the latitude of the customer included in the customer information stored in the data storage unit 15. Specifically, for example, the feature generating unit 12 calculates S, which is the "theoretical solar radiation intensity at the time of noon," by the following formula described in "Meteorology of Water Environments" (edited by Junsei Kondo, Asakura Publishing, 1994): t Calculate I00 is the solar constant (1365 W / m 2 ), θ is the zenith angle, φ is latitude, δ is the solar declination, h is the hour angle from the solar noon, M is the month, DAY is the day, HOUR is the hour, H n is the time of noon. S t =I 00 (d0 / d) 2 cosθ cosθ=sinφsinδ+cosφcosδcosh (d0 / d) 2 =1.00011+0.034221cosη +0.00128sinη+0.000719cos2η +0.000077sin2η δ=sin -1 (0.398sina2) a2=4.871+η+0.033sinη η=(2π / 365)·i i=30.36·(M-1)+DAY h=(HOUR-H n )·15°

[0033] At noon, HOUR=H n Therefore, h=0. In the above equation, I 00 is a constant, and (d0 / d) 2 is a value determined by the date, and cosθ is a value determined by the date and latitude. Therefore, by inputting the date and latitude to be calculated, the "theoretical solar radiation intensity at noon" can be calculated.

[0034] FIG. 4 is a diagram showing an example of the "theoretical solar irradiance intensity at meridian time." In FIG. 4, the horizontal axis represents the date and the vertical axis represents the solar irradiance intensity. As such, the "theoretical solar irradiance intensity at meridian time" changes over a one-year cycle, being approximately the same in spring and autumn, and being higher in summer than in spring and autumn and lower in winter than in spring and autumn. In this embodiment, the "theoretical solar irradiance intensity at meridian time" can be calculated using the latitude of the consumer and the month and date, which are information stored internally. This eliminates the need to obtain information from an external source or install a pyranometer. During inference, which will be described later, feature values ​​can be calculated from the inference target date and time information and latitude in a similar manner. This allows power demand to be inferred at low cost.

[0035] Furthermore, instead of using the latitude based on the address of each consumer, the same latitude may be used for each area. For example, a representative latitude may be determined for each area, such as a prefecture or a city, town, or village, and the same representative latitude may be used for consumers within the same area. The representative latitude may be the geometric center of the area or the location of a government building, such as a prefectural office. Furthermore, if the geographical range in which the consumers to be inferred are located is defined and the difference in latitude between the consumers is small, a predetermined representative latitude may be used regardless of the consumer, rather than using the latitude for each consumer. For example, when estimating the power demand of consumers within the same prefecture, a representative latitude, such as the prefectural capital, may be used.

[0036] Furthermore, when the total value of the power demands of multiple consumers is the target of inference, the feature generator 12 may calculate the "theoretical irradiance intensity at the time of meridian" using the centroid of the latitudes of the multiple consumers. Alternatively, the feature generator 12 may calculate the "theoretical irradiance intensity at the time of meridian" using the centroid calculated by weighting the latitudes of the multiple consumers by the contracted power of each consumer.

[0037] In this embodiment, an example will be described in which the feature quantity includes the "theoretical solar irradiance intensity at meridian time." However, the feature quantity is not limited to the "theoretical solar irradiance intensity at meridian time." A quantity other than the "theoretical solar irradiance intensity at meridian time" may be used instead of the "theoretical solar irradiance intensity at meridian time" as long as its annual change is similar to the "theoretical solar irradiance intensity at meridian time." That is, it is sufficient to use an index related to the altitude of the sun as the feature quantity. The index related to the altitude of the sun may be the "theoretical solar irradiance intensity at meridian time," the theoretical solar irradiance at a time offset by a certain time from meridian time, the solar altitude at meridian time, the length of time from sunrise to sunset, or at least one of the sunrise time and the sunset time. These values ​​may also be used in combination. For example, the feature quantity may include the "theoretical solar irradiance intensity at meridian time," the sunrise time, and the sunset time.

[0038] In addition, here, an example is described in which the “theoretical solar irradiance at meridian time,” “month,” “day of the week (holidays are treated separately),” and “weekday / Saturday / Sunday / holiday type” are used as feature quantities. However, the feature quantities may include an index related to the solar altitude and information related to the day of the week, and “month” need not be used as a feature quantity. Both “day of the week (holidays are treated separately)” and “weekday / Saturday / Sunday / holiday type” are information related to the day of the week, and the information related to the day of the week may include both “day of the week (holidays are treated separately)” and “weekday / Saturday / Sunday / holiday type,” or either one of them. For example, “theoretical solar irradiance at meridian time” and “day of the week (holidays are treated separately)” may be used as feature quantities. That is, the feature generator 12 may generate feature quantities including an index related to the solar altitude and information indicating the day of the week using a timestamp. The feature quantities may further include “month” extracted from the timestamp, and may further include at least one of the temperature and humidity at the date and time corresponding to the actual power demand value. If the features generated during learning include at least one of the temperature and humidity of the date and time corresponding to the actual value of electricity demand, the features generated during inference will also similarly include at least one of the temperature and humidity of the date and time indicated by the inference target date and time information.

[0039] Furthermore, the feature quantities may further include at least one of temperature and humidity. Because temperature and humidity are easier to obtain than the amount of solar radiation, adding at least one of these is expected to result in only a minor increase in cost.

[0040] Returning to the explanation of Figure 3, the learning device 1 then executes machine learning processing, i.e., generation of a trained model (step S3), and terminates the processing. In detail, in step S3, the trained model generation unit 13 generates a trained model through machine learning using the feature amounts stored in the data storage unit 15 and the power demand data that is the correct data corresponding to the feature amounts, and stores the generated trained model in the model storage unit 14. As described above, a trained model is generated for each time slot. For example, 48 trained models corresponding to each time slot in 30-minute increments are generated.

[0041] Through the above process, a trained model for inferring power demand is generated. For example, the learning device 1 can generate a trained model that can accurately infer power demand for any season by performing training using actual power demand values ​​with time stamps for one year or more.

[0042] Next, a description will be given of the inference in the power demand inference device 2 of this embodiment. Fig. 5 is a flowchart showing an example of a processing procedure when inferring power demand in the power demand inference device 2 of this embodiment.

[0043] The power demand inference device 2 generates features (step S11). More specifically, the acquisition unit 22 acquires inference target date and time information indicating the date and time of the inference target, and outputs the acquired inference target date and time information to the feature generation unit 23. The feature generation unit 23 uses the inference target date and time information and the latitude of the consumer information contained in the data storage unit 26 to calculate the "theoretical solar radiation intensity at noon," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday" as features, similar to the feature generation unit 12 of the learning device 1.

[0044] Next, the power demand inference device 2 executes a power demand inference process (step S12). In detail, the inference unit 24 acquires an inference value, which is an inference result of the power demand, by inputting the feature into the trained model generated by the learning device 1 and stored in the model storage unit 21, i.e., the trained model for inferring power demand from the feature including an index related to the solar altitude and information indicating the day of the week, and stores the acquired inference value as inference data (power demand inference data) in the data storage unit 26. More specifically, the inference unit 24 selects, from the 48 trained models stored in the model storage unit 21, a trained model that corresponds to the inference target date and time information input from the acquisition unit 22, and acquires the inference value by inputting the feature into the selected trained model.

[0045] Next, the power demand inference device 2 transmits the power demand inference data (step S13) and ends the process. More specifically, in step S13, the output unit 25 transmits the inference data stored as the inference result in the data storage unit 26 to the control device. When predicting power demand, inference target date and time information indicating the date and time to be predicted is input in step S11, and a predicted value corresponding to the date and time to be predicted is obtained as the inference result in step S12. When estimating a value to replace a missing value, inference target date and time information indicating the date and time corresponding to the missing value is input in step S11, and estimated data corresponding to the missing value is obtained as the inference result in step S12. In this embodiment, even if there is missing data for one week, for example, the power demand inference result can be used as a substitute for the missing value for one week.

[0046] Next, the effects of this embodiment will be described. FIGS. 6 and 7 are diagrams showing verification results of inference using the trained model of this embodiment. In FIGS. 6 and 7, the horizontal axis represents date, and the vertical axis represents active power. In addition, in FIGS. 6 and 7, the dashed lines represent predicted values ​​predicted by inference using the trained model of this embodiment, and the solid lines represent actual measured values ​​at the corresponding dates and times. FIG. 6 shows a comparison between predicted values ​​and actual values ​​in autumn, and FIG. 7 shows a comparison between predicted values ​​and actual values ​​in winter. As shown in FIGS. 6 and 7, the predicted values ​​predicted by inference using the trained model of this embodiment are nearly identical to the actual measured values. In this way, this embodiment can generate a highly accurate trained model for inferring power demand. This makes it possible to achieve highly accurate inference of power demand without using actual values ​​of solar radiation.

[0047] Next, the hardware configurations of the learning device 1 and the power demand inference device 2 of this embodiment will be described. In the learning device 1 of this embodiment, a computer program describing the processing to be performed by the learning device 1 is executed on a computer system, causing the computer system to function as the learning device 1. Similarly, in the power demand inference device 2 of this embodiment, a computer program describing the processing to be performed by the power demand inference device 2 is executed on a computer system, causing the computer system to function as the power demand inference device 2. FIG. 8 is a diagram showing an example configuration of a computer system that realizes the learning device 1 and the power demand inference device 2 of this embodiment. As shown in FIG. 8, this computer system includes a control unit 101, an input unit 102, a memory unit 103, a display unit 104, a communication unit 105, and an output unit 106, which are connected via a system bus 107.

[0048] In FIG. 8 , the control unit 101 is a processor such as a CPU (Central Processing Unit) that executes a program describing the processing of the learning device 1 or the power demand inference device 2 according to this embodiment. Note that a portion of the control unit 101 may be realized by dedicated hardware such as a GPU (Graphics Processing Unit) or an FPGA (Field-Programmable Gate Array). The input unit 102 is configured, for example, by a keyboard, a mouse, etc., and is used by a user of the computer system to input various pieces of information. The storage unit 103 includes various types of memory, such as RAM (Random Access Memory) and ROM (Read Only Memory), and a storage device, such as a hard disk, and stores programs to be executed by the control unit 101, necessary data obtained during processing, etc. The storage unit 103 is also used as a temporary storage area for programs. The display unit 104 is configured by a display, an LCD (Liquid Crystal Display Panel), etc., and displays various screens to the user of the computer system. The communication unit 105 is a receiver and transmitter that executes communication processing. The output unit 106 is a printer, a speaker, etc. Note that FIG. 8 is just an example, and the configuration of the computer system is not limited to the example of FIG.

[0049] Here, an example of the operation of the computer system until the program of this embodiment is ready to be executed will be described. In the computer system having the above configuration, for example, a computer program is installed in storage unit 103 from a CD-ROM or DVD-ROM inserted in a CD (Compact Disc)-ROM drive or DVD (Digital Versatile Disc)-ROM drive (not shown). Then, when the program is executed, the program read from storage unit 103 is stored in the main storage area of ​​storage unit 103. In this state, control unit 101 executes processing as learning device 1 or power demand inference device 2 of this embodiment in accordance with the program stored in storage unit 103.

[0050] In the above explanation, a program describing the processing in the learning device 1 or the power demand inference device 2 is provided on a CD-ROM or DVD-ROM as a recording medium, but this is not limited to this. Depending on the configuration of the computer system, the capacity of the program to be provided, etc., it is also possible to use a program provided via a transmission medium such as the Internet via the communication unit 105.

[0051] The program of this embodiment causes a computer system to execute, for example, the steps of acquiring an actual value of power demand and date and time information indicating a date and time corresponding to the actual value, generating features including an index related to the solar altitude and information indicating the day of the week using the date and time information, and generating a trained model for inferring power demand from the features by machine learning using the actual value, which is ground truth data, and the features corresponding to the actual value. Furthermore, the program of this embodiment causes a computer system to execute, for example, the steps of acquiring inference target date and time information indicating a date and time for inferring power demand from the features including the index related to the solar altitude and the information indicating the day of the week using the inference target date and time information, and inferring power demand by inputting the generated features into a trained model for inferring power demand from the features including the index related to the solar altitude and the information indicating the day of the week.

[0052] The feature generation unit 12 and the learning model generation unit 13 shown in FIG. 1 are realized by the control unit 101 shown in FIG. 8 executing a computer program stored in the storage unit 103 shown in FIG. 8. The storage unit 103 shown in FIG. 8 is also used to realize the feature generation unit 12 and the learning model generation unit 13 shown in FIG. 1. The acquisition unit 11 shown in FIG. 1 is realized by the communication unit 105 shown in FIG. 8. Furthermore, when the acquisition unit 11 shown in FIG. 1 receives input from an operator, the acquisition unit 11 is realized by the input unit 102 shown in FIG. 8. The model storage unit 14 and the data storage unit 15 shown in FIG. 1 are part of the storage unit 103 shown in FIG. 8. Furthermore, the learning device 1 may be realized by multiple computer systems. For example, the learning device 1 may be realized by a cloud computer system.

[0053] The feature generation unit 23 and the inference unit 24 shown in FIG. 1 are realized by the control unit 101 shown in FIG. 8 executing a computer program stored in the storage unit 103 shown in FIG. 8. The storage unit 103 shown in FIG. 8 is also used to realize the feature generation unit 23 and the inference unit 24 shown in FIG. 1. The acquisition unit 22 and the output unit 25 shown in FIG. 1 are realized by the communication unit 105 shown in FIG. 8. If the acquisition unit 22 shown in FIG. 1 accepts input from an operator, the acquisition unit 22 is realized by the input unit 102 shown in FIG. 8. If the output unit 25 has a display function, the output unit 25 is realized by the display unit 104 shown in FIG. 8. The model storage unit 21 and the data storage unit 26 shown in FIG. 1 are part of the storage unit 103 shown in FIG. 8. The power demand inference device 2 may be realized by multiple computer systems. For example, the power demand inference device 2 may be realized by a cloud computer system. The learning device 1 and the power demand inference device 2 may be realized by a single computer system.

[0054] As described above, the learning device 1 of this embodiment uses an index related to the solar altitude and information related to the day of the week as features and actual values ​​of power demand as ground truth data to generate a trained model for inferring power demand from the features through machine learning. The power demand inference device 2 of this embodiment generates features using inference target date and time information indicating the date and time of the inference target, and infers power demand by inputting the generated features into the trained model. This makes it possible to accurately infer power demand without using actual values ​​of solar radiation.

[0055] Embodiment 2 Next, a second embodiment will be described. FIG. 9 is a diagram showing an example of the configuration of a power system control system including a learning device according to the second embodiment. A power system control system 4a according to the present embodiment includes a learning device 1, a power demand inference device 2a, and a control device 3. The control device 3 is the same as that of the first embodiment. The configuration of the learning device 1 is the same as that of the first embodiment. The configuration of the power demand inference device 2a is the same as that of the first embodiment, except that it includes a data storage unit 26a instead of the data storage unit 26. The purpose of using the inference results obtained by the learning device 1 and the power demand inference device 2a is also the same as that of the first embodiment. Below, differences from the first embodiment will be mainly described, and explanations that overlap with those of the first embodiment will be omitted.

[0056] In this embodiment, the actual power demand value of the day before the target inference date is added to the feature values ​​of the first embodiment. FIG. 10 is a diagram showing an example of the actual power demand value used as a feature value in this embodiment. In the example shown in FIG. 10, at the time of inference by the power demand estimation device 2a, i.e., at the time of prediction, 48 actual power demand values ​​(hereinafter also referred to as actual demand) of the day before the target prediction date, which is the target inference date, are added as feature values. Therefore, at the time of learning by the learning device 1, the actual power demand value of the day before the day corresponding to the power demand used as the correct answer data is added as a feature value. That is, in this embodiment, the feature values ​​generated during learning further include the actual power demand value of the day before the day corresponding to the timestamp added to the actual power demand value. Note that, here, an example is described in which the actual power demand value of the day before is used, but the actual power demand values ​​of two days, the day before and the day before that, may also be used. In this case, the actual power demand values ​​for N days (N is an integer equal to or greater than 1) including the day before may be used as feature values. Furthermore, the actual power demand value for a period shorter than one day may also be used as feature values. In this way, the length (period) of the actual value of power demand used as a feature amount is not limited to one day, but may be longer or shorter than one day.

[0057] In this embodiment, during learning, in step S2 of the process shown in FIG. 3, the feature generator 12 generates, as features, "theoretical solar radiation intensity at meridian time," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday," as well as "actual value of power demand on the previous day" (the actual value of power demand on the previous day before the actual value of power demand used as correct answer data). In detail, the feature generator 12 generates the "actual value of power demand on the previous day" as a feature by reading the power demand data for the previous day from the data storage unit 15. Furthermore, in step S3 shown in FIG. 3, the learning model generator 13 generates a trained model by machine learning using, as features, "theoretical solar radiation intensity at meridian time," "month," "day of the week (holidays are treated separately)," and "type of weekday / weekend / holiday," as well as "actual value of power demand on the previous day." Operations during learning other than those described above are the same as those in the first embodiment.

[0058] FIG. 11 is a flowchart showing an example of a processing procedure for inferring power demand in the power demand inference device 2a of this embodiment. The power demand inference device 2a acquires power demand data (step S21). Specifically, the acquisition unit 22 acquires not only the estimation target date and time information but also the power demand data for the day before the estimation target date and time, and stores the acquired power demand data in the data storage unit 26a. In step S11, the feature generation unit 23 of the power demand inference device 2a generates features. At this time, the generated features are the features described in embodiment 1 plus the power demand data for the day before the estimation target date. In step S12, as in embodiment 1, the inference unit 24 inputs the features into the trained model stored in the model storage unit 21 to calculate a predicted value, which is the inference result. At this time, the features have been added with the power demand data for the day before the estimation target date as described above. Step S13 is the same as in embodiment 1.

[0059] In the above example, the power demand for the next day (the day to be inferred) is predicted by adding the power demand of the previous day as a feature, but the processing of this embodiment can also be applied to missing value imputation. When performing missing value imputation, the power demand inference device 2a can estimate the power demand corresponding to the missing value by using the actual value of the power demand of the previous day corresponding to the missing value as a feature.

[0060] 12 and 13 are diagrams showing the verification results of inference using the trained model of this embodiment. In FIGS. 12 and 13, the horizontal axis represents the date, and the vertical axis represents the active power. In addition, in FIGS. 12 and 13, the dashed lines represent the predicted values ​​predicted by inference using the trained model of this embodiment, and the solid lines represent the actual measured values ​​at the corresponding dates and times. FIG. 12 shows a comparison between the predicted values ​​and actual values ​​in autumn, and FIG. 13 shows a comparison between the predicted values ​​and actual values ​​in winter. As shown in FIGS. 12 and 13, the predicted values ​​predicted by inference using the trained model of this embodiment are nearly identical to the actual measured values. In this way, this embodiment can generate a highly accurate trained model for inferring power demand without using actual values ​​of solar radiation.

[0061] As described above, in this embodiment, the actual value of the power demand of the previous day is added to the feature quantity. This not only provides the same effect as in the first embodiment, but also makes it possible to improve the accuracy of the inference result compared to the first embodiment when the demand of the previous day is related to or similar to the demand at the target estimation date and time. Furthermore, it is possible to improve the accuracy of the inference result compared to the first embodiment when there is a trend in power demand that is unique to the year to be inferred.

[0062] Embodiment 3 Next, a third embodiment will be described. The configurations of the learning device 1 and the power demand inference device 2a of this embodiment are the same as those of the second embodiment. The purpose of using the inference results obtained by the learning device 1 and the power demand inference device 2a is also the same as that of the first embodiment. Below, differences from the first or second embodiment will be mainly described, and explanations that overlap with the first or second embodiment will be omitted.

[0063] In the second embodiment, the actual value of the power demand of the previous day is added to the feature quantities in the first embodiment. However, in the present embodiment, the actual value of the power demand up to a certain time before the inference target time on the current day is added to the feature quantities in the first embodiment. The current day is the day corresponding to the timestamp added to the power demand data corresponding to the correct answer data during learning, and is the inference target day during inference. FIG. 14 is a diagram showing an example of the actual value of the power demand used as a feature quantity in the present embodiment. In the example shown in FIG. 14, during inference by the power demand estimation device 2a, i.e., during prediction, the actual value of the power demand from the point corresponding to midnight on the prediction target day to n points (n is an integer equal to or greater than 1) before the prediction target time is added as a feature quantity. In the example shown in FIG. 14, since the prediction target time is 12 o'clock, the actual demand from the point corresponding to midnight on the prediction target day to n points before 12 o'clock is used as a feature quantity. During learning by the learning device 1, the actual demand from the point corresponding to midnight on the day corresponding to the actual demand used as the correct answer data to n points before the prediction target time is added as a feature quantity. That is, in this embodiment, the features generated during learning further include the actual value of power demand from midnight to a certain time before the time corresponding to the timestamp added to the actual value of power demand used as the correct answer data on the day corresponding to the timestamp.

[0064] The learning process in the learning device 1 of this embodiment is the same as that of embodiment 2, except that the "actual value of power demand n points before the target time from the point corresponding to midnight on the current day" is used instead of the "actual value of power demand on the previous day."

[0065] In addition, the processing of the power demand inference device 2a of this embodiment is the same as that of embodiment 2, except that in the processing shown in Figure 11 of embodiment 2, the "actual value of power demand n points before the target time from the point corresponding to midnight on the day" is used instead of the "actual value of power demand on the previous day."

[0066] In the above example, the feature includes "the actual value of power demand from the point corresponding to midnight on the current day to n points before the target time," but is not limited to this. The feature may include "the actual value of power demand from point m (m is an integer equal to or greater than n) before point n before the target time." Point m before the target time may be the day before the target day, or may be after midnight on the target time. In other words, the feature may include the actual value of power demand on the current day, and the start point of the period corresponding to the actual value of power demand may be the current day or the previous day.

[0067] Furthermore, in the above example, an example was described in which the power demand was predicted by adding, as a feature, "the actual value of the power demand n points before the target time from the point corresponding to midnight on the current day," but the processing of this embodiment can also be applied to missing value imputation. When performing missing value imputation, the power demand inference device 2a can estimate the power demand corresponding to the missing value by using, as a feature, "the actual value of the power demand n points before the target time from the point corresponding to midnight on the current day."

[0068] 15 and 16 are diagrams showing the verification results of inference using the trained model of this embodiment. In FIGS. 15 and 16, the horizontal axis represents the date, and the vertical axis represents the active power. In addition, in FIGS. 15 and 16, the dashed lines represent the predicted values ​​predicted by inference using the trained model of this embodiment, and the solid lines represent the actual measured values ​​at the corresponding dates and times. FIG. 15 shows a comparison between the predicted values ​​and actual values ​​in autumn, and FIG. 16 shows a comparison between the predicted values ​​and actual values ​​in winter. As shown in FIGS. 15 and 16, the predicted values ​​predicted by inference using the trained model of this embodiment are nearly identical to the actual measured values. In this way, this embodiment can generate a highly accurate trained model for inferring power demand without using actual values ​​of solar radiation.

[0069] As described above, in this embodiment, the actual value of the power demand up to a certain time before the inference target time on the current day is added to the feature. This not only provides the same effect as in embodiment 1, but also makes it possible to improve the accuracy of the inference result compared to embodiment 1 when, for example, there is a trend in power demand that is unique to the inference target day. In other words, in this embodiment, by adding the actual value of the power demand up to a certain time before the inference target time on the current day as a feature, this is equivalent to correcting the inference value by looking at the immediately preceding power demand (performing intraday correction), and thereby makes it possible to improve the accuracy of the inference result compared to embodiment 1.

[0070] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other. It is also possible to omit or modify parts of the configurations as long as they do not deviate from the gist of the invention. [Explanation of symbols]

[0071] 1 Learning device, 2,2a Power demand inference device, 3 Control device, 4,4a Power system control system, 11,22 Acquisition unit, 12,23 Feature generation unit, 13 Learning model generation unit, 14,21 Model memory unit, 15,26,26a Data storage unit, 24 Inference unit, 25 Output unit.

Claims

1. an acquisition unit that acquires a performance value of the power demand and date and time information indicating a date and time corresponding to the performance value; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether or not the date and time information is a holiday, using the date and time information; a learning model generation unit that generates a learned model by machine learning using the actual value, which is correct data, and the feature value corresponding to the actual value, for inferring electricity demand from the feature value; A learning device comprising:

2. The learning device according to claim 1 , wherein the holidays are region-specific holidays or consumer-specific holidays.

3. an acquisition unit that acquires inference target date and time information indicating a date and time of an inference target for power demand; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday, using the inference target date and time information; an inference unit that infers power demand by inputting the feature generated by the feature generation unit into a trained model for inferring power demand from feature values ​​including an index related to the altitude of the sun and information indicating whether or not it is a holiday; An electric power demand inference device comprising:

4. The power demand inference device according to claim 3 , wherein the holidays are local holidays or consumer holidays.

5. an acquisition unit that acquires a performance value of the power demand and date and time information indicating a date and time corresponding to the performance value; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether or not the date and time information is a holiday, using the date and time information; a learning model generation unit that generates a learned model by machine learning using the actual value, which is correct data, and the feature value corresponding to the actual value, for inferring electricity demand from the feature value; an inference unit that infers power demand using the trained model; Equipped with the acquisition unit acquires inference target date and time information indicating a date and time of an inference target for power demand; the feature generating unit generates an inference feature including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday, using the inference target date and time information; The inference unit is an electric power demand inference device that infers the electric power demand by inputting the inference feature into the trained model.

6. an electric power demand inference device that infers electric power demand; a control device that controls a voltage of a power grid using the power demand inferred by the power demand inferring device; Equipped with The power demand inference device an acquisition unit that acquires inference target date and time information indicating a date and time of an inference target for power demand; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday, using the inference target date and time information; an inference unit that infers power demand by inputting the feature generated by the feature generation unit into a trained model for inferring power demand from feature values ​​including an index related to the altitude of the sun and information indicating whether or not it is a holiday; A grid control system equipped with:

7. an electric power demand inference device that infers electric power demand; a supply and demand control device that controls supply and demand of electricity using the electricity demand inferred by the electricity demand inference device; Equipped with The power demand inference device an acquisition unit that acquires inference target date and time information indicating a date and time of an inference target for power demand; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday, using the inference target date and time information; an inference unit that infers power demand by inputting the feature generated by the feature generation unit into a trained model for inferring power demand from feature values ​​including an index related to the altitude of the sun and information indicating whether or not it is a holiday; A supply and demand control system equipped with:

8. an electric power demand inference device that infers electric power demand; a state management device that presents a state of the power system using the power demand inferred by the power demand inference device; Equipped with The power demand inference device an acquisition unit that acquires inference target date and time information indicating a date and time of an inference target for power demand; a feature generating unit that generates a feature including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday, using the inference target date and time information; an inference unit that infers power demand by inputting the feature generated by the feature generation unit into a trained model for inferring power demand from feature values ​​including an index related to the altitude of the sun and information indicating whether or not it is a holiday; A facility formation support system equipped with:

9. In the computer system, acquiring a performance value of the power demand and date and time information indicating a date and time corresponding to the performance value; generating a feature amount including an index relating to the altitude of the sun and information indicating whether or not the date and time information is a holiday, using the date and time information; generating a trained model for inferring power demand from the feature amounts by machine learning using the actual value, which is correct data, and the feature amounts corresponding to the actual value; A program that executes the following.

10. In the computer system, acquiring inference target date and time information indicating a date and time of an inference target for power demand; generating a feature amount including an index relating to the altitude of the sun and information indicating whether the date and time is a holiday using the inference target date and time information; a step of inferring power demand by inputting the generated feature into a trained model for inferring power demand from feature including an index related to the altitude of the sun and information indicating whether it is a holiday; A program that executes the following.

Images