Electricity consumption forecasting device and electricity consumption forecasting method

Abstract

Even when there is a discrepancy between the operational plan and actual operational results, it is possible to accurately predict electricity consumption at a specific point in the future. [Solution] The power consumption forecasting device (1) includes a data acquisition unit (131) that acquires input data including the factory's past and future operation plan, past operation results, and past power consumption results, and a power consumption inference unit (134) that infers the amount of power consumption expected to be used in the factory during a first period by inputting the input data into a trained model (MD).

Description

[Technical Field]

[0001] This disclosure relates to an electricity consumption forecasting device and an electricity consumption forecasting method, etc. [Background technology]

[0002] Large factories that use a lot of electricity, such as steel mills, generate their own electricity and also enter into electricity purchase agreements with power companies to purchase electricity that cannot be covered by their own generation. Generally, in electricity purchase agreements, the contracted amount of electricity is determined based on the maximum amount of electricity used in a specified period (for example, every 30 minutes), and a penalty is imposed if the actual amount of electricity used exceeds the contracted amount. On the other hand, even if the actual amount of electricity used is significantly less than the contracted amount, the amount equivalent to the contracted amount is still paid. Under such contracts, it is economically important for factories to adjust their demand so that their electricity usage is always kept stable, and to do so, it is necessary to accurately predict the amount of electricity the factory will use and the amount of electricity supplied by its own generation for a specified period of time in advance.

[0003] Patent Document 1 discloses a method for predicting power consumption, which involves calculating the final predicted power consumption for a demand cycle from the amount of power consumed within the demand cycle and a preset planned load.

[0004] In factories with high electricity consumption, such as steel mills, where electricity consumption varies greatly depending on the manufacturing conditions of the rolled material, electricity consumption changes significantly in a short period of time. Therefore, it is important to accurately grasp information about the rolled material to be rolled within a predetermined time, i.e., the factory's operating plan, and predict electricity consumption. However, the method disclosed in Patent Document 1 does not predict electricity consumption based on the operating plan.

[0005] On the other hand, Patent Document 2 describes a method for predicting electricity consumption based on an operation plan. Specifically, the method disclosed in Patent Document 2 modifies the future operation plan based on the degree of change in actual operation compared to the operation plan within a predetermined period in the past, and predicts the future electricity consumption of the factory based on the modified operation plan. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 55-136832 [Patent Document 2] Japanese Patent Publication No. 2024-24914 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] When using operational plans to forecast factory electricity consumption, discrepancies between actual operations and the operational plan degrade the accuracy of electricity consumption forecasts. Therefore, it is important to appropriately revise the operational plan, for example, using the method disclosed in Patent Document 2. However, even after revising the operational plan, it can be difficult to perfectly match it with the actual operation plan.

[0008] One aspect of this disclosure aims to realize a power consumption forecasting device that can accurately predict power consumption at a future point in time, even when there is a discrepancy between the operational plan and the actual operational results. [Means for solving the problem]

[0009] A power consumption forecasting device according to one aspect of the present disclosure is a power consumption forecasting device that forecasts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period later, and comprises: a data acquisition unit that acquires input data including the factory's past and future operational plans, past operational results, and past power consumption results; and a power consumption inference unit that infers the amount of power consumption expected to be used in the factory during the first period by inputting the input data into a trained model.

[0010] A power consumption forecasting method according to one aspect of the present disclosure is a power consumption forecasting method that forecasts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period later, comprising: a data acquisition step of acquiring input data including the factory's past and future operational plan, past operational performance, and past power consumption performance; and a power consumption inference step of inputting the input data into a trained model to infer the amount of power consumption expected to be used in the factory during the first period. [Effects of the Invention]

[0011] According to one aspect of this disclosure, even when there is a discrepancy between the operational plan and the actual operational results, it is possible to accurately predict the amount of electricity used at a certain point in the future. [Brief explanation of the drawing]

[0012] [Figure 1] Block diagram showing an example of a power consumption prediction device according to the embodiments of this disclosure. [Figure 2] This table shows an example of input data. [Figure 3] This flowchart shows an example of the process for generating a pre-trained model. [Figure 4] A flowchart illustrating an example of the process for predicting electricity usage. [Modes for carrying out the invention]

[0013] [Embodiment 1] <Electricity Consumption Prediction Device> Figure 1 is a block diagram showing an example of a power consumption forecasting device 1 according to an embodiment of the present disclosure. The power consumption forecasting device 1 is a device that forecasts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period in the future. Examples of factories include various types of factories such as rolling mills installed in steel mills. As shown in Figure 1, the power consumption forecasting device 1 comprises an input unit 11, an output unit 12, a forecasting unit 13, and a storage unit 14.

[0014] The input unit 11 is a device for users of the power consumption prediction device 1 to input data and signals used by the prediction unit 13, and consists of at least one of a communication module that receives data or signals from other devices, a terminal that connects to other devices, a drive that reads information from a recording medium, and an operating device operated by the user. Examples of operating devices include a keyboard, mouse, and touch panel. The output unit 12 is a device for presenting the calculation results, such as power consumption inferred by the prediction unit 13, to users of the power consumption prediction device 1, and consists of at least one of a communication module that transmits data or signals to other devices, a terminal that connects to other devices, a drive that writes information to a recording medium, and a display device such as a display that shows images. The storage unit 14 is a device that stores various programs and machine learning models used by the prediction unit 13, as well as data used or calculated by the prediction unit 13, in a read-write manner, and stores various programs and information used by the prediction unit 13. For example, the storage unit 14 stores a trained model MD that infers power consumption during the first period.

[0015] The prediction unit 13 is a device that predicts power consumption. In this embodiment, the prediction unit 13 comprises a data acquisition unit 131, a data calculation unit 132, a model acquisition unit 133, and a power consumption inference unit 134.

[0016] The data acquisition unit 131 is a functional unit that acquires input data including the factory's operational plan from the past to the future, past operational performance, and past power usage performance.

[0017] The operational plan includes information such as the planned start time and end time of operations at the factory, and planned values ​​for the products to be processed per given period. The operational plan includes both past and future operational plans. The operational results include information such as the planned start time and end time of operations at the factory, and actual values ​​for the products processed per given period. The power usage results include the amount of power used per given period to process products.

[0018] If the factory is a rolling mill, the product information shown in the operational plan and operational results includes, for example, information on the coil of the rolled material, such as coil width, coil weight, coil thickness, coil length, coil stretching ratio, and carbon equivalent. The coil stretching ratio is the ratio of the coil length after rolling to the coil length before rolling. The carbon equivalent is an indicator of the hardness of the steel sheet being manufactured. The harder the steel sheet being manufactured, the greater the power consumption.

[0019] The input data includes the operational plan for the factory whose electricity consumption is to be predicted, covering the past and future, as well as the factory's past operational performance and past electricity consumption performance. The data acquisition unit 131 acquires the data necessary for the electricity consumption inference unit 134, described later, to infer electricity consumption in the first period using the trained model MD. Therefore, the operational plan, past operational performance, and past electricity consumption performance included in the input data are data with the time of electricity consumption inference in the first period, i.e., "present", as the reference time.

[0020] The data calculation unit 132 is a functional unit that calculates the amount of change per predetermined period for each of the operation plan, past operation results, and past power usage results when inferring power consumption during the first period. Therefore, in this embodiment, the data acquisition unit 131 acquires data per predetermined period for each of the operation plan, past operation results, and past power usage results as input data.

[0021] The operational plan included in the input data consists of values ​​corresponding to the products planned to be processed at the factory during the first period, and data corresponding to the products planned to be processed at the factory during the period from the present to a point in time one or more units prior to that point in time, up to a point in time one unit later than that point in time.

[0022] The operational plan included in the input data may be expressed as data (change amount) representing the difference between the value corresponding to the product planned to be processed at the factory in the first period and the value corresponding to the product planned to be processed at the factory in the predetermined period immediately preceding the first period, or as data (change amount) representing the difference between the value corresponding to the product planned to be processed at the factory in the period from the present to a point in time one predetermined period prior or several predetermined periods prior, up to a point in time one predetermined period later than that point, and the value corresponding to the product planned to be processed at the factory in the predetermined period immediately preceding that period. The former data expressed as a difference indicates the change in the future operational plan, and the latter data expressed as a difference indicates the change in the past operational plan.

[0023] The past operational data included in the input data corresponds to values ​​for products processed at the factory during the period from one or more units prior to the present to one unit later than that point in time.

[0024] Furthermore, the past operational performance data included in the input data may be expressed as the difference between the value corresponding to the products processed at the factory during the period from one predetermined period prior to the present to one predetermined period later, and the value corresponding to the products processed at the factory during the period immediately preceding that predetermined period.

[0025] The historical power usage data included in the input data represents the amount of power used by the factory during the period from one predetermined period prior to the present, or several predetermined periods prior, to one predetermined period after that point in time.

[0026] Furthermore, the past electricity usage data included in the input data may be expressed as the difference between the amount of electricity used by the factory during the period from one predetermined period prior to the present to one predetermined period later, and the amount of electricity used by the factory during the period immediately preceding that predetermined period.

[0027] Furthermore, as a value corresponding to the product processed at the factory, the data acquisition unit 131 may acquire data such as the average value per unit time of the rolled material rolled over a predetermined period, or the cumulative value of the rolled material.

[0028] Depending on the products processed at the factory, the data acquisition unit 131 may calculate cumulative values ​​for coil length and coil weight, taking into account the impact on power consumption, and acquire data by calculating average values ​​per unit time for other information such as coil width. For example, if one coil is rolled at 3-minute intervals, then 3 coils and 1 / 3 of a coil will be rolled in 10 minutes. In this case, the data acquisition unit 131 may calculate the coil width x as the average value per unit time, which is the value obtained by dividing the coil width of each coil after rolling by 10 minutes. p,w or coil width x a,wObtain the data calculated as the value of. On the other hand, for the coil weight, as the above integrated value, the total weight of three coils and one-third of a coil is the coil weight x p,wt or the coil weight x a,wt Obtain the data calculated as the value of. Note that, as will be described later, the coil width x p,w and the coil weight x p,wt is an example of an operation plan, and the coil width x a,w and the coil weight x a,wt is an example of past operation results.

[0029] Figure 2 is a table showing an example of input data. In Figure 2, as an example of the product information shown in the operation plan and operation results, the coil width and coil weight are illustrated. In Figure 2, t indicates the current time (now), and t + 1 indicates 10 minutes after the current time (for example). Also, t - n (n is a natural number) indicates n × 10 minutes before the current time. That is, in Figure 2, the predetermined period is 10 minutes. Note that, hereinafter, for the sake of simplicity of explanation, mainly the case where the difference value (change amount) is used as the operation plan, past operation results, and past power consumption results will be taken as an example for explanation, but it is not limited to such a case.

[0030] In Figure 2, the change amount (difference value) of the coil width x p,w , which is an example of a future operation plan, is Δx p,w (t + 1)=x p,w (t + 1)-x p,w (t) and is calculated as. x p,w (t + 1) is the coil width planned to be processed in the period from now to 10 minutes later (the first period). x p,w (t) is the coil width planned to be processed in the period from 10 minutes before now to now. The said period is the period of one predetermined period before the first period, and is the period from the time point one predetermined period before now to the time point one predetermined period after that time point.

[0031] Also, the coil width xp,w The amount of change (difference value) is, Δx p,w (t) = x p,w (t)-x p,w (t-1) Δx p,w (t-1)=x p,w (t-1)-x p,w (t-2) Δx p,w (t-2)=x p,w (t-2)-x p,w (t-3) x is calculated as follows: p,w (t-1) is the coil width that was planned to be processed during the period from 20 minutes ago to 10 minutes ago. This period is from two predetermined periods prior to the present to one predetermined period after that point. p,w (t-2) is the coil width that was planned to be processed during the period from 30 minutes ago to 20 minutes ago. This period is from three predetermined time units prior to the present to one predetermined time unit after that point. p,w (t-3) is the coil width that was planned to be processed during the period from 40 minutes ago to 30 minutes ago from the present. This period is from four predetermined periods before the present to one predetermined period after that point.

[0032] Coil weight x is an example of future and past operational plans. p,wt The amount of change (difference value) is also shown in the coil width x shown in future and past operation plans. p,w The amount of change over the same period is calculated. That is, the coil weight x in the future operation plan. p,wt The amount of change is Δx p,wt (t+1)=x p,wt (t+1)-x p,wt (t) This is calculated as follows. Also, the coil weight x in past operating plans p,wt The amount of change is Δx p,wt (t) = xp,wt (t)-x p,wt (t-1) Δx p,wt (t-1)=x p,wt (t-1)-x p,wt (t-2) Δx p,wt (t-2)=x p,wt (t-2)-x p,wt (t-3) This is the result.

[0033] Furthermore, coil width x is an example of past operational performance. a,w and coil weight x a,wt The change (difference value) is the coil width x shown in past operation plans. p,w and coil weight x p,wt The amount of change over the same period is calculated. That is, the coil width x in past operating records. a,w The amount of change is Δx a,w (t) = x a,w (t)-x a,w (t-1) Δx a,w (t-1)=x a,w (t-1)-x a,w (t-2) Δx a,w (t-2)=x a,w (t-2)-x a,w (t-3) This is calculated as follows. Also, the coil weight x from past operating records a,wt The amount of change is Δx a,wt (t) = x a,wt (t)-x a,wt (t-1) Δx a,wt (t-1)=x a,wt (t-1)-x a,wt (t-2) Δx a,wt (t-2)=x a,wt (t-2)-x a,wt (t-3) This is the result.

[0034] Furthermore, the change in past electricity usage (difference) is used as the amount of electricity used by the factory during the period corresponding to past operating performance. ΔW(t)=W(t)-W(t-1) ΔW(t-1)=W(t-1)-W(t-2) ΔW(t-2)=W(t-2)-W(t-3) The calculation is as follows: W(t) represents the current power usage. W(t-1) represents the power usage 10 minutes ago, W(t-2) represents the power usage 20 minutes ago, and W(t-3) represents the power usage 30 minutes ago. In other words, the change in past power usage (difference value) represents the change in power usage (difference value) for the period corresponding to past operating performance.

[0035] Therefore, ΔW(t) represents the difference between the amount of electricity used from 10 minutes ago to the present and the amount of electricity used from 20 minutes ago to 10 minutes ago. Also, ΔW(t-1) represents the difference between the amount of electricity used from 20 minutes ago to 10 minutes ago and the amount of electricity used from 30 minutes ago to 20 minutes ago, and ΔW(t-2) represents the difference between the amount of electricity used from 30 minutes ago to 20 minutes ago and the amount of electricity used from 40 minutes ago to 30 minutes ago.

[0036] The model acquisition unit 133 is a functional unit that acquires a trained model MD. In this embodiment, the model acquisition unit 133 uses a point in time prior to the present as the reference time, and acquires a trained model MD by performing machine learning on the following: the operational plan for the period from a point in time prior to the reference time to a point in time after the reference time, the operational results prior to the reference time, and the power usage results prior to the reference time as explanatory variables, and the power usage results for a second period from the reference time to a point in time a predetermined period later as the objective variable. As a result, the power usage inference unit 134 can infer the power usage for the first period using the trained model MD generated by its own device.

[0037] Specifically, the point in time referred to here as "a point in time prior to the present" is the point in time before the creation of the trained model MD (before training is complete). The second period is also the period before the creation of the trained model MD. Therefore, the model acquisition unit 133 generates the trained model MD by performing machine learning using training data that includes changes in the operation plan, changes in the actual operation, and changes in the actual power consumption for the period in which factory operations have already been completed at the point in time before the creation of the trained model MD.

[0038] During the generation of the trained model MD, the data acquisition unit 131 acquires the data necessary to generate the trained model MD.

[0039] Furthermore, when generating the trained model MD, the data calculation unit 132 calculates the amount of change per predetermined period at a point in time prior to the present for each of the operation plan, operation results, and power usage results as training data. In other words, in this embodiment, the data acquisition unit 131 acquires data for each of the operation plan, operation results, and power usage results at a point in time prior to the present for a predetermined period. Hereafter, "a point in time prior to the present" will be referred to as "a point in the past."

[0040] The change in the operational plan included in the training data is represented as the difference between the value corresponding to the products planned to be processed at the factory during the period from a certain point in the past to a point one unit later, and the value corresponding to the products planned to be processed at the factory during the period one unit earlier, which is one unit prior to that period. Furthermore, the change in the operational plan is represented as the difference between the value corresponding to the products planned to be processed at the factory during the period from a certain point in the past to a point one unit earlier or more units earlier, to a point one unit later, which is one unit prior to that point, and the value corresponding to the products planned to be processed at the factory during the period one unit earlier, which is one unit prior to that period.

[0041] The change in operational performance included in the training data is represented by the difference between the value corresponding to the products processed at the factory during the period from a certain point in the past to a point one or more predetermined periods prior to that point, and the value corresponding to the products processed at the factory during the predetermined period immediately preceding that period.

[0042] The change in power usage data included in the training data is represented by the difference between the power usage data for the period from one predetermined period prior to a certain point in the past to one predetermined period later than that point in time, and the power usage data for the predetermined period immediately preceding that period; and by the difference between the power usage data for the second period and the power usage data for the predetermined period immediately preceding that second period.

[0043] The input data shown in Figure 2 is an example of training data, obtained by replacing "present" with "a point in the past" in the explanation of Figure 2 above. In this case, the actual power usage during the second period can be represented by W(t+1).

[0044] In this embodiment, the data calculation unit 132 calculates four change amounts (for different periods) as the operation plan included in the training data, and three change amounts (for different periods) as the actual operation and actual power usage, respectively, but is not limited to this. The data calculation unit 132 can appropriately change the number of change amounts calculated as the operation plan (number of periods used in the calculation) and the number of change amounts calculated as the actual operation and actual power usage (number of periods used in the calculation). In this case, the number of change amounts calculated as the operation plan must be at least one greater than the number of change amounts calculated as the actual operation and actual power usage, but the number itself is not limited as long as it is at least one greater. Therefore, the data calculation unit 132 may also appropriately change the input data, such as 10 change amounts and 7 change amounts, rather than limiting it to four change amounts and three change amounts for the operation plan, past operation and past power usage, respectively.

[0045] One example of a machine learning method for generating a trained model MD is random forest regression. However, the method is not limited to this; trained model MDs may also be generated using machine learning methods such as neural networks or support vector regression. A trained model MD can be generated using known machine learning methods with the above-mentioned explanatory variables and target variables as training data. That is, a large amount of data consisting of the above-mentioned explanatory variables and target variables as training data is prepared, and the trained model MD is generated by inputting this data into a machine learning model and optimizing the weights.

[0046] The power consumption inference unit 134 is a functional unit that infers the amount of power consumption (relative value) expected to be used in the factory during the first period by inputting input data into a trained model MD. In this embodiment, the power consumption inference unit 134 inputs the amount of change calculated by the data calculation unit 132 as input data into the trained model MD. The power consumption inference unit 134 is also a functional unit that uses the inferred power consumption (relative value) to calculate a predicted value (absolute value) of power consumption at a point in time one period later from the present. The power consumption inference unit 134 uses the following equation (1); W p (t+1)=W(t)+ΔW p (t+1)…Formula (1) Using this method, we calculate the amount of electricity used at a point in time one period later than the present.

[0047] W p (t+1) is the predicted value of electricity consumption at a point in time one period after the present. W(t) is the actual electricity consumption at the present time. ΔW p (t+1) is the predicted value of the power consumption (relative value) in the first period, inferred by inputting the input data into the trained model MD.

[0048] Furthermore, the power consumption inference unit 134 infers the amount of power consumption (relative value) that is expected to be used in the factory during a new first period, which is the period from the end of the first period to a predetermined period later, by inputting new input data into the trained model MD.

[0049] In this case, the data acquisition unit 131 replaces the above-mentioned operation plan with an operation plan that considers the last point in the first period as the reference time and the present as the current time, replaces the above-mentioned past operation results with past operation results that consider the last point in the first period as the reference time and the present as the current time, replaces at least one value in the replaced past operation results with a preset value, and acquires data obtained by adding the amount of electricity used in the first period inferred by the electricity usage inference unit 134 to the past electricity usage results as new input data.

[0050] Specifically, the data acquisition unit 131 acquires an operation plan in which the reference time is shifted to a predetermined period later than the present, as an operation plan included in the new input data. As a result, the data acquisition unit 131 acquires new input data in which the operation plan included in the most recently input data to the trained model MD has been replaced with an operation plan that uses the last point in the first period as the reference time.

[0051] In Figure 2, the data acquisition unit 131 replaces "t" with "t+1" to set "t+1" (10 minutes from the present) as the reference time. The data acquisition unit 131 acquires the operation plan in accordance with this change in reference time, replacing "t+1" with "t+2", "t-1" with "t", "t-2" with "t-1", and "t-3" with "t-2".

[0052] For example, the data acquisition unit 131 will determine the coil width x in the future operation plan. p,w (t+2) is obtained. Then, the data calculation unit 132 calculates the operation plan included in the new input data as follows: Δx p,w (t+2)=x p,w (t+2)-x p,w (t+1) The data acquisition unit 131 calculates the coil width x in past operation plans.p,w is obtained, and in the data calculation unit 132, Δx p,w (t + 1)=x p,w (t + 1)-x p,w (t) Δx p,w (t)=x p,w (t)-x p,w (t - 1) Δx p,w (t - 1)=x p,w (t - 1)-x p,w (t - 2) is calculated.

[0053] Also, the data acquisition unit 131 acquires, as the past operation results included in the new input data, the past operation results with the reference time shifted by a predetermined period from the current time to a later time for the past operation results. Thereby, the data acquisition unit 131 acquires new input data in which the past operation results included in the input data most recently input to the learned model MD are replaced with the past operation results with the last time point of the first period as the reference time. However, the data acquisition unit 131 acquires, as the new input data, the operation results in which the operation results that do not exist at the time of inferring the power consumption are replaced with a preset value in the past operation results after replacement. The preset value may be, for example, the operation results at the time of inferring the power consumption.

[0054] In the example of FIG. 2, the data calculation unit 132 calculates, as the change amount of the coil width x a,w in the past operation results included in the new input data, Δx a,w (t + 1)=x a,w (t + 1)-x a,w (t) Δx a,w (t)=x a,w (t)-x a,w (t - 1) Δx a,w (t - 1)=x a,w (t - 1)-x a,w (t - 2) However, at the time of inferring the power consumption, the coil width x at the time point of t + 1 a,w(t + 1) has no actual performance. Therefore, the data calculation unit 132 cannot calculate Δx a,w (t + 1). Thus, the data calculation unit 132 replaces the actual performance of the coil width x a,w (t + 1 at the time of t + 1 with a preset value. The preset value is, for example, the actual coil width performance x a,w (t) at the time of inferring the power consumption.

[0055] Also, the data acquisition unit 131 excludes the power consumption performance of the oldest period among the past power consumption performances included in the input data most recently input to the learned model MD, and adds the power consumption amount in the first period inferred by the power consumption inference unit 134 to the past power consumption performances. Thereby, new input data including new past power consumption performances is acquired.

[0056] In the example of FIG. 2, the data acquisition unit 131 excludes W(t - 3) and acquires, as the past power consumption performance, the data obtained by adding W p (t + 1), which is the power consumption amount in the first period inferred by the power consumption inference unit 134.

[0057] By thus acquiring new input data by the data acquisition unit 131, the power consumption inference unit 134 infers the power consumption amount in a new first period, which is a period from the last time point of the first period to a predetermined period later. Then, the power consumption inference unit 134 calculates a predicted value (absolute value) of the power consumption amount at a time point a predetermined period later from the last time point of the first period using the inferred power consumption amount. In the example of FIG. 2, the power consumption inference unit 134 W p (t + 2)=W p (t + 1)+ΔW p (t + 2) to calculate. W p (t + 2) is the predicted value of the power consumption amount at a time point a predetermined period later from the last time point of the first period. W p (t + 1) is the predicted value of the power consumption amount at a time point one first period later from the present (the last time point of the first period) obtained by the above formula (1). ΔW p(t+2) is the power consumption (relative value) for the new first period, inferred by inputting new input data into the trained model MD.

[0058] Furthermore, the power consumption inference unit 134 can infer power consumption for a newer first period, which is a predetermined period after the previous first period, by inputting new input data into the trained model MD.

[0059] In this case, the data acquisition unit 131 acquires an operation plan in which the reference time is shifted by a predetermined period from the end of the first period, as an operation plan included in the new input data. As a result, the data acquisition unit 131 acquires new input data in which the operation plan included in the most recently input data to the trained model MD (an operation plan with the end of the first period as the reference time) has been replaced with an operation plan with the end of the new first period as the reference time.

[0060] Similarly, the data acquisition unit 131 acquires new input data in which past operating records included in the most recently inputted input data to the trained model MD are replaced with past operating records based on the last point in time of the new first period. However, the data acquisition unit 131 acquires even newer input data in which, in the replaced past operating records, any operating records that do not exist during the power consumption inference are replaced with pre-set values.

[0061] Furthermore, the data acquisition unit 131 acquires data as new input data, which is the power consumption for the new first period inferred by the power consumption inference unit 134, plus the past power consumption data used when inferring the power consumption for the new first period.

[0062] In Figure 2, the data acquisition unit 131 replaces "t" with "t+2" to set "t+2" (20 minutes from the present) as the reference time. The data acquisition unit 131 acquires data in which "t+1" is replaced with "t+3", "t-1" with "t+1", "t-2" with "t", and "t-3" with "t-1" in accordance with this change in reference time.

[0063] The data acquisition unit 131 acquires the operation plan and past operation results, respectively, by shifting the reference time to a predetermined period later than the last point in the new first period. However, at the time of power consumption estimation, there is no operation result at time t+2 in addition to time t+1. Therefore, the data calculation unit 132 replaces the operation results at time t+1 and time t+2 with a predetermined value (for example, the operation result at the time of power consumption estimation).

[0064] Furthermore, the data acquisition unit 131 excludes the oldest period of past power usage from the new input data input to the trained model MD, and adds the power usage for the new first period inferred by the power usage inference unit 134 to the past power usage. In the example in Figure 2, the data acquisition unit 131 excludes W(t-2) and adds W, which is the power usage for the new first period inferred by the power usage inference unit 134. p The data with (t+2) added is obtained as historical electricity usage data.

[0065] As the data acquisition unit 131 acquires new input data in this manner, the power consumption inference unit 134 infers the power consumption for a new first period. Then, using the inferred power consumption, the power consumption inference unit 134 predicts the power consumption (absolute value, W) at a point in time a predetermined period after the end of the new first period. p Calculate (t+3).

[0066] Therefore, by having the data acquisition unit 131 and the power consumption inference unit 134 repeat the above-described process, the power consumption inference unit 134 can successively infer power consumption for a period from a certain reference time to a predetermined period after that time. As a result, the power consumption inference unit 134 can successively calculate predicted values ​​of power consumption at a point in time a predetermined period after a certain reference time.

[0067] Specifically, the data acquisition unit 131 acquires new input data obtained by shifting the reference time in the most recently inputted input data to the trained model MD by a predetermined period. For operational records that no longer exist due to the shift from the reference time by a predetermined period, the data acquisition unit 131 acquires new input data by inputting a pre-set value. Furthermore, for past power usage records, the data acquisition unit 131 acquires new input data by excluding the oldest power usage record and adding the most recently inferred power usage record. The power usage inference unit 134 then inputs the new input data to the trained model MD, and is able to successively infer power usage for a period from a certain reference time to a predetermined period later (a period corresponding to a new first period).

[0068] <Process for generating a pre-trained model> Next, we will describe an example of the process flow for generating a trained model MD. Figure 3 is a flowchart showing an example of the process flow for generating a trained model MD.

[0069] The data acquisition unit 131 acquires data including the operation plan, operation results, and power usage results, with a past point in time as the reference time, as data necessary to generate the trained model MD (S1). In this embodiment, the data acquisition unit 131 acquires data for each of the operation plan, operation results, and power usage results for a predetermined period with a past point in time as the reference time. Then, the data calculation unit 132 calculates the amount of change for each of the acquired operation plan, operation results, and power usage results for the predetermined period.

[0070] The model acquisition unit 133 generates a trained model MD that infers power consumption during the first period by performing machine learning using the amount of change per predetermined period calculated from the data acquired in S1 (S2), and stores the trained model MD in the storage unit 14 (S3).

[0071] <Electricity usage forecasting process> Next, we will explain an example of the power consumption forecasting process. Figure 4 is a flowchart of an example of the power consumption forecasting process. This flowchart shows an example of a power consumption forecasting method by the power consumption forecasting device 1.

[0072] The data acquisition unit 131 acquires input data including the factory's operational plan from the past to the future, past operational results, and past power usage results (S11; data acquisition step). In this embodiment, the data acquisition unit 131 acquires the operational plan, past operational results, and past power usage results as quantities per predetermined period with the present as the reference time. The data calculation unit 132 then calculates the amount of change per predetermined period for each of the acquired operational plan, operational results, and power usage results.

[0073] Furthermore, the model acquisition unit 133 acquires the trained model MD generated by the model acquisition unit 133 from the storage unit 14 (S12). Note that the processes in S11 and S12 may be performed in parallel or in reverse order.

[0074] The power consumption inference unit 134 infers the amount of power consumption expected to be used in the factory during the first period by inputting the input data acquired in S11 into the trained model MD acquired in S12 (S13; power consumption inference step). In this embodiment, the power consumption inference unit 134 inputs the amount of change calculated by the data calculation unit 132 into the trained model MD as input data.

[0075] The power consumption inference unit 134 uses the power consumption (relative value) inferred in S13 to calculate a predicted value of power consumption at the end of the first period (S14). The power consumption inference unit 134 then outputs the calculated predicted value of power consumption to the output unit 12 (S15). As a result, if the output unit 12 is a display device, or if the output destination of the output unit 12 is a display device, the user can understand the predicted value of power consumption calculated by the power consumption inference unit 134 via the display device. The power consumption inference unit 134 may also output the power consumption inferred in S13 to the output unit 12.

[0076] The prediction unit 13 determines whether it has calculated a predicted value for power consumption up to the desired time. Information indicating the time point for which the user desires a predicted value for power consumption is received via the input unit 11. For example, if the desired time is 30 minutes from now, and the power consumption inference unit 134 has calculated a predicted value for power consumption 10 minutes from now, the prediction unit 13 determines that it has not calculated a predicted value for power consumption up to the desired time.

[0077] If it is determined that the predicted value of power consumption up to the desired point in time has not been calculated (NO in S16), the data acquisition unit 131 acquires new input data (S17). The power consumption inference unit 134 inputs the new input data acquired in S17 into the trained model MD acquired in S12 to infer the amount of power consumption that is expected to be used in the factory during the new first period (S18).

[0078] Subsequently, the power consumption inference unit 134 performs the process in S14. That is, the power consumption inference unit 134 uses the new power consumption inferred in S18 to calculate a predicted value of power consumption at the end of the new first period (S14). Then, the processes from S14 to S18 are performed until it is determined in S16 that a predicted value of power consumption up to the desired time has been calculated (YES in S16). As a result, the power consumption prediction device 1 can output a predicted value of power consumption up to the time desired by the user (for example, 30 minutes from now).

[0079] <Variation> In the example described above, the data acquisition unit 131 acquires data calculated as the amount per predetermined period with the present or a later point in time as the reference time for each of the operation plan, past operation performance, and past power usage performance, as input data, but is not limited to this. For example, the data acquisition unit 131 may acquire the amount of change in the planned value or actual value obtained for each of the operation plan, past operation performance, and past power usage performance, with the present or a later point in time as the reference time, as input data.

[0080] Furthermore, the data acquisition unit 131 may acquire data calculated as the amount per predetermined period with a past point in time as the base time for each of the operation plan, operation results, and power usage results, as training data. In addition, the data acquisition unit 131 may acquire the amount of change in the planned value or actual value obtained for each of the operation plan, operation results, and power usage results for each of the past point in time as the base time for each of the predetermined period.

[0081] In the above example, for the sake of simplicity, we mainly used differential values ​​(changes) for the operational plan, past operational performance, and past power usage performance, but the explanation is not limited to this case. For example, it is possible to achieve a similar embodiment by using actual values ​​for the corresponding period instead of differential values ​​for the operational plan, past operational performance, and past power usage performance. In other words, actual values ​​may be used for the operational plan, past operational performance, and past power usage performance as input data to the trained model MD, and as training data used in machine learning to generate the trained model MD.

[0082] Referring to the example in Figure 2, for example, coil width x is an example of a future operation plan. p,w as x p,w Using (t+1), an example of a past operation plan is the coil width x p,w as x p,w (t), xp,w (t-1), x p,w (t-2) may also be used. Additionally, coil weight x is an example of a future operational plan. p,wt as x p,wt Using (t+1), we have an example of a past operation plan: coil weight x p,wt as x p,wt (t), x p,wt (t-1), x p,wt (t-2) may also be used.

[0083] An example of past operating performance is coil width x a,w and coil weight x a,wt For example, the coil width x shown in past operating plans p,w and coil weight x p,wt The same period value may be used. That is, referring to the example in Figure 2, coil width x a,w as x a,w (t), x a,w (t-1), x a,w Using (t-2), coil weight x a,wt as x a,wt (t), x a,wt (t-1), x a,wt (t-2) may also be used.

[0084] Past electricity usage data can be the amount of electricity used by the factory during the period corresponding to past operating records. In the example above, W(t), W(t-1), and W(t-2) may be used.

[0085] Furthermore, it is not limited to using four values ​​(for different periods) as the operational plan and three values ​​(for different periods) each for past operational performance and past electricity usage performance; the number of these values ​​may be changed as appropriate. The number of values ​​used as the operational plan should be at least one greater than the number of values ​​used for operational performance and electricity usage performance. For example, ten values ​​may be used as the operational plan, and seven values ​​each for operational performance and electricity usage performance.

[0086] Furthermore, in the above example, the input data and training data as change amounts are calculated in the data calculation unit 132 using the data acquired by the data acquisition unit 131, but this is not limited to this. The input data and training data may be calculated by an external device having the functions of the data calculation unit 132, and the data acquisition unit 131 may acquire the input data and training data calculated from the external device.

[0087] Furthermore, in the example described above, the predetermined period was assumed to be 10 minutes. That is, the power consumption inference unit 134 was described as calculating a predicted value of power consumption at a point 10 minutes after a certain reference time (e.g., the present). For example, if the demand cycle of the power purchase contract is 30 minutes, and the desired point in time is 30 minutes from the present, in the example above, the power consumption inference unit 134 will sequentially calculate the predicted value of power consumption three times. That is, the power consumption inference unit 134 will sequentially infer the power consumption for the period from the present to 10 minutes later, the power consumption for the period from 10 minutes later to 20 minutes later, and the power consumption for the period from 20 minutes later to 30 minutes later. As a result, the power consumption inference unit 134 will calculate a predicted value of power consumption at a point 30 minutes from the present.

[0088] However, the predetermined period can be set to any period. For example, if the demand period is 30 minutes, the predetermined period may be set to 30 minutes. In this case, data with a predetermined period of 30 minutes is prepared as both training data and input data. Therefore, the power consumption inference unit 134 can calculate a predicted value of power consumption 30 minutes after a certain reference time (e.g., the present) simply by inputting the input data once into the trained model MD.

[0089] <Effects> The power consumption forecasting device 1 uses the following as explanatory variables: the operational plan for the period from a point in the past before a reference time to a point in the past after that reference time; the operational results for the period from a point in the past before that reference time to that reference time; and the power consumption results for the period from a point in the past before that reference time to that reference time. The power consumption forecasting device 1 then uses the resulting trained model MD as input data, which includes the operational plan from the past to the future, past operational results, and past power consumption results, to infer the power consumption for the first period.

[0090] Furthermore, the power consumption forecasting device 1 uses the change in the operational plan during the period from a point in the past (a reference time) to a point in the past (a reference time) to a point in the past (a reference time), the change in the operational performance during the period from a point in the past (a reference time) to the reference time, and the change in the power consumption performance during the period from a point in the past (a reference time) to the reference time as explanatory variables, and the change in the power consumption performance during the second period as the dependent variable. The power consumption forecasting device 1 may also infer the change in power consumption during the first period by inputting input data, which includes the change in the operational plan from the past to the future, the change in past operational performance, and the change in past power consumption performance, into the trained model MD generated as a result.

[0091] By using such input data and trained data MD, the power consumption forecasting device 1 can accurately predict power consumption at a future point in time, even when there is a discrepancy between the operational plan and the actual operational results.

[0092] As mentioned above, in factories with high electricity consumption, such as steel mills, where electricity consumption varies greatly depending on the manufacturing conditions of the rolled material, it is important to predict electricity consumption using an operational plan. However, if there is a discrepancy between the operational plan and actual operation, the accuracy of electricity consumption prediction deteriorates.

[0093] The inventors hypothesized that the above discrepancy could be represented based on the history of changes in the operational plan and actual operational results, i.e., the change pattern. Based on this hypothesis, they set the explanatory and dependent variables as described above and generated a trained model MD. As a result, they found that even if there was some discrepancy between the operational plan and actual operational results, it was possible to accurately predict electricity consumption without modifying the operational plan. Specifically, they found that by generating a trained model MD using data showing the amount of change per predetermined period with a certain past point in time as the base time for each of the operational plan, actual operational results, and actual electricity consumption, electricity consumption could be predicted with greater accuracy.

[0094] Here, by inputting the input data shown in the example in Figure 2 above into the trained model MD, we inferred the amount of electricity used from the present to 10 minutes later, and calculated a predicted value for electricity use at 10 minutes later. We also calculated the error between this predicted value of electricity use and the actual electricity use 10 minutes later as the mean squared error (RMSE).

[0095] On the other hand, as a comparative example, a trained model was generated to infer electricity consumption during the first period using data with operational performance excluded from the explanatory variables described above. That is, a trained model was generated using training data that does not include operational performance, as shown in the example in Figure 2. Then, by inputting the input data that does not include operational performance, as shown in the example in Figure 2, into this trained model, electricity consumption for the period from the present to 10 minutes later was inferred, and a predicted value of electricity consumption at 10 minutes later was calculated. The error between this predicted value of electricity consumption and the actual electricity consumption 10 minutes later was also calculated as RMSE.

[0096] As a result, it was found that when predicting power consumption using the trained MD model, the RMSE was reduced by approximately 15% compared to when predicting power consumption using the trained model in the comparative example. In other words, it was found that by predicting power consumption using the operational plan and actual operational results, power consumption can be predicted accurately without modifying the operational plan.

[0097] [Embodiment 2] Other embodiments of this disclosure are described below. For the sake of clarity, components having the same function as those described in the above embodiments are denoted by the same reference numerals, and their descriptions are not repeated.

[0098] In the power consumption forecasting device 1 of this embodiment, the model acquisition unit 133 obtains a trained model MD by machine learning, using a point in time prior to the present as the reference time, and using the operation plan, operation results prior to the reference time, and power consumption results prior to the reference time as explanatory variables, and the power consumption results for a second period from the reference time to a point in time a predetermined period later as the objective variable. In other words, the model acquisition unit 133 of this embodiment obtains a trained model MD that is generated by a model generation device different from the power consumption forecasting device 1, without generating the trained model MD itself.

[0099] Therefore, the power consumption prediction device 1 of this embodiment does not need to have the processing power to generate the trained model MD. Thus, an inexpensive power consumption prediction device 1 can be provided.

[0100] [Examples of implementation using software] The function of the power consumption forecasting device 1 (hereinafter referred to as "the device") is a program that causes a computer to function as the device, and can be realized by a program that causes a computer to function as each control block of the device (in particular, each part included in the forecasting unit 13).

[0101] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the program using this control device and storage device, the functions described in each of the embodiments are realized.

[0102] The above program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.

[0103] Furthermore, some or all of the functions of each of the above control blocks can also be implemented by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of this disclosure. In addition, it is also possible to implement the functions of each of the above control blocks by, for example, a quantum computer.

[0104] Furthermore, each process described in the above embodiments may be performed by AI (Artificial Intelligence). In this case, the AI ​​may operate on the control device described above, or it may operate on other devices (for example, an edge computer or a cloud server).

[0105] 〔summary〕 A power consumption forecasting device according to Embodiment 1 of the present disclosure is a power consumption forecasting device that forecasts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period later, and comprises: a data acquisition unit that acquires input data including the factory's past and future operational plans, past operational results, and past power consumption results; and a power consumption inference unit that infers the amount of power consumption expected to be used in the factory during the first period by inputting the input data into a trained model.

[0106] The power consumption forecasting device according to aspect 2 of the present disclosure, in aspect 1, has a model acquisition unit that acquires the trained model, and the model acquisition unit acquires the trained model by generating a trained model through machine learning with the following variables: the operational plan for the period from a point before the reference time to a point after the reference time, the operational results for the period from a point before the reference time to a point after the reference time, the power consumption results for the period before the reference time and the power consumption results for the period before the reference time and the power consumption results for the period from the reference time to a point after a predetermined period, with the power consumption results for the period from a point before the reference time to a point after the reference time, with the power consumption results for the period from a point before the reference time to a point after the reference time and the power consumption results for the period before the reference time and the power consumption results for the period before the reference time and the power consumption results for the period before the reference time and the power consumption results for the period after the reference time to a point after a predetermined period, with the power consumption model acquired through machine learning, or the model acquisition unit acquires the trained model by obtaining a trained model through machine learning with the following variables: the operational plan for the period from a point before the reference time to a point after the reference time, the operational results for the period from a point before the reference time to a point after the reference time, the operational results for the period before the reference time and the power consumption results for the period before the reference time and the power consumption results for the period after the reference time, with the power consumption results for the period after the reference time to a point after a predetermined period, with the power consumption model acquired through machine learning, with the following variables:

[0107] The power consumption forecasting device according to Embodiment 3 of the present disclosure, in Embodiment 1 or 2, is data in which the operation plan is expressed as the difference between a value corresponding to the products planned to be processed at the factory during the first period and a value corresponding to the products planned to be processed at the factory during the period of one predetermined period preceding the first period, and data in which the difference between a value corresponding to the products planned to be processed at the factory during the period from the present to one predetermined period prior or more predetermined periods prior to that point in time and a value corresponding to the products planned to be processed at the factory during the period of one predetermined period preceding that period.

[0108] The power consumption forecasting device according to Embodiment 4 of the present disclosure, in any of Embodiments 1 to 3, is data in which the past operating performance is represented by the difference between a value corresponding to the products processed at the factory during the period from the present to a point in time one or more predetermined periods prior to that point, up to a point in time one predetermined period later than that point, and a value corresponding to the products processed at the factory during the period one predetermined period prior to that period, and the past power consumption performance is data in which the amount of power used at the factory during the period from the present to a point in time one or more predetermined periods prior to that point, up to a point in time one predetermined period later than that point, and a value corresponding to the amount of power used at the factory during the period one predetermined period prior to that period.

[0109] In any of embodiments 1 to 4, the power consumption forecasting device according to embodiment 5 of the present disclosure replaces the operation plan with an operation plan that is considered to be the present with the last moment of the first period as the reference time, replaces the past operation record with a past operation record that is considered to be the present with the last moment of the first period as the reference time, replaces at least one value in the replaced past operation record with a preset value, acquires data obtained by adding the power consumption for the first period inferred by the power consumption inference unit to the past power consumption record as new input data, and the power consumption inference unit infers the amount of power consumption that is expected to be used in the factory in a new first period that is the period from the last moment of the first period to a predetermined period later by inputting the new input data into the trained model.

[0110] A power consumption forecasting method according to aspect 6 of the present disclosure is a power consumption forecasting method that forecasts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period later, and comprises: a data acquisition step of acquiring input data including an operation plan from the past to the future, past operation results, and past power consumption results of the factory; and a power consumption inference step of inputting the input data into a trained model to infer the amount of power consumption expected to be used in the factory during the first period.

[0111] [Additional Notes] This disclosure is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of this disclosure. [Explanation of Symbols]

[0112] 1. Power consumption forecasting device 131 Data Acquisition Unit 133 Model Acquisition Section 134 Power usage inference section MD pre-trained model

Claims

1. A power consumption forecasting device that predicts the amount of power consumption expected to be used in a factory during a first period from the present to a predetermined period later, A data acquisition unit acquires input data including the past and future operational plans, past operational results, and past power usage results of the aforementioned factory. A power consumption inference unit that infers the amount of power consumption expected to be used in the factory during the first period by inputting the input data into a trained model, A power consumption prediction device having the following features.

2. The unit includes a model acquisition unit that acquires the aforementioned trained model, The model acquisition unit uses a point in time prior to the present as the reference time, and generates a trained model by machine learning using the following as explanatory variables: the operational plan for the period from a point in time prior to the reference time to a point in time after the reference time, the operational results prior to the reference time, and the power usage results prior to the reference time, and the power usage results for a second period from the reference time to a point in time a predetermined period later as the dependent variable. Alternatively, the trained model can be acquired by: The power consumption prediction device according to claim 1, wherein the model acquisition unit obtains a trained model by machine learning, using a point in time prior to the present as the reference time, and using the operation plan for the period from a point in time prior to the reference time to a point in time after the reference time, the operation results prior to the reference time, and the power consumption results prior to the reference time as explanatory variables, and the power consumption results for a second period from the reference time to a point in time a predetermined period later as the objective variable.

3. The aforementioned operational plan is, Data expressed as the difference between a value corresponding to the product planned to be processed at the factory during the first period and a value corresponding to the product planned to be processed at the factory during the period of one predetermined period immediately preceding the first period, and The power consumption prediction device according to claim 1 or 2, wherein the data is expressed as the difference between a value corresponding to a product planned to be processed at the factory during the period from the present to a point in time one or more predetermined periods prior to that point in time, up to a point in time one predetermined period later than that point in time, and a value corresponding to a product planned to be processed at the factory during the period one predetermined period prior to that period.

4. The aforementioned past operational record is, This data is represented by the difference between a value corresponding to the products processed at the factory during the period from the present to a point in time one or more units prior to the predetermined period, up to a point in time one unit later than that point in time, and a value corresponding to the products processed at the factory during the period immediately preceding that predetermined period. The aforementioned past electricity usage record is, The power consumption prediction device according to claim 1 or 2, wherein the data is expressed as the difference between the amount of electricity used by the factory during the period from the present to a point in time one or more predetermined periods prior to that point in time, and the amount of electricity used by the factory during the period one predetermined period prior to that period.

5. The data acquisition unit, The aforementioned operational plan is replaced with an operational plan that considers the last point in the first period as the reference time and the present as the current time. The aforementioned past operational performance is replaced with past operational performance that is considered to be the present, with the last point in the first period as the reference time, and at least one value in the replaced past operational performance is replaced with a predetermined value. The power consumption amount for the first period inferred by the power consumption inference unit is added to the past power consumption record, and this data is acquired as new input data. The aforementioned power consumption inference unit, The power consumption prediction device according to claim 1 or 2, wherein by inputting new input data into the trained model, it infers the amount of power consumption expected to be used in the factory during a new first period, which is the period from the end of the first period to a predetermined period later.

6. A method for predicting electricity consumption, which predicts the amount of electricity that is expected to be used in a factory during a first period from the present to a predetermined period later, A data acquisition step involves acquiring input data including the past and future operational plans, past operational performance, and past power usage performance of the aforementioned factory. A power consumption inference step in which the power consumption expected to be used in the factory during the first period is inferred by inputting the input data into a trained model, A method for predicting electricity consumption, comprising the following characteristics.

Images