Training device, prediction device, training method, and program

The learning device extracts customer characteristics to build a demand prediction model, addressing the challenge of inaccurate power demand forecasting by considering business status and temperature sensitivity, achieving reduced prediction errors and improved forecasting accuracy.

WO2024247070A1PCT designated stage expired Publication Date: 2024-12-05NT T INC
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2023/019986
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-29
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

Conventional power demand estimation technologies for retail electricity businesses fail to accurately account for customer business status, leading to significant prediction errors due to insufficient consideration of business days and hours, which affects the distribution and forecasting of electricity demand.

Method used

A learning device and method that extracts customer characteristics, including business type, temperature sensitivity, and demand clusters, using weather information, customer information, and demand history to build a demand prediction model, enabling precise power demand forecasting by preprocessing data into regional meshes and clusters.

Benefits of technology

The solution reduces prediction errors by accurately identifying business types and hours, analyzing temperature sensitivity, and performing demand cluster analysis, resulting in highly accurate demand forecasting that matches or exceeds traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2023019986_05122024_PF_FP_ABST
    Figure JP2023019986_05122024_PF_FP_ABST
Patent Text Reader

Abstract

This training device includes: a characteristic extraction unit that extracts consumer characteristics including a predetermined profile, temperature sensitivity, and cluster of a consumer by using weather information, consumer information, and demand records as inputs; and a training unit that trains a demand prediction model by a machine training method using, as inputs, training data including the weather information, the consumer information, the demand records, and the consumer characteristics.
Need to check novelty before this filing date? Find Prior Art

Description

Learning device, prediction device, learning method, and program

[0001] The disclosed technology relates to a learning device, a prediction device, a learning method, and a program.

[0002] Conventionally, as a technology for estimating power demand, particularly for new entrants to retail electricity businesses, there is a technology for classifying consumers into multiple clusters and estimating power demand (see Non-Patent Document 1).

[0003] Nobuhiro Gotoda, Ikuo Shigemori, Yoko Sakikubo, Toru Watanabe, and Kengo Uegaki, "Electricity System Reform / Hitachi's IT Solutions for Creating Value in a Competitive Market amid Full Retail Deregulation," Hitachi Review, December 2015. URL: "https: / / www.hitachihyoron.com / jp / pdf / 2015 / 12 / 2015_12_02.pdf"

[0004] Conventional technologies have taken into account air conditioning control, i.e., temperature, which is a major factor in fluctuations in electricity demand. However, conventional technologies have not taken into account another factor, the business status of consumers, which is another factor in fluctuations. Even if attempts are made to consider business status from calendar information, it is not always the case that holidays (Saturdays, Sundays, and public holidays) on the calendar are closed days for some consumers, so simply identifying business status from calendar information is not sufficient.

[0005] The disclosed technology has been made in consideration of the above points, and aims to provide a learning device, a prediction device, a learning method, and a program that enable highly accurate demand forecasting.

[0006] The learning device of the present disclosure includes a characteristic extraction unit that uses weather information, consumer information, and actual demand as input to extract consumer characteristics including a predetermined profile, temperature sensitivity, and cluster of the consumer, and a learning unit that uses learning data including the weather information, consumer information, actual demand, and the consumer characteristics as input to learn a demand forecasting model using machine learning techniques.

[0007] The prediction device disclosed herein includes a pre-processing unit that divides each data item of weather information, consumer information, and consumer characteristics into regional mesh and cluster units as pre-processing, and a prediction unit that inputs learning data including pre-processed weather information, consumer information, actual demand, and consumer characteristics, and inputs each data item of weather information, consumer information, and consumer characteristics using a demand prediction model to predict electricity demand in a predetermined unit corresponding to the pre-processing.

[0008] The learning method disclosed herein is a process in which a computer executes a process in which weather information, consumer information, and actual demand are used as inputs to extract consumer characteristics including a predetermined profile, temperature sensitivity, and cluster of the consumer, and learning data including the weather information, consumer information, actual demand, and the consumer characteristics is used as inputs to learn a demand forecasting model using machine learning techniques.

[0009] The disclosed technology enables highly accurate demand forecasting.

[0010] FIG. 1 is a block diagram showing the hardware configuration of a learning device and a prediction device of this embodiment. FIG. 2 is a block diagram showing the configuration of the learning device of this embodiment. FIG. 3 is a block diagram showing the configuration of the prediction device of this embodiment. FIG. 4 is an example of a data flow when extracting existing consumer characteristics. FIG. 5 is an example of a data flow when extracting new consumer characteristics. FIG. 6 is a diagram showing an example of a data structure of consumer information and actual demand. FIG. 7 is an example of a process flow for identifying business types. FIG. 8 is an example of a microtremor fluctuation extracted in the microtremor fluctuation detection process. FIG. 9 is a schematic diagram showing an example of extraction of the median of electricity demand. FIG. 10 is a schematic diagram of profile generation. FIG. 11 is an example of a graph identifying business hours. FIG. 12 is an example of a graph showing temperature and electricity demand in a temperature sensitivity analysis. FIG. 13 is an image showing the flow of demand cluster analysis. FIG. 14 is a flowchart showing the flow of learning processing by the learning device. FIG. 15 is a flowchart showing the flow of consumer characteristic extraction processing. FIG. 16 is a flowchart showing the flow of prediction processing by the prediction device. FIG. 17 shows an example of the business type (business status) identification results and demand curves for office customers. FIG. 18 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 19 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 20 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 21 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 22 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 23 shows an example of the business type (business status) identification results and demand curves for each industry. FIG. 24 shows an example of a list of business type identification results for each customer. FIG. 25 shows the proportion of business days and holidays for each rate menu. FIG. 26 shows the proportion of holidays for weekdays, Saturdays, Sundays, and holidays for each rate menu, and the proportion of holidays for weekdays and Sundays for each rate menu. FIG. 27 shows the proportion of business days and holidays for each industry. Fig. 28 shows the distribution of business hours for each industry. Fig. 29 shows the distribution of business hours for each industry. Fig. 30 is an example of a graph showing the relationship between temperature and electricity demand at each time. Fig. 31 shows the setting of the number of clusters using the elbow method.Figure 32 shows the matching rate of clusters for each consumer in 2020 and 2021. Figure 33 shows demand curves by cluster on business days and demand curves by cluster on holidays. Figure 34 shows the proportion of consumers by cluster in each industry. Figure 35 is an example of the importance of features. Figure 36 is a graph showing a comparison of forecast performance for each month. Figure 37 is a diagram showing the correlation between temperature and electricity demand for each regional mesh. Figure 38 is a diagram showing the correlation between temperature and electricity demand for each regional mesh.

[0011] An example of an embodiment of the disclosed technology will be described below with reference to the drawings. Note that the same or equivalent components and parts in each drawing are given the same reference numerals. Also, the dimensional proportions in the drawings are exaggerated for the convenience of explanation and may differ from the actual proportions.

[0012] First, the background and overview of the technology of this embodiment will be described.

[0013] The background to this is the emergence of many new retail electricity providers (so-called new power providers, hereafter referred to as new power providers) following the full liberalization of the electricity retail market. New power providers have fewer customers than the major former general electricity providers, and electricity demand is not uniform. Furthermore, new power providers are prone to large errors in their demand forecasts, as customer turnover (new contracts and cancellations) has a significant impact on overall electricity demand. In order to reduce forecast errors, new power providers are conducting individual business surveys and adjusting forecast values ​​for customers with large electricity demand, but there is a limit to how much forecast error can be reduced by responding individually, and there are risks to business expansion.

[0014] Regarding the prediction of power demand, for example, there is a demand cluster analysis method using past power demand. It has been reported that the demand cluster analysis method reduces the prediction error (MAPE: Mean Absolute Percentage Error) by aggregating consumers into clusters and predicting the power demand for each cluster. However, as described above, it does not take into account the different business days and business hours of each consumer.

[0015] Recently, the construction of regional microgrids has been progressing with the aim of local production and consumption of electricity and improving resilience. As mentioned above, reducing forecast errors is an issue when adjusting supply and demand for specific consumers or areas. When constructing regional microgrids, forecast errors not only pose an imbalance risk, but also lead to capital investment costs and maintenance costs for securing the necessary adjustment capacity to compensate for the forecast errors. Examples of ways to secure adjustment capacity include battery storage capacity.

[0016] Therefore, there is a need for demand forecasts with small prediction errors that can reduce the imbalance risk for new power suppliers and reduce the required adjustment capacity in regional microgrids. In this embodiment, we aim to reduce prediction errors by extracting latent characteristics not found in existing consumer information from past demand records in terms of each consumer's business type and power demand amount, and applying these to power demand forecasts. We comprehensively consider the business types of consumers and then perform demand cluster analysis to achieve power demand forecasts with smaller prediction errors than conventional methods. This embodiment focuses on the following first to third points.

[0017] First, the business type (e.g., business days and holidays) of each consumer is identified, and regular holidays are identified to consider changes in electricity demand. For example, weekdays are not necessarily business days and holidays are not necessarily closed days; many consumers are open on holidays. Furthermore, the correlation between weather and electricity demand is weaker on holidays than on business days, and electricity demand tends to be generally low on holidays. Regular holidays are identified for each consumer based on past business days and holidays, and changes in electricity demand are considered in future predictions. This makes it possible to automatically identify the business type of consumers based on past electricity demand, while previously only consumers with large electricity demand were individually surveyed, and to comprehensively consider the business type of consumers. Note that in this embodiment, "business type" refers to the trend of business conditions on a daily basis, including business days and holidays, which tend to vary depending on the industry.

[0018] Second, accurate temperature sensitivity is analyzed by taking into account the business hours of each consumer. When it comes to business days and business hours, there is a correlation between weather (especially temperature) and electricity demand. If temperature sensitivity is analyzed without taking business hours into account, a spurious correlation between business characteristics and electricity demand that is unrelated to weather will appear. With regard to business characteristics, it is hypothesized that changes in electricity demand due to fixture operation have a major impact, rather than changes in demand due to air conditioning operation. An appropriate population is formed by identifying business hours for each consumer, and accurate temperature sensitivity is analyzed. This makes it possible to automatically identify business hours on business days for each consumer, and accurately derive each consumer's temperature sensitivity.

[0019] Third, a demand cluster analysis is performed on business days, taking into account the demand curve for each consumer, i.e., the trend in electricity demand. For example, it is expected that consumer information such as industry and contracted power cannot uniquely identify the demand curve pattern for each consumer. Furthermore, when electricity demand is aggregated by industry, the demand curve pattern for each industry may change due to changes in consumers. Therefore, a demand cluster analysis that takes business days into account is performed to identify the demand curve pattern for each consumer, thereby reducing the impact of consumer changes in future predictions. In this way, a demand cluster analysis is performed using the electricity demand for each consumer on business days, making it possible to understand the trend in electricity demand (demand curve) on business days.

[0020] It was confirmed that applying the above-mentioned first to third viewpoints to demand forecasting reduces forecasting errors. Furthermore, when the results were reported to electricity retailers, it was confirmed that the performance achieved was equal to or better than that of demand forecasts that had previously been manually surveyed and adjusted individually.

[0021] The configuration of this embodiment will be described below.

[0022] FIG. 1 is a block diagram showing the hardware configuration of a learning device 100 and a prediction device 200 according to this embodiment. The learning device 100 will be described as an example. As shown in FIG. 1 , the learning device 100 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other. The hardware configuration of the prediction device 200 can also be configured in a similar manner.

[0023] The CPU 11 is a central processing unit that executes various programs and controls each component. That is, the CPU 11 reads programs from the ROM 12 or storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs stored in the ROM 12 or storage 14. In this embodiment, a learning program is stored in the ROM 12 or storage 14 of the learning device 100. Note that a prediction program is stored in the ROM 12 or storage 14 of the prediction device 200.

[0024] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured by a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores various programs including an operating system and various data.

[0025] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0026] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may be a touch panel type and function as the input unit 15.

[0027] The communication interface 17 is an interface for communicating with other devices such as terminals, etc. For this communication, for example, a wired communication standard such as Ethernet (registered trademark) or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi (registered trademark) is used.

[0028] Next, the functional components of the learning device 100 will be described. Fig. 2 is a block diagram showing the configuration of the learning device 100 of this embodiment. Fig. 3 is a block diagram showing the configuration of the prediction device 200 of this embodiment. Each functional component is realized by the CPU 11 reading various programs stored in the ROM 12 or storage 14, expanding them in the RAM 13, and executing them.

[0029] 2, the learning device 100 includes a characteristic extraction unit 102, a data storage unit 110, a model storage unit 112, a preprocessing unit 120, and a learning unit 122. As shown in Fig. 3, the prediction device 200 includes a data storage unit 210, a model storage unit 212, a preprocessing unit 220, and a prediction unit 222.

[0030] In this embodiment, four main types of data are used: weather information, customer information, customer performance, and customer characteristics (such as business days and holidays), and a demand forecasting model is constructed to predict, as an example, a 30-minute value of electricity demand. Regarding weather information, past weather performance is used when constructing the model, and weather forecasts are used when making predictions.

[0031] First, we will explain each unit of the learning device 100. The characteristic extraction unit 102 receives weather information, customer information, and actual demand data as input and extracts customer characteristics including profiles (regular holiday profiles), temperature sensitivities, and clusters. Details will be described later.

[0032] The data storage unit 110 stores weather information, customer information, and actual demand as pre-stored information. The data storage unit 110 also stores customer characteristics extracted by the characteristic extraction unit 102.

[0033] The model storage unit 112 stores a demand forecasting model learned by the learning unit 122. The demand forecasting model is constructed as a model that, for example, predicts the amount of power demand in regional mesh and cluster units, and then sums and forecasts the amount of power demand in power area units.

[0034] Fig. 4 shows an example of a data flow when extracting existing customer characteristics, and Fig. 5 shows an example of a data flow when extracting new customer characteristics.

[0035] 6 shows an example of the data structure of customer information and actual demand. Columns in the data structure of customer information include, for example, a user ID, a supply point identification number, a contract start date, a contract end date, an address, a metered / negotiated system, a contracted power, a self-generation supply contract, a rate menu, a start date of the rate menu, and an end date of the rate menu. Columns in the data structure of actual demand include, for example, a supply point identification number, a supply date and time, and an amount of electricity demand.

[0036] The extraction of existing consumer characteristics will be described. The processes performed by the characteristic extraction unit 102 when extracting existing consumer characteristics can be broadly divided into (1) business type identification, (2) profile generation, (3) temperature sensitivity derivation, and (4) demand cluster analysis. When extracting new consumer characteristics, similar existing consumers are determined, and the consumer characteristics of consumers determined to be similar are transcribed. Note that "transcribing" in this embodiment corresponds to copying the consumer characteristics of similar existing consumers as the consumer characteristics of a new consumer.

[0037] (1) In business type identification, the typical power demand and the trend of power demand on the reference date are evaluated, and business days and holidays are identified on a daily basis. The typical power demand is defined by extracting the median power demand for each time period for two weeks before and after the reference date.

[0038] FIG. 7 shows an example of the flow of the business type identification process. In the business type identification process, actual demand data is input and the processes (A1) to (A5) are performed. The flow of the business type identification process is as follows: (A1) smoothing, (A2) detecting slight fluctuations, (A3) extracting representative values, (A4) evaluating and aggregating (R 2 (A5) Evaluation, demand aggregation, and (A6) Judgment.

[0039] FIG. 8 shows an example of tremor fluctuations extracted in the tremor fluctuation detection process. (A2) In tremor fluctuation detection, the following first to fourth processes are performed. First, the standard deviation is calculated on a daily basis using the smoothed power demand. Second, outliers in the standard deviation are detected for each consumer, and the maximum value of the standard deviation is defined. Third, tremor fluctuations are determined based on the condition that (standard deviation / maximum value) is less than 5%. Fourth, if tremor fluctuations continue for more than two weeks, "holiday" is treated as "other" for that period. Note that "holiday" is an identification result of a business type that indicates one of the business statuses: "business day," "holiday," or "other."

[0040] (A3) In extracting the representative value, first, data for two weeks before and after the reference date (a total of four weeks + one day) is extracted. Here, data for "holiday" ≠ "other" is targeted. Second, the median value of the power demand for each time period for each reference date is extracted. Figure 9 is a schematic diagram showing an example of extracting the median value of the power demand. It is extracted for each reference date a1 to a3. The extracted median value of the power demand is the representative value.

[0041] (A4) Evaluation and aggregation (R 2 In (A5) and (A6), a judgment is made using an evaluation function. 2 |Σb t ≦Σr t The case where r is satisfied. t is the smoothed data, b t is the representative value data. The solid line in the graph (A4) of FIG. t , the dashed line is Σr t "Holidays" refers to other days other than business days. "Other" refers to cases where the tremor fluctuations last for more than two weeks, i.e., when they continue for 15 days or more.

[0042] (2) When generating profiles, two types of regular holiday profiles are created for each consumer so that future business days and holidays can be taken into account. For new consumers, profiles of consumers with similar contract power, rate plans, etc. can be copied and referenced. Figure 10 is a schematic diagram of profile generation. The number (hash) is the supply point identification number. In the regular holiday profile, the business status column is set to 0 for business days and 1 for holidays. Regular holiday profiles are created by day of the week and by public holiday. As described above, the characteristic extraction unit 102 detects slight fluctuations in actual demand and identifies business days and holidays in the consumer profile using a predetermined evaluation function that compares the data.

[0043] (3) When deriving temperature sensitivity, business hours are identified before the temperature sensitivity is analyzed. Business hours on business days are identified based on the amount of change in electricity demand. Days identified as non-business days are not included in the business hours identification.

[0044] FIG. 11 is an example of a graph in which business hours have been identified. First, the amount of electricity demand is smoothed over 24 frames before and after the identified time. Second, the maximum and minimum values ​​are derived from the smoothed values ​​over 20 business days before and after the identified date, excluding outliers, and the difference is calculated. Third, the time when the value is equal to or greater than "minimum value + 50% x difference" is defined as business hours.

[0045] In analyzing temperature sensitivity, the correlation between temperature and power demand is calculated using data from each customer during business hours, and this is set as the temperature sensitivity for each customer. Note that since there is no past data for new customers, the temperature sensitivity of customers with similar contract power, rate menus, etc. is referenced.

[0046] FIG. 12 is an example of a graph showing temperature and power demand in a temperature sensitivity analysis. The analysis flow is as follows: (B1) input, (B2) business-hour data extraction, (B3) seasonal data extraction, (B4) temperature sensitivity feature calculation, and (B5) output. (B1) In the input step, the customer's past demand and temperature are input. The graph shows power demand on the vertical axis and temperature on the horizontal axis. (B2) In the business-hour data extraction step, only data from business hours, when temperature sensitivity is expected to be high, is extracted. (B3) In the seasonal data extraction step, data is extracted for summer (June, July, August) and winter (December, January, February), when the correlation between temperature and power demand is expected to differ. (B4) In the calculation of temperature sensitivity feature, the magnitude of the correlation between temperature and power demand is calculated as a coefficient for each customer, in summer and winter. Differences in contracted power are also taken into consideration. (B5) In the output, the summer temperature sensitivity and winter temperature sensitivity for each consumer are output. For each consumer, an approximation formula capable of reproducing the relationship between summer and winter temperatures and power demand is set, and output as temperature sensitivity. As described above, the characteristic extraction unit 102 receives the consumer's actual demand record and the temperature in the weather information as input, extracts business data, extracts seasonal data, and calculates the feature value of temperature sensitivity, thereby deriving temperature sensitivity.

[0047] (4) In the demand cluster analysis, representative demand curves are calculated for business days and holidays from the demand records of each customer, and groups with similar demand curve patterns over 48 24-hour periods are created through clustering. Note that since there are no past records for new customers, cluster information for existing customers with similar contract power, rate menus, etc. is referenced.

[0048] FIG. 13 is an illustration showing the flow of demand cluster analysis. The flow of demand cluster analysis is as follows: (C1) input, (C2) business day / holiday data extraction, (C3) representative value calculation, (C4) cluster creation, and (C5) output. In (C1) input, the consumer's past demand performance is input. In (C2) business day / holiday data extraction, data is extracted separately for business days and holidays to make it easier to grasp trends. In (C3) representative value calculation, a representative demand value (median) of the consumer is calculated based on the demand performance over one month. These are the representative demand curves for business days and the representative demand curves for holidays. In (C4) cluster creation, consumers are classified by cluster analysis based on each consumer's representative demand value. In (C5) output, a cluster ID classified by business days and holidays is assigned to each consumer and output. The cluster classification result (illustration) is the demand curve corresponding to each cluster. As described above, the characteristic extraction unit 102 inputs the demand performance of the customer, calculates representative values ​​based on the representative demand curve for business days and the representative demand curve for non-business days, creates clusters based on the representative values, and classifies each customer into a cluster.

[0049] The learning device 100 uses the consumer characteristics extracted as described above to learn a demand forecasting model.

[0050] The preprocessing unit 120 processes the data stored in the data storage unit 110 into learning data for input to the learning unit 122 for machine learning, and outputs the created learning data to the learning unit 122. The learning data includes weather information, consumer information, actual demand, and consumer characteristics. As preprocessing, the preprocessing unit 120 processes the data into learning data that includes the actual demand item as the target variable and the other items as explanatory variables. As preprocessing, each piece of data is divided into regional meshes and cluster units. Furthermore, for actual demand, feature quantities summarized in predetermined units, such as the day before, the day before, or the most recent business day / holiday, are used.

[0051] The learning unit 122 uses the learning data to learn a demand forecasting model by a machine learning technique. The learning device that learns the demand forecasting model by machine learning may use any regression analysis technique.

[0052] Next, each unit of the prediction device 200 shown in Fig. 3 will be described. The data storage unit 210 is the same as the data storage unit 110 except for the demand record. The model storage unit 212 stores a trained demand prediction model trained by the learning device 100. The prediction device 200 may be provided with a function similar to the characteristic extraction unit 102 of the learning device 100 to update the consumer characteristics.

[0053] The pre-processing unit 220 divides each data of weather information, consumer information, and consumer characteristics into regional mesh and cluster units as pre-processing, and outputs them to the prediction unit 222 .

[0054] The prediction unit 222 receives the preprocessed weather information, consumer information, and consumer characteristic data as input, and uses a demand forecasting model to predict the amount of power demand in regional mesh and cluster units, and outputs the final power demand forecast result summed up in power area units. The regional mesh and cluster units are examples of predetermined units corresponding to the preprocessing of the present disclosure.

[0055] The function of each device will now be described. Fig. 14 is a flowchart showing the flow of the learning process by the learning device 100. The learning process is performed by the CPU 11 reading out a learning program from the ROM 12 or storage 14, expanding it into the RAM 13, and executing it.

[0056] In step S100, the CPU 11 extracts consumer characteristics by inputting weather information, consumer information, and actual demand data as the characteristic extraction unit 102. Details of this step will be described later.

[0057] In step S102, the CPU 11 functions as the pre-processing unit 120 to generate learning data including weather information, consumer information, demand records, and consumer characteristics.

[0058] In step S104, the CPU 11 functions as the learning unit 122 and learns a demand forecasting model using the learning data.

[0059] The details of the extraction of customer characteristics in step S100 will be described below with reference to Fig. 15, which is a flowchart of the process of extracting customer characteristics.

[0060] In step S110, the actual demand is input, and a representative amount of power demand and the trend of the amount of power demand on a reference day are evaluated, and business days and holidays are identified as business types on a daily basis, resulting in the business type identification result.

[0061] In step S112, the customer information and the results of the business type identification are input to generate a profile for each customer.

[0062] In step S114, the customer information, actual demand, and the results of the business type identification are input to identify the business hours on business days for each customer.

[0063] In step S116, the customer information, actual demand, weather information, and business hours identification results are input to derive temperature sensitivity for each customer.

[0064] In step S118, clusters are generated using the customer information, actual demand, and the results of business type identification as input, and each customer is classified into a cluster.

[0065] As described above, the learning device 100 of this embodiment can learn a demand estimation model that can reduce prediction errors by using the extracted consumer characteristics.

[0066] 16 is a flowchart showing the flow of prediction processing by the prediction device 200. The CPU 11 reads out a prediction program from the ROM 12 or the storage 14, loads it into the RAM 13, and executes it, thereby performing the prediction processing.

[0067] In step S200, the CPU 11 functions as the pre-processing unit 220 to divide the meteorological information, consumer information, and consumer characteristic data into regional mesh and cluster units as pre-processing.

[0068] In step S202, the CPU 11, as the prediction unit 222, inputs the preprocessed weather information, consumer information, and consumer characteristic data and predicts the amount of power demand using a demand prediction model. The amount of power demand is predicted for each regional mesh and cluster, and the final prediction result of the amount of power demand summed for each power area is output.

[0069] Next, examples of each data extracted as customer characteristics will be described.

[0070] FIG. 17 shows an example of the business type (business status) identification results and demand curves for office customers. (d1) corresponds to the demand curve used as a representative value, (d2) corresponds to the demand curve for the identification target day, and (d3) corresponds to the identification of non-working days. In this example, not only Saturdays and Sundays but also public holidays are identified as non-working days. The public holidays correspond to Showa Day and Golden Week in Japan.

[0071] Figures 18 to 23 show examples of business type (business status) identification results and demand curves for each industry. Each consumer has a different demand curve pattern, and the amplitude, period, and phase of the demand curve can change over time. Each consumer has different business days and holidays, such as a regular holiday every Wednesday, a regular holiday every Saturday, or a regular holiday every Saturday and Sunday. Some consumers have zero or fixed demand for a certain period of time, and these periods are automatically classified as "other." Figure 18 shows an example of a restaurant consumer, whose regular holiday is Wednesday. Figure 19 shows an example of a school education consumer, whose regular holiday is according to the calendar. Figure 20 shows an example of an information and communications industry consumer, whose regular holiday is Sunday. In this example, the system adapts to the increase in demand starting on May 1st and identifies the holiday. Figure 21 shows an example of a manufacturing consumer, whose holiday can be identified from changes in the demand curve. Figure 22 shows an example of an office consumer, whose electricity demand remained unchanged in April and was classified as "other." After that, when a change in the amount of power demand appears, the days are identified as holidays. Figure 23 shows an example of an office consumer, and since there has been no change in the amount of power demand since the middle of May, it is identified as "other."

[0072] Figure 24 shows an example of a list of the business type identification results for each customer. The vertical axis represents each customer, and the horizontal axis represents each date in the identified period. Figure 24 shows an example of an office customer, but it can also be listed by business type.

[0073] Figure 25 shows the percentage of business days and holidays for each rate menu. Figure 26 shows the percentage of holidays for weekdays, Saturdays, Sundays, and holidays for each rate menu, as well as the percentage of holidays for weekdays and Sundays for each rate menu. From left to right, the graphs show "standard," "seasonal time," and "high holiday operation." Seasonal time represents the period when premium rates are charged. Compared to "standard," "high holiday operation" has a smaller percentage of holidays, and days other than holidays are relatively open days. "High holiday operation" tends to reflect the large number of customers who set Tuesdays as their regular closing days.

[0074] Figure 27 shows the percentage of business days and holidays for each industry. The horizontal axis corresponds to each industry, but the industry is not shown here. According to the calendar, the percentage of weekdays is approximately 67% of the total, but there are consumers in some industries who are open for more than 90% of the time. It was also found that while "school education" consumers have a business day percentage according to the calendar, the percentage of weekdays for other industries is higher than the calendar percentage. From the above, it can be seen that weekdays do not necessarily mean business days, and weekends and holidays do not necessarily mean non-business days, so it is important to consider the business status of each consumer in order to accurately grasp electricity demand.

[0075] Figures 28 and 29 show the distribution of business hours for each industry. The distribution shows that, when counting the customers identified as open for business in each time slot, there is a strong overall tendency for business hours to be between 8:00 and 18:00. Commercial (supermarket) customers also tend to have relatively long business hours, from 7:30 to 20:00. Although not shown in the figures, different trends were also evident for customers in the real estate rental and management industry, religious organizations, and other sectors. From the above, it can be seen that business hours (time periods when electricity is used most) vary by industry, even on days identified as business days.

[0076] An example of temperature sensitivity will be given below. FIG. 30 is an example of a graph showing the relationship between temperature and power demand at each time. Since the correlation between temperature and power demand basically appears during business hours, it is important to identify the business hours of each consumer in order to derive accurate temperature sensitivity. In this example, it was confirmed that by deriving temperature sensitivity taking business hours into consideration, the error rate of power demand was reduced to about one-third.

[0077] An example of demand cluster analysis is given below. Figure 31 shows the setting of the number of clusters using the elbow method. Figure 32 shows the matching rate of clusters for each consumer in 2020 and 2021. Figure 33 shows the demand curves by cluster on business days and by cluster on holidays.

[0078] Figure 34 shows the proportion of consumers by cluster in each industry. The trend shows that a high proportion of consumers in the information and communications industry belonged to Cluster 2, indicating a strong relationship between industry and cluster. However, no particular relationship was observed between industry and cluster in other industries. Therefore, it is thought that it is not possible to uniquely identify the demand curve pattern by industry. In other words, as explained in the issue, if electricity demand is aggregated by industry (demand by industry) and a demand forecast is carried out, it was confirmed that there is a possibility that demand cannot be accurately forecast because the demand curve for demand by industry changes due to changes in consumers.

[0079] When the method of this embodiment was applied to short-term demand forecasting and its accuracy was verified using a MAPE evaluation, it was confirmed that the error rate improved by approximately 17% compared to the conventional method. FIG. 35 shows an example of the importance of features. Features (variables) that are particularly important for predicting the next day's electricity demand are (1) the actual demand data from the day before and the day before that, (2) the actual demand data from the most recent business day and holiday, and (3) the actual demand data from the most recent weekday and holiday, and the influence of meteorological factors such as temperature and solar radiation tends to be relatively small. Furthermore, it is expected that improving the discrimination performance of (2), which is one of the three features mentioned above and has a variable factor, will likely have a significant impact on the accuracy of the overall demand forecast.

[0080] Figure 36 is a graph comparing forecast performance for each month. Figures 37 and 38 are diagrams showing the correlation between temperature and power demand for each regional mesh. In central Tokyo, the correlation between temperature and power demand throughout all time periods tends to be weak, while the correlation tends to be relatively high during business hours. For office users, some demand trends are observed that appear to be influenced by air conditioning such as district cooling and geothermal energy, but overall there is a correlation. For users in the information and communications industry, there is a weak correlation between temperature and power demand, regardless of whether it is a weekday, holiday, or business hours. In the information and communications industry, it is presumed that data centers and communication stations in the information and communications industry do not need to heat their equipment even when the room temperature drops, so air conditioning is not operated and increases in power demand are suppressed.

[0081] In the above embodiment, the processing performed by the CPU after reading the software (program) may be performed by various processors other than the CPU. Examples of processors in this case include dedicated electrical circuits, such as programmable logic devices (PLDs) whose circuit configuration can be changed after manufacture, such as field-programmable gate arrays (FPGAs), graphics processing units (GPUs), and application-specific integrated circuits (ASICs), which are processors having a circuit configuration designed specifically to perform specific processing. Furthermore, the processing may be performed by one of these various processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, a combination of a CPU and an FPGA, etc.). Furthermore, the hardware structure of these various processors is, more specifically, an electric circuit that combines circuit elements such as semiconductor elements.

[0082] In the above embodiment, various programs are pre-stored (installed) in the storage 14, but the present invention is not limited to this. The programs may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The programs may also be downloaded from an external device via a network.

[0083] The following additional notes are provided regarding the above-described embodiments.

[0084] (Supplementary Item 1) A learning device including: a memory; and at least one processor connected to the memory, wherein the processor is configured to: extract consumer characteristics including a predetermined profile, temperature sensitivity, and cluster of a consumer using weather information, consumer information, and demand history as input; and learn a demand forecasting model using a machine learning technique using learning data including the weather information, consumer information, demand history, and the consumer characteristics as input.

[0085] (Supplementary Item 2) A non-transitory storage medium storing a program executable by a computer to perform a learning process, which extracts consumer characteristics including a predetermined profile, temperature sensitivity, and cluster of a consumer using weather information, consumer information, and actual demand as input, and learns a demand forecasting model using machine learning techniques using learning data including the weather information, consumer information, actual demand, and the consumer characteristics as input.

[0086] 100 Learning device 102 Feature extraction unit 120 Preprocessing unit 122 Learning unit 200 Prediction device 220 Preprocessing unit 222 Prediction unit

Claims

1. A learning device including: a characteristic extraction unit that uses weather information, consumer information, and demand history as input to extract consumer characteristics including a consumer's specified profile, temperature sensitivity, and cluster; and a learning unit that uses learning data including the weather information, consumer information, demand history, and the consumer characteristics as input to learn a demand forecasting model using a machine learning technique.

2. The learning device according to claim 1, wherein the characteristic extraction unit detects minute fluctuations in the actual demand data and identifies business days and holidays as part of the profile of the consumer using a predetermined evaluation function for comparing data.

3. The learning device described in claim 1, wherein the characteristic extraction unit inputs the demand record of the customer and the temperature in the weather information, extracts business time data, extracts seasonal data, and calculates features of temperature sensitivity, thereby deriving the temperature sensitivity.

4. The learning device of claim 1, wherein the characteristic extraction unit uses the demand record of the consumer as input, calculates representative values ​​based on a representative demand curve for business days and a representative demand curve for non-business days, creates clusters based on the representative values, and classifies each consumer into the cluster.

5. A prediction device including: a pre-processing unit that divides each data of meteorological information, consumer information, and consumer characteristics into regional mesh and cluster units as pre-processing; and a prediction unit that takes as input learning data including pre-processed meteorological information, consumer information, demand record, and said consumer characteristics, and takes as input each data of meteorological information, consumer information, and consumer characteristics, and predicts electricity demand in predetermined units corresponding to said pre-processing using a demand prediction model.

6. A learning method in which a computer executes the following processes: extracting consumer characteristics including a consumer's specified profile, temperature sensitivity, and cluster using meteorological information, consumer information, and actual demand as input; and learning a demand forecasting model using machine learning techniques using learning data including the meteorological information, consumer information, actual demand, and the consumer characteristics as input.

7. A program for causing a computer to function as each part of the learning device according to any one of claims 1 to 4 or the prediction device according to claim 5.

Citation Information

Patent Citations

  • Energy demand prediction system

    JP2018088742A

  • Demand prediction device and demand prediction method and computer program

    JP2019117601A