JP7828763B2 - Program, power demand forecasting method, and information processing device - Google Patents

Description

This invention relates to a technology for predicting electricity demand at target consumers.

Techniques for predicting electricity demand are known. For example, Patent Document 1 describes generating a demand fluctuation pattern from past demand data for the same day of the week as the prediction date, and then concatenating this generated demand fluctuation pattern with the current actual value to calculate a basic fluctuation pattern, which is the predicted pattern for electricity demand from the present onward.

Japanese Patent Publication No. 2016-19358

The technology described in Patent Document 1 predicts electricity demand based on information such as the day of the week on which electricity was consumed. However, for consumers that produce goods, there are factors other than the day of the week that influence electricity consumption. Therefore, if electricity demand for such consumers is predicted using the technology described in Patent Document 1, the accuracy of the electricity demand prediction may be low.

One of the objectives of this invention is to improve the accuracy of predicting electricity demand at target consumers.

One aspect of this disclosure provides a program for causing a computer to perform the following steps: acquiring first calendar information indicating a forecast period; acquiring production schedule information for a target consumer producing the product during the forecast period; and predicting the electricity demand of the target consumer during the forecast period by inputting the first calendar information and the production schedule information into a machine learning model generated by machine learning, using second calendar information indicating the date and time when electricity was consumed by at least one consumer and the actual production information of the product at the at least one consumer at the date and time as explanatory variables, and the amount of electricity consumed by the at least one consumer at the date and time as the dependent variable.

Another aspect of this disclosure provides a power demand forecasting method comprising: acquiring first calendar information indicating a forecast period; acquiring production schedule information for a target consumer producing the product during the forecast period; and forecasting by inputting the first calendar information and the production schedule information into a machine learning model generated by machine learning, using second calendar information indicating the date and time when power was consumed by at least one consumer and the actual production information of the product at the at least one consumer at the date and time as explanatory variables, and the power consumption of the at least one consumer at the date and time as the dependent variable.

A further aspect of this disclosure provides an information processing device comprising: a first acquisition means for acquiring first calendar information indicating a forecast period; a second acquisition means for acquiring production schedule information of a target consumer producing the product during the forecast period; and a forecast means for predicting the electricity demand of the target consumer during the forecast period by inputting the first calendar information and the production schedule information into a machine learning model generated by machine learning, using second calendar information indicating the date and time when electricity was consumed by at least one consumer and the actual production information of the product at the at least one consumer at the aforementioned date and time as explanatory variables, and the amount of electricity consumed by the at least one consumer at the aforementioned date and time as the dependent variable.

According to the present invention, the accuracy of predicting the amount of electricity demand at a target customer can be improved.

A diagram showing an overview of a power control system according to one embodiment. A diagram illustrating the functional configuration of a power control system. A diagram illustrating the server's hardware configuration. A sequence chart illustrating the operation of a power control system that generates machine learning models. A flowchart illustrating the operation of a server that predicts electricity demand.

1. Configuration Figure 1 is a diagram showing an overview of a power control system 1 according to one embodiment. The power control system 1 is a so-called FEMS (Factory Energy Management System). The power control system 1 predicts the amount of electricity demand for each customer H in order to optimize energy management at each customer H. A customer H refers to a person who receives and uses electricity. In one example, customer H is a specific customer that receives electricity at high voltage or extra-high voltage, specifically a facility such as a factory that produces goods. The goods may be chemical products, industrial products, or agricultural products. At customer H, goods are produced by consuming electricity. Therefore, the amount of electricity consumed at customer H fluctuates depending on the production status of the goods. The power control system 1 includes a power generation device 10, a plurality of power loads 20, an energy meter 30, a measuring instrument 40, a control device 50, and a server 60 (an example of an information processing device).

Of these components, the power generation device 10, multiple power loads 20, electricity meter 30, measuring instrument 40, and control device 50 are installed indoors or outdoors at customer H. For simplicity, Figure 1 shows only a single customer H, but the power control system 1 may include multiple customers H. Each customer H is also supplied with power from a commercial power source (not shown in Figure 1). This commercial power source is sold (or provided) by a power retailer. The server 60 is connected to customer H's control device 50 via a network 90. The control device 50 is connected to the power generation device 10, multiple power loads 20, electricity meter 30, and measuring instrument 40 via a network 95 such as a LAN (Local Area Network).

The power generation device 10 is a device that generates electricity. In one example, the power generation device 10 is a device that generates electricity using natural energy, specifically a solar power generation device or a wind power generation device. Alternatively, the power generation device 10 may be an emergency power generation device using a gas turbine or diesel engine. The power load 20 is a device that consumes electricity at the consumer H. The power load 20 includes a storage battery 21 and a general power load 22.

The battery 21 (an example of the first power load) is a device that is charged using power supplied from the power generation device 10 or commercial power source, and discharges the stored power as needed. The operation of the battery 21 is controlled by the control device 50 based on the operation plan created by the server 60. Since the battery 21 stops charging when fully charged, there is an upper limit to its power consumption. In one example, the battery 21 is an industrial battery, specifically including NAS batteries (registered trademark), lithium-ion batteries, lead-acid batteries, and nickel-metal hydride batteries.

The general power load 22 (an example of the second power load) is the power load 20 other than the battery 21 installed at customer H, and includes machinery and equipment used in the production of goods, heating and cooling systems, and lighting equipment, etc. The operation of the general power load 22 is not controlled by the control device 50, but operates, for example, according to user operation or a PLC (Programmable Logic Controller).

The electricity meter 30 is a device that measures the total amount of electricity consumed at customer H. The total amount of electricity consumed at customer H is the sum of the amount of electricity consumed by the battery 21 and the amount of electricity consumed by the general power load 22. In this example, the electricity meter 30 is a so-called smart meter. The electricity meter 30 supplies the measured total amount of electricity consumed to the control device 50.

The measuring instrument 40 is a device for measuring the power consumption of the storage battery 21. The power consumption of the storage battery 21 refers to, for example, the amount of electricity consumed by the storage battery 21. In one example, if the power consumption of the storage battery 21 per unit time is approximately constant, the measuring instrument 40 measures the power consumption of the storage battery 21 using the operating time of the storage battery 21 and the power consumption per unit time. The measuring instrument 40 supplies the measured power consumption of the storage battery 21 to the control device 50.

In this example, the measuring instrument 40 measures power consumption, but the measuring instrument 40 may also measure the operating time, and the control device 50 may calculate power consumption using the operating time and power consumption per unit time measured by the measuring instrument 40.

The control device 50 controls the power generation device 10, power load 20, energy meter 30, and measuring instrument 40 at customer H. For example, the control device 50 controls the operation of the battery 21 according to commands from the server 60. The control device 50 also transmits the history of power consumption at customer H to the server 60. The server 60 controls the power control system 1. For example, the server 60 creates an operation plan for the battery 21 and commands the control device 50 to operate the battery 21 according to this operation plan. Furthermore, the server 60 predicts the power demand at customer H based on the power consumption history received from the control device 50.

Figure 2 illustrates the functional configuration of the power control system 1. The power control system 1 includes a storage means 51, a transmission means 52, a storage means 61, an acquisition means 62, an acquisition means 63, an acquisition means 64, an acquisition means 65, a generation means 66, and a prediction means 67. Of these functional elements, the storage means 51 and the transmission means 52 are implemented in the control device 50. The storage means 61, the acquisition means 62, the acquisition means 63, the acquisition means 64, the acquisition means 65, the generation means 66, and the prediction means 67 are implemented in the server 60.

In the control device 50, the storage means 51 stores the history of the total power consumption at customer H as measured by the power meter 30, and the history of the power consumption of the storage battery 21 as measured by the measuring instrument 40. Each history includes the date and time when power was consumed (hereinafter also referred to as "power consumption date and time") and the amount of power consumed at that time. The transmission means 52 transmits the history of the total power consumption at customer H and the history of the power consumption of the storage battery 21, stored in the storage means 51, to the server 60.

In server 60, storage means 61 stores various types of data, including a customer database and calendar information. The customer database stores various information about each customer H. The customer database is used, for example, to identify the region to which customer H belongs and the input terminal (not shown) installed at customer H. In one example, the customer database stores the identification information of each customer H, the address of that customer H, and the address of the input terminal installed at that customer H, in association with each other. Calendar information includes, for example, information for this year's calendar. Calendar information includes, for example, month, day, day of the week, weekday, holiday, public holiday, beginning of the month, middle of the month, end of the month, week number, season, sunrise, sunset, and major events. Weekdays, holidays, and public holidays are examples of whether or not a day is a public holiday. Public holidays indicate that it is a public holiday, while weekdays and holidays indicate that it is not a public holiday. Beginning of the month, middle of the month, and end of the month are examples of whether or not it is the end of the month. End of the month indicates that it is the end of the month, while beginning and middle of the month indicate that it is not the end of the month.

The acquisition means 62 acquires the history of total power consumption at each customer H and the history of power consumption of the storage battery 21. The acquisition means 63 (an example of the first acquisition means) acquires calendar information indicating the target date and time. This calendar information includes calendar information indicating the date and time of power consumption at customer H (an example of the second calendar information) and calendar information indicating the forecast period (an example of the first calendar information). The forecast period is the period that serves as the unit of forecasting. Each piece of calendar information includes, for example, at least one of the following: month, day, time, day of the week, weekday, holiday, public holiday, beginning of the month, middle of the month, end of the month, week number, season, sunrise, sunset, and major events. In one example, the acquisition means 63 acquires the date and time of power consumption included in each history acquired by the acquisition means 62. The acquisition means 63 also acquires additional calendar information indicating this date and time of power consumption from the storage means 61. The additional calendar information includes calendar information other than date and time, such as at least one of the following: day of the week, weekdays, holidays, public holidays, beginning of the month, middle of the month, end of the month, week number, season, sunrise, sunset, and major events. Furthermore, the acquisition means 63 acquires the prediction period. Furthermore, the acquisition means 63 acquires additional calendar information indicating the prediction period from the storage means 61.

The acquisition means 64 acquires weather information for the region to which each customer H belongs at the target date and time. The weather information includes weather information for the region to which customer H belongs at the time of power consumption (an example of second weather information) and weather information for the region to which customer H belongs during the forecast period (an example of first weather information). Since the forecast period is a future period, the weather information for the forecast period does not represent actual weather, but rather a weather forecast. Each weather information includes at least one of the following: weather, temperature, humidity, precipitation, solar radiation, wind direction, and wind speed. In one example, the acquisition means 64 acquires weather information for the region to which customer H belongs at the time of power consumption, included in the history acquired by the acquisition means 62, from an external server operated by a weather information provider or organization. The acquisition means 64 also acquires weather information for the region to which customer H belongs during the forecast period from an external server operated by a weather information provider or organization.

The acquisition means 65 (an example of the second acquisition means) acquires production information for each customer H's products at the target date and time. Production information refers to information indicating the production status of the products. This production information includes actual production information for the products at the time of customer H's power consumption and planned production information for customer H's products during the forecast period. Each piece of production information includes at least one of the following: product type, production quantity, and production process. If the product is produced through multiple production processes, this production process refers to the production process actually performed at the target date and time. In one example, customer H creates actual production information and planned production information and inputs it into an input terminal installed at customer H. The acquisition means 65 acquires the actual production information and planned production information from this input terminal. Alternatively, customer H creates actual production information and planned production information and provides it to the administrator of server 60. The acquisition means 65 acquires the actual production information and planned production information input into server 60 through the administrator's operation.

The generation means 66 generates a machine learning model for each customer H by machine learning training data in which the power consumption amount of the general power load 22 at each customer H is the objective variable, and the calendar information indicating the date and time of power consumption for that customer H, weather information at the date and time of power consumption, and production performance information acquired by the acquisition means 63 to 65 are the explanatory variables. The generation means 66 uses the power consumption amount of the general power load 22 obtained by subtracting the power consumption amount of the storage battery 21 from the total power consumption amount at customer H acquired by the acquisition means 62 as the objective variable. The generation means 66 then learns the relationship between the power consumption amount of the general power load 22 at each customer H and the calendar information indicating the date and time of power consumption for that customer H, weather information at the date and time of power consumption, and production performance information. Through this machine learning, a machine learning model is generated that takes the calendar information, weather information, and production information as input and outputs a prediction of the power demand amount of the general power load 22 at customer H. The electricity demand for general power load 22 is the amount of electricity expected to be consumed by general power load 22, excluding the battery 21, from the power load 20 installed at customer H.

The reason for using the power consumption of the general power load 22 as the target variable is to improve the accuracy of power demand prediction. As described above, the battery 21's operation is controlled based on the operation plan created by the server 60. Since the battery 21's power demand is the amount of power consumed by operating according to this operation plan, prediction based on the actual operation plan yields higher accuracy than prediction using machine learning. Therefore, the generation means 66 uses the power consumption of the general power load 22 as the target variable to generate a machine learning model that predicts the power demand of the general power load 22 excluding the battery 21's power demand. This allows, for example, the power demand of the general power load 22 to be predicted using machine learning, while the power demand of the battery 21 is predicted individually based on the operation plan. As a result, the accuracy of the power demand prediction for customer H is improved.

The reason for using calendar information indicating the date and time of power consumption, weather information at the time of power consumption, and production performance information as explanatory variables is that this information influences the amount of power consumed by the general power load 22. For example, the amount of power consumed by the general power load 22 differs depending on the month, day, time, day of the week, whether it is a public holiday or not, and whether it is the end of the month or not. Furthermore, the amount of power consumed by the general power load 22 differs depending on the weather, temperature, etc. In addition, the amount of power consumed by the general power load 22 differs depending on the production volume, type of product, and production process of the product produced by customer H. Therefore, by performing machine learning using calendar information indicating the date and time of power consumption, weather information at the time of power consumption, and production performance information as explanatory variables, the accuracy of predicting the power demand of the general power load 22 can be improved.

The prediction means 67 inputs the calendar information indicating the prediction period, the second weather information, and the production schedule information acquired by the acquisition means 63-65 into the machine learning model for each customer H generated by the generation means 66, and predicts the amount of electricity demand for each customer H during the prediction period. However, the amount of electricity consumed by the storage battery 21 is not included in the amount of electricity demand for each customer H predicted by the prediction means 67.

Figure 3 illustrates the hardware configuration of server 60. Server 60 is a computer device comprising a CPU (Central Processing Unit) 601, memory 602, storage 603, and a communication interface (IF) 604. The CPU 601 is a control device that executes programs, performs various calculations, and controls other hardware elements of server 60. Memory 602 is the main memory, functioning as a work area when the CPU 601 executes programs. Storage 603 is a non-volatile auxiliary storage device that stores various programs and data. The communication interface 604 is a communication device that communicates with other devices according to a predetermined communication standard (e.g., Ethernet®). Furthermore, when production performance information and production schedule information are input to server 60 by an administrator, server 60 also has input devices such as a mouse and keyboard.

In this example, storage 603 stores a program (hereinafter referred to as the "server program") that causes the computer device to function as a server 60 in the power control system 1. The CPU 601 executes the server program, thereby implementing the functions shown in Figure 2 in the computer device. While the CPU 601 is executing the server program, at least one of memory 602 and storage 603 is an example of storage means 61. Furthermore, the CPU 601 is an example of acquisition means 62-65, generation means 66, and prediction means 67.

Although detailed illustrations are omitted, the control device 50 includes a microcomputer comprising a processor, memory, storage, and a communication interface. This storage stores a program (hereinafter referred to as the "control program") that causes the microcomputer to function as the control device 50 in the power control system 1. The microcomputer implements the functions shown in Figure 2 by executing the control program. While the processor is executing the control program, at least one of the memory and storage is an example of the storage means 51. The communication interface is an example of the transmission means 52.

2. Operation 2.1 Generation of Machine Learning Models Figure 4 is a sequence chart illustrating the operation of a power control system 1 that generates a machine learning model according to one embodiment. The sequence in Figure 4 is started, for example, when a reference time such as midnight arrives. In the following, functional elements such as the acquisition means 62 are described as the main processing elements, which means that hardware elements such as the CPU 601 that executes a program such as a server program work together with other hardware elements to perform the processing.

Prior to this process, the server 60 determines the target customer H (hereinafter referred to as the "target customer") from among the multiple customers H included in the power control system 1. The target customers are determined one by one in order according to a predetermined rule. Here, it is assumed that customer H with identification information "001" is determined as the target customer.

In step S101, the transmitting means 52 of the target consumer's control device 50 transmits the history of total power consumption stored in the storage means 51 and the history of power consumption of the storage battery 21 to the server 60, along with the identification information of the transmitting control device 50 (or target consumer). Each history entry shows, for example, the power consumption in 30-minute intervals. Each history entry includes the date and time when power was consumed and the amount of power consumed at that date and time. The acquiring means 62 of the server 60 receives these histories and identification information from the control device 50.

In step S102, the acquisition means 63 acquires calendar information indicating the date and time of power consumption included in the history acquired in step S101. Here, it is assumed that the date and time of power consumption included in the power consumption history for the target consumer is from 0:00 to 24:00 on December 1, 20XX. For example, the acquisition means 63 acquires the date and time of power consumption from 0:00 to 24:00 on December 1, 20XX as calendar information from the history acquired in step S101. Furthermore, the acquisition means 63 acquires additional calendar information from the storage means 61, indicating the day of the week for December 1, 20XX, whether it is a public holiday, or the end of the month if it is the end of the month.

In step S103, the acquisition means 64 acquires weather information for the region to which the target customer belongs at the time of power consumption included in the history acquired in step S101. For example, the acquisition means 64 identifies the address associated with the identification information "001" in the customer database stored in the storage means 61. Here, it is assumed that this address includes City X. In this case, the acquisition means 64 acquires weather information such as the weather, temperature, etc., for City X from an external server from 0:00 to 24:00 on December 1, 20XX.

In step S104, the acquisition means 65 acquires production performance information of the target customer's products at the power consumption date and time included in the history acquired in step S101. For example, the acquisition means 65 identifies the address of the input terminal associated with the identification information "001" in the customer database stored in the storage means 61. Using the identified address, the acquisition means 65 acquires production performance information of the target customer's products from the input terminal installed at the target customer for the period from 0:00 to 24:00 on December 1, 20XX. The production performance information includes the type of product produced at customer H at the power consumption date and time, the quantity produced, and the production process of the product carried out at that time.

In step S105, the generation means 66 calculates the power consumption of the general power load 22 for each time period by subtracting the power consumption of the storage battery 21 from the total power consumption for each time period included in the history acquired in step S101. For example, let's assume that the total power consumption for the time period from 0:00 to 0:30 on December 1, 20XX, is 10 kW, and the power consumption of the storage battery 21 is 5 kW. In this case, the power consumption of the general power load 22 for this time period is 10 kW - 5 kW = 5 kW. The power consumption of the general power load 22 for other time periods is calculated in a similar manner.

In step S106, the generation means 66 generates a machine learning model by machine learning training data in which the power consumption amount calculated in step S105 is the target variable and the calendar information, weather information, and production performance information acquired in steps S102 to S104 are the explanatory variables. The generation means 66 learns the relationship between the power consumption amount and calendar information at the target consumer. For example, if the power consumption tends to be high in a particular month and low in other months, the generation means 66 learns that trend. For example, if the power consumption tends to be high at the end of the month and low during other periods, the generation means 66 learns that trend. For example, if the power consumption tends to be high on the days of the week when products are produced at the target consumer and low on other days of the week, the generation means 66 learns that trend. For example, if the power consumption tends to be high on weekdays and low on weekends and holidays, the generation means 66 learns that trend. The reason for using whether or not it is the end of the month as an explanatory variable is that even on the same day of the week, the power consumption trend may differ between the end of the month and other days. The reason for using whether or not it is a public holiday as an explanatory variable is that the trend in power consumption may differ between public holidays and weekdays, even on the same day of the week.

Furthermore, the generation means 66 learns the relationship between the amount of electricity consumed by the target consumer and weather information through machine learning. Therefore, for example, if electricity consumption tends to be low on sunny days and high on cloudy and rainy days, the generation means 66 learns this trend. Similarly, if electricity consumption tends to be high on hot and cold days and low on moderately hot days, the generation means 66 learns this trend.

Furthermore, the generation means 66 learns the relationship between the power consumption and production performance information of the target customer through machine learning. Therefore, for example, if power consumption tends to be high when production volume is high and low when production volume is low, the generation means 66 learns this trend. For example, if power consumption tends to be high when producing a specific type of product and low when producing other types of products, the generation means 66 learns this trend. For example, if power consumption tends to be high when a specific production process for producing a product is performed and low when other production processes are performed, the generation means 66 learns this trend. Furthermore, if power consumption tends to fluctuate due to at least two combinations of calendar information, weather information, and production performance information, the generation means 66 may also learn this trend. The generation means 66 stores the machine learning model thus generated in the storage means 61, associating it with the identification information of the target customer.

Server 60 repeatedly executes steps S101 to S106, sequentially changing the target customers. In this way, machine learning models are generated for all customers H belonging to the power control system 1. Furthermore, even after machine learning models have been generated for all customers H, steps S101 to S106 may be repeatedly executed. In this case, in step S106, further machine learning is performed using newly acquired data.

2.2 Prediction of Power Demand Figure 5 is a flowchart illustrating the operation of a server 60 that predicts power demand according to one embodiment. The flow in Figure 5 is initiated, for example, when a reference time such as midnight arrives. In the following, functional elements such as the acquisition means 62 are described as the main processing element, which means that hardware elements such as the CPU 601 that executes a program such as a server program work together with other hardware elements to perform the processing.

Prior to this process, the server 60 determines the target customer H (hereinafter referred to as the "target customer") from among the multiple customers H included in the power control system 1. The target customers are determined one by one in order according to a predetermined rule. Here, it is assumed that customer H with identification information "001" is determined as the target customer.

In step S111, the prediction means 67 sets the prediction period. For example, if the prediction is to forecast the amount of electricity demand for the 24 hours from midnight to midnight of the following day, the prediction means 67 sets the period from midnight to midnight of the following day as the prediction period. Here, the prediction period is assumed to be from midnight to midnight on December 14, 20XX.

In step S112, the acquisition means 63 acquires calendar information indicating the prediction period set in step S111. For example, the acquisition means 63 acquires the date and time indicating the prediction period set in step S111, such as December 14, 20XX, from 0:00 to 24:00, as calendar information. The acquisition means 63 also acquires additional calendar information from the storage means 61, such as the day of the week for December 14, 20XX, whether it is a public holiday, or the end of the month if it is the end of the month.

In step S113, the acquisition means 64 acquires weather information for the area to which the target customer belongs during the forecast period set in step S112. For example, the acquisition means 64 identifies an address associated with the identification information "001" in the customer database stored in the storage means 61. Here, it is assumed that this address includes City X. In this case, the acquisition means 64 acquires weather information such as the weather and temperature for City X from 0:00 to 24:00 on December 14, 20XX, from an external server.

In step S114, the acquisition means 65 acquires production schedule information for the target customer's products during the forecast period. For example, the acquisition means 65 identifies the address of an input terminal associated with the identification information "001" in the customer database stored in the storage means 61. Using the identified address, the acquisition means 65 acquires production schedule information for the target customer's products from the input terminal installed at the target customer from 0:00 to 24:00 on December 14, 20XX. The production schedule information includes the type of product to be produced during the forecast period, its production quantity, and the production process for the product to be carried out during the forecast period.

In step S115 (an example of a prediction step), the prediction means 67 inputs the calendar information, weather information, and production schedule information acquired in steps S112 to S114 into the machine learning model of the target customer stored in the storage means 61 to predict the power demand of the target customer during the prediction period. For example, the prediction means 67 uses a machine learning model stored in the storage means 61 associated with the identification information "001". The prediction means 67 inputs the calendar information, weather information, and production schedule information acquired in steps S112 to S114 into this machine learning model to predict the power demand of the target customer in 30-minute increments from 0:00 to 24:00 on December 14, 20XX. However, this power demand prediction is for the general power load 22 excluding the battery 21. In other words, this power demand prediction does not include the power demand of the battery 21.

Server 60 repeatedly executes steps S111 to S115, sequentially changing the target consumers. In this way, the power demand for the forecast period is predicted for all consumers H belonging to the power control system 1.

Customer H uses the electricity demand forecast obtained in this manner to prepare a power generation device 10 and a storage battery 21 corresponding to this electricity demand. When there is high electricity demand in society or during disasters, Customer H can supply electricity generated by the power generation device 10 or stored in the storage battery 21 to the surrounding area. By supplying electricity to society in this way, Customer H can receive benefits such as monetary compensation. However, in order to supply electricity to society, Customer H needs to generate and store more electricity than it consumes. Therefore, Customer H determines the necessary amount of electricity to be generated and stored based on the electricity demand forecast, and installs a power generation device 10 with the necessary amount of electricity and a storage battery 21 with the necessary amount of electricity to be stored at Customer H. For example, if the amount of electricity generated or stored is insufficient, Customer H may add more power generation devices 10 or storage batteries 21 to cover the shortfall.

Furthermore, the server 60 creates an operation plan for the battery 21 so that energy management is optimized. This operation plan includes the start and end times of operation for the battery 21. The server 60 transmits commands for the operation of the battery 21 to the control device 50 according to the created operation plan. The control device 50 controls the operation of the battery 21 according to the commands received from the server 60. If the power consumption of the battery 21 per unit time is approximately constant, the server 60 can accurately predict the power demand of the battery 21 using the power consumption per unit time and the operation plan for the battery 21. If the power demand of all power loads 20, including the battery 21, were predicted by machine learning, the accuracy of the prediction would be lower, at least for the power demand of the battery 21, compared to predicting the power demand based on its operation plan. As described above, by predicting the power demand of the general power loads 22 excluding the battery 21 at customer H using machine learning, the accuracy of the power demand prediction is higher compared to predicting the power demand including the power demand of the battery 21 using machine learning.

According to the embodiments described above, since production information of products that affect the amount of electricity consumed at customer H is included in the explanatory variables of the machine learning system, the accuracy of electricity demand prediction is higher compared to when production information is not included in the explanatory variables. Furthermore, since calendar information and weather information that affect the amount of electricity consumed at customer H are included in the explanatory variables of the machine learning system, the accuracy of electricity demand prediction is higher compared to when this information is not included in the explanatory variables. Moreover, since the electricity demand of the general power loads 22, excluding the battery 21 whose operation is controlled by the control device 50, is predicted by machine learning, the accuracy of electricity demand prediction is higher compared to when the electricity demand including the battery 21 is predicted by machine learning. Furthermore, since the electricity demand of the general power loads 22 at customer H is predicted collectively by machine learning, the amount of computation required for prediction is less compared to when the electricity demand of each general power load 22 is predicted separately.

3. Modifications The present invention is not limited to the embodiments described above, and various modifications are possible. Several modifications are described below. Two or more of the following modifications may be used in combination.

Variation 1
Production information is not limited to the type of product, its production volume, or the production process. Any information that indicates the production status of a product is acceptable. Furthermore, explanatory variables used in machine learning do not necessarily have to include all of the calendar information, weather information, and production information. For example, one or two of the calendar information, weather information, and production information may be used as explanatory variables.

Variation 2
The explanatory variables used in machine learning are not limited to those exemplified in the embodiments described above. The explanatory variables can be any information that affects the amount of electricity consumed at customer H. For example, the explanatory variables may include situational information that indicates the situation at customer H. Situational information may include, for example, at least one of the following: production information, the number of people in customer H, economic indicators such as stock prices and exchange rates related to customer H, raw material prices, and disaster information. The number of people in customer H can be determined for each predetermined time period, for example, by using the start and end times of work when employees working at customer H clock in and out using time cards. Economic indicators can be provided, for example, by businesses or organizations that provide economic indicators. Raw material prices can be provided, for example, by businesses or organizations that provide raw material prices. Disaster information can be provided, for example, by external servers operated by businesses or organizations that provide weather information.

The acquisition means 65 acquires status information of customer H at the target date and time. This status information includes status information for customer H at the date and time of power consumption and status information for the forecast period. Since the forecast period is a future period, the status information for the forecast period will not represent actual results, but rather plans or predictions. In one example, customer H inputs status information into an input terminal installed at customer H. The acquisition means 65 acquires status information for the date and time of power consumption and status information for the forecast period from this input terminal or an external server. The generation means 66 further performs machine learning using the status information for the date and time of power consumption acquired by the acquisition means 65 as explanatory variables.

For example, if the amount of electricity consumed tends to be high when the number of people in customer H is large, and low when the number of people is small, the generation means 66 learns this trend. For example, if the amount of electricity consumed tends to be high when economic indicators show a good economy, and low when they show a bad economy, the generation means 66 learns this trend. For example, if the amount of electricity consumed tends to be high when raw material prices are low, and low when raw material prices are high, the generation means 66 learns this trend. For example, if the amount of electricity consumed tends to be high when a disaster occurs, and low when no disaster occurs, the generation means 66 learns this trend.

The prediction means 67 further inputs the situational information acquired by the acquisition means 65 during the prediction period into a machine learning model to predict electricity demand. According to this modified version, the situational information of customer H, which affects the amount of electricity consumed by customer H, is used as an explanatory variable, thus improving the accuracy of the machine learning-based prediction of electricity demand.

Furthermore, in this modified example, only the situational information acquired by the acquisition means 65 that correlates with the power consumption of customer H may be used. Here, it is assumed that multiple pieces of situational information are acquired from customer H. The generation means 66 extracts situational information correlated with the power consumption of customer H from the multiple pieces of situational information using a known method, and uses the extracted situational information as explanatory variables. In one example, the generation means 66 calculates the correlation coefficient between each piece of situational information acquired by the acquisition means 65 and the power consumption of customer H. The generation means 66 extracts situational information whose correlation coefficient is above a threshold from the multiple pieces of situational information and uses it as an explanatory variable. The generation means 66 does not use situational information whose correlation coefficient is below the threshold as an explanatory variable. This threshold is preferably a positive value. According to this modified example, since only situational information correlated with the power consumption of customer H is used as an explanatory variable, the accuracy of the power demand prediction by machine learning is improved.

Variation 3
Machine learning for predicting electricity demand at customer H does not necessarily require the use of information about customer H itself. For example, if there is little history of electricity consumption at customer H, the history of electricity consumption at other customers H similar to customer H may be used as the history of electricity consumption at customer H. For example, if production information for products at customer H cannot be obtained, the production information for products at other customers H similar to customer H may be used as the production information for products at customer H. It is preferable that the other customers H are customers H whose profiles are similar to customer H, such as the organization to which they belong, products, production volume, size, and number of employees. In one example, if multiple customers H belong to a certain organization, a machine learning model generated by performing machine learning using the history of electricity consumption and production information of products at one customer H belonging to this organization may be used to predict electricity demand at other customers H belonging to this organization. In short, the at least one customer H that provides data used in machine learning may be the target customer itself, a set of the target customer and other customers H, or one or more other customers H that do not include the target customer. This modified configuration also allows for the prediction of electricity demand at customer H.

Variation 4
Any learning algorithm can be used in machine learning. Examples of machine learning learning algorithms include decision trees, support vector machines, neural networks, and deep learning.

Variation 5
The power load 20 whose operation is controlled by the control device 50 is not limited to the storage battery 21. The power load 20 whose operation is controlled by the control device 50 can be any power load 20 as long as the control device 50 can control its operation based on the operation plan created by the server 60 and the amount of power consumed can be measured individually. However, it is preferable that the power load 20 whose operation is controlled by the control device 50 has an upper limit on its power consumption.

Variation 6
The correspondence between functional elements and hardware elements in the power control system 1 is not limited to those exemplified in the embodiments. For example, some of the functions described as those of the server 60 in the embodiments may be implemented in the control device 50. Alternatively, some of the functions described as those of the server 60 in the embodiments may be implemented in other devices on the network. In one example, another device on the network, different from the server 60, may have a storage means 61, acquisition means 62-65, and generation means 66, and generate a machine learning model. The server 60 may have a storage means 61, acquisition means 62-65, and prediction means 67, acquire a machine learning model from the other device, and use the acquired machine learning model to predict power demand. In this case, the server 60 and the other device may be managed and operated by different operators.

Variation 7
The operation of the power control system 1 is not limited to the examples described above. The processing steps of the power control system 1 may be rearranged as long as they do not contradict each other. In addition, some of the processing steps of the power control system 1 may be omitted.

Variation 8
Another embodiment of the present invention may provide a power demand forecasting method having processing steps performed in at least one of the following: a power control system 1, a power generator 10, a power load 20, an energy meter 30, a measuring instrument 40, a control device 50, and a server 60.

Variant 9
The various programs exemplified in the embodiments may be provided via download over a network such as the Internet, or they may be provided recorded on a recording medium such as a CD-ROM (Compact Disc Read Only Memory).

1: Power control system, 10: Power generator, 20: Power load, 21: Storage battery, 22: General power load, 30: Energy meter, 40: Measuring instrument, 50: Control device, 51: Storage means, 52: Transmission means, 60: Server, 61: Storage means, 62: Acquisition means, 63: Acquisition means, 64: Acquisition means, 65: Acquisition means, 66: Generation means, 67: Prediction means, 90: Network, 95: Network, 601: CPU, 602: Memory, 603: Storage, 604: Communication interface

Claims (6)

On the computer,
A step of indicating a forecast period and obtaining first calendar information that includes at least one of the following types within the forecast period: day of the week, whether or not it is a public holiday, and whether or not it is the end of the month .
A step of obtaining production schedule information for the product of the target consumer that produces the product during the forecast period,
A program for performing a prediction step of predicting the amount of electricity demand of a target customer during the prediction period by inputting the first calendar information and the production schedule information into a machine learning model generated by machine learning, in which the date and time of electricity consumption of at least one customer is indicated, and
At least one of the aforementioned consumers is equipped with a first power load, a second power load, and a control device for controlling the first power load.
The objective variable is the amount of power consumed, obtained by subtracting the amount of power consumed by the first power load from the sum of the power consumed by the first power load and the power consumed by the second power load at the aforementioned date and time.
program. The program according to claim 1, wherein each of the production performance information and the production schedule information includes at least one of the production volume of the product, the type of product, and the production process of the product. A program for further performing the step of obtaining first weather information indicating the weather in the area to which the target customer belongs during the forecast period,
The explanatory variables include a second meteorological information indicating the weather in the area to which the at least one customer belongs at the date and time,
The program according to claim 1 or 2 , wherein in the prediction step, the amount of electricity demand is predicted by further inputting the first weather information into the machine learning model. The program according to claim 1 , wherein the first power load includes a storage battery. Computers
A step of indicating a forecast period and obtaining first calendar information that includes at least one of the following types within the forecast period: day of the week, whether or not it is a public holiday, and whether or not it is the end of the month .
A step of obtaining production schedule information for the product of the target consumer that produces the product during the forecast period,
A power demand forecasting method comprising: a forecasting step of predicting the amount of power demand of a target customer during the forecast period by inputting the first calendar information and the production schedule information into a machine learning model generated by machine learning, in which the date and time of power consumption of at least one customer is indicated, and the date and time of power consumption of at least one customer is indicated, and the date and time of power consumption of at least one customer is indicated, and the date and time of power consumption of at least one customer is indicated, and the date and time of power consumption of at least one customer is indicated, and the date and time of power consumption of at least one customer is indicated, and the production schedule information of the product of at least one customer at the date and time is used as explanatory variables;
At least one of the aforementioned consumers is equipped with a first power load, a second power load, and a control device for controlling the first power load.
The objective variable is the amount of power consumed, obtained by subtracting the amount of power consumed by the first power load from the sum of the power consumed by the first power load and the power consumed by the second power load at the aforementioned date and time.
Methods for forecasting electricity demand . A first acquisition means that indicates a period to be predicted and acquires first calendar information that includes at least one of the following types within the period to be predicted: day of the week, whether or not it is a public holiday, and whether or not it is the end of the month .
A second acquisition means for acquiring production schedule information of the product for the target consumer that produces the product during the forecast period,
An information processing device comprising: a prediction means for predicting the amount of electricity demand of a target customer during the prediction period, by inputting the first calendar information and the production schedule information into a machine learning model generated by performing machine learning, which uses second calendar information indicating the date and time when electricity was consumed by at least one customer, and at least one of the following types for the date and time: day of the week, public holiday or not, and end of the month or not, and production performance information of the at least one customer at the date and time, as explanatory variables and the amount of electricity consumed by the at least one customer at the date and time as the objective variable;
At least one of the aforementioned consumers is equipped with a first power load, a second power load, and a control device for controlling the first power load.
The objective variable is the amount of power consumed, obtained by subtracting the amount of power consumed by the first power load from the sum of the power consumed by the first power load and the power consumed by the second power load at the aforementioned date and time.
Information processing device .