JP2019201440A - Demand prediction server and demand prediction method - Google Patents

Abstract

To provide a demand prediction server and a demand prediction method which select data necessary and available to predict a prediction target for each prediction.SOLUTION: A demand prediction server 100 comprises a data selection meta model learning part which learns a data selection meta model used to create a data selection rule using past data including an actual value of a prediction target, a data selection part which creates the data selection rule based on the data selection meta model, dynamically selects prediction input data from future data about a demand value of the prediction target as a prediction input according to the data selection rule, and dynamically selects learning data from past data according to the data selection rule to be used in order to predict the demand value of the prediction target, a prediction model learning part which learns a prediction model for predict the demand value of the prediction target using the learning data, and a prediction part which predicts the demand value of the prediction target by means of inputting prediction input data in the prediction model.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a demand prediction server and a demand prediction method.

  There is a demand prediction service as one of services using IoT (Internet of Things) and big data. Demand prediction service, for example, is a service that predicts demand values such as power demand and heat demand that are valuable to the user by analyzing data that is constantly accumulated in the IoT platform configured on the cloud, and provides it to the user It is.

  By using the demand prediction service, for example, the generated power can be bought and sold in the power trading market in accordance with the predicted value of heat demand that changes every moment in the combined heat and power plant. However, the data stored in the IoT platform is enormous and there are many types of data. The IoT platform also stores a large amount of unintended data and data with low reliability. For this reason, in order to effectively use data accumulated in the IoT platform by applying techniques such as machine learning and artificial intelligence, it is necessary to select necessary data according to the prediction target.

  As a technique for selecting data necessary for the prediction, for example, Patent Document 1 discloses that “the demand demand target value based on the point prediction and the management value excess probability based on the section prediction, that is, the demand management target value is the actual demand. A technique for calculating the probability of exceeding is disclosed. Further, the literature 1 describes that “by using only weather forecast data and not using actual weather values, it is possible to include uncertainty of weather forecasts, and even in areas where measured data does not exist, demand A prediction model can be built. "

JP 2017-147792 A

  The demand prediction method using the technique disclosed in Patent Document 1 corresponds to the time at which a demand value is predicted from a weather forecast value database in which data including past weather forecast values is accumulated (hereinafter referred to as “prediction time”). It is necessary to extract past weather forecast values. If the forecast time is changed or there are multiple forecast times, the weather forecast value database that is different from the database accumulated on the IoT platform so far, or multiple sets of weather forecast values according to the forecast time A database must be prepared.

  However, the data created by weather forecasts made in the past is enormous, and not all data is accumulated in the weather forecast value database. Also, it is not known when the data accumulated in the weather forecast value database was forecasted and created. In addition, for example, the power demand changes with time, and may suddenly increase when the temperature falls below a certain temperature at the turn of the season. For this reason, for example, there is a case where the demand value to be predicted cannot be accurately obtained unless the latest data is used at one time and the data of one year ago is used at another time.

  If it is difficult to secure a sufficient amount of data without data loss, the technique disclosed in Patent Document 1 cannot be used. Furthermore, because the demand value to be forecasted has characteristics such as an increasing / decreasing tendency and seasonal changes, the forecasted demand is selected by uniformly selecting past data (weather forecast value) corresponding to the forecast time from the weather forecast value database. The value did not provide the expected prediction accuracy.

  The present invention has been made in view of such a situation, and an object thereof is to select necessary and usable data for each prediction in order to perform prediction of a prediction target.

  A demand prediction server according to the present invention includes a data selection metamodel learning unit that learns a data selection metamodel used to create a data selection rule using past data including actual values to be predicted, a data selection meta Create a data selection rule based on the model, dynamically select the prediction input data that corresponds to the data selection rule from the future data related to the demand value to be predicted, and use the past data as the data selection rule Corresponding and prediction that learns the prediction model for predicting the demand value of the prediction target using the data selection part that dynamically selects the learning data used to predict the demand value of the prediction target and the learning data A model learning unit; and a prediction unit that inputs prediction input data to the prediction model and predicts a demand value to be predicted.

According to the present invention, according to the prediction target, the past data and the future data necessary and usable for predicting the demand value of the prediction target are dynamically selected for each prediction target, and the demand value of the prediction target is highly accurate. Can be predicted. Thereby, for example, the data accumulated on the IoT platform can be effectively used for the demand prediction of the prediction target.
Problems, configurations, and effects other than those described above will be clarified by the following description of embodiments.

It is a block diagram which shows the structural example of the demand prediction server which concerns on one embodiment of this invention. It is explanatory drawing which shows the definition of the phrase which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the past data which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the future data which concerns on one embodiment of this invention. It is a flowchart which shows the example of the demand prediction process performed with the demand prediction server which concerns on one embodiment of this invention. It is a flowchart which shows the example of the data selection metamodel learning process which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the data classification after setting which concerns on one embodiment of this invention. It is explanatory drawing which shows the example of the data selection rule and search range which concern on one embodiment of this invention. It is explanatory drawing which shows the structural example of the data selection metamodel which concerns on one embodiment of this invention. It is a flowchart which shows the example of the data selection process which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the learning data which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the prediction input data which concerns on one embodiment of this invention. It is explanatory drawing which shows the structural example of the demand forecast data which concerns on one embodiment of this invention. It is a block diagram which shows the hardware structural example of the computer which concerns on one embodiment of this invention.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the accompanying drawings. In the present specification and drawings, components having substantially the same function or configuration are denoted by the same reference numerals, and redundant description is omitted.

FIG. 1 is a block diagram illustrating a configuration example of the demand prediction server 100.
The demand prediction server 100 is, for example, a device that predicts values after several hours or days after a heat demand that is a prediction target. The demand prediction server 100 is constructed as an IoT platform, for example, so that a user can access the demand prediction server 100 through a network and acquire a demand value of a prediction target predicted by the demand prediction server 100. Composed. The demand prediction server 100 includes a past database 101A, a data selection metamodel learning unit 102, a data selection unit 103, future data 104, a prediction model learning unit 105, and a prediction unit 106.

In the past database 101A, past data 101 including the actual value to be predicted is accumulated.
The data selection metamodel learning unit 102 uses the past data 101 read from the past database 101A to create a data selection rule 131 (see FIG. 8 to be described later) and a data selection metamodel 120 (to be described later). 9). The learning of the data selection metamodel 120 is, for example, for creating a plurality of nodes 121a and links 121b between the plurality of nodes 121a constituting the neural network 121 shown in FIG. This is a process for obtaining necessary coefficients and the like.

  The data selection unit 103 creates a data selection rule 131 based on the data selection metamodel 120. The data selection rule 131 is a rule created by inputting actual values such as the predicted time and future data 104 to the learned data selection metamodel 120. And the data selection part 103 corresponds to the data selection rule 131 from the future data 104 regarding the demand value of prediction object, and dynamically inputs the prediction input data 112 (refer to FIG. 12 described later) to be input to the prediction part 106. The learning data 110 (see FIG. 11 described later) that is selected and corresponds to the data selection rule 131 from the past data 101 and is used to predict the demand value to be predicted is dynamically selected. That is, the learning data 110 is a part of data selected from the past data 101, and the predicted input data 112 is a part of data selected from the future data 104.

  The future data 104 stores the current value of an optional item excluding the heat demand that is the prediction target at the prediction time that is the future time viewed from the prediction time when the demand value of the prediction target is predicted. The current value is the latest value that can be acquired at the time of performing a demand prediction process shown in FIG. The future data 104 is created based on, for example, forecast values received from the Japan Meteorological Agency that forecasts weather, information acquired from calendar information, or the like.

  The prediction model learning unit 105 uses the learning data 110 selected by the data selection unit 103 to learn a prediction model 111 for predicting a demand value to be predicted. For example, a general machine learning method such as a support vector machine or a Gaussian process is used as the prediction model 111 learned by the prediction model learning unit 105 and the learning method of the prediction model 111.

  The prediction unit 106 predicts a demand value to be predicted by inputting the prediction input data 112 to the prediction model 111 learned by the prediction model learning unit 105. At this time, the demand value predicted by the prediction unit 106 is the demand value of the heat demand, and this demand value is used as the demand prediction value.

  The demand prediction data 107 is data including a demand prediction value predicted by the prediction unit 106. The demand prediction data 107 is displayed on a PC (Personal Computer) (not shown) connected to the demand prediction server 100 or supplied to a power system (not shown).

Next, the definition of the phrase used in this specification will be described.
FIG. 2 is an explanatory diagram showing the definition of words.
In the following description, the time when the demand prediction server 100 predicts the heat demand (that is, the prediction time) is 3:00 pm on January 30, 2018, and the day when the demand prediction server 100 predicts the heat demand (that is, the demand) The predicted date is January 30, 2018. And the demand prediction server 100 shall estimate the heat demand which is a prediction object by the hour unit of the next day (January 31, 2018). For this reason, for example, January 31, 2018 is referred to as a “prediction target date”. When the predicted time is the current time, past data based on the predicted time is represented as past data 101, and future data is represented as future data 104.

  For example, the time when the Japan Meteorological Agency or the like made a weather forecast at a future forecast target time is called “forecast time”. A weather forecast for a certain forecast target time may be performed at a different forecast time. Therefore, a plurality of different forecast times can exist for a certain forecast target time. Therefore, in the following description, the difference between the forecast time, which is the forecast time of the forecast value of the forecast target time, and the forecast target time is called a “forecast step”. For example, a time that is back by the forecast step A from the forecast target time is the forecast time A, and a time that is traced back by the forecast step B from the forecast target time is the forecast time B.

Next, configuration examples of the past data 101 and the future data 104 will be described with reference to FIGS.
FIG. 3 is an explanatory diagram illustrating a configuration example of the past data 101. An example of past data 101 expressed in a comma separated CSV (Comma Separated Values) format is shown on the upper side of FIG. 3, and an example of past data 101 being rewritten into a table on the lower side of FIG.

As shown in FIG. 1, the past data 101 is stored in the past database 101A. Each item constituting the past data 101 represents a past performance value of an arbitrary item related to demand prediction such as date and time, forecast step, date type, outside air temperature, wind speed, humidity, solar radiation amount, event ID, and heat demand.
Here, the content of each item which comprises the past data 101 is demonstrated. The phrase in parentheses added to the name of each item represents the field name of each item constituting past data 101.

The date and time (Date Time) is an item representing the forecast target time shown in FIG.
The forecast step (ForecastStep) is an item that represents the difference between the forecast time that indicates when the forecast value is forecasted by the weather forecast and the forecast target time, as shown in FIG. For example, in a data record in which “2015-1-1” is stored in the forecast time and “9” is stored in the forecast step, December 31, 2014, which is 9 hours before midnight on January 1, 2015 The forecast value (outside temperature, wind speed, humidity) obtained at 15:00 on January 1, 2015 is stored. In the data record stored as “33” in the forecast step, the forecast is obtained at 15:00 on December 30, 2014, which is 33 hours before January 1, 2015 (24 hours and 9 hours before). The predicted value (outside temperature, wind speed, humidity) at 0:00 on January 1, 2015 is stored.

  The past data 101 can hold a plurality of data records with different forecast times even if the forecast time is the same for each forecast step. For example, as shown in FIG. 2, the past data 101 holds a plurality of data records at a forecast time A that is traced back by the forecast step A and a forecast time B that is traced back by the forecast step B with respect to one forecast target time. it can.

  The past data 101 can also hold data records having actual values such as weather observation values by setting the forecast step to “0”, for example. As shown in FIG. 3, the past data 101 holds three data records with the prediction steps set to “0”, “9”, and “33”. As a result, conventionally, only one data record with the prediction step set to “0” could be held, but in the present embodiment, a sufficient amount of forecast data necessary for predicting the heat demand is stored. It becomes possible to secure the past data 101.

The “Holiday” is an item for distinguishing, for example, only weekdays, only closed days, only public holidays, closed days and public holidays. For example, if “1” is stored in the day type, it indicates that the school is closed and is a public holiday.
The outside air temperature (Temperature) is an item for storing the outside air temperature at a certain point.
The wind speed (Wind100m, Wind2m) is an item for storing the wind speed at a certain point. Here, the wind speed (Wind100m) stores the wind speed at the ground surface 100m, and the wind speed (Wind2m) stores the wind speed at the ground surface 2m.

Humidity is an item that stores the humidity at a certain point. Since humidity takes a value in the range of “0” to “1”, for example, a value of “0.2” indicates that the humidity is 20%.
The solar radiation amount (SolarRadiation) is an item for storing the solar radiation amount at a certain point. In this data record, since it is 0:00 (midnight), “0” is stored as the amount of solar radiation.

  The event ID (Event_ID) is an item for storing an identifier of a special event (for example, sports or outdoor event) held at a certain point. When special events are held, heat demand is also expected to increase. If no special event is held, “0” is stored in the event ID.

  The heat demand (HeatDemand) is an item for storing the heat demand (Wh) at a certain point. In the past data 101, the value of the heat demand actually observed at a certain date and time is stored as the same value regardless of the value of the forecast step. For example, the value of the heat demand of all data records whose date is expressed as “2015-1-1 00:00:00” is “11.24”.

  FIG. 4 is an explanatory diagram illustrating a configuration example of the future data 104. An example of future data 104 expressed in CSV format is shown on the upper side of FIG. 4, and an example of rewriting future data 104 into a table on the lower side of FIG.

  As shown in FIG. 1, the future data 104 stores the current value of an arbitrary item at the predicted time. For example, when the prediction target date and time is set to January 31, 2018 00:00:00, it was predicted at 15:00 on January 30, 2018, which is 9 hours before indicated in the prediction step. The latest forecast value at 00:00:00 on January 31 of the year is stored as the current value. The configuration of each item of the future data 104 is the same as the configuration of the past data 101, but only the item of heat demand (HeatDemand) that is a prediction target is excluded.

FIG. 5 is a flowchart illustrating an example of a demand prediction process performed by the demand prediction server 100.
First, the data selection metamodel learning unit 102 learns the data selection metamodel 120 (see FIG. 10 described later) (S1). Details of processing in which the data selection metamodel learning unit 102 learns the data selection metamodel 120 in step S1 will be described with reference to FIG.

  Next, the data selection unit 103 selects the learning data 110 to be input to the prediction model 111 from the past data 101 using the data selection metamodel 120 learned in step S1, and sets the prediction input data 112 in the future. A selection is made from the data 104 (S2). Details of the process in which the data selection unit 103 selects the past data 101 and the future data 104 in step S2 will be described with reference to FIG.

  Next, the prediction model learning unit 105 learns the prediction model 111 using the learning data 110 selected by the data selection unit 103 (S3).

  Next, the prediction unit 106 performs a prediction process of calculating a demand prediction value to be predicted using the prediction model 111 learned by the prediction model learning unit 105 and the prediction input data 112 selected by the data selection unit 103. (S4) The demand forecast data 107 is created and this process is terminated. The demand forecast data 107 storing the demand forecast value is displayed on a display screen provided in a PC or the like (not shown) connected to the demand forecast server 100. However, when the demand forecast data 107 is used by internal processing, the demand forecast data 107 may not be displayed.

Next, the data selection metamodel learning process in step S1 and the data selection process in step S2 will be described in order.
FIG. 6 is a flowchart illustrating an example of the data selection metamodel learning process. This data selection metamodel learning process represents the detailed contents of the process in step S1 of the demand prediction process shown in FIG.

  First, the data selection metamodel learning unit 102 sets a data type used for data selection (hereinafter referred to as “post-setting data type 301”) (S11).

Here, a configuration example of the data type set in step S11 will be described with reference to FIG.
FIG. 7 is an explanatory diagram illustrating a configuration example of the post-setting data type 301.

  The post-setting data type 301 is a part of data items of the past data 101 and future data 104. The post-setting data type 301 is used to extract only necessary and usable data for prediction of heat demand from various past data 101 accumulated in the past database 101A. In the post-setting data type 301, items are set by reading from an external file or by an input operation by an operator.

  Among the items included in the past data 101 as the post-setting data type 301, for example, date / time (Date Time), forecast step (ForecastStep), date type (Holiday), outside temperature (Temperature), wind speed (Wind2m), humidity ( Humidity), solar radiation (SolarRadiation), event ID (Event_ID), and heat demand (HeatDemand) are set. However, even if it is included in the past data 101, items that are considered unnecessary for prediction of heat demand, for example, wind speed (Wind 100m), are not included in the post-setting data type 301.

Returning to the description of the processing in FIG.
After step S11, the data selection metamodel learning unit 102 sets a search range for data selection rules (S12). As the search range of the data selection rule, for example, searching the search range 132 defined in “train-data” and “wheather” shown in FIG. 8 to be described later can be read from an external file or input operation by an operator Is set by

Here, examples of the data selection rule 131 and the search range 132 will be described with reference to FIG.
FIG. 8 is an explanatory diagram illustrating an example of the data selection rule 131 and the search range 132.

The data selection rule 131 includes a learning data selection rule 131A and a data item selection rule 131B (see FIG. 9 described later) regarding the post-setting data type 301.
The learning data selection rule 131 </ b> A is a rule that is referred to when the data selection unit 103 selects the learning data 110. As the learning data selection rule 131A, for example, the number of days going back from the prediction target date, whether to select only when the same day type, whether to select only when the event ID is the same, the outside temperature of the data record to be selected There is a lower and upper limit for the forecast step of the data record to be selected.
The data item selection rule 131B is a rule for selecting a data item common to the learning data 110 and the prediction input data 112, which is referred to when the data selection unit 103 selects the learning data 110 and the prediction input data 112. is there. The data item selection rule 131B includes, for example, weather forecast, time information, and prediction target.

  The search range 132 is dynamically changed according to the prediction target in the data selection rule 131. FIG. 8 shows a setting for the data selection metamodel learning unit 102 to search for the data selection rule 131 based on “train-data” and “wheather” of the data selection rule 131 as the search range 132, for example.

Here, the contents of each item in the data selection rule 131 will be described.
First, “train-data” will be explained.
“N-days-ahead” is a parameter that indicates how many days ago data is selected from the prediction target date.
“Same-workday” is a parameter indicating whether or not the day type specified in “n-days-ahead” is the same as the day type of the prediction target day. If the day specified in “n-days-ahead” is the same as the day type of the prediction target day, “true” is stored, and if different, “false” is stored. In this example, if “n-days-ahead” is defined as “35” and the day 35 days before the prediction target date is the same day type “true” as the prediction target date, data can be selected. Indicated.

  Similarly, “same-event” is a parameter indicating whether the event held on the day specified in “n-days-ahead” is the same as the event held on the prediction target day. . “True” is stored if the event defined by “n-days-ahead” and the prediction target date are the same, and “false” is stored if they are different.

In addition, “temperature-range” is a parameter indicating that, for example, data on the day when the minimum temperature is −5 ° C. and the maximum temperature is 20 ° C. and the outside temperature is selected from the past data 101 can be selected. It is.
“Available-forecast-steps” is a parameter that limits the range of available forecast steps. For example, it is indicated that data having a prediction step within the range of 0 hours to 33 hours before the prediction target time can be selected from the past data 101.

Next, “weather” will be described. “Weather” is a parameter that specifies items related to weather forecasts that can be used for demand forecasting.
“Temperature” indicates that the item added to the learning data 110 (see FIG. 11 described later) and the predicted input data 112 (see FIG. 12 described later) is “Temperature”. In addition to “Temperature”, an item such as “Wind2m” may be specified.

  “Time” is a parameter indicating what the data item 112A (see FIG. 12 described later) in the predicted input data 112 is. For example, in addition to “Date Time”, “Holiday”, and “Forecast Step” defined in “time”, “Temperature” defined in “weather” is selected as the data item 112A.

  “Target” is a parameter indicating what the prediction target is. In the present embodiment, “Heat Demand” is set as “target”. Since the heat demand is not known at the prediction stage, “Heat Demand” is excluded from the data item 112A.

  On the other hand, the learning data 110 is data selected from the past data 101. For this reason, the data item 110A (see FIG. 11 described later) in the learning data 110 includes the past data 101 as “Date Time”, “Holiday”, “Forecast Step”, and “weather” defined as “time”. It has “Temperature” defined and “Heat Demand” defined in “target”.

Returning to the description of FIG.
The data selection metamodel learning unit 102 starts an iterative process for searching for corresponding data from the past data 101 based on the search range 132 of the data selection rule set in step S12 (S13), and from step S14. The process of step S17 is repeated. The processing from step S14 to step S17 is processing for searching the past data 101 for data necessary for the data selection unit 103 to create the data selection rule 131 in the data selection processing shown in FIG.

  In the iterative process of the past data 101, the data selection metamodel learning unit 102 starts the iterative process based on the data selection rule 131 (S14), and repeats the processes of step S15 and step S16. Therefore, the data selection metamodel learning unit 102 searches for the data selection rule 131 (S15).

  In step S15, the data selection metamodel learning unit 102 changes each item in the search range 132 as follows. For example, the data selection metamodel learning unit 102 adds or deletes the number of days going back from the prediction target date, adds or deletes the selection rule of the day type and the event ID for the learning data selection rule 131A of the data selection rule 131. Change the upper and lower limits of the outside air temperature and the upper and lower limits of the forecast step. In addition, the data selection metamodel learning unit 102 adds or deletes a weather forecast data item for the data item selection rule 131B of the data selection rule 131.

  The data selection metamodel learning unit 102 evaluates the prediction generalization performance using the data selection rule 131 and the past data 101 after searching in step S15 (S16). As the prediction generalization performance, for example, the marginal likelihood obtained when the prediction model 111 is learned by the same procedure as the demand prediction process shown in FIG. 5 is multiplied by the probability of the actual demand value in the prediction distribution of the prediction model 111. Calculate the calculated value. Thereafter, returning to step S14, the data selection metamodel learning unit 102 repeats the processing.

  After the iterative processing in steps S15 and S16 is completed, the data selection metamodel learning unit 102 outputs the data selection metamodel 120 so that the data selection rule 131 having the best predicted generalization performance calculated in step S16 is output. Is updated (S17). By evaluating the prediction generalization performance, a data selection meta that can predict not only the predicted value of heat demand using only the past data 101 but also the predicted value of heat demand that can be acquired in the future with high accuracy. A model 120 can be obtained.

  Thereafter, the processing returns to step S13, and the processing between steps S14 to S17 is repeated until the processing of all past data is completed. When the processing of all past data is completed, the data selection metamodel learning unit 102 outputs the data selection metamodel 120 and ends this processing.

Next, the data selection metamodel 120 will be described with reference to FIG.
FIG. 9 is an explanatory diagram illustrating a configuration example of the data selection metamodel 120.

  The data selection metamodel 120 is a model that outputs a data selection rule 131 from input data. As input data, there are a prediction time 130 when the demand value of the prediction target on the prediction target date is predicted, and future data 104. The data selection metamodel 120 includes a neural network 121 that outputs a data selection rule 131 based on input data. For example, the neural network 121 can update the weight of the node 121a, the configuration of the link 121b connecting the plurality of nodes 121a, the coefficient, and the like by deep learning based on input data.

  The data selection rule 131 output from the neural network 121 includes a learning data selection rule 131A and a data item selection rule 131B. As described above, the data selection metamodel learning unit 102 learns the data selection metamodel 120, so that the relationship between the prediction target and the data selection rule 131 can be learned.

  FIG. 10 is a flowchart illustrating an example of data selection processing. This data selection process represents the detailed contents of the process in step S2 of the demand prediction process shown in FIG.

  First, the data selection unit 103 generates a data selection rule 131 based on the predicted time 130, the future data 104, and the data selection metamodel 120 (S21). At this time, as shown in FIG. 9, the data selection rule 131 is output based on the prediction time 130 and the future data 104 input to the data selection metamodel 120.

  Next, the data selection unit 103 selects the learning data 110 by performing the non-determination of the past data 101 using the data selection rule 131 (S22). That is, the past data 101 corresponding to the data selection rule 131 is selected as the learning data 110. For example, when the data selection rule 131 for selecting the past data 101 from the predicted time one day before as the learning data 110 is created, the previous past data 101 is not selected as the learning data 110. Thus, the past data 101 not corresponding to the data selection rule 131 is not used in the subsequent processing.

  In addition, the data selection unit 103 selects the predicted input data 112 by performing the non-determination of the future data 104 using the data selection rule 131 (S22). That is, the future data 104 corresponding to the data selection rule 131 is selected as the predicted input data 112. Future data 104 not corresponding to the data selection rule 131 is not used in the subsequent processing.

  Then, as shown in FIG. 5 described above, the learning data 110 selected by the data selection unit 103 is used for learning the prediction model 111 (step S3 in FIG. 5), and the prediction input data 112 uses the prediction model 111. It is used for the demand forecast which had been (step S4 of FIG. 5).

FIG. 11 is an explanatory diagram illustrating a configuration example of the learning data 110.
The data item 110A and the data record 110B constituting the learning data 110 are selected from the past data 101 based on the learning data selection rule 131A and the data item selection rule 131B constituting the data selection rule 131. Here, it is shown that the date and time (Date Time), date type (Holiday), forecast step (ForecastStep), outside temperature (Temperature), and heat demand (HeatDemand) are selected from the past data 101.

FIG. 12 is an explanatory diagram illustrating a configuration example of the predicted input data 112.
The data item 112A constituting the predicted input data 112 is selected from the future data 104 based on the data item selection rule 131B. Here, it is shown that the date and time (Date Time), date type (Holiday), forecast step (ForecastStep), and outside temperature (Temperature) are selected from the past data 101.

FIG. 13 is an explanatory diagram illustrating a configuration example of the demand prediction data 107.
The data item 107A of the demand prediction data 107 includes a demand prediction value output from the prediction process (S4) of the prediction unit 106 based on the prediction model 111 and the prediction input data 112. This data item 107A has date and time (Date Time) and heat demand (HeatDemand).

  Here, when the prediction input data 112 selected from the future data 104 is input to the prediction model 111, the heat demand (HeatDemand) at the prediction target date and time not included in the future data 104 is predicted by the prediction model 111. The For this reason, the data item 107A stores the value of heat demand (HeatDemand) predicted every hour on the prediction target date by the prediction model 111 together with the date and time (Date Time) selected from the prediction input data 112. .

Next, the hardware configuration of the computer C configuring the demand prediction apparatus 100 will be described.
FIG. 14 is a block diagram illustrating a hardware configuration example of the computer C.

  The computer C is hardware used as a so-called computer. The computer C includes a CPU (Central Processing Unit) C1, a ROM (Read Only Memory) C2, and a RAM (Random Access Memory) C3 connected to the bus C4. Furthermore, the computer C includes a nonvolatile storage C5 and a network interface C6.

  The CPU C1 reads out the program code of software that realizes each function according to the present embodiment from the ROM C2, and executes it. In the RAM C3, variables, parameters and the like generated during the arithmetic processing are temporarily written. The data selection metamodel learning unit 102, the data selection unit 103, the prediction model learning unit 105, and the prediction unit 106 illustrated in FIG. 1 represent functions executed by the CPU C1.

  As the nonvolatile storage C5, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), a flexible disk, an optical disk, a magneto-optical disk, a CD-ROM, a CD-R, a magnetic tape, a nonvolatile memory, or the like is used. It is done. In the nonvolatile storage C5, in addition to the OS (Operating System) and various parameters, a program for causing the computer C to function is recorded. The ROM C2 and the non-volatile storage C5 permanently record programs and data necessary for the operation of the CPU C1, and are computer-readable non-transitory storing the programs executed by the computer C. Used as an example of a recording medium. For example, each data such as the past database 101A, future data 104, and demand forecast data 107 is stored in the nonvolatile storage C5.

  For example, a network interface card (NIC) or the like is used as the network interface C6, and various types of data can be transmitted and received between devices via a LAN (local area network) connected to a terminal, a dedicated line, or the like. is there. The demand prediction server 100 can communicate with a PC terminal (not shown) used by an operator or the like through an interface C6.

  In the demand prediction server 100 according to the embodiment described above, the prediction input data 112 and the learning data 110 necessary and usable for prediction can be dynamically selected from the past data 101 for each prediction according to the prediction target. It becomes possible. Then, by applying the selected prediction input data 112 and learning data 110 to the prediction model 111, the demand value of the prediction target at the future prediction target date and time can be predicted with high accuracy. As described above, it is possible to select the past data 101 necessary for the prediction from the enormous past data 101 accumulated on the IoT platform.

  The demand prediction server 100 according to the present embodiment can be applied to a process for predicting a demand value in a power system such as a VPP (Virtual Power Plant), a power transaction support system, and a demand response device. Furthermore, the demand prediction server 100 can be applied to a general energy supply system such as thermal energy, biomass energy, and renewable energy to predict demand values of various energy.

Further, the present invention is not limited to the above-described embodiments, and it is needless to say that other various application examples and modifications can be taken without departing from the gist of the present invention described in the claims.
For example, the above-described embodiment is a detailed and specific description of the configuration of the demand prediction server in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to one having all the configurations described. Moreover, it is also possible to add, delete, or replace another configuration with respect to part of the configuration of the present embodiment.
Further, the control lines and information lines indicate what is considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. Actually, it may be considered that almost all the components are connected to each other.

  DESCRIPTION OF SYMBOLS 100 ... Demand prediction server 101 ... Past data 102 ... Data selection metamodel learning part 103 ... Data selection part 104 ... Future data 105 ... Prediction model learning part 106 ... Prediction part 107 ... Demand prediction data 110 ... Learning data, 111 ... Predictive model, 112 ... Predicted input data, 120 ... Data selection metamodel, 131 ... Data selection rule

Claims (6)

A data selection metamodel learning unit that learns a data selection metamodel that is used to create a data selection rule using past data that includes actual values to be predicted;
Creating the data selection rule based on the data selection metamodel, dynamically selecting prediction input data corresponding to the data selection rule from the future data regarding the demand value of the prediction target, and serving as a prediction input; and A data selection unit that dynamically selects learning data that corresponds to the data selection rule from the past data and is used to predict the demand value of the prediction target;
A prediction model learning unit that learns a prediction model for predicting the demand value of the prediction target using the learning data;
A demand prediction server comprising: a prediction unit that inputs the prediction input data to the prediction model and predicts the demand value of the prediction target. The item selected from the past data according to the data selection rule includes a difference between a forecast time which is a forecast time of a forecast value at a forecast target time and the forecast target time,
The demand prediction server according to claim 1, wherein the past data and the future data include a data record including the forecast value for each difference. The data selection metamodel is composed of a neural network that inputs the prediction time when the demand value of the prediction target on the prediction target date is predicted, and the future data, and outputs the data selection rule.
The data selection metamodel learning unit updates the data selection metamodel using a data type in which an item to be selected from the past data is defined and a value searched from the past data according to the data selection rule. Demand forecast server described in. The data selection metamodel learning unit evaluates a predicted generalization performance of a value searched from the past data according to the data selection rule, and outputs the data selection rule with the best predicted generalization performance. The demand prediction server according to claim 3, wherein the selected metamodel is updated. The demand prediction server according to any one of claims 1 to 4, wherein the prediction target is heat demand. Learning a data selection metamodel used to create a data selection rule using past data including actual values to be predicted;
Creating the data selection rule based on the data selection metamodel, dynamically selecting prediction input data corresponding to the data selection rule from the future data regarding the demand value of the prediction target, and serving as a prediction input; and Dynamically selecting learning data corresponding to the data selection rule from the past data and used to predict the demand value of the prediction target;
Learning a prediction model for predicting the demand value of the prediction target using the learning data;
A step of inputting the prediction input data to the prediction model and predicting the demand value of the prediction target.

Images