JPH0518995A - Device for forecasting total power demand - Google Patents

Abstract

PURPOSE:To forecast very accurately the total electric power demand. CONSTITUTION:A data take-in means A for taking in and store the meteorological variables including the past air temperature and humidity, and the data of the total electric power demand; and a statistically forecasting means B for preparing a statistical forecast model about values of the meteorological variables and the total electric power demand on the basis of the data of the meteorological variables and the total electric power demand stored in the data take-in means A, and for forecasting the total electric power demand on the day of forecast using this statistical forecast model are provided. Further, a corrected power estimating means C for computing the error of the total electric power demand, forecasted by the statistical forecasting means B, using the values of meteorological variables and the calendar day information till the day of forecast to learn the forecast error about the past data by a neural network, and for correcting the error of the statistical model forecast on the basis of it to estimate the total electric power demand is provided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a total electric power demand predicting apparatus for predicting total electric power demand from the statistical properties of meteorological variables and total electric power demand.

[0002]

2. Description of the Related Art In power system operation, it is indispensable to predict the daily total power demand in order to maintain supply reliability and improve operation efficiency. As a conventional prediction method, a method using time series analysis of total power demand and weather variables,
There are methods such as using correlation with economic indicators. In particular, with the recent increase in temperature sensitivity of total electric power demand due to the spread of air conditioning equipment, the multiple regression method of past meteorological variables and total electric power demand has become popular as a relatively high-precision method. There is.

[0003]

In all of the above-mentioned conventional forecasting methods, although a good forecast can be made during a period in which the daily total power demand is small, it is possible to obtain a daily forecast When it is applied to a period such as early summer or early winter when there is a large number of problems, there are the following problems.

[0004] Although the total power demand is roughly correlated with the weather variables, the relationship is not necessarily linear. Taking air temperature as an example, if the air temperature rises, the power demand due to the use of the air conditioning equipment increases, and conversely, if the air temperature becomes too low, the power demand due to the use of the air conditioning equipment increases. This is also clear from the relationship between the temperature and the total power demand (maximum value) shown in FIG. FIG. 4 is a diagram showing the relationship between the average maximum temperature (degree) and the maximum power demand (MW) in one year. As is clear from this, the relationship between the temperature and the total power demand is moderate V. It has a glyph. Therefore, when the temperature changes drastically, a large error occurs when the linear regression based on the temperature is simply applied.

Moreover, when the base of electric power demand itself such as early summer increases daily, the regression forecast is based on past data, and therefore tends to be lower than the actual result. Therefore, even if a non-linear regression that approximates the correlation diagram between the temperature and the electric power demand with a curve is performed, the prediction error is unavoidable due to the influence of the increase of the electric power demand base. Further, since the increase amount of the power demand base varies from day to day, a simple detrending operation such as a difference cannot escape the influence of the increase of the power demand base.

According to the present invention, the error tendency of the statistical model prediction is estimated / corrected in advance to accurately predict the total power demand even in the situation where an error occurs in the prediction based on the statistical model such as regression. It is an object of the present invention to provide a total power demand forecasting device capable of performing the above.

[0007]

In order to achieve the above object, the present invention is configured as follows. That is,
The invention corresponding to claim 1 is based on the past meteorological data and the data fetching means for fetching and storing the data of the total power demand, and the meteorological data and the data of the total power demand stored by the data fetching means. , A statistical model for the value of meteorological data and total power demand, and using this statistical model, a statistical forecasting means for forecasting the total power demand on the forecast day and a forecast total power demand for the forecast day predicted by this statistical forecasting means Quantity and the difference calculation means for calculating the difference between the actual total power demand up to the forecast day, and the predicted difference amount calculated by this difference calculation means and the value of the meteorological data up to the forecast day and the total power demand / calendar A correction amount estimating means for estimating the total power demand by correcting the difference based on the past data with respect to the relationship with the day information by the statistical model is provided.

Further, the invention according to claim 2 is the data fetching means for fetching and storing the past meteorological data and the data of the total power demand, and the meteorological data and the total power demand stored by the data fetching means. A statistical model for the value of meteorological data and total electric power demand is created based on the data of, and using this statistical model, a statistical prediction means for predicting the total electric power demand on the day of the forecast and the forecast predicted by this statistical prediction means Difference calculating means for calculating the difference between the total electric power demand on the day and the actual total electric power demand up to the forecast day, and the predicted difference amount calculated by this difference calculating means and the value of the meteorological data up to the forecast day and the electric power The relationship between the total demand and calendar date information is learned by using a neural network for past data, and based on that, the difference due to the statistical model prediction is corrected and the total power demand is corrected. It is provided with a correction amount estimating means for estimating a.

[0009]

According to the present invention, a statistical model relating to the value of the meteorological data and the total electric power demand is created based on the past meteorological data and the total electric power demand data, and the statistical model is used to calculate the electric power for the forecast day. The total amount of demand is predicted, and the difference amount between the predicted total power demand amount on the forecast day and the actual total power demand amount up to the forecast date is calculated, and the difference amount is calculated as the first or second correction amount. The total electric power demand is estimated by making corrections by any of the estimation means. That is, the first correction amount estimating means corrects the difference between the value of the meteorological data and the total electric power demand / calendar day information up to the forecast day, based on the past data, by the difference by the statistical model prediction. Estimate the total power demand. The second correction amount estimating means learns the relationship between the value of the meteorological data up to the forecast date and the total power demand / calendar day information by using a neural network for past data, and based on that, the statistics are obtained. The total power demand is estimated by correcting the difference due to model prediction. Therefore, according to the present invention, the error tendency of the prediction by the statistical model is estimated and corrected even in the situation where an error occurs in the prediction only by the statistical model, and thus the highly accurate prediction of the total power demand is performed. be able to.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of the present invention, which is composed of a prediction device that predicts the actual total power demand and a correction device that corrects statistical prediction. , A correction amount estimating means C and a predicted value storing means D.

FIG. 2 is a block diagram showing a specific embodiment of FIG. The apparatus of this embodiment has two operation modes, one is a mode in which actual prediction is performed, and the other is a mode in which a corrected portion of regression prediction is created.

The meteorological data / electric power demand data storage unit (hereinafter simply referred to as a data storage unit) 1 stores sampled average temperature / humidity data up to the forecast day and power demand amount data up to the forecast day before. is there. Specifically, as shown in D1, 1 of the data creation date (the date when data is fetched)
Electric power demand J (t-1), ..., J (t-11) from one day ago to the previous day
Then, as shown in D2, the average temperature x (t-10) and humidity x2 (t), ..., x2 from 10 days before the data creation date to the forecast day
(t-10) is stored. The data fetching unit 2 is able to fetch the period data required for prediction from the data stored in the storage unit 1.

The data classification unit 3 classifies the data captured by the data capturing unit 2 into the following four types. That is, 1) a data string of demand, average temperature, humidity, etc. from 10 days before the data creation date, 2) average temperature, humidity data of the data creation date, 3) demand amount of the data creation date,
4) Demand amount data string from 11 days before to the day before the data creation date for obtaining the comparative demand difference.

Here, the average temperature, humidity, and demand data from 10 days before the data creation date to the previous day are used to set the coefficient of the regression equation in relation to the power demand amount from the data string of 1) described above. . The average temperature and humidity data on the data creation date shown in 2) are used for regression prediction of the data creation date. In order to verify the forecast demand by regression forecast on the data creation date, 11 of the data creation date shown in 3)
Power demand data from the day before to the day before, and average temperature data from 10 days before the forecast day to the forecast day are used. 4)
The demand data from 11 days before the data creation date to the previous day, and the average temperature data from 10 days before the data creation date to the forecast day are used to create the correction amount calculation data. The storage unit 1, the data acquisition unit 2 described above,
The data classification unit 3 constitutes data acquisition means A.

The regression equation calculation unit 4 performs a process of setting a coefficient of a regression equation that describes the relationship between the average temperature and humidity and the power demand. Power demand J from average temperature x1 and humidity x2
The regression equation for obtaining is J = a * x1 + b * x2 + c (1), and the regression coefficients a, b, c are calculated. The regression coefficient storage unit 5 stores the regression coefficients a, b, calculated by the regression equation calculation unit 4,
Store c.

The regression prediction unit 6 calculates the average temperature x1 on the day of prediction.
From (t), humidity data x2 (t), and the regression coefficient values a, b, and c stored in the regression coefficient storage unit 5, the regression forecast value J ₁ (t) of the power demand is calculated by the equation (2). calculate. J ₁ (t) = a · x1 (t) + b · x2 (t) + c (2) The regression prediction value storage unit 7 stores the regression prediction value J ₁ (t) calculated by the regression prediction unit 6. . The regression equation calculation unit 4, the regression coefficient storage unit 5, the regression prediction unit 6, and the regression prediction value storage unit 7 described above constitute a statistical prediction unit B that predicts the demand amount by regression calculation. The correction amount calculation data creation unit 8
Δt1, Δt required to calculate the correction amount for regression prediction
Calculate 2, ΔJ1 and ΔJ2 by the following formulas. Δt1 (t) = x1 (t) -x1 (t-1) (3) Δt2 (t) = x1 (t)-{Σ (i = 1,10) x1 (ti)} / 10 (4) ΔJ1 (t) = J (t-1) -J (t-2) (5) ΔJ2 (t) = J (t-1)-{Σ (i = 2,11) J (ti)} / 10 … (6)

Here, Δt1 is the difference between the average temperature of the data creation date and the average temperature of the day before the data creation date, and Δt2 is the average temperature of the data creation date and the average temperature of the past 10 days from the data creation date. The difference, ΔJ1 is the ratio of the total power demand on the data creation date and the power demand on the day before the data creation date, and ΔJ2 is the total power demand on the data creation date and the past 10 days from the data creation date.
It is the ratio to the average value of the total daily electric power demand.

In the actual prediction mode, the correction amount calculation / storage unit 9 uses the correction function F described later from the data Δt1, Δt2, ΔJ1, and ΔJ2 created by the correction amount calculation data creation unit 8. (7) Equation E
To calculate. E = F (Δt1, Δt2, ΔJ1, ΔJ2) (7)

This correction amount function F is a function that outputs the correction amount E of regression prediction when the data Δt1, Δt2, ΔJ1, ΔJ2 created by the correction amount calculation data creation unit 8 are input, and the correction part of the prediction In a mode for creating the above, a neural network described later is created from the past data prior to the prediction.

The regression prediction error rate calculation unit 10 operates only in the mode in which the correction portion of the prediction is created, and from the regression prediction value J ₁ (t) on the data creation date and the actual demand amount J (t),
The error rate E (t) on the regression prediction side is calculated from the equation (8). E (t) = {J ₁ (t) -J (t)} / J (t) (8) Here, the calculated error rate E (t) is Δt 1, Δt 2, ΔJ 1 , ΔJ2, and used to create a correction function F for regression prediction.

The predicted value correction unit 11 calculates the regression predicted value J ₁ (t)
Is corrected using the correction amount E of the regression prediction calculated in the correction amount calculation / storage unit 9 and the equation (9), and the final predicted value J
Find ₂ (t). J ₂ (t) = J ₁ (t) / (1 + E) (9) The correction amount calculation data creation section 8 described above, the correction amount calculation /
Storage unit 9, regression prediction error rate calculation unit 10, prediction value correction unit 1
1 constitutes the correction amount estimating means C. The predicted value storage means D, that is, the predicted value storage unit 12 stores the final predicted value J ₂ (t) obtained by the predicted value correction unit 11.

In addition to the above-mentioned configuration, the device of this embodiment is
It is configured so that two operation modes can be switched, that is, one mode in which actual prediction is performed and the other mode in which a correction portion of regression prediction is created. That is,
Mode setter M, data period setter T, OR circuit O
1, O2, AND circuits A1, A2, inverters I1, I
2. The changeover switches CH1 and CH2 are provided. The mode setter M is capable of selecting the mode in which the demand prediction is performed as one of the modes in which the correction amount for the statistical prediction result is controlled. In the mode in which the demand prediction is performed, "H", that is, "1" is set. In the mode of outputting a signal and controlling the correction amount for the statistical prediction result, "L", that is,
A signal of "0" is output. When the signal from the mode setter M is "H", the logical condition of the AND circuit A1 is satisfied and the data stored in the storage unit 1 is input to the data fetching unit 2 via the OR circuit O2. . When the signal from the mode setter M is "L", the output of the inverter I1 becomes "H", the logical condition of the AND circuit A2 is satisfied, and the timing signal from the data period setter T fetches the data. Input to part 2.

The data period setter T takes in the signals Δt1, Δt2, ΔJ1 and ΔJ2 from the correction amount calculation data creating section 8 and sets the creation period of the correction function. Set from the day before to 3 months before. Each of the changeover switches CH1 and CH2 is switched to the L side when the signal from the mode setter M is “L”, and the output signal of the regression prediction value storage unit 7 is the regression prediction error rate calculation unit 10
Is input, and the output signal from the correction amount calculation / storage unit 9 is input to the OR circuit O1. Changeover switch CH
Both 1 and CH2 are switched to the H side when the signal from the mode setter M is "H", the output signal of the regression prediction value storage unit 7 is input to the prediction value correction unit 11, and the correction amount is set. The output signal from the calculation / storage unit 9 is input to the predicted value correction unit 11.

When the changeover switch CH1 is on the "L" side, the comparator C1 compares the power demand J (t) from the data classification unit 3 on the date of data creation with the output signal of the regression prediction value storage unit 7, This comparison result is input to the regression prediction error rate calculation unit 10. In the comparator C2, the changeover switch CH1 is "L".
On the side, the output signal from the correction amount calculation / storage unit 9 and the output signal from the regression prediction error rate calculation unit 10 are compared, and the comparison result is sent to the data acquisition unit 2 via the inverter I2.
Entered in.

Here, a method of learning the structure of the correction amount calculation / storage unit 9 by a neural network will be described with reference to FIG. In FIG. 3, the input layer has units corresponding to each information item related to meteorological variables, total power demand, calendar date, etc. in the data retrieved from the correction amount calculation / storage unit 9 in FIG. Has a link with each unit in the middle layer. Each unit in the middle layer has a link with a unit in the output layer. The unit of the output layer corresponds to the error of the regression prediction calculated by the regression prediction error rate calculation unit 10 in the data extracted from the correction amount calculation / storage unit 9.

The neural network adjusts the weighting of the links between the layers to calculate the correction amount / storage unit 9
Learn the input / output relationship of the data retrieved from. Figure 3
In the above example, the fluctuation amount of the temperature and the fluctuation rate of the past power demand are given to the input layer as the input variables, and the error rate of the regression prediction due to the meteorological variables is given to the output layer as the output variable. I am learning. Specifically, the difference Δt1 between the average temperature of the forecast day and the average temperature of the day before the forecast, the difference Δt2 of the average temperature of the forecast day and the average temperature of the past 10 days from the forecast day, and the power of the day before the forecast. Ratio of total power demand to total power demand two days before forecast ΔJ1, Ratio of total power demand to one day before forecast and average value of total power demand for the past 10 days from the forecast day
J2 is given as an input variable. As a result, NeuralNet has a tendency to predict regression errors such as "regression forecasts tend to be negative when the growth in total electricity demand is large," or "regression forecasts tend to be negative when temperature fluctuations are large." learn. In the example of FIG. 3, the learning function is realized by using the learning algorithm of the back propagation algorithm.

Next, the operation of the total electric power demand predicting apparatus configured as described above will be described. When the apparatus of this embodiment is operated in the mode for predicting demand, the signal from the mode setting unit M is set to be "H".
Then, the changeover switches CH1 and CH2 are both switched to the H side, and the meteorological variables before the forecast day and the past total power demand are fetched from the storage unit 1 to the data fetching unit 2. These data are sorted by the data sorting unit 3 into 1), 2), 3) and 4) of FIG. Of the classified data, 1) is sent to the regression equation calculation unit 4, where the regression coefficient a,
b and c are calculated, and the result is stored in the regression coefficient storage unit 5. In addition, 2) of the classified data is sent to the regression prediction unit 6, and here, the values of the regression coefficients a, b, and c stored in the regression coefficient storage unit 5 are expressed by the equation (2). By substituting into, the regression prediction value of the power demand is calculated, and the calculation result is stored in the regression prediction value storage unit 7.

On the other hand, 4) of the data sorted by the data sorting unit 3 is sent to the correction amount calculation data creating unit 8 where the equations (3), (4), (5) and (6) are used. Thus, Δt1 (t), Δt2 (t), ΔJ1 (t) and ΔJ2 (t) are calculated. Then, in the correction amount calculation / storage unit 9, Δt1 (t) and Δt2 calculated by the correction amount calculation data creation unit 8 are calculated.
The correction amount E for regression prediction is calculated from (t), ΔJ1 (t), and ΔJ2 by the formula (7) using the correction function F, and the calculation result is stored. The predicted value correction unit 11 obtains the final predicted value J ₂ (t) from the regression predicted value J ₁ (t) stored in the regression predicted value storage unit 7 by the equation (9), and the predicted value storage unit Stored in 12.

Next, when the apparatus of this embodiment is operated in the mode for controlling the correction amount for the statistical prediction result, the signal from the mode setter M is set to "L".

Then, the changeover switches CH1 and CH2 are
Both switch to the "L" side. Regression prediction error rate calculation unit 10
In, the regression prediction value J ₁ (t) stored in the regression prediction value storage unit 7 and the data 3) sorted by the data sorting unit, that is, the actual power demand J (t) on the data creation date are input. Then, the error rate E (t) of the regression prediction is calculated by the equation (8). The error rate E (t) output from the regression prediction error rate calculation unit 10 and the output Δt1 (t), Δt2 (t), ΔJ1 (t), ΔJ2 (t) from the correction amount calculation data creation unit 8 Based on the above, the predicted value correcting unit 11 corrects the predicted value by regression prediction.

In the embodiment described above, the neural network by the correction amount calculating / storing unit 9 is used, but an expert system is used in this part, and for example, the difference between the average temperature of the past month and the average temperature of the present month is calculated. When the value exceeds the predetermined value, the total power demand forecast value may be corrected by some percentage. Further, although backpropagation is used as the learning method in the embodiment, the invention is not limited to this, and a vector quantization method, a competitive learning method, or a method in which the vector quantization method and backpropagation are combined may be used. Furthermore, in the embodiment, the regression model of the meteorological variable and the total power demand is given as an example of the statistical model used for the prediction.
The method is not limited to this, and other methods such as an autoregressive model may be used.

[0032]

According to the present invention, even in a situation where an error occurs in prediction based on a statistical model such as regression,
It is possible to provide a total power demand prediction device that can estimate and correct the error tendency of the statistical model prediction in advance and accurately predict the total power demand.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of the present invention.

FIG. 2 is a block diagram showing an embodiment of the present invention.

FIG. 3 is a diagram for explaining a neural network forming the correction amount calculation / storage unit of FIG.

FIG. 4 is a diagram showing a relationship between average temperature and total power demand.

[Explanation of symbols]

1 ... Meteorological data / electric power demand data storage unit, 2 ... Data capturing unit, 3 ... Data classification unit, 4 ... Regression formula calculation unit, 5 ...
Regression coefficient storage unit, 6 ... Regression prediction unit, 7 ... Regression predicted value storage unit, 8 ... Correction amount calculation data creation unit, 9 ... Correction amount calculation
Storage unit, 10 ... Regression prediction error rate calculation unit, 11 ... Prediction value correction unit, 12 ... Prediction value storage unit, M ... Mode setting device, T ... Data period setting device.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Keiichi Nakamura             1-3-1, Uchisaiwaicho, Chiyoda-ku, Tokyo East             Inside Kyoden Electric Co., Ltd. (72) Inventor Masaya Ono             70 Yanagicho, Saiwai-ku, Kawasaki City, Kanagawa Prefecture             Toshiba Yanagimachi Factory

Claims (2)

[Claims] 1. Data acquisition means for acquiring and storing past meteorological data and total electric power demand data, and meteorological data based on the meteorological data and total electric power demand data stored by the data acquisition means. Create a statistical model for values and total power demand, and use this statistical model to predict the total power demand on the forecast day, and the total power demand on the forecast day and the forecast date predicted by this statistical forecast unit. Difference calculation means for calculating the difference amount of the actual total electric power demand up to and the predicted difference amount calculated by this difference calculation means, the value of the meteorological data up to the forecast day, and the total electric power demand / calendar day information. A total electric power demand forecasting device comprising: a correction amount estimating means for estimating the total electric power demand by correcting the difference by the statistical model based on past data. 2. Meteorological data based on the meteorological data and the total power demand data stored by the data capturing means, which stores and stores past meteorological data and total power demand data. Create a statistical model for values and total power demand, and use this statistical model to predict the total power demand on the forecast day, and the total power demand on the forecast day and the forecast date predicted by this statistical forecast unit. Difference calculation means for calculating the difference amount of the actual total electric power demand up to and the predicted difference amount calculated by this difference calculation means, the value of the meteorological data up to the forecast day, and the total electric power demand / calendar day information. A correction amount estimation unit that learns the relationship with respect to past data by a neural net and corrects the difference based on the statistical model prediction based on it to estimate the total power demand, An apparatus for predicting total power demand, which comprises: