Training data recommendation system and training data recommendation method

Abstract

Select training data that will increase profits. [Solution] A learning data recommendation system comprising a processor and memory, wherein the processor performs data collection processing to acquire training data used by the estimation model for training, model accuracy estimation processing to estimate the accuracy of the estimation results of the estimation model when the estimation model is trained using the training data, training cost calculation processing to calculate the cost incurred when the estimation model is trained using the training data, processing to calculate the revenue obtained from providing the estimation results output when the estimation model is trained using the training data, and a learning data recommendation system that outputs recommendations for learning data that can be obtained from the profit obtained from training the estimation model based on the cost and revenue.

Description

Technical Field

[0001] The present invention relates to a system and method for recommending learning data.

Background Art

[0002] In recent years, in order to optimize power utilization in a region, when power demand is tight, the suppression of power consumption by power consumers such as households and factories is promoted. Conversely, when power supply becomes excessive due to power generation from renewable energy or the like, a mechanism called demand response (DR) that promotes an increase in the power consumption of power consumers is used.

[0003] During the DR implementation period, methods such as granting incentives according to the amount of adjustment of power consumption by power consumers are being considered.

[0004] Conventionally, technologies for predicting power demand in a target region are known. Also known are technologies for estimating which devices generated the power demand by operating at which times with respect to the power demand.

[0005] Also, in a target region, technologies for predicting power supply from renewable energy based on the introduction rate of renewable energy power generation devices, meteorological data, building layout information, etc. are known.

[0006] It is considered that by combining these technologies, it becomes possible to predict the power demand for each device, the power supply from renewable energy, and the excess or deficiency between these power demand and power supply amounts in a target region.

[0007] Thereby, by predicting the power demand of devices whose operating time zone can be changed and its operating time zone, it is possible to estimate the adjustable amount (adjustment power potential) of the power consumption of consumers during the DR implementation period.

[0008] For power generators, estimated adjustment potential is crucial information for adjusting power generation and reducing wasteful generation. A possible business model involves providing power generators with estimated adjustment potential in a given area and receiving compensation for it. In this business, improving the accuracy of the estimation model is essential for estimating adjustment potential.

[0009] The business model of providing estimation results using estimation models and receiving compensation for them is not unique to the power industry; it is also common in other industries such as logistics, retail, and service industries. Improving the accuracy of estimation results using estimation models is a common challenge across all industries.

[0010] Patent Document 1 discloses a system for "predicting the accuracy of a retraining model when retraining is performed using retraining data including newly collected data." The system disclosed in Patent Document 1 prevents unnecessary retraining and reduces the processing cost of retraining the learning model by performing a determination process to determine whether or not retraining is necessary based on the predicted accuracy of the retraining model. [Prior art documents] [Patent Documents]

[0011] [Patent Document 1] Japanese Patent Publication No. 2021-184139 [Overview of the project] [Problems that the invention aims to solve]

[0012] For estimation models that estimate the adjustment capacity potential in a target region, it is preferable to train them using training data related to that region. Examples of training data related to a target region include smart electricity meter data, although obtaining such data may be subject to a fee.

[0013] Furthermore, training an estimation model using this data generally involves computational costs. Therefore, there is a trade-off between the improvement in the estimation accuracy of the model through training and the cost of training.

[0014] In a business model that earns revenue by providing estimates using estimation models, profit cannot be generated if the cost of training the estimation model exceeds the revenue. It is desirable to be able to select training data that maximizes profit by considering both the cost of training the model and the revenue obtained from the model's estimation results.

[0015] The system disclosed in Patent Document 1 proposes reducing the learning cost by avoiding unnecessary learning based on predictions of model accuracy improvement through learning, but it does not mention the cost of acquiring training data or the compensation obtained from the model's estimation results.

[0016] This invention has been made in view of the above circumstances, and one of its objectives is to provide a training data recommendation system that recommends training data that maximizes profit, taking into account both the cost of training the model and the reward obtained from the model's estimation results. [Means for solving the problem]

[0017] The above problem is solved by a system comprising a processor and memory, the processor performing data collection processing to acquire training data used by the estimation model for training, model accuracy estimation processing to estimate the accuracy of the estimation results of the estimation model when the estimation model is trained using the training data, training cost calculation processing to calculate the cost incurred when the estimation model is trained using the training data, processing to calculate the revenue obtained from providing the estimation results output when the estimation model is trained using the training data, and a training data recommendation system that outputs recommendations for training data that can be obtained from the profits obtained from training the estimation model based on the cost and revenue. [Effects of the Invention]

[0018] According to the present invention, it is possible to select learning data that increases profits.

Brief Description of the Drawings

[0019] [Figure 1] It is an example of a block diagram showing a configuration example of a network to which the learning data recommendation system in the embodiment is applied. [Figure 2] It is an example of a diagram showing the hardware configuration of the learning data recommendation system in the embodiment. [Figure 3] It is an example of a block diagram showing the functional configuration of the learning data recommendation system in Embodiment 1. [Figure 4] It is an example of a diagram for explaining the operation of the learning data recommendation system in Embodiment 1. [Figure 5] It is a diagram showing an example of smart meter data stored in the data storage unit in the embodiment. [Figure 6] It is a diagram showing an example of estimated model information in the embodiment. [Figure 7] It is a diagram showing an example of learning data candidate information in the embodiment. [Figure 8] It is an example of a flowchart showing the processing of the learning data recommendation system in Embodiment 1. [Figure 9] It is a diagram for explaining a procedure for calculating a change in consideration obtained from the business system in Embodiment 1. [Figure 10] It is a diagram showing an example of a procedure for selecting learning data that increases profits in Embodiment 1. [Figure 11] It is an example of a screen for presenting the learning data recommendation result in Embodiment 1. [Figure 12] It is a diagram showing an example of a procedure for calculating a consideration obtained when future regional information is input in Embodiment 2. [Figure 13] It is an example of a screen for displaying the learning data recommendation result in units obtained by dividing the area in Embodiment 3. [Figure 14]This is an example of a screen that displays the learning data recommendation results created based on the input of the customer's electricity demand adjustment history in Embodiment 4. [Modes for carrying out the invention]

[0020] [Embodiment 1] Embodiments will be described with reference to the drawings. Note that the embodiments described below are not intended to limit the invention as defined in the claims, and not all of the components and combinations thereof described in the embodiments are necessarily essential to the solution of the present invention.

[0021] In this embodiment, we will explain the case where the learning data recommendation system is applied to an estimation model that estimates the adjustment capacity potential in electricity demand, but it can also be applied to systems that provide results estimated by estimation models in any industry, such as logistics, retail, or service industries. [network] Figure 1 is an example block diagram showing an example of the network configuration to which the learning data recommendation system in the embodiment is applied. In Figure 1, the learning data recommendation system 110 is connected to the network 120. The network 120 is also connected to one or more data generators 130, one or more data storage units 140, an estimation system 150, and a business system 160.

[0022] Furthermore, the learning data recommendation system 110 is connected to an operating terminal 170 for operating the learning data recommendation system 110. The operating terminal 170 does not necessarily need to be directly connected to the learning data recommendation system 110; it may be accessed via a web browser over a network.

[0023] The data generator 130 may be, for example, a power smart meter data generator that obtains the electricity usage history of consumers, satellite image data, a sensor that observes local weather information, or a sensor that measures local pedestrian traffic. The data collected or generated by these data generators 130 is transmitted to the data storage unit 140 via the network 120.

[0024] The data storage unit 140 is, for example, a storage device such as a server or memory, and stores data collected or generated by the data generator 130. It may also store metadata such as geographical information of where the data generator 130 is installed and the operating status of the data generator 130.

[0025] These metadata may be provided by the data generator 130, by the business operator owning the data generator 130, or by the business operator owning the data storage unit 140.

[0026] The estimation system 150 has an estimation model for estimating the adjustment force potential. Based on the data stored in the data storage unit 140, the estimation system 150 can estimate the adjustment force potential, which is the output data of the estimation model.

[0027] The business system 160 operates using the estimated value of the adjustment capacity, which is output data estimated by the estimation system 150. The operator that owns the business system 160 may be a power generation company, a power distribution company, or a renewable energy company.

[0028] The business operator possessing the learning data recommendation system 110 may be the same as the business operator possessing the estimation system 150. Furthermore, the business operator possessing the data storage unit 140 may be the same as the business operator possessing the learning data recommendation system 110 or the business operator possessing the estimation system 150.

[0029] Furthermore, the data storage unit 140 may be located at the same location as the data generating device 130, or at the location of the business operator that owns the estimation system 150. Also, the business operator that owns the business system 160 may be the same as the business operator that owns the estimation system 150, the business operator that owns the data generating device 130, and the business operator that owns the data storage unit 140.

[0030] [System Configuration] Figure 2 is an example of a diagram showing the hardware configuration of a learning data recommendation system in an embodiment.

[0031] The learning data recommendation system 110 is implemented on a computer that includes a processor 301 such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit) that controls the learning data recommendation system 110 as a whole, memory 302 such as ROM (Read Only Memory) or RAM (Random Access Memory) that stores various processing programs for realizing the functions of the learning data recommendation system 110, external storage devices 303 such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), input / output devices 304 such as a keyboard, mouse, or touch panel, and a network interface 305 such as a NIC (Network Interface Card).

[0032] The functions of the learning data recommendation system are realized when the processor 301 executes various processing programs stored in memory 302 while referring to data stored in external storage device 303.

[0033] Some or all of these programs may be introduced from other devices via non-temporary storage media or communication lines, or they may be pre-stored. In this embodiment, an example of implementation using a standalone computer is described, but it may also be implemented using a cloud service that provides computing resources.

[0034] Figure 3 is an example of a block diagram showing the functional configuration of the learning data recommendation system in Embodiment 1. As shown in Figure 3, the learning data recommendation system 110 includes a model information acquisition unit 210, a data collection unit 220, a model accuracy estimation unit 230, a learning cost calculation unit 240, a learning data recommendation unit 250, and an input / output unit 260.

[0035] The learning data recommendation system 110 is connected to the data storage unit 140 via the network 120.

[0036] The learning data recommendation system 110 is connected to the operation terminal 170 via the input / output unit 260. The operation terminal 170 provides the user with an environment for inputting predetermined information into the model information acquisition unit 210, the data collection unit 220, the model accuracy estimation unit 230, the learning cost calculation unit 240, and the learning data recommendation unit 250.

[0037] The learning data recommendation system 110 is connected to the estimation system 150 via the network 120. The estimation system 150 is connected to the data storage unit 140 via the network 120.

[0038] The training data recommendation system 110 estimates the accuracy of the estimation results of the trained model of the estimation system 150 when it is trained with data acquired from the data storage unit 140. Based on the cost of training and the compensation obtained for providing the estimation results of the trained model, it provides the estimation system 150 with a recommendation result for data to be trained (training data recommendation result). The estimation system 150 provides the estimation results from the estimation model to the business system 160.

[0039] The training data recommendation system 110 receives a training data recommendation system usage fee based on the profits derived from the estimation results obtained from the trained model included in the recommendation results for data to be trained, and the costs incurred in training. [Operating Procedure] Figure 4 is an example of a diagram illustrating the operation of the learning data recommendation system in Embodiment 1.

[0040] A user of the training data recommendation system 110 uses the operation terminal 170 to issue instructions to the training data recommendation system 110 for training data recommendation (S301).

[0041] Next, the training data recommendation system 110 obtains information about the estimation model that recommends the training data from the estimation system 150 (S302).

[0042] Next, the learning data recommendation system 110 obtains learning data candidate information from the data storage unit 140, which is information about data that will be used as candidate learning data by the estimation system 150 for learning (S303).

[0043] In this case, conditions may be imposed on the data to be acquired, such as limiting the selection of data to only those generated during a specific period or region, or the user may specify these conditions on the operating terminal 170.

[0044] Next, the training data recommendation system 110 selects data from the training data candidate information that the estimation model in the estimation system 150 should learn from, and as a result creates a training data recommendation result, which is output to the operation terminal 170 (S304).

[0045] Users of the training data recommendation system 110 refer to the training data recommendation results presented in S304 using the operation terminal 170 and confirm the training data to be used in the estimation system 150 (S305). At this time, the training data recommendation system 110 may output multiple training data candidates to the operation terminal 170 and accept the user's selection of training data.

[0046] The learning data recommendation results presented by the learning data recommendation system 110 may be reviewed automatically by setting conditions in advance. If the conditions are not met, the learning data recommendation results may be recreated.

[0047] The user may specify conditions to be imposed on the training data recommendation results using the operating terminal 170. Furthermore, the creation of the training data recommendation results may be redone based on user operations from the operating terminal 170.

[0048] Next, the training data recommendation system 110 provides the training data recommendation results to the estimation system 150 (S306).

[0049] The estimation system 150 acquires training data from the data storage unit 140 based on the training data recommendation results provided by the training data recommendation system 110 (S307).

[0050] The estimation system 150 uses the training data acquired from the data storage unit 140 to train the estimation model (S308).

[0051] Here, as long as the learning data recommendation system 110 creates learning data recommendation results for the data acquired from the data storage unit 140, and the estimation system 150 trains the estimation model based on the learning data recommendation results, the order of steps S301 to S308 in Figure 4 may be changed. [Data Contents] Figure 5 shows an example of smart meter data stored in the data storage unit in the embodiment. The data registered in the data storage unit 140 includes smart meter data 410, etc. The smart meter data 410 consists of two parts: electricity usage history information 410a and geographic information 410b.

[0052] 410a contains data ID 411, which uniquely identifies the data source, and information such as power usage history 412 from smart meter data corresponding to data ID 411. Power usage history 412 is time-series data that records the amount of power used 414 at time 413.

[0053] 410b contains information such as a data ID 411 that uniquely identifies the source of the smart meter data, and geographic information 415 that indicates the location where the smart meter data corresponding to data ID 411 exists.

[0054] Furthermore, the geographic information 415 contains coordinates 416 relating to the location where the smart meter data exists, and land use information 417 relating to the location where the smart meter data exists. The coordinates 416 consist of latitude and longitude information. The coordinates 416 may also represent information that shows a polygon shape composed of positional information of multiple points.

[0055] Figure 6 shows an example of estimated model information in an embodiment. The estimated model information 510 includes information such as a model ID 511 that uniquely identifies the estimated model and model information 512 that shows the details of the estimated model.

[0056] The model information 512 includes a creation date and time 513 indicating the date and time the model was created, input parameters 514 indicating the format of the input data to the model, output parameters 515 indicating the format of the output data from the model, previously trained data 516 indicating information on training data that has been used to train the model in the past, and information on payment acquisition 517 indicating the relationship between the payment obtained by the estimation system 150 from the business system 160 and the estimated value of the model.

[0057] Figure 7 shows an example of training data candidate information in an embodiment. The training data candidate information 610 includes information such as a training data candidate ID 611 that uniquely identifies a training data candidate, and candidate data information 612 that indicates information about the training data candidate.

[0058] The candidate data information 612 contains information such as the creation date and time 613, which indicates the date and time the training data candidate was created; the data ID 411, which uniquely identifies the data source; the acquisition cost 615, which indicates the cost of acquiring the data; and the data features 616, which indicates the characteristics of the training data candidate. In this example, information indicating the region from which the candidate data was acquired is stored. [Process for generating recommendation results from training data] Figure 8 is an example flowchart showing the processing of the learning data recommendation system in Embodiment 1. The steps shown in Figure 8 correspond to steps S302 to S304 of the operation procedure shown in Figure 4.

[0059] In S302, the model information acquisition unit 210 acquires estimated model information from the estimation system 150, which is information about the estimated model in the estimation system 150 (S701).

[0060] In S303, the data collection unit 220 acquires candidate training data information from the data storage unit 140 as information about data that will be used as candidate training data for the estimation system 150 to learn (S702).

[0061] In S304, the model accuracy estimation unit 230 uses the information on the estimated model of the estimation system 150 acquired in S701 and the candidate training data information acquired in S702 to predict the improvement in the accuracy of the estimation result of the estimated model when it is trained using a specific candidate training data from the candidate training data information (S703).

[0062] In S304, the learning cost calculation unit 240 uses the candidate learning data information acquired in S702 to calculate the cost incurred when training the estimation model of the estimation system 150 with a specific candidate learning data in the candidate learning data information (S704).

[0063] In S304, the learning data recommendation unit 250 calculates the change in compensation obtained from the business system 160 by improving the accuracy of the estimation results of the estimation model predicted in S703 (S705).

[0064] In this case, regarding the compensation obtained by providing the output data of the estimation model of the estimation system 150 to the business system 160, information showing the relationship between the estimation results output by the model and their accuracy and the compensation obtained may be provided by the estimation system 150 as compensation acquisition information 517 included in the estimation model information 510, or it may be provided by the business system 160.

[0065] The procedure by which the training data recommendation unit 250 calculates the change in compensation resulting from the improvement in the accuracy of the output data of the estimation model is described later in Figure 9.

[0066] The processing from S703 to S705 for a specific training data candidate in the training data candidate information may be repeated until a pre-set condition is met. Alternatively, a search may be performed on all training data candidates in the training data candidate information, or the search for training data candidates may be repeated until the number of training data candidates whose accuracy exceeds a pre-set accuracy in the output data of the trained estimation model reaches a pre-specified threshold.

[0067] Furthermore, the user may specify the termination condition for the search process for training data candidates in the training data candidate information using the operation terminal 170 (S706). In S706, when a search is performed on all training data candidates in the training data candidate information, the training data recommendation unit 250 can select the training data candidate that has the lowest training cost in order for the accuracy of the estimation result of the trained estimation model to reach a preset accuracy.

[0068] Then, in S304, the training data recommendation unit 250 selects the training data candidate that maximizes profit from the cost calculated in S704 and the compensation calculated in S705 as the training data recommendation result (S707). [Calculating the benefits gained from improved accuracy] Figure 9 illustrates the procedure for calculating the change in compensation obtained from the business system in Embodiment 1. The estimation model 810 of the estimation system 150 is given smart meter data of the target area and land use information, which is characteristic information of the area, as input information (820) to estimate the adjustment capacity potential in the target area (830).

[0069] The business system 160 pays the estimation system 150 a fee (840) for providing the estimation system 150 with an estimate of the adjustment potential. The fee paid by the business system 160 to the estimation system 150 is calculated from the value and accuracy of the estimation of the adjustment potential.

[0070] For example, if the estimation result of the pre-training estimation model 810 shows that the adjustment force potential falls within the range of 100 ± 20 MWh with 90% accuracy, a rule may be set to calculate the payment from the lowest value in the range, 80 MWh, and pay 80k yen. The following explanation will be based on this rule.

[0071] Here, we will explain using an example where the estimation model 810 is trained using a candidate training data set. The estimation model 850, after training with the estimation system 150, takes smart meter data and land use information for the target area as input information (820) and estimates the adjustment capacity potential in the target area (860). The business system 160 pays the estimation system 150 a fee (870) in exchange for the estimation system 150 providing the estimated value of the adjustment capacity potential.

[0072] If the estimated result from the trained estimation model 850 shows that the adjustment potential falls within the range of 100 ± 5 MWh with 90% accuracy, the consideration paid by the business system 160 to the estimation system 150 will be 95,000 yen.

[0073] Therefore, due to the improved accuracy of the estimation model in estimation system 150, the compensation paid from business system 160 to estimation system 150 will increase from 80,000 yen to 95,000 yen. [Creating recommendation results from training data] Figure 10 shows an example of the procedure for selecting training data that increases profit in Embodiment 1. It compares the cases in which the estimation model 910 of the estimation system 150 is trained with training data candidate A (920), with training data candidate B (930), and with training data candidate C (940).

[0074] If the system is trained using candidate training data A, the training cost is 50,000 yen, and the compensation received by the estimated system 150 increases by 60,000 yen.

[0075] If the system is trained using candidate training data B, the training cost is 30,000 yen, and the compensation received by the estimated system 150 increases by 80,000 yen.

[0076] If training is done using training data candidate C, the training cost is 60,000 yen, and the estimated system 150 receives an increased reward of 50,000 yen. In this case, training with training data candidate B yields the greatest profit, so training data candidate B is selected as the training data candidate recommendation result.

[0077] Thus, when the estimation model 910 is trained on one or more training data candidates, the training data candidate that maximizes the profit calculated using the training cost and the compensation received by the estimation system 150 is designated as the training data recommendation result.

[0078] Here, we can prioritize the learning process based on the level of profit, and present multiple training data candidates to the user along with their priorities.

[0079] Alternatively, the system may be trained using multiple training datasets, and the combination of training dataset candidates that maximizes profit may be presented to the user as a training dataset recommendation result. [Screen display] Figure 11 shows an example of a screen displaying the training data recommendation results in Embodiment 1. In Figure 11, screen 1010 includes a shaded area 1020 for setting the region for which the adjustment potential is to be estimated, an area 1030 for displaying the estimated adjustment potential and its accuracy, an area 1040 for displaying the training data recommendation results, and an area 1050 for displaying a list of training data candidates.

[0080] Area 1020 displays map information with the distribution area superimposed. The user may also be able to select one or more distribution areas. Area 1020 displays information about the adjustment potential estimated by the estimation model of the estimation system 150 for the selected distribution area. The information about the adjustment potential may be displayed as a numerical value, or the value may be represented by color.

[0081] Region 1020 may display information about the adjustment potential, specifically the estimation accuracy of the adjustment potential. It may also display the estimated result of accuracy improvement due to training.

[0082] Region 1030 has the function of displaying time-series information about the adjustment force potential estimated by the estimation model of estimation system 150. In region 1030, the user can set the period to be displayed as time-series information.

[0083] Region 1030 may display both the numerical value of the adjustment force potential and the estimation accuracy of the adjustment force potential, or it may display other information as well.

[0084] Area 1040 has the function of displaying the training data recommendation results created by the training data recommendation unit 250. The training data recommendation results display the type of data in question, the profit obtained by subtracting costs from the value obtained by training using the data, etc.

[0085] The training data recommendation results may include multiple training data candidates, and these may be displayed along with their training priority. Furthermore, the system may have a function to instruct training based on user input using the displayed training data recommendation results.

[0086] In domain 1040, the estimation model of the estimation system 150 may output only training data candidates in the training recommendation results that satisfy a predetermined accuracy for the estimation accuracy of the adjustment power potential output by the trained estimation model.

[0087] Region 1040 may be sorted and rearranged based on the learning cost, according to the order in which the candidate training data are presented to the user.

[0088] Area 1050 has the function of displaying a list of training data candidates acquired by the data acquisition unit 220 from the data storage unit 140. The display of the list of training data candidates may also include information such as the type of data, the cost and benefit of acquiring the data, and the lead time required to acquire the data.

[0089] Screen 1010 may interactively provide information to the user using artificial intelligence, such as generative AI. [Embodiment 2] An example of the procedure by which the learning data recommendation system 110 according to Embodiment 2 presents the learning data recommendation results to the user will be described.

[0090] Figure 12 shows an example of the procedure for calculating the compensation obtained when future regional information is used as input in Embodiment 2. For the estimation model 1110 of the estimation system 150 before training, the adjustment capacity potential in the target region is estimated (1130) using smart meter data and land use information, which is characteristic information of the region, as input information (1120).

[0091] The business system 160 pays the estimation system 150 a fee (1140) for providing the estimation system 150 with an estimate of the adjustment potential. The fee paid by the business system 160 to the estimation system 150 is calculated from the value and accuracy of the estimation of the adjustment potential.

[0092] For example, if the estimation result of the pre-training estimation model 1110 falls within the range of 100 ± 20 MWh with 90% accuracy for the adjustment potential, a rule may be set to calculate the price to be paid from the lowest value in the range, 80 MWh, and pay 80k yen. The following explanation will be based on this rule.

[0093] Let's explain using an example where the estimation model 1110 is trained using a candidate set of training data. After the estimation system 150 has been trained, the estimation model 1150 may predict the future state based on smart meter data and land use information for the target area, and use the prediction results as input information (1160) to estimate the adjustment capacity potential in the target area (1170).

[0094] The information on the future state of the target region used as input information by the estimation model 1150 may be information predicted by the learning data recommendation system 110. The future state may be predicted from information such as land development plans, changes in weather conditions, and policy guidelines for the target region.

[0095] Business system 160 pays the estimation system 150 a fee (1180) in exchange for providing the estimation of the adjustment potential from the estimation system 150. For example, if the estimation result from the trained estimation model 1150 shows that the adjustment potential falls within the range of 150 ± 10 MWh with 90% accuracy, the fee that business system 160 pays to the estimation system 150 will be 140k.

[0096] The learning data recommendation system 110 may calculate profits based on the future revenue that the estimation system 150 will receive from the business system 160, and create learning data recommendation results. [Embodiment 3] An example of the procedure by which the learning data recommendation system 110 according to Embodiment 3 presents the learning data recommendation results to the user will be described.

[0097] Figure 13 shows an example of a screen displaying the learning data recommendation results in units of area divisions in Embodiment 3. Data such as smart meter data and land use information may have specified area units for acquisition. For example, smart meter data may be available in units of several square kilometers.

[0098] Since power distribution areas typically span one or more municipalities, there may be multiple sets of data, such as smart meter data and land use information, for the distribution area targeted for estimation of adjustment capacity potential. The estimation system 150 may divide the power distribution area into units for which smart meter data can be acquired, and perform the estimation of adjustment capacity potential for each divided unit.

[0099] For each zone within the distribution area, the estimated regulating capacity differs depending on land use patterns, weather conditions, etc. In mountainous areas where there are no buildings, the regulating capacity is small and can be estimated with high accuracy.

[0100] In areas with many commercial facilities that have installed thermal storage equipment, the adjustment capacity is estimated to be high but with low accuracy. The learning data recommendation system 110 may display the numerical value of the adjustment capacity potential for each category (1210), or it may display the estimation accuracy of the adjustment capacity potential (1220).

[0101] The learning data recommendation system 110 may estimate the benefits obtained from learning for each category based on the estimated adjustment power value, the accuracy of the estimated adjustment power, the improvement in the accuracy of the estimated model when learning with data available in that category, and the learning cost when learning with data available in that category. It may also display a learning priority according to the benefits obtained from learning (1230).

[0102] In this case, for each area within the power distribution zone, information such as land use patterns and weather conditions may be similar across different zones. If training is performed using data available for a particular zone, an improvement in estimation accuracy can be expected even for zones similar to that zone.

[0103] The learning data recommendation system 110, when trained using data obtainable in a certain area, estimates the improvement in estimation accuracy for the entire target area for estimation of adjustment potential, and presents the user with the benefits that can be obtained through training.

[0104] The training data recommendation system 110 may, when a user selects a candidate training data, display to the user the degree of improvement in estimation accuracy for each category and the degree of improvement in the benefits obtained by training, when the system is trained using the candidate training data (1240).

[0105] In this case, if the degree of improvement in estimation accuracy is similar, the user may be presented with information such as the land use patterns of the areas to which the categories belong that are considered similar. [Embodiment 4] An example of the procedure by which the learning data recommendation system 110 according to Embodiment 4 presents the learning data recommendation results to the user will be described.

[0106] Figure 14 shows an example of a screen displaying the learning data recommendation results created based on the input of the customer's electricity demand adjustment history in Embodiment 4.

[0107] The estimation system 150, for example, divides the distribution area into units for which smart meter data can be acquired, and estimates the adjustment potential for each divided unit.

[0108] For each division of the power distribution area, candidate learning data is acquired for each division that has a large track record of adjusting electricity demand by consumers belonging to that division. This adjustment performance data is acquired and stored by the business operator that has business system 160.

[0109] The learning data recommendation system 110 displays the numerical values ​​of the power demand adjustment performance for each category (1310), and displays the categories with the largest power demand adjustment performance values. The learning data recommendation system 110 learns the learning data of the categories specified by the user. By accepting such user specifications, the adjustment capacity potential can be estimated using the learning data of categories with large adjustment performance.

[0110] Alternatively, the system may select training data for categories with high performance values ​​without accepting user specifications.

[0111] This function allows for situations where, for example, in a large supermarket in Category A, the operating time of a thermal storage unit cannot be changed due to hardware limitations, and data on adjustment performance reveals that Category A actually has almost no adjustment history. In such cases, the system can be instructed to prioritize learning data from Category B.

[0112] As described above, according to the embodiment described above, the learning data recommendation system 110 can provide the user with learning data that maximizes the profit calculated from the cost of training the estimation model of the estimation system 150 and the compensation that the estimation system 150 receives from the business system 160.

[0113] In the embodiment described above, an example was explained in which training data is recommended using an estimation model that estimates the adjustment capacity potential in the power industry. However, the training data recommendation system 110 can be applied to any industry, not just the power industry.

[0114] The present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are described in detail for the purpose of clearly illustrating the present invention, and are not necessarily limited to those having all the configurations described.

[0115] Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment.

[0116] Furthermore, some of the configurations in each embodiment can be added to, deleted from, and / or replaced with other configurations. Also, some or all of the above configurations, functions, and processing units may be implemented in hardware, for example, by designing them as integrated circuits. [Explanation of Symbols]

[0117] 110 Learning Data Recommendation System 130 Data Generator 140 Data Storage Unit 150 Estimation Systems 160 Business Systems 170 Operating terminals 210 Model Information Acquisition Unit 220 Data Acquisition Unit 230 Model Accuracy Estimation Unit 240 Learning Cost Calculation Unit 250 Learning Data Recommendation Department

Claims

1. It has a processor and memory, The aforementioned processor includes a data collection process to acquire training data used by the estimation model for training, A model accuracy estimation process that estimates the accuracy of the estimation result of the estimation model when the estimation model is trained using the aforementioned training data, A learning cost calculation process that calculates the cost incurred when the estimation model is trained using the aforementioned training data, A process for calculating the compensation obtained for providing the estimation results output when the estimation model is trained using the aforementioned training data, A learning data recommendation system that outputs a recommendation result for learning data, which calculates the benefit obtained by learning the estimation model based on the aforementioned cost and consideration.

2. In the learning data recommendation system according to claim 1, A training data recommendation system that, based on the aforementioned costs and considerations, outputs training data with lower costs in order for the estimation model to achieve a predetermined estimation accuracy of the estimation result.

3. In the learning data recommendation system according to claim 1, A learning data recommendation system that outputs learning data to maximize the profit obtained by learning the estimation model, based on the aforementioned costs and considerations.

4. A learning data recommendation system according to claim 3, The aforementioned estimation result represents the potential for adjusting electricity demand, as determined by the learning data recommendation system.

5. A learning data recommendation system according to claim 4, The estimation model is a learning data recommendation system that estimates the adjustment potential of an area by taking characteristic information of the area as input.

6. A learning data recommendation system according to claim 5, A learning data recommendation system that provides estimated values ​​based on the adjustment potential and estimation accuracy estimated by the aforementioned estimation model, and calculates profit.

7. A learning data recommendation system according to claim 5, The estimation model is a learning data recommendation system that accepts predictions of future characteristic information of the area and estimates and outputs the future adjustment potential of the area.

8. A learning data recommendation system according to claim 5, The estimation model is a learning data recommendation system that estimates the adjustment potential in units of the divided area and outputs the estimated value and estimation accuracy of the adjustment potential in the divided unit.

9. A learning data recommendation system according to claim 8, The estimation model is a learning data recommendation system that outputs the estimation accuracy of the adjustment potential and the degree of improvement in the obtained benefits when the estimation model is trained with learning data of divisions that have similar regional information to the divisions of the area.

10. A learning data recommendation system according to claim 5, A learning data recommendation system that estimates the adjustment capacity of an area based on the power demand adjustment performance of the divisions belonging to the area to be estimated, using learning data of the divisions with a large adjustment performance among the divisions.

11. The processor acquires the training data that the estimation model will use for training, The accuracy of the estimation result of the estimation model is predicted when the estimation model is trained using the aforementioned training data. The cost incurred when the estimation model is trained using the aforementioned training data is calculated. The compensation obtained for providing the estimation results output when the estimation model is trained using the aforementioned training data is calculated. A method for recommending training data that outputs a recommendation result for training data, which calculates the benefit obtained by training the estimation model based on the aforementioned cost and consideration.

Images