Power demand prediction method, power demand prediction device, and power demand prediction program

Abstract

This power demand prediction device outputs a period prediction value including a power demand prediction value in a prediction period for each of a plurality of unit times into which the prediction period is divided by inputting, to a trained model for each of the plurality of unit times, input data including an actual weather value in the past prior to the prediction period, an actual power demand value in the past prior to the prediction period, and a weather forecast value in the prediction period. The weather forecast value is acquired for each of the plurality of unit times and, for the power demand prediction value having a unit time overlapping with that between the period prediction value output in a first unit time and the period prediction value output in a second unit time after the first unit time, adopts the power demand prediction value output in the second period.

Description

Electricity demand forecasting method, electricity demand forecasting device, and electricity demand forecasting program

[0001] The present disclosure relates to a technique for predicting power demand of a facility.

[0002] Patent Document 1 discloses a technology for extracting at least one power consumption element that affects power consumption, generating a relationship model showing the total power consumption for each extracted power consumption element and the relationship between the power consumption elements, and calculating the power consumption based on the generated relationship model.

[0003] However, the above-mentioned conventional technology does not disclose a trained model that predicts power demand forecast values ​​using actual weather values ​​as explanatory variables during training and weather forecast values ​​as explanatory variables during forecasting. Therefore, the above-mentioned conventional technology cannot suppress prediction errors that occur when calculating power demand forecast values ​​using such a trained model.

[0004] JP 2016-185059 A

[0005] The present disclosure has been made to solve such problems, and aims to provide a technology that can more accurately calculate a power demand forecast value.

[0006] In one aspect of the present disclosure, a computer provides a power demand forecasting method in which, for each of a plurality of unit times that divide a prediction period, a computer inputs input data into a trained model, the input data including actual weather values ​​from prior to the prediction period, actual power demand values ​​from prior to the prediction period, and weather forecast values ​​for the prediction period, thereby outputting a period forecast value including a power demand forecast value for each of the unit times in the prediction period, the weather forecast value being a value obtained for each of the plurality of unit times, and for the power demand forecast value that overlaps between the period forecast value output in a first unit time and the period forecast value output in a second unit time that is after the first unit time, the power demand forecast value output in the second unit time is used.

[0007] According to the present disclosure, it is possible to more accurately calculate a power demand forecast value.

[0008] 1 is a block diagram showing an example of the overall configuration of a power demand forecasting system according to an embodiment of the present disclosure; FIG. 2 is an explanatory diagram of interpolation processing; FIG. 3 is a diagram showing an overview of processing by a power demand forecasting device; FIG. 4 is a graph showing the relationship between a plurality of forecasting periods; FIG. 5 is a flowchart showing an example of processing by a power demand forecasting device; FIG. 6 is a flowchart showing an example of correction processing by a correction unit; FIG. 7 is an explanatory diagram of correction processing; FIG. 8 is a graph explaining correction processing; FIG. 9 is a flowchart showing an example of learning processing of an hourly predictor;

[0009] (Findings underlying the present disclosure) Research is underway to calculate a forecast value of power demand for a facility for a future prediction period by inputting weather forecast values, past actual power demand values ​​for the facility, etc. into a trained model. In this case, the trained model is generated by machine learning learning data in which the actual power demand value before the base time point and the actual weather value before the base time point are used as explanatory variables, and the actual power demand value after the base time point is used as a target variable.

[0010] However, when making predictions, the trained model uses weather forecast values ​​for the prediction period, rather than actual weather values, as explanatory variables. As a result, the electricity demand forecast calculated by the trained model includes errors in the weather forecast values, resulting in inaccurate calculations of the electricity demand forecast values.

[0011] In Patent Document 1, power consumption is calculated using a relational model, which is generated by expressing the logarithmic transformation value of power consumption as a polynomial function for power consumption elements and calculating the coefficients for the polynomial function using the least squares method.

[0012] In this way, the relational model shown in Patent Document 1 is not a model in which actual weather values ​​are used as explanatory variables during learning and forecast weather values ​​are used as explanatory variables during prediction, so the problem of errors in the forecast weather values ​​being added does not arise.

[0013] Generally, weather forecast values ​​for the near future are considered to be more accurate, so it is desirable to input the most recent weather forecast values ​​into the trained model.

[0014] Therefore, the inventor discovered that if weather forecast values ​​for the prediction period are obtained for each unit time that divides the prediction period and power demand forecast values ​​for the prediction period are calculated for each unit time, the latest prediction period will overlap with past prediction periods, and as a result, the latest power demand forecast values ​​can be used for the overlapping periods, and as a result, power demand forecast values ​​can be calculated more accurately, which led to the present disclosure.

[0015] (1) In one aspect of the present disclosure, a computer provides a power demand forecasting method in which, for each of a plurality of unit times that define a forecast period, a computer inputs input data to a trained model, the input data including actual weather values ​​from the past before the forecast period, actual power demand values ​​from the past before the forecast period, and weather forecast values ​​for the forecast period, thereby outputting a period forecast value including a power demand forecast value for each of the unit times in the forecast period, the weather forecast value being a value obtained for each of the plurality of unit times, and, for the power demand forecast value that overlaps between the period forecast value output for a first unit time and the period forecast value output for a second unit time that is later than the first unit time, the power demand forecast value output for the second unit time is used.

[0016] According to this configuration, input data including the latest weather forecast values ​​is acquired for each unit time that divides the prediction period, and each time the input data is acquired, the input data is input into the trained model to calculate a power demand forecast value for each unit time in the prediction period. As a result, the latest prediction period overlaps with a previous prediction period. The latest power demand forecast value is then used as the power demand forecast value for the overlapping period. This reduces prediction errors that occur when calculating a power demand forecast value using a trained model that uses actual weather values ​​as explanatory variables during learning and weather forecast values ​​for the prediction period as explanatory variables during prediction. As a result, a more accurate power demand forecast value can be obtained.

[0017] (2) In the power demand forecasting method described in (1) above, if there is a missing time in the period forecast value, which is a unit time during which the calculation of the power demand forecast value failed, the most recent power demand forecast value calculated in the same time period as the missing time may be adopted as the power demand forecast value for the missing time.

[0018] For example, the period forecast value may include a missing time period when the calculation of the power demand forecast value failed due to a missing portion of input data. With this configuration, when there is a missing time period, the most recent power demand forecast value calculated for the same time period as the missing time period is used as the power demand forecast value for the missing time period. This makes it possible to prevent missing power demand forecast values ​​from being included in the period forecast value.

[0019] (3) In the power demand forecasting method described in (1) or (2) above, a first error may be calculated, which is the difference between the power demand forecast value calculated for a first period in the past from the time of prediction and the actual power demand value for the first period, and if the first error exceeds a threshold, the first error may be used to correct the power demand forecast value for the second period so that a second error, which is the difference for a second period in the future from the time of prediction, becomes smaller.

[0020] When a first error, which is the difference between the predicted power demand value and the actual power demand value for a first period in the past from the time of prediction, is large, a second error, which is the difference between the predicted power demand value and the actual power demand value for a second period in the future from the time of prediction, is also estimated to be large. With this configuration, when the first error is larger than a threshold, the power demand prediction value is corrected so that the second error is reduced. Therefore, the power demand prediction value can be calculated with higher accuracy in such a situation.

[0021] (4) In the power demand forecasting method described in (3) above, when the first period extends over to the previous day, the correction of the power demand forecast value may not be performed.

[0022] Since power demand can vary significantly from day to day, if the first period extends into the previous day, power demand can vary significantly between the previous day and the current day. Therefore, even if the first error exceeds the threshold, the second error does not necessarily increase in conjunction with the magnitude of the first error. With this configuration, if the first period extends into the previous day, the power demand forecast value is not corrected, so that it is possible to prevent the correction from worsening the accuracy of the power demand forecast value.

[0023] (5) In the power demand forecasting method described in (3) or (4) above, the second period may be determined in advance based on a full charge time of a storage battery included in the facility for which the power demand is to be forecast.

[0024] According to this configuration, the second period is set based on the full charge time of the storage battery provided in the facility, so that it is possible to avoid a situation where there is a shortage of power when power from the storage battery is needed.

[0025] (6) In the electricity demand forecasting method described in any one of (1) to (5) above, the input data may further include actual operation information of the equipment for which electricity demand is to be forecast and planned operation information of the equipment during the forecast period.

[0026] According to this configuration, the actual power demand value can be calculated taking into consideration the operation record and operation schedule of the equipment.

[0027] (7) In the power demand forecasting method described in any one of (1) to (6) above, the trained model may include a plurality of time-specific trained models that calculate the power demand forecast value for each of the plurality of unit times.

[0028] The range of fluctuation of the actual power demand value for each unit time is smaller than the range of fluctuation of the actual power demand value for the entire prediction period. With this configuration, the power demand forecast value is calculated using multiple time-based trained models that calculate power demand forecast values ​​for each of multiple unit times, so the power demand forecast value can be calculated more accurately.

[0029] (8) In the power demand forecasting method described in (7) above, outputting the power demand forecast value may include generating features of the input data by classifying the input data according to categories, dividing the features into a plurality of features corresponding to each of the plurality of time-based trained models, inputting the plurality of features into a time-based trained model corresponding to the plurality of features to calculate the power demand forecast value for each unit time, and calculating the period forecast value by integrating the power demand forecast values ​​calculated for each unit time.

[0030] According to this configuration, input data is classified according to category to generate features of the input data, the generated features are divided according to each of multiple time-based trained models, and the divided features are input into the time-based trained models, thereby calculating the power demand forecast value for each unit time, thereby enabling more accurate calculation of the power demand forecast value.

[0031] (9) In the power demand forecasting method described in (6) above, the input data may include actual power demand values ​​for each unit time on the day before the prediction time, and the output of the power demand forecast value may include determining whether the prediction time falls on an operating day based on the operation schedule information, determining whether the previous day falls on an operating day based on the operation performance information, and, if the prediction time falls on an operating day and the previous day falls on a non-operating day, inputting the actual power demand values ​​for each unit time on the most recent operating day into the trained model.

[0032] According to this configuration, if the day before the prediction time falls on a non-operating day and the prediction time falls on an operating day, the power demand forecast value is calculated using the actual power demand value per unit time on the most recent operating day instead of the actual power demand value per unit time on the previous day, thereby allowing the power demand forecast value to be calculated more accurately by taking the operating day into consideration.

[0033] (10) In the power demand forecasting method described in (6) above, the input data may include an actual power demand value for each unit time on the day before the prediction time, and calculation of the power demand forecast value may include determining whether the prediction time falls on an operating day based on the operation schedule information, determining whether the previous day falls on an operating day based on the operation performance information, and, if the prediction time falls on a non-operating day and the previous day falls on an operating day, inputting the actual power demand value for each unit time on the most recent non-operating day into the trained model.

[0034] According to this configuration, if the prediction time falls on a non-working day and the day before the prediction time falls on a working day, the power demand forecast value is calculated using the actual power demand value per unit time on the most recent non-working day instead of the actual power demand value per unit time on the previous day, thereby allowing the power demand forecast value to be calculated more accurately by taking the non-working day into consideration.

[0035] (11) In the power demand forecasting method described in (3) above, the first error may be a linear weighted average error between the power demand forecast value and the power demand actual value.

[0036] According to this configuration, the first error can be accurately estimated.

[0037] (12) In the electricity demand forecasting method described in any one of (1) to (11) above, the trained model may be a model obtained by machine learning learning data in which actual weather values ​​before a reference time point and actual electricity demand values ​​before the reference time point are used as explanatory variables, and actual electricity demand values ​​after the reference time point corresponding to the explanatory variables are used as objective variables.

[0038] According to this configuration, a trained model can be obtained that can predict the power demand forecast value for the prediction period from actual weather values ​​and past actual power demand values.

[0039] (13) In another aspect of the present disclosure, an electric power demand forecasting device includes a processor, wherein the processor inputs input data including actual weather values ​​from the past before the prediction period, actual electric power demand values ​​from the past before the prediction period, and weather forecast values ​​for the prediction period into a trained model for each of a plurality of unit times that divide the prediction period, thereby outputting a period forecast value including an electric power demand forecast value for each of the unit times in the prediction period, wherein the weather forecast value is a value obtained for each of the plurality of unit times, and wherein, among the plurality of unit times that divide the prediction period, the period forecast value output for a first unit time and the period forecast value output for a second unit time that is after the first unit time, for the electric power demand forecast value whose unit times overlap, the electric power demand forecast value output for the second unit time is adopted.

[0040] According to this configuration, it is possible to provide a power demand forecasting device that can more accurately calculate a power demand forecast value.

[0041] (14) In yet another aspect of the present disclosure, an electricity demand forecasting program causes a computer to execute a process of inputting input data to a trained model for each of a plurality of unit times that divide a forecast period, the input data including actual weather values ​​from before the forecast period, actual electricity demand values ​​from before the forecast period, and weather forecast values ​​for the forecast period, thereby outputting a period forecast value including an electricity demand forecast value for each of the unit times in the forecast period, wherein the weather forecast value is a value obtained for each of the plurality of unit times, and among the plurality of unit times that divide the forecast period, the period forecast value output for a first unit time and the period forecast value output for a second unit time that is after the first unit time, for the electricity demand forecast value whose unit times overlap, adopting the electricity demand forecast value output for the second unit time.

[0042] According to this configuration, it is possible to provide a power demand forecasting program that can more accurately calculate a power demand forecast value.

[0043] The present disclosure can also be realized as an electricity demand forecasting system that operates by this electricity demand forecasting program. Needless to say, such a computer program can be distributed on a computer-readable non-transitory recording medium such as a CD-ROM or via a communication network such as the Internet.

[0044] Note that each of the embodiments described below represents a specific example of the present disclosure. The numerical values, shapes, components, steps, and step orders shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Furthermore, among the components in the following embodiments, components that are not described in the independent claims that represent the highest concept are described as optional components. Furthermore, in all of the embodiments, the respective contents can be combined.

[0045] (Embodiment) Fig. 1 is a block diagram showing an example of the overall configuration of a power demand forecasting system 10 according to an embodiment of the present disclosure. The power demand forecasting system 10 is a system that predicts power demand at a facility. An example of the facility is a factory equipped with a storage battery. The power demand forecasting system 10 includes a power demand forecasting device 1, a terminal device 2, and a management device 3. The power demand forecasting device 1, the terminal device 2, and the management device 3 are communicatively connected via a network NT. A wide area communication network including the Internet, for example, can be used as the network NT.

[0046] The power demand prediction device 1 is configured, for example, by a cloud server including one or more computers. However, this is just one example, and the power demand prediction device 1 may be configured by an edge-side computer implemented in a terminal device 2. Alternatively, the power demand prediction device 1 may be implemented in a management device 3. The terminal device 2 may be configured, for example, by a mobile information terminal such as a smartphone or a desktop computer. The management device 3 is configured, for example, by a computer, and creates a charging schedule for a storage battery installed in a facility, for example.

[0047] The power demand prediction device 1 includes a processor 11, a memory 12, and a communication unit 13. The processor 11 is configured, for example, by a central processing unit (CPU), and includes an acquisition unit 111, a prediction unit 112, a correction unit 113, an adoption unit 114, and a learning unit 115. The acquisition unit 111 to the learning unit 115 are realized by the processor 11 executing a power demand prediction program that causes a computer to function as the power demand prediction device 1. However, this is just an example, and the acquisition unit 111 to the learning unit 115 may also be configured by dedicated hardware circuits.

[0048] The acquisition unit 111 acquires input data for each of a plurality of unit times that divide the prediction period. The prediction period is a period for which the power demand prediction device 1 predicts the power demand forecast value. The prediction period is, for example, 24 hours. However, this is just an example, and any appropriate value such as 48 hours, 12 hours, or 36 hours can be used for the prediction period. The unit time is the time obtained by dividing the prediction period into predetermined time intervals. Examples of unit times are 1 hour, 2 hours, and 30 minutes. Therefore, if the unit time is 1 hour, input data will be acquired every hour.

[0049] The input data may include actual power demand values ​​prior to the prediction period, weather forecast values ​​for the prediction period, and actual weather values ​​prior to the prediction period. The input data may include timer timing information. The actual power demand value is the actual value of power actually consumed in the facility at a point in time prior to the prediction period. The actual weather value is the actual value of weather information around the facility at a point in time prior to the prediction period. Weather information includes temperature, humidity, precipitation, weather, etc. The weather forecast value for the prediction period is the forecast value of weather information around the facility for the prediction period. For example, the temperature, humidity, precipitation, etc. around the facility are weather forecast values. Note that the weather forecast value for the prediction period acquired by the acquisition unit 111 is the latest value acquired every multiple unit time.

[0050] The acquisition unit 111 may acquire, for example, an actual power demand value from a power meter installed in the facility. The acquisition unit 111 may acquire, for example, actual weather values ​​and weather forecast values ​​from a weather information server connected via the network NT. The actual power demand value, actual weather value, and weather forecast value are each assigned a timestamp indicating the measurement time. For example, each time the acquisition unit 111 acquires an actual power demand value, actual weather value, and weather forecast value, the acquisition unit 111 stores the acquired data in the input data storage unit 121, and when it is time to calculate a forecast power demand value, the acquisition unit 111 may acquire the required actual power demand value, actual weather value, and weather forecast value from the input data storage unit 121 according to the timestamp.

[0051] The input data may further include operation record information and operation schedule information. The operation record information is information indicating whether the facility was actually operating on a daily basis. The operation schedule information is information indicating the operation schedule of the facility on a daily basis.

[0052] The acquisition unit 111 may acquire operation schedule information and operation record information from the management device 3 that manages the facility.

[0053] The prediction unit 112 inputs input data for each of a plurality of unit times to the predictor 120 (an example of a trained model) to calculate a period prediction value including a power demand prediction value for each unit time in a prediction period. For example, if the prediction period is 24 hours and the unit time is 1 hour, the period prediction value is composed of 24 power demand prediction values ​​for the next 24 hours from the prediction point in time.

[0054] The predictor 120 is a model obtained by machine learning of learning data in which actual weather values ​​before a reference time point and actual power demand values ​​before the reference time point are used as explanatory variables, and actual power demand values ​​after the reference time point corresponding to the explanatory variables are used as response variables. The reference time point refers to the point in time that is the starting point of the prediction period set during learning. The actual power demand values ​​after the reference time point corresponding to the explanatory variables are actual power demand values ​​for the prediction period set during learning. The explanatory variables may include actual operation information and operation schedule information. The predictor 120 is machine-learned to minimize the error between the power demand forecast value per unit time that is output when the explanatory variables are input and the actual power demand value for the prediction period, which is the response variable corresponding to the input explanatory variables.

[0055] The predictor 120 may include a plurality of hourly predictors (an example of an hourly trained model) that calculate a power demand forecast value for each of a plurality of unit times. If the prediction period is 24 hours and the unit time is one hour, the predictor 120 includes 24 hourly predictors, such as an hourly predictor that calculates a power demand forecast value for a unit time of midnight, an hourly predictor that calculates a power demand forecast value for a unit time of 1 a.m., ..., an hourly predictor that calculates a power demand forecast value for a unit time of 11 p.m.

[0056] Let Tx be the unit time in the X-hour range. An hourly predictor that calculates a predicted power demand value for the unit time Tx is let Px be the hourly predictor. The hourly predictor Px is created in advance by machine learning learning data including explanatory variables input to the hourly predictor Px and an actual power demand value for the unit time Tx, which is an objective variable corresponding to the explanatory variables. In detail, the hourly predictor Px is machine-learned to minimize the error between the predicted power demand value for the unit time Tx that is output when the explanatory variables are input, and the actual power demand value for the unit time Tx, which is an objective variable corresponding to the input explanatory variables.

[0057] The machine learning model employed by the predictor 120 and the hourly predictor Px may be, for example, a random forest, a LightGBM, a deep neural network, or any other suitable model.

[0058] When there is a missing time, which is a unit of time in the period prediction value where the calculation of the power demand prediction value failed, the prediction unit 112 performs an interpolation process in which the most recent power demand prediction value calculated at the same time as the missing time is used as the power demand prediction value for the missing time.

[0059] FIG. 2 is an explanatory diagram of the interpolation process. A prediction period 401 is a prediction period for the next 24 hours starting from midnight on January 1st. A prediction period 402 is a prediction period for the next 24 hours starting from 1:00 on January 1st. A prediction period 403 is a prediction period for the next 24 hours starting from 2:00 on January 1st. In the prediction period 403, a power demand forecast value for the 3:00 hour period was not calculated due to reasons such as a lack of necessary data. Therefore, the prediction unit 112 identifies the 3:00 hour period as a missing time. In this example, the power demand forecast value for the 3:00 hour period in the prediction period 402 corresponds to the most recent power demand forecast value calculated in the same unit time as the missing time period. Therefore, the prediction unit 112 interpolates the power demand forecast value for the 3:00 hour period in the prediction period 403 using the power demand forecast value for the 3:00 hour period in the prediction period 402. This makes it possible to prevent missing power demand forecast values ​​in the period prediction values.

[0060] 1 , the prediction unit 112 determines whether the prediction time point corresponds to an operating day based on the operation schedule information, determines whether the day before the prediction time point corresponds to an operating day based on the operation result information, and if the prediction time point corresponds to an operating day and the previous day corresponds to a non-operating day, may input to the predictor 120 the actual power demand value per unit hour on the most recent operating day instead of the actual power demand value per unit hour on the previous day.

[0061] On the other hand, if the prediction time falls on a non-working day and the previous day was a working day, the prediction unit 112 may input the actual power demand value per unit time on the most recent non-working day to the predictor 120 instead of the actual power demand value per unit time on the previous day.

[0062] The correction unit 113 calculates a first error, which is an error between a power demand forecast value for a first period from the time of prediction and an actual power demand value for the first period. If the first error exceeds a threshold, the correction unit 113 corrects the power demand forecast value for the second period so that a second error, which is an error for a second period from the time of prediction, becomes smaller.

[0063] If the first period extends into the previous day, the correction unit 113 does not need to correct the power demand forecast value. For example, if the first period is five hours and the prediction time point is between 3:00 and 3:00, the first period extends into the previous day, so the correction unit 113 does not correct the power demand forecast value. On the other hand, if the prediction time point is between 5:00 and 5:00, the first period does not extend into the previous day, so the correction unit 113 corrects the power demand forecast value.

[0064] Furthermore, if the second period extends into the next day, the correction unit 113 does not need to correct the power demand forecast value. For example, if the second period is three hours long and the prediction time point is between 10:00 PM and 10:00 PM, the second period extends into the next day, so the correction unit 113 does not correct the power demand forecast value. On the other hand, if the second period is three hours long and the prediction time point is 7:00 PM, the second period does not extend into the next day, so the correction unit 113 corrects the power demand forecast value.

[0065] The second period may be determined in advance based on the full charge time of the storage battery included in the facility. The full charge time is, for example, the time it takes for the storage battery to charge from 0% to 100% SOC. If the full charge time is three hours, the second period may be, for example, three hours or three hours plus a small margin. This makes it possible to avoid a situation where the storage battery runs out of power when it is needed.

[0066] The prediction unit 112 may generate features by classifying input data according to categories, divide the generated features into a plurality of features corresponding to each of a plurality of hourly predictors Px, input the plurality of features to the hourly predictors Px corresponding to the plurality of features, thereby calculating a power demand forecast value for each unit time, and calculate a period forecast value by integrating the power demand forecast values ​​calculated for each unit time.

[0067] In addition, when a mode is adopted in which the feature amounts of input data are input to the predictor 120 to calculate the actual power demand value, the hourly predictor Px is machine-learned using the feature amounts.

[0068] The features are generated by classifying the input data into the categories of power, weather, operation, and date. The features by category are called power features, weather features, operation features, and date features.

[0069] The power feature quantities are generated by classifying the actual power demand values. The power feature quantities include a first power feature quantity indicating the actual power demand value Y days ago, a second power feature quantity indicating the actual power demand value at H:M on the previous day, a third power feature quantity indicating the rolling average value of the actual power demand values ​​classified into working days and non-working days, and a fourth power feature quantity indicating the cumulative average value of the actual power demand values ​​classified into working days and non-working days. Y is, for example, an integer value from 1 to 14. In this case, the first power feature quantity is composed of the actual power demand values ​​from 1 day ago to 14 days ago. H is, for example, an integer value from 1 to 24. M is, for example, one or more integer values ​​from 1 to 60.

[0070] The weather feature values ​​include a first weather feature value indicating the temperature in the prediction period, a second weather feature value indicating the precipitation in the prediction period, a third weather feature value indicating the temperature Y days before, and a fourth weather feature value indicating the temperature at H:M on the previous day. The first weather feature value and the second weather feature value are generated by classifying the weather forecast value. The third weather feature value and the fourth weather feature value are generated by classifying the actual weather value.

[0071] The operation feature amounts include a first operation feature amount indicating whether the prediction time point corresponds to an operating day or a non-operating day and a second operation feature amount indicating whether Y days before corresponds to an operating day or a non-operating day. The first operation feature amount is generated by classifying operation schedule information, and the second operation feature amount is generated by classifying operation result information.

[0072] The date feature includes a first date feature indicating the month to which the prediction time belongs, a second date feature indicating the week to which the prediction time belongs, and a third date feature consisting of flags indicating Saturdays, Sundays, and holidays. The second date feature is expressed as a number from 0 to 6 indicating the day of the week, with Monday being 0, ..., Sunday being 6. The first to third date feature are identified using, for example, timing information from a timer.

[0073] The adopting unit 114 adopts the power demand forecast value output in the second unit time for power demand forecast values ​​that overlap the unit time between a first unit time and a second unit time that is later than the first unit time among the multiple unit times that divide the prediction period. One example of the adopted process is a process of updating, among the period forecast values ​​stored in the period forecast value storage unit 123, the power demand forecast value that overlaps the unit time with the currently calculated period forecast value with the currently calculated power demand forecast value. Another example of the adopted process is a process of updating, among the period forecast values ​​stored in the memory of the management device 3, the power demand forecast value that overlaps the unit time with the currently calculated period forecast value with the currently calculated power demand forecast value. Another example of the adopted process is a process of updating, among the period forecast values ​​displayed on the display of the terminal device 2, the power demand forecast value that overlaps the unit time with the currently calculated period forecast value with the currently calculated power demand forecast value.

[0074] The learning unit 115 uses the learning data to train the hourly predictor Px that constitutes the predictor 120. The hourly predictor Px that has been trained by machine learning is stored in the input data storage unit 121.

[0075] The memory 12 is configured with a rewritable non-volatile storage device such as a solid state drive or a hard disk drive, and includes an input data storage unit 121, a predictor storage unit 122, and a period forecasted value storage unit 123. The input data storage unit 121 stores the input data acquired by the acquisition unit 111. The predictor storage unit 122 stores the predictor 120. The period forecasted value storage unit 123 stores period forecasted values ​​including the power demand forecast value adopted by the adoption unit 114.

[0076] The communication unit 13 is configured with a communication circuit that connects the power demand prediction device 1 to the network NT.

[0077] FIG. 3 is a diagram illustrating an overview of the processing performed by the power demand prediction device 1. First, the acquisition unit 111 reads input data from the input data storage unit 121. The input data includes actual weather values ​​from the past prediction period, weather forecast values ​​for the prediction period, actual power demand values ​​from the past prediction period, timing information, actual operation information from the past prediction period, and operation schedule information at the time of prediction. Next, the prediction unit 112 generates feature quantities for the input data by classifying this information by category. This generates the first to fourth power feature quantities, first to fourth weather feature quantities, first and second operation feature quantities, and first to third date feature quantities described above.

[0078] Next, the prediction unit 112 classifies these feature quantities into a plurality of feature quantities corresponding to each of the hourly predictors Px. In this example, there are 24 hourly predictors P0 to P23 corresponding to the time periods from midnight to 11:00 p.m., so the feature quantities are classified into each of the hourly predictors P0 to P23.

[0079] For example, the hourly predictor P1 receives, as the first power feature, the actual power demand value for the 1:00 hour hour on each day from 14 days ago to 1 day ago, and receives, as the second power feature, the actual power demand value for each hour of the previous day. Furthermore, the hourly predictor P1 receives, as the third power feature, the rolling average value of the actual power demand value for the working day if today is a working day, or the rolling average value of the actual power demand value for the non-working day if today is a non-working day. Furthermore, the hourly predictor P1 receives, as the fourth power feature, the cumulative average value of the actual power demand value for the working day if today is a working day, or the cumulative average value of the actual power demand value for the non-working day if today is a non-working day.

[0080] In addition, the hourly forecaster P1 receives as input the latest temperature at 1 o'clock indicated by the weather forecast value as the first weather feature, the latest humidity at 1 o'clock indicated by the weather forecast value as the second weather feature, the temperature at 1 o'clock for each day from 14 days ago to 1 day ago indicated by the weather actual value as the third weather feature, and the temperature for each hour of the previous day indicated by the weather actual value as the fourth weather feature.

[0081] Furthermore, information indicating whether the prediction time point is a working day or a non-working day is input to the hourly predictor P1 as the first operation characteristic, and information indicating whether each day from 14 days ago to 1 day ago is a working day or a non-working day is input to the hourly predictor P1 as the second operation characteristic.

[0082] In addition, the hourly predictor P1 receives as input the month to which the prediction time falls as the first date feature, the day of the week to which the prediction time falls as the second date feature, and a flag indicating whether the prediction time falls on a weekend or public holiday as the third date feature.

[0083] Next, the prediction unit 112 causes the hourly predictor Px to calculate a power demand prediction value for the unit time Tx.

[0084] Next, the prediction unit 112 integrates the power demand forecast values ​​calculated by the hourly predictors P0 to P23 to calculate a period forecast value. Each point on graph 301 represents 24 power demand forecast values ​​calculated by the hourly predictors P0 to P23. The vertical axis of graph 301 represents the actual power demand value, and the horizontal axis represents time. The prediction unit 112 calculates a period forecast value consisting of power demand forecast values ​​for the next 24 hours by arranging these 24 power demand forecast values ​​in chronological order from the time of prediction.

[0085] Next, the adoption unit 114 adopts the power demand forecast value calculated by the prediction unit 112 for the power demand forecast value that overlaps with the unit time of the period forecast value calculated by the prediction unit 112 among the period forecast values ​​stored in the period forecast value memory unit 123.

[0086] FIG. 4 is a graph showing the relationship between multiple prediction periods. The prediction periods 401 to 403 are the same as the prediction periods 401 to 403 shown in FIG. 2 . The predictor 120 uses actual weather values ​​as explanatory variables during learning, and uses weather forecast values ​​for the prediction period as explanatory variables during prediction. As a result, prediction errors occur in the power demand prediction values ​​calculated by the predictor. Generally, weather forecast values ​​for the nearer future are considered to be more accurate. Therefore, the power demand prediction device 1 calculates power demand prediction values ​​for the next 24 hours, for example, every hour, for prediction periods 401 to 403 while acquiring the latest weather forecast values. As a result, the latest prediction period 403 overlaps with the previous prediction periods 401 and 402. The power demand prediction value for the overlapping period is the power demand prediction value for the prediction period 403. As a result, the power demand prediction device 1 can calculate power demand prediction values ​​more accurately than a configuration in which power demand prediction values ​​for 24-hour prediction periods are calculated all at once without overlapping prediction periods.

[0087] 5 is a flowchart showing an example of the processing of the power demand prediction device 1. This flowchart is started, for example, at a predetermined prediction time point. The prediction time point is, for example, any time point within a unit time Tx. Therefore, this flowchart is executed every unit time Tx.

[0088] First, in step S1, the acquisition unit 111 acquires input data from the input data storage unit 121. Here, input data to be input to the hourly predictor Px is acquired from the input data stored in the input data storage unit 121. This input data also includes timekeeping information from a timer.

[0089] Next, in step S2, the prediction unit 112 performs a filtering process on the input data. For example, the filtering process may involve averaging the input data so that the time intervals between timestamps added to the input data are aligned to a fixed time interval. An example of the fixed time interval is one hour. However, this is just an example, and any other appropriate value such as 30 minutes may be used as the fixed time interval.

[0090] Next, in step S3, the prediction unit 112 classifies the input data according to categories as described above, and generates features using the classified input data.

[0091] Next, in step S4, the prediction unit 112 divides the feature quantity generated in step S3 into a plurality of time-based feature quantities to be input to the time-based predictors P0 to P23, respectively.

[0092] Next, the prediction unit 112 executes a process of calculating a power demand forecast value for each unit time Tx (steps S5 to S11). Here, the process of calculating a power demand forecast value for Tx=T0 to T23, i.e., from midnight to 11:00 p.m., is executed in order.

[0093] First, in step S5, the prediction unit 112 sets a unit time Tx.

[0094] Next, in step S6, the prediction unit 112 determines whether the input data is missing. For example, if the actual power demand value for the 1:00 hour on any day from 14 days ago to 1 day ago is missing in the first power feature input to the hourly predictor P1, the prediction unit 112 determines that the input data is missing. For example, if the temperature for the 1:00 hour is missing in the first weather feature input to the hourly predictor P1, the prediction unit 112 determines that the input data is missing. If it is determined that the input data is missing (YES in step S6), the process proceeds to step S9. If it is determined that the input data is not missing (NO in step S6), the process proceeds to step S7.

[0095] Next, in step S7, the prediction unit 112 calculates a power demand forecast value for a unit time Tx by inputting the hourly feature amounts sorted in step S4 to an hourly predictor Px. For example, the hourly feature amounts sorted in the hourly predictor P1 are input to the hourly predictor P1, and a power demand forecast value for the 1 o'clock hour is calculated.

[0096] In step S7, if the prediction time point is an operating day and the day before the prediction time point is a non-operating day, the prediction unit 112 may adopt the actual power demand value for H hours and M minutes on the most recent operating day as the second power feature amount, and input it to the hourly predictor Px. In step S7, if the prediction time point is a non-operating day and the day before the prediction time point is an operating day, the prediction unit 112 may adopt the actual power demand value for H hours and M minutes on the most recent non-operating day as the second power feature amount, and input it to the hourly predictor Px.

[0097] Next, in step S8, the prediction unit 112 determines whether or not the prediction of the power demand prediction value has been successful. For example, if the power demand prediction value is not output from the hourly predictor Px due to some cause such as a processing error, or if the power demand prediction value output from the hourly predictor Px indicates a numerical value that clearly indicates an error, the prediction unit 112 may determine that the prediction of the power demand prediction value has failed.

[0098] If it is determined that the prediction of the power demand forecast value was successful (YES in step S8), the processing proceeds to step S11, and if it is determined that the prediction of the power demand forecast value was unsuccessful (NO in step S8), the processing proceeds to step S9.

[0099] Next, in step S9, the prediction unit 112 executes the interpolation process described with reference to FIG. 2, and adopts the most recent measured power demand value calculated at the same time as the missing time as the power demand prediction value for the missing time.

[0100] Next, in step S10, the prediction unit 112 generates notification information indicating that the period forecast value output in step S13 includes the interpolated power demand forecast value, and notifies the notification information. For example, the prediction unit 112 may notify the facility manager of the notification information by outputting the notification information to the management device 3. When the processing of step S10 ends, the processing proceeds to step S11.

[0101] In step S11, if the process of calculating the power demand forecast value for the final unit time T23, i.e., the 11 p.m. time period, is completed, the process proceeds to step S12. On the other hand, in step S11, if the process of calculating the power demand forecast value for the final unit time T23 is not completed, the process returns to step S5, and the process of calculating the power demand forecast value for the next unit time Tx is executed.

[0102] Next, in step S12, the prediction unit 112 integrates the power demand forecast values ​​by arranging the power demand forecast values ​​calculated for each unit time Tx in a time series from the prediction time point, and generates a period forecast value including the power demand forecast values ​​for the next 24 hours from the prediction time point.

[0103] Next, in step S13, the employing unit 114 employs, among the power demand forecast values ​​included in the period forecast values ​​calculated for the unit time before the prediction time point, the power demand forecast values ​​whose unit time overlaps with the power demand forecast values ​​included in the currently calculated period forecast values. Referring to Fig. 4, it is assumed that the currently calculated period forecast values ​​are period forecast values ​​calculated for a prediction period 403. In this case, among the period forecast values ​​stored in the period forecast value storage unit 123, the employing unit 114 employs the power demand forecast values ​​calculated for the prediction period 403 for the power demand forecast values ​​from 2:00 on January 1st to 1:00 on January 2nd.

[0104] 6 is a flowchart showing an example of the correction process performed by the correction unit 113. This flowchart is executed every time a period predicted value is calculated.

[0105] First, in step S21, the correction unit 113 acquires actual power demand values ​​for a first period from the prediction time point to the past from the input data storage unit 121. FIG. 7 is an explanatory diagram of the correction process. In this example, the prediction time point is 10:15 on December 15, 2022. In this example, the first period TM1 is five hours long. Therefore, five actual power demand values ​​are calculated from 6:00 to 10:00 on the same day.

[0106] Next, in step S22, the correction unit 113 obtains the power demand forecast value for the first period TM1 from the period forecast value storage unit 123. See FIG. 7 . In this case, five power demand forecast values ​​from 6:00 to 10:00 on the same day are obtained. The most recent power demand forecast value is used as the power demand forecast value obtained here. For example, the power demand forecast value calculated at the time of prediction for the 6:00 hour is used as the power demand forecast value for the 7:00 hour, and the power demand forecast value calculated at the time of prediction for the 7:00 hour is used as the power demand forecast value for the 8:00 hour.

[0107] Next, in step S23, the correction unit 113 calculates the linear weighted average error between the actual power demand value and the forecast power demand value for the first time period TM1. The linear weighted average error is a value obtained by linearly averaging five errors between the five actual power demand values ​​and the five forecast power demand values ​​included in the first time period TM1 for each unit time Tx.

[0108] Linear weighted average error μ ｔ is expressed by the following formula (1). ｔ is expressed by equation (2).

[0109]

[0110] Here, t is an index for specifying the unit time Tx, and n is an index for determining the first period TM1, and in the example of FIG. 7, n=5. ｔ ' is the power demand forecast value for the t hour, and y ｔ is the actual power demand value at time t. ｔ ´、y ｔ ) is y ｔ ´ and y ｔ The larger of these values.

[0111] Next, in step S24, the correction unit 113 calculates the linear weighted mean error μ ｔ Determine whether or not is greater than the threshold. ｔ If is greater than the threshold value (YES in step S24), the process proceeds to step S25, where the linear weighted mean error μ ｔ If is equal to or smaller than the threshold value (NO in step S24), no correction is necessary, and the correction process ends.

[0112] Next, in step S25, the correction unit 113 calculates the linear weighted mean error μ ｔ The power demand forecast value included in the second period TM2 is corrected using the formula (3). ｔ ｍｏｄ is the corrected power demand forecast value for the t-th hour. The linear weighted mean error μ ｔ－１ is the linear weighted average error in unit time t-1. ｔ－１ is the predicted power demand value y ｔ ´ is the actual power demand value y ｔ If the power demand forecast value y is higher than the ｔ ´ is the actual power demand value y ｔ If the power demand forecast value y is below the ｔ ´ is the actual power demand value y ｔ If the power demand forecast value y ｔ ' is revised downward, and the power demand forecast value y ｔ ´ is the actual power demand value y ｔ If the power demand forecast value y ｔ ´ is revised upward.

[0113] In the example of FIG. 7, the second period TM2 is three hours, so the three predicted power demand values ​​y ｔ ´ is corrected.

[0114] Next, in step S26, the corrector 113 calculates the corrected power demand forecast value y ｔ ｍｏｄ Then, the updated period forecast value is output.

[0115] Note that the correction unit 113 does not need to perform the correction process if the first period TM1 extends into the previous day. For example, if the prediction time is 3:15, the first period TM1 will be between 23, 0, 1, 2, and 3, so it will extend into the previous day. On the other hand, if the prediction time is 4:15, the first period TM1 will be between 0, 1, 2, 3, and 4, so it will not extend into the previous day. Therefore, the correction unit 113 does not perform the correction process if the prediction time is before 4:00.

[0116] Furthermore, the correction unit 113 does not need to perform the correction process if the second period TM2 extends into the next day. For example, if the prediction time is 9:15 PM, the second period TM2 will be between 10:00 PM, 11:00 PM, and midnight, so it will extend into the next day. On the other hand, if the prediction time is 8:15 PM, the second period TM2 will be between 9:00 PM, 10:00 PM, and 11:00 PM, so it will not extend into the next day. Therefore, the correction unit 113 does not perform the correction process if the prediction time is after 9:00 PM. In summary, the correction unit 113 performs the correction process if the prediction time belongs to a period after 4:00 AM and before 9:00 PM.

[0117] Fig. 8 is a graph illustrating the correction process. In Fig. 8, the vertical axis represents power, and the horizontal axis represents time. For example, on the time axis, "07_12 00" represents midnight on July 12th, and "07_12 06" represents 6:00 on July 12th. A solid line graph G1 represents actual power demand values, a dashed-dotted line graph G2 represents corrected power demand forecast values, and a dotted line graph G3 represents uncorrected power demand forecast values.

[0118] In this example, the moving average error between the power demand forecast value and the actual power demand value was larger than the threshold value during the period from midnight to 6:00 on July 12, and therefore the power demand forecast value shown in graph G3 is corrected as shown in graph G2 after 6:00 on July 12. Therefore, after 6:00 on July 12, graph G2 showing the corrected power demand forecast value almost matches graph G1 showing the actual power demand value, and it can be confirmed that the prediction accuracy of the power demand forecast value has been improved.

[0119] Next, the learning process will be described. FIG. 9 is a flowchart showing an example of the learning process of the hourly predictor Px. This flowchart is executed, for example, in the initial state and at the update timing of the hourly predictor Px. The initial state refers to a state in which the hourly predictor Px has not yet been generated. Assuming that the hourly predictor Px is updated once a day, the update timing is executed, for example, at a predetermined time of the day (for example, midnight).

[0120] First, in step S201, the learning unit 115 acquires, as learning data, input data required for learning from the input data storage unit 121. In the initial state, for example, input data for the past one year is acquired, and at the update timing, taking into consideration that data for the first power feature amount and the third weather feature amount up to the past 14 days is required, input data for the past 14 days is acquired as learning data.

[0121] Next, in step S202, the learning unit 115 performs a filtering process on the learning data. For example, the filtering process may be a process of averaging the learning data so that the time intervals of the timestamps added to the learning data are aligned to a constant time interval.

[0122] Next, in step S203, the learning unit 115 classifies the filtered learning data by category and calculates feature quantities from the classified learning data. As a result, the first to fourth power feature quantities, the first to fourth weather feature quantities, the first and second operation feature quantities, and the first to third date feature quantities are calculated. Note that, since weather forecast values ​​are not used during learning, actual weather values ​​for the prediction period are used as the first and second weather feature quantities, rather than weather forecast values ​​for the prediction period.

[0123] Next, in step S204, the learning unit 115 divides the feature amount calculated in step S203 into a plurality of time-based feature amounts corresponding to the respective time-based predictors P0 to P23.

[0124] Next, in step S205, the learning unit 115 performs machine learning on the hourly predictors P0 to P23. For example, the learning unit 115 inputs an hourly feature corresponding to the hourly predictor Px and outputs a power demand prediction value for a unit time Tx. The learning unit 115 corrects the parameters of the hourly predictor Px so as to minimize the error between the power demand prediction value output by the hourly predictor Px and the actual power demand value for the unit time Tx.

[0125] The time-based predictor Px is trained in this manner. The time-based predictor Px that has undergone the training process is stored in the predictor storage unit 122.

[0126] As described above, according to this embodiment, input data is acquired for each unit time that divides the prediction period, and each time the input data is acquired, the input data is input to the predictor 120 to calculate a power demand forecast value for each unit time Tx in the prediction period. As a result, the latest prediction period and the past prediction period overlap, and the latest power demand forecast value can be used for the overlapping period. This makes it possible to reduce prediction errors that occur when calculating a power demand forecast value using the predictor 120 that uses actual weather values ​​during learning and weather forecast values ​​during prediction, and makes it possible to obtain a more accurate power demand forecast value.

[0127] The present embodiment can employ the following modifications.

[0128] In the above embodiment, 23 power demand forecast values ​​are calculated by 23 hourly forecasters Px, but this is just an example, and 23 power demand forecast values ​​may be calculated by one forecaster 120.

[0129] INDUSTRIAL APPLICABILITY The present disclosure is useful in the technical field of predicting power demand in a facility.

Claims

1. A power demand forecasting method, comprising: a computer inputting input data for each of multiple unit times dividing a prediction period, the input data including actual weather values ​​prior to the prediction period, actual power demand values ​​prior to the prediction period, and weather forecast values ​​for the prediction period into a trained model, and outputting a period forecast value including a power demand forecast value for each of the unit times in the prediction period, the weather forecast values ​​being values ​​obtained for each of the multiple unit times; and for the power demand forecast value output in a first unit time among the multiple unit times dividing the prediction period and the period forecast value output in a second unit time after the first unit time, for the power demand forecast value whose unit times overlap, the power demand forecast value output in the second unit time is adopted.

2. The method for forecasting electricity demand according to claim 1, further comprising the steps of: if there is a missing time in the period forecast value, which is a unit time during which calculation of the electricity demand forecast value has failed, adopting the most recent electricity demand forecast value calculated in the same time zone as the missing time as the missing time, as the electricity demand forecast value for the missing time.

3. The power demand forecasting method according to claim 1 or 2, further comprising: calculating a first error, which is an error between the power demand forecast value calculated for a first period in the past from the prediction time point and the actual power demand value for the first period; and, if the first error exceeds a threshold value, correcting the power demand forecast value for the second period using the first error so that a second error, which is the error for a second period in the future from the prediction time point, becomes smaller.

4. The method for forecasting electric power demand according to claim 3, wherein when the first period extends over the previous day, the correction of the electric power demand forecast value is not executed.

5. The method for forecasting electric power demand according to claim 3, wherein the second period is determined in advance based on a full charge time of a storage battery provided in the facility for which the electric power demand is to be predicted.

6. The method for forecasting electricity demand according to claim 1 or 2, wherein the input data further includes operation record information of equipment for which electricity demand is forecast, and operation schedule information of the equipment during the forecast period.

7. The electricity demand forecasting method according to claim 1 or 2, wherein the trained model includes a plurality of time-specific trained models that calculate the electricity demand forecast value for each of the plurality of unit times.

8. The power demand forecasting method of claim 7, wherein output of the power demand forecast value includes: generating features of the input data by classifying the input data according to categories; dividing the features into a plurality of features corresponding to each of the plurality of time-based trained models; calculating the power demand forecast value for each unit time by inputting the plurality of features into a time-based trained model corresponding to the plurality of features; and calculating the period forecast value by integrating the power demand forecast values ​​calculated for each unit time.

9. The power demand forecasting method of claim 6, wherein the input data includes an actual power demand value for each unit time on the day before the prediction time, and the output of the power demand forecast value includes: determining whether the prediction time falls on an operating day based on the operation schedule information; determining whether the previous day falls on an operating day based on the actual operation information; and if the prediction time falls on an operating day and the previous day falls on a non-operating day, inputting the actual power demand value for each unit time on the most recent operating day into the learned model.

10. The power demand forecasting method of claim 6, wherein the input data includes an actual power demand value for each unit time on the day before the prediction time, and calculation of the power demand forecast value includes: determining whether the prediction time falls on an operating day based on the operation schedule information; determining whether the previous day falls on an operating day based on the actual operation information; and if the prediction time falls on a non-operating day and the previous day falls on an operating day, inputting the actual power demand value for each unit time on the most recent non-operating day into the learned model.

11. The method for forecasting electric power demand according to claim 3, wherein the first error is a linear weighted average error between the electric power demand forecast value and the actual electric power demand value.

12. The electricity demand forecasting method of claim 1 or 2, wherein the trained model is a model obtained by machine learning learning data in which actual weather values ​​prior to a reference time point and the actual electricity demand values ​​prior to the reference time point are explanatory variables, and actual electricity demand values ​​after the reference time point corresponding to the explanatory variables are objective variables.

13. An electricity demand forecasting device including a processor, wherein the processor inputs input data including actual weather values ​​from the past before the prediction period, actual electricity demand values ​​from the past before the prediction period, and weather forecast values ​​for the prediction period into a learned model for each of multiple unit times that divide the prediction period, to output a period forecast value including an electricity demand forecast value for each of the unit times in the prediction period, the weather forecast values ​​being values ​​obtained for each of the multiple unit times, and for the electricity demand forecast value that overlaps in unit time between the period forecast value output in a first unit time and the period forecast value output in a second unit time after the first unit time, the electricity demand forecast value output in the second unit time is adopted.

14. A power demand forecasting program that causes a computer to execute a process of inputting input data including actual weather values ​​from the past before the prediction period, actual power demand values ​​from the past before the prediction period, and weather forecast values ​​for the prediction period into a trained model for each of multiple unit times that divide the prediction period, thereby outputting a period forecast value including a power demand forecast value for each of the unit times in the prediction period, wherein the weather forecast value is a value obtained for each of the multiple unit times, and for the power demand forecast value that overlaps in unit times between the period forecast value output in a first unit time and the period forecast value output in a second unit time after the first unit time, the power demand forecast value output in the second unit time is adopted.

Images