Prediction device, prediction method, and non-transitory computer readable medium

Patent metadata

Field	Value
Publication number	US-12033100-B2
Application number	US-201916567604-A
Country	US
Kind code	B2
Filing date	Sep 11, 2019
Priority date	Mar 12, 2019
Publication date	Jul 9, 2024
Grant date	Jul 9, 2024

Abstract

According to one embodiment, a prediction device includes: first processing circuitry and second processing circuitry. The first processing circuitry corrects a demand result value on a particular date on a basis of an adjustment coefficient depending on the particular date; and generates a prediction model of a demand on a basis of the corrected demand result value on the particular date and a demand result value on a non-particular date. The second processing circuitry calculates a demand predicted value on a date of a prediction target on a basis of demand result values on one or more dates, an adjustment coefficient depending on a particular date among the one or more dates, and the prediction model. The second processing circuitry inversely corrects the demand predicted value using an adjustment coefficient depending on the date of the prediction target if the date of the prediction target is a particular date.

First claim

The invention claimed is:

1. A prediction device comprising:
a hardware storage configured to store demand result data including a plurality of demand result values on a plurality of dates and calendar data including information identifying whether the plurality of dates are particular dates or not, a date that is a particular date including a holiday, a non-particular date that is not the particular date including a weekday, the demand result values being result values of electric power supply amounts supplied in a supply area of an electric power company, and a number of dates with the non-particular date being greater than a number of dates with the particular date;
adjustment coefficient calculation circuitry configured to:
identify the particular dates based on the calendar data,
calculate a number of elapsed days in a year, corresponding to each of the particular dates,
calculate an adjustment coefficient for each of the particular dates based on the demand result data, and
generate an adjustment coefficient database that maps the number of elapsed days in the year to the adjustment coefficient;

adjustment coefficient determination circuitry configured to:
receive a request to obtain an adjustment coefficient for a specified date,
calculate a number of elapsed days in a year for the specified date,
identify, in the adjustment coefficient database, a number of elapsed days closest to the calculated elapsed days,
obtain from the adjustment coefficient database an adjustment coefficient corresponding to the identified elapsed days, and
return the obtained adjustment coefficient to a source of the request;

model learner circuitry configured to:
judge whether each date in the demand result data is the particular date or not based on the calendar data, and if the date is the particular date, send a request to the adjustment coefficient determination circuitry to obtain the adjustment coefficient for the date,
divide one of the demand result values on a date corresponding to the particular date among the plurality of dates by the obtained adjustment coefficient to acquire a first normalized demand result value, and
generate, by machine learning, a prediction model for predicting a demand value on any date on a basis of the first normalized demand result value, which is received as learning data for the machine learning, on the date corresponding to the particular date and ones of the demand result values on dates corresponding to the non-particular date among the plurality of dates, the prediction model having input variables each assigned one of the first normalized demand result or one of the demand result values on one or more dates before the any date, having an output variable of a predicted demand value;

input circuitry configured to receive information of a target date for demand prediction specified by a user; and
predictor circuitry configured to:
specify dates before the target date for demand prediction, wherein a number of the dates before the target date is identical to a number of the input variables in the prediction model,
judge whether the specified date is the particular date or not, and if the specified date is the particular date, send a request to the adjustment coefficient determination circuitry to obtain the adjustment coefficient for the specified date,
divide one of the demand result values on the date corresponding to the particular date among the demand result values of the specified dates by the obtained adjustment coefficient to acquire a second normalized demand result value,
predict a demand value on the target date based on the prediction model in which the input variables are assigned the second normalized demand result value on the date corresponding to the particular date and the demand result values on the dates corresponding to the non-particular date among the specified dates,
judge whether the target date is the particular date or not based on the calendar data, and if the target date is the particular date, send a request to the adjustment coefficient determination circuitry to obtain the adjustment coefficient for the target date, and
multiply the predicted demand value by the obtained adjustment coefficient to acquire an inversely-normalized predicted demand value if the target date corresponds to the particular date, the inversely-normalized predicted demand value being a prediction demand value of the particular date, wherein

the predicted demand value is a prediction demand value of the non-particular date if the target date corresponds to the non-particular date,
the predicted demand value is a predicted value of the electric power supply amount,
electric power supply on the target date to the supply area by the electric power company is controlled based on the predicted demand value on the target date, and
the adjustment coefficient calculation circuitry is configured to:
calculate a mean value of first ones of the demand result values and second ones of the demand result values for each of a plurality of particular dates,
the first ones of the demand result values being demand result values on dates corresponding to the non-particular date among one or more dates before the date corresponding to each of the particular dates based on the demand result data, and
the second ones of the demand result values being demand result values on dates corresponding to the non-particular date among one or more dates after the date corresponding to each of the particular dates; and

calculate the adjustment coefficient for each of the particular dates based on a ratio between one of the demand result values on the date corresponding to each of the particular dates and the mean value.

2. The prediction device according to claim 1 , wherein
the model learner circuitry is configured to generate the prediction model based on a weather predicted value on the date corresponding to the particular date and weather predicted values on dates corresponding to the non-particular date; and
the predictor circuitry is configured to predict the demand value on the target date based on the prediction model in which the input variables are further assigned weather predicted values on the dates before the target date.

3. The prediction device according to claim 1 , wherein
the demand result data includes demand result values at a plurality of times in the plurality of dates,
the model learner circuitry is configured to divide one of the demand results value at a first time on the date corresponding to the particular date by an adjustment coefficient associated with the first time of the particular date to acquire the first normalized demand result value,
the model learner circuitry is configured to generate a prediction model for predicting a demand value at the first time on the target date on a basis of the first normalized demand result value at the first time on the date corresponding to the particular date and the demand result values at the first time on the dates corresponding to the non-particular date,
the predictor circuitry is configured to predict a predicted demand value at the first time on the target date based on the prediction model, the demand result values at the first time on the dates before the target date, and the adjustment coefficient associated with the first time of the particular date corresponding to at least one of the dates before the target date, and
the predictor circuitry is configured to multiply the predicted demand value by the adjustment coefficient associated with the first time of the particular date to acquire an inversely-normalized predicted demand value if the target date correspond to the particular date.

4. The prediction device according to claim 1 , wherein the prediction model is a regression model.

5. The

Assignees

• Toshiba Kk
• Toshiba Energy Systems & Solutions Corp

Inventors

• Shin Hiromasa
• Shiga Yoshiaki
• Ichikawa Ryoichi
• Toyoshima Ichiro

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N20/00

  Machine learning · CPC title

• G06Q10/04 Primary

  Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" (market predictions or forecasting for commercial activities G06Q30/0202) · CPC title

• G06Q10/109

  Time management, e.g. calendars, reminders, meetings or time accounting · CPC title

• G06N3/08

  Learning methods · CPC title