JP2681421B2 - Demand forecasting method and device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention is designed to meet future demand for electric power.
The present invention relates to a power demand forecasting device for forecasting .

[0002]

2. Description of the Related Art It is known that demand for electric power and water varies greatly depending on factors such as differences in weather, types of days of the week, social events, and temperature (hereinafter collectively referred to as influential factors). Similarly, for certain products, it is known that the demand volume fluctuates due to these influencing factors. In order to supply electricity, water, products, etc. smoothly and without excess or deficiency according to the fluctuations in demand, forecast future demand, and operate electricity, water supply, product production and distribution according to the forecast. It is important to plan and manage such things.
For example, as a conventional power demand forecasting method, Japanese Patent Laid-Open No.
The one described in Japanese Patent Laid-Open No. 102822 has been proposed. According to this, one day is divided into a plurality of time zones, the actual value of the power demand in each time zone is collected in association with the mode of the influencing factors that are the fluctuation factors of the demand, and the collected data is collected. Is stored as the occurrence frequency in the corresponding element of the actual data table in which the relationship between the
Using the actual data table, the demand frequency vector is calculated from the elements that match the prediction target time zone and the mode of the prediction influencing factor in that time zone, and based on this, the maximum frequency in the time zone to be predicted The power demand is decided as a predicted value.

According to the one disclosed in Japanese Patent Laid-Open No. S60-106325, the power demand characteristic with respect to the daily maximum temperature is calculated by using the least squares method based on the past record of the maximum daily temperature and the amount of power demand. It is expressed by the following formula, and by substituting the predicted maximum temperature of the prediction target day into this relational expression, the power demand is predicted.

[0004]

However, according to the above prior art, there are the following problems to be solved. First, according to the technique disclosed in Japanese Patent Laid-Open No. 60-102822, when the actual data table is created and updated, the actual data is stored cumulatively, so that the influence of the past data pattern is relatively small. There is a problem of getting big. In particular, there is a tendency that the prediction accuracy around the sudden change of the weather or the specific day deteriorates.

Further, according to the technique disclosed in Japanese Patent Laid-Open No. S60-106325, the predicted maximum daily temperature of the predicted target day is simply substituted into the quadratic expression representing the relationship between the maximum daily temperature and the demand amount, and the predicted target day is calculated. There is a problem in that there is a limit to the accuracy of the prediction because it predicts the electric power demand of the above and does not consider the fluctuation of the electric power demand due to the weather change such as the past temperature.

[0006] Such a problem also applies to the demand forecast of water and products.

An object of the present invention, using the latest demand result data as possible, and power demand that can improve the prediction accuracy
It is to provide a quantity predicting device .

[0008]

In order to achieve the above-mentioned object, the power demand forecasting apparatus of the present invention is a power demand track record.
There are several impacts that the data will have
It is stored in the storage means of the computer
The database and the multiple influencing factors
Continuity impact that power demand continuously increases and decreases
Factors and discontinuity influencing factors that increase and decrease discontinuously,
Each predetermined discontinuity influence factor or discontinuity influence
Actual data for a certain recent period that corresponds to each combination of factors
Data sequentially read from the database and read
Discontinuity based on actual data
For each combination of continuity influencing factors,
First function deriving means for obtaining a first correlation function of the power demand amount
And about the area deviating from the read actual data
Pseudo data at the time of prediction based on past actual data
Means for creating a pseudo data, the pseudo data and
Factors affecting discontinuity based on the read actual data
And the continuity for each combination of discontinuity influencing factors.
The second correlation function of the influential factors and the power demand
The function deriving means and the first and second function deriving means
Storing the obtained first and second functions in the storage means
From the function table and the input means of the computer
Corresponding to the prediction influencing factors of the input prediction target time
Reading the first or second function from the function table,
Electric power demand at the time of the target of prediction using the read function
Prediction execution means for predicting
I did it.

[0009]

[0010]

Further, it is desirable that the new function data of the demand amount corresponding to the prediction result be added to the performance data of the certain period to update the function. In this case, the function is initialized when the period of the new actual result data related to the update reaches a predetermined additional period, and the actual result data including the new actual result data is used by using the latest actual result data of the certain period. It is desirable to find a new function.

A multiple regression model or a neural network model can be used as a means for obtaining the function. At that time, the convergence condition of the model is changed according to the reliability of the actual data to be used, or the convergence condition of the model when using the pseudo data is used rather than when using the actual data when obtaining the second function. It is desirable to set it gently.

Further, if the actual data used for deriving the function is data obtained by removing the base component that does not change from the actual actual data, the sensitivity of the function becomes high. Furthermore, after deriving the function, the actual data that deviates from the derived function by a certain amount is rejected, the function is re-derived using the remaining actual data, or the derived function and the actual data are compared. In order to improve the prediction accuracy, it is desirable to correct the function according to the degree of deviation of the actual result data.

[0014]

[0015]

[0016]

[0017]

With this structure, the present invention can achieve the object of the present invention by the following operations.

First, the correlation function between the continuity influencing factors and the demand amount is obtained for each discontinuity influencing factor or each combination thereof, and the demand amount at the prediction target time is predicted based on the function, so that the day of the week Since the function based on the actual data can be used to predict even a singular day when a difference or social event occurs, the prediction accuracy is improved. Also, since the correlation function for only the continuity influencing factors is obtained and predicted excluding the discontinuity influencing factors, the reliability of the function can be increased even if the number of actually measured data is small, so The actual data can be used, and the influence of past old actual results can be minimized. Further, if a multiple regression model or a neural network is applied to the function model, the target period of the actual data can be further shortened.

According to the method of generating pseudo data, the prediction accuracy is improved even when the actual data is small or even when the actual data corresponding to the continuity influencing factor at the prediction target time point does not exist.

If the function is updated by adding new performance data, the prediction accuracy is further improved. In particular, when the period of the new record data related to the update reaches a predetermined additional period, if the function is initialized and redistributed, the influence of the past record of the record data can be reduced.

As a means for obtaining the function, a multiple regression model or a neural net model is used, the convergence condition of the model is changed according to the reliability of the actual data used, or the convergence condition of the model is relaxed when pseudo data is used. By setting to, reasonable prediction can be realized.

[0022]

The present invention will be described below with reference to the illustrated embodiments. FIG. 1 is a block diagram of a power demand forecasting apparatus of one embodiment in which the present invention is applied to power demand forecasting. As shown in the figure, the actual data of the electric power demand is input to the database 2 from the actual data input means 1. The input performance data includes new performance data obtained by operation, in addition to the past performance data accumulated. The function deriving means (I) 3 appropriately reads the actual data in the database 2, derives a function used for prediction according to a preset procedure, and stores it in the function table (I) 4. The function updating means (I) 5 has basically the same configuration as the function deriving means (I) 3, except that new performance data collected according to the execution of the prediction is read out from the database 2 and new. The function in the function table is replaced with a new one by obtaining a function that takes into account such data.

On the other hand, the prediction execution command is input from the prediction command input means 6. The prediction command includes designation of a prediction target time point such as a prediction target date, prediction conditions such as weather, a day of the week, a temperature condition, and other influencing factors related to the prediction target time point. The prediction executing means 6 reads the function corresponding to the input prediction condition from the function table (I) 4,
The function is used to obtain the power demand at the time of the prediction target and output it to the prediction result output means 8 as a prediction value.

The above configuration is the basic configuration of the present invention, but in this embodiment, in order to further improve the accuracy of prediction, the pseudo data creating means 11, the function deriving means (II) 12, and the function table (II) 13 are provided. And function updating means (II) 14
Is added.

Next, the detailed configuration of each unit will be described.
The performance data input from the performance data input means 1 includes
There are several influential factors that affect the increase and decrease of electricity demand. Here, the influencing factors refer to various factors that influence the increase and decrease of the power demand, and an example in which these factors are classified into a homogeneous factor group is shown below. Further, the influential factor group is further divided into a continuous influential factor in which the change between the influential factor and the change in the demand amount are continuous, and a discontinuous influential factor in which the relationship is discontinuous.

(Examples of factors affecting discontinuity) -Type of weather (clear, rain, cloudy, snow, etc.)-Type of day of the week (weekdays, holidays, Saturday, medium day of stepping stone holidays, factory holidays, etc.)-One Time of day ・ Social events (sports events, etc.) (Examples of factors affecting continuity) ・ Maximum daily temperature ・ Minimum daily temperature ・ Others: sensible temperature, discomfort index, humidity, solar radiation amount In order to improve the accuracy, first, the cases are classified according to the discontinuity influencing factors or combinations thereof, and the correlation between the continuity influencing factors and the electric power demand is made into a function in the case classification and the prediction is performed. ing.

Therefore, the database 2 is constructed such that the actual data is stored in the form of a table as shown in the conceptual diagram of FIG. Note that, in the present embodiment, a case will be described below in which the power demand amount is predicted in consideration of the type of weather, the type of day of the week, the time, the maximum daily temperature, and the minimum daily temperature.

The function deriving means (I) 3 is adapted to obtain a correlation function between the continuity influencing factors and the demand amount for each discontinuity influencing factor or each combination of the discontinuity influencing factors, and is required. The actual performance data is extracted from the database 2. In this example, data is extracted for each type of weather and for each type of day of the week, and for each of these combinations, a function of the power demand amount and the daily maximum temperature and the minimum daily temperature with respect to the time of the prediction target day is calculated. Ask. It should be noted that the actual data to be extracted is limited to the data for a certain fixed period from the time of deriving the function.

Here, FIG. 3 shows an example of changes in the power demand amount during a day, and FIG. 4 shows an example of the relationship between the power demand amount and the daily maximum and minimum temperatures at certain times. In FIG. 3, the vertical axis represents power demand and the horizontal axis represents time. When predicting the power demand that changes in this way every hour, the function deriving means (I) 3 derives each time, that is, 24 functions. It is also possible to obtain a function for each year, day, minute, second, etc. On the other hand, as is clear from the relationship curve H between the daily maximum temperature and the power demand and the relationship curve L between the daily minimum temperature and the power demand shown in FIG. Some kind of non-linear relationship can be observed. In the figure, the vertical axis represents the power demand and the horizontal axis represents the temperature. The reason for this tendency is considered to be as follows. That is,
In the summer when the temperature rises, demand for cooling such as coolers tends to cause a rapid increase in power demand as the temperature rises. On the other hand, the electric power demand is almost constant during the period when the temperature does not change so much from early spring to early summer or from early autumn to early winter. Further, in winter when the temperature is low, the demand for heating increases, so that the demand for electricity tends to increase as the temperature decreases. A functional relationship is derived based on such a relationship between the daily maximum temperature and the minimum daily temperature, and is stored in the function table (I) 4 for each discontinuity influencing factor or each combination thereof.

A well-known multiple regression model or a model using a neural network can be applied as a calculation model related to the function derivation of the function derivation means (I) 3. Here, a specific example using a neural network will be described with reference to FIG. This is because, in general, when the object is non-linear, the use of the neural net in obtaining the functional relationship is great. As shown in FIG. 5, a neural net is a device in which elements 51 called neurons are coupled to each other by connecting lines 52a, 52b. . .
Is a model for simulating the operation learned by a human brain cell by connecting the input layer, one or more intermediate layers, and an output layer in a hierarchical manner. In this case, data for learning, for example, input influential factors such as daily maximum temperature and minimum daily temperature are α (° C) and β, respectively.
Data that the output power demand is γ (MW) at (° C) is called teacher data. Learning with a neural network means that when the temperature data, which is the teacher data, is input, the connecting lines 52a, 52b ,. . . Is to find the weighting factor for. There are various methods for determining the weight coefficient, and the back propagation method will be described below in the present embodiment, but the present invention is not limited to this.

As an example, a backpropagation method for a three-layer neural network will be described for simplicity. The output value from the input layer neuron is Xk, the output value of the intermediate layer neuron shell is Xj, and the output value from the output layer neuron is Xi. The output values Xi and Xj of the intermediate layer and the output layer are obtained by the following mathematical expression 1.

[0032]

## EQU1 ## Xi = f (ΣWij · Xj−θi) xj = f (ΣWjk · Xk−θj) where Wpq: coupling lines 52a, 52b ,. . . Weighting coefficient for p, q: i, j, k i, j, k: 1, 2, ... θp: threshold value at which neuron starts firing f: sigmoid function f (n) = (1 / (1 + exp ( -N)) Now, when a certain teacher data is input, it is assumed that the output Xi is different from the expected Xi _{0. The} end of learning is equivalent to the following inequality.

[0033]

Σ (Xi−Xi ₀ ) ² <ε where Σ: symbol representing the sum ε: preset error threshold Therefore, in order to satisfy Expression 2, Wij is set to the error on the left side of Expression 2. Inversely correct in proportion. The function deriving means (I) 3 is configured using such a neural network.

Now, with regard to the embodiment configured as described above, the more detailed structure and operation of the basic components will be described with reference to FIGS. 6 to 8. Also, the detailed configuration of the additional portion will be described together with the operation. Function deriving means (I) 3
The processing procedure of is as shown in FIG. First, in step 61, in order to obtain all the functions for the combinations of discontinuity influencing factors in the conditions to be predicted, the combinations are sequentially set / changed. In step 62, the function corresponding to the combination of the set / changed discontinuity affecting factors is initialized. That is, it makes a meaningless function. Next, step 6
At 3, the actual data corresponding to the set combination is extracted from the data table 2. For example, from the daily actual data in the recent fixed period, the actual data in which the weather is “fine” and the day of the week is “weekday” is read. Based on the read data, in step 64, the correlation function between the daily maximum temperature or the minimum daily temperature and the electric power demand for each time is derived. The function thus derived is stored in the function table (I) 4. Then, in step 65, it is judged whether or not the derivation of the functions related to all the combinations is completed, and if the derivation is not completed, the process returns to step 61 to change the combinations, and until the derivation of all the functions is completed, The process of is repeated.

Here, the process in step 64 will be described in detail. Now, the correlation function between the influencing factor and the power demand is expressed by Expression 3, and the function is derived by setting the weighting coefficient of the neural network so as to satisfy the condition of Expression 4.

[0036]

## EQU00003 ## P (i, h) = g (X1, X2, ..., Xn)

[0037]

## EQU4 ## | P (i, h) -P '(i, h) | <ε However, the meanings of the symbols are as follows.

P (i, h): Electric power demand (MW) at time h on day i P '(i, h): Output of the neural network at time h on day i g: Symbol indicating a functional relationship Xj: Influencing factors representing climate, social events, etc. (j = 1, 2, ...) ε: Threshold of output error To increase the reliability of the function derivation by the neural network, as is generally done. , It is desirable to repeat the learning with teacher data. For example, the learning is repeated 20 times for one combination of the daily maximum temperature and the demand amount, and the combination is further changed and repeated 100 times as a whole. This is because the influence of the previous learning result remains, and the influence is removed by repetition. In this way, the demand quantity prediction function relating to the factors that influence continuity can be obtained with high accuracy. The reason why the discontinuity influencing factor is not made into a function is that even if the weather is the same, for example, the tendency of the demand amount may be completely different depending on the type of day of the week, and the accuracy of the function cannot be improved. is there.

FIG. 7 shows the processing procedure of the prediction executing means 7. As shown, in step 71, the prediction condition of the prediction target is input from the prediction command means 6. As the prediction condition, for example, it is assumed that the day-to-day change in the power demand on what day is predicted, the expected weather of the day is fine, the weekday is weekday, the expected maximum daily temperature of the day and the minimum daily temperature are input. Next, at step 72, the initial value of the prediction target time is set. In the next step 73, as will be described later, it is determined whether the condition to be predicted is inside or outside the region where the actual data exists, that is, whether or not the condition is interpolation. In the case of interpolation, the process proceeds to step 74, the neural network of the corresponding function shown in Formula 3 is read from the function table 3, the prediction condition is input to the input layer, and the output is the power demand amount at that time. Use as a predicted value. Subsequently, in step 75, it is determined whether or not the prediction has been completed for all times, and the process returns to step 72 until the end, and the above-described prediction process is repeated. When all predictions have been completed, the routine proceeds to step 76, where the prediction result is output.

The function updating means (I) 5 updates the function in the function table (I) 4 by the procedure shown in FIG. 8 to reflect the new performance data. The actual data at the time corresponding to the above prediction result is stored in the database 2 via the input means 1.
Will be input at any time. The function updating means (I) 5 judges in step 81 of FIG. 8 whether or not new record data has been input. When the result is affirmative, in step 82, the new data is added to derive a corresponding function. , Function table (I) 4 is updated. That is, the weighting factor of the neural network is changed. The processing content of this step 82 is shown in FIG.
Since it corresponds to steps 62, 63, 64, 66, and 67 of, the details are omitted. Hereinafter, each time new performance data is input, the function is repeatedly updated, but it is judged whether or not the repetition period exceeds a preset period (learning addition period) (step 83), and the result is affirmative. If so, the process proceeds to step 84 to initialize the function relation, and newly finds the function of the influence factor of the initial learning period and the power demand amount by using the neural net at each time point, and renews the function. Reset it.

An example of the result of power demand forecasting according to the above procedure will be described below.・ Conditions affecting the continuity: maximum daily temperature, minimum daily temperature ・ Conditions affecting the discontinuity: weather-clear, day-weekday ・ Initial learning period: 1 month ・ Additional learning period: 1 month The results of the prediction are shown. In the neural network, ○ is the actual amount one month before the prediction execution, which is the teacher data, ● is the predicted amount obtained by this demand forecasting method, and * is the actual amount corresponding to ●. When the temperature, which is an influencing factor, is in the range of existence of the circle, the predicted amount is almost the same as the actual amount, and accurate prediction can be performed (interpolation case). However, the prediction amount in the other range has a very low degree of coincidence, resulting in poor accuracy (case of extrapolation).

Based on these facts, the accuracy of the demand forecast is evaluated only in the case of interpolation, and the case of extrapolation and the case of interpolation are grouped together.
Criteria for accuracy evaluation of demand forecast are roughly classified into two, and in the case of next-day forecast, there are the following two points.

Reducing the daily average error Reducing the variation in the daily average error Here, the daily average error is calculated by the accuracy evaluation formula of the following formula 5.

[0044]

(Equation 5)

However, abs: a symbol representing an absolute value Σ: a symbol representing a sum i = 1, 2 ,. . . , 24 The above formula calculates the relative error by dividing the absolute value of the difference between the demand forecast value and the demand actual value at each time point i by the actual value, and the total relative error at each time point is divided by the number of time points. By doing so, the error of the daily average is obtained.

Various modifications can be considered as a method of evaluating such an error. Further, in the present embodiment, the evaluation was performed using the relative error, but it is of course possible to use the absolute value of the predicted value and the demand value as the evaluation standard.

As an example of the error, a prediction result based on the past 1 year and 5 months of actual data is shown in FIG. As shown in the figure, the average value of the daily mean error is 2.18% when only the interpolation case is included, and 2.84% when the extrapolation case is included, including the extrapolation case. It can be observed that the case is less accurate than the case of interpolation alone.

Here, the additional portions 11 to 14 of the embodiment shown in FIG.
Will be described in detail. This added portion is added to improve the prediction accuracy, and is characterized in that the pseudo data creating means 11 is provided. The reason for providing the pseudo data creating means 11 will be described. It was mentioned above that it is not desirable to make a demand forecast under the extrapolation case in order to make an accurate demand forecast. Therefore, in order to reduce the extrapolation case, in order to change the influencing factor at the time of prediction, which originally corresponds to the extrapolation case, to the extrapolation case, the pseudo teacher data is used as the function deriving means (II) 1
I input it in 2. An example of the pseudo data will be described with reference to FIG. FIG. 11 is an example of a case where each daily power demand in the next month is calculated based on the actual data of a certain month in early spring. In the figure, the ellipse filled with black is the range of the actual data of a certain month, and the ellipse hatched is the range of the actual data of the next month which is the prediction target. The vertical axis shows the power demand and the horizontal axis shows the temperature. Since the actual data of the predicted execution month does not exist in the temperature range in which the actual data serving as the learning data exists, this is an extrapolation case when executing the prediction, and the accuracy of the prediction deteriorates as described above. Generally, in order to eliminate such an extrapolation case, there is a method of accumulating teacher data without clearing it. However, this method has the following two drawbacks.

1) It takes a lot of time to derive the functional relationship 2) The influence of past data is great, and the prediction accuracy is lowered. The data represented by the range of an open ellipse is added, and the method of reducing the case of extrapolation is introduced. In principle, pseudo data is created automatically using a computer. However, it is also possible to generate pseudo data by the operator manually inputting the know-how possessed by humans, if necessary.

Various methods are conceivable as a method of creating the pseudo data in the pseudo data creating means 11. In this embodiment, a typical example among them is shown below. Now, Tmin (i): minimum daily temperature of the same month last year of the forecast month Tmax (i): minimum daily temperature of the same month last year of the forecast month P (i, h): time of the same month last year of the forecast month It is the power demand in h.

First, from the data (Tmin (i), Tmax (i), P (i, h)) at the same time point one year ago,
(Tmin (i), Tmax (i), P ″ (i, h)) is obtained, where

[0052]

(Equation 6)

Pmean (h): average value of power demand at time h in the previous month of the forecast month P'mean (h): average value of power demand at time h in the same month of the previous year of the forecast month Create the data. In the above formula, Pmean
Instead of (h) / P'mean (h), it is also possible to use an annual load growth rate or the like. In addition, factors such as year-on-year changes in sales of electrical products, which are one of the influential factors, or factors such as social events can also be taken into consideration.

When the function deriving means (II) 12 derives a non-linear function using the above pseudo data,
Since the reliability of the pseudo data is not 100%, the method of deriving the function is changed according to the reliability of the data in this embodiment. Specifically, when a neural network is used as a model of a function, the weighting coefficient of the neural network is obtained under a gentle convergence condition, while when actual data exists, the weighting coefficient of the neural network is obtained under a severe convergence condition. And That is, a function derivation method is used in which the error between the output value of the neural network and the pseudo data output value is set higher than when the actual data is used. That is, the weight coefficient of the neural network is updated so that the condition of Expression 8 is satisfied with respect to the function type of Expression 7.

[0055]

P ″ (i, h) = g (X1, X2, ..., Xn)

[0056]

## EQU8 ## | P "(i, h) -P '(i, h) | <[epsilon]' However, the meanings of the symbols are as follows.

P ″ (i, h): Pseudo data of electric power demand at time h on day i P ′ (i, h): Output of neural network at time h on day i g: Symbol Xj indicating functional relationship : Influential factors representing climate, social events, etc. (j = 1, 2, ...) ε ': Output error threshold for pseudo data (ε <
ε ′) The pseudo data generated in this way is added to the actual data, and the function deriving means (II) 12 obtains the function II. The structure of the function deriving means (II) 12 is the same as that of the function deriving means (I) 3, and the description thereof will be omitted. Further, the processing procedure of pseudo data creation and function derivation is inserted in steps 66 and 67 of FIG. 6, and the function I and the function II are obtained in parallel.

In place of the processing procedure of FIG. 6, the judgment result (step 73) of whether or not the condition of the prediction object in the prediction executing means (I) 5 is out of the range of the actual result data which is the basis of the function I, The pseudo data may be output to the pseudo data creating means 11, the pseudo data may be created in response to this, and the function deriving means (II) 12 may be activated to derive the function II.

In the embodiments of FIGS. 1, 6 and 7, the processing of the function deriving means (I) 3 and 12 and the prediction executing means 7 are not directly related to each other, and the processing is independently executed. Explained. However, the present invention is not limited to this, and the prediction command and the condition input from the prediction command input means 6 are input to the function deriving means (I) 3 and 12, and only the actual data that matches the prediction condition is input. It is possible to extract the data from the database 2 and to generate pseudo data as needed, and to derive only the function that responds to the input of the prediction command and is related to the prediction target. According to this, the derivation of unnecessary functions can be omitted.

A modification of the above embodiment will be described below. An embodiment for further improving the prediction accuracy will be described with reference to FIG. In this example, as shown in FIG. 12 (a), the daily change in the power demand amount is based on a certain minimum demand power amount B and does not become lower than that.
The demand amount after subtracting the base removal amount B (Fig. 12
By making (b)) the prediction target, the prediction sensitivity is increased, and thereby the prediction accuracy is improved. That is, the above-mentioned prediction performance data (Tmin (i), Tmax (i), P (i, h))
(Tmin (i), Tmax (i), P (i, h)
-B), and the function deriving means 4 and 12 are activated. Here, Tmin (i): minimum daily temperature of i-th teacher data Tmax (i): maximum daily temperature of i-th teacher data P (i, h): power demand at time h of the i-th teacher data B: Base Removal Part Also, with reference to FIGS. 13 and 14, an example will be described in which the accuracy of prediction is improved by modifying the function derived in the above example. FIG. 13 expresses the concept of function correction in a two-dimensional plane for convenience. As shown in FIGS. 13 and 14, in the function deriving means (I) 3 and 12, based on the result of deriving the function, the derived function and the actual data are compared with each other. The data is detected based on preset criteria (steps 101 and 111), and the inappropriate data is rejected to derive the function again using only the appropriate data (steps 102 and 112). ).
This method can be applied to the derivation of both functions I and II. As a specific method of discarding data, it is possible to substantially reject it by gently adjusting the threshold value of the output error of the neural network at the time of re-deriving the function.

An embodiment in which the function is corrected in order to further improve the prediction accuracy will be described with reference to FIGS. FIG. 15 illustrates the concept of this embodiment, in which an error E defined next is used as a correction value to correct the derived function (steps 103 and 104 in FIG. 16).

[0062]

(Equation 9)

Here, g (X1, ... Xn): derived function X1 ,. . . , Xn: Influencing factors Yi: Electric power demand m: Number of actual data n: Number of influencing factors By the correction value E obtained by the above equation 9, as shown in FIG.
The function g before correction is translated by E and the function g ′ after correction is obtained.
Ask for. This can improve the prediction accuracy.

Although not considered in the above embodiment,
When making a demand forecast, a singular day, for example
In the forecast such as the first weekday after Obon, there is often a large error between the predicted value and the actual value. In order to deal with this point, the error between the predicted value and the actual value in the past singular day is stored in the database 2 or input through the actual data input means 1, and the data is used to derive the function deriving means (I) 3, By correcting the functions I and II obtained in 12, accurate prediction can be performed.

Although the present invention has been described based on the embodiment in which it is applied to the prediction of electric power demand, as described above, the present invention is not limited to electric power, and the demand for water supply, the season such as bread, and the like. It is obvious that it can be applied to the target where the demand amount changes due to the influence factors such as temperature.

[0066]

As described above, according to the present invention,
It has the following effects. For each discontinuity influencing factor or each combination thereof, obtain the correlation function between the continuity influencing factor and the demand amount,
Since the demand amount at the time of the prediction target is predicted based on the function, it is possible to predict a specific day when a day of the week is different, a social event or the like is performed by the function based on the actual data, so that the prediction accuracy is improved.

Further, since the correlation function for the continuity influencing factors is obtained and predicted, the reliability of the function is high even if the number of actually measured data is small, so that it is possible to use the actual result data for the recent fixed period. The influence of past past achievements can be minimized. In particular, if a multiple regression model or a neural network is applied to the function model, the target period of the actual data can be further shortened.

Further, according to the method of generating the pseudo data, the prediction accuracy is improved even when the actual data is small or the actual data corresponding to the continuity influencing factor at the prediction target time point does not exist.

Further, according to the one in which the function is updated by adding new performance data, the prediction accuracy is further improved. In particular, when the period of the new record data related to the update reaches a predetermined additional period, if the function is initialized and redistributed, the influence of the past record of the record data can be reduced.

As a means for obtaining the function, a multiple regression model or a neural network model is used, the convergence condition of the model is changed according to the reliability of the actual data, or the convergence condition of the model when using pseudo data is loosened. According to what is set, a reasonable prediction can be realized.

Further, the actual value obtained by removing the base removed portion from the actual data is used, the actual data that greatly deviates from the derived function is rejected, and the function is again derived, or the degree of deviation between the derived function and the actual data According to the one in which the function is corrected according to, the prediction accuracy can be further improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a demand forecasting apparatus according to an embodiment of the present invention.

FIG. 2 is a table configuration diagram of an embodiment of a database.

FIG. 3 is a diagram illustrating an example of a daily change in power demand.

FIG. 4 is a diagram showing an example of a correlation between air temperature and power demand.

FIG. 5 is a configuration diagram of a neural network as an example of a function derivation model.

FIG. 6 is a flowchart of a processing procedure of a function deriving unit.

FIG. 7 is a flowchart of a processing procedure of a prediction executing unit.

FIG. 8 is a flowchart of a processing procedure of a function updating unit.

FIG. 9 is a diagram for comparing and explaining the prediction results in the case of the interpolation case and the case of the extrapolation case of the range of the actual data of the function derived by the influence factor of the prediction target.

FIG. 10 is a diagram comparing errors in prediction results of the interpolation case and the extrapolation case.

FIG. 11 is a diagram illustrating a method of creating pseudo data.

FIG. 12 is an explanatory diagram of an example in which a function is derived by using a value obtained by subtracting the base removal amount from the actual demand amount as a demand actual value, (a) is a demand actual value including the base removal amount, (b)
Is a diagram of actual demand value after subtracting the base removal amount.

FIG. 13 is a diagram for explaining an example in which performance data that deviates from the derived function is rejected.

FIG. 14 is a flowchart of a main part of the processing procedure of the embodiment in which the actual data deviating from the derived function is rejected.

FIG. 15 is a diagram for explaining an example in which a function is corrected according to the degree of performance data that deviates from the derived function.

FIG. 16 is a main part flowchart of a processing procedure of an embodiment in which a function is corrected according to the degree of actual data that deviates from the derived function.

[Explanation of symbols]

 1 Actual Data Input Means 2 Database 3 Function Derivation Means (I) 4 Function Table (I) 5 Function Update Means (I) 6 Prediction Command Input Means 7 Prediction Execution Means 8 Prediction Result Output Means 11 Pseudo Data Creation Means 12 Function Derivation Means (II) 13 function table (II) 14 function updating means (II)

 ─────────────────────────────────────────────────── --- Continuation of the front page (72) Inventor Shigeru Tamura 4026 Kuji-machi, Hitachi City, Ibaraki Prefecture Hitachi Research Laboratory, Hitachi Ltd. (56) References Japanese Patent Publication Sho 60-53904 (JP, B2)

Claims (6)

(57) [Claims] 1. A performance data of the power demand is denoted by the plurality of influencing factors affecting the increase or decrease of the power demand said competent
A database stored in the storage means of the computer, and the plurality of influencing factors continuously increase or decrease the power demand amount with respect to changes in the influencing factors Separating the discontinuous impact factors, sequentially reads the performance data of the recent predetermined period corresponding to each combination of pre-per-determined discontinuities influence factors or discontinuities influence factors from said database, said read-out performance for each combination of or discontinuities influencing factors for each discontinuity influence factors on the basis of the data, the first function deriving means for obtaining a first correlation function of the continuity effect factor and the power demand, said read the basis of the random data generating means for generating a pseudo data of the prediction target time based on historical data for the area out of the actual data, to the actual data read said the該疑similar data For each combination of or discontinuities influencing factors for each discontinuity contributors, a second function deriving means for obtaining the second correlation function of the continuity effect factor and the power demand, the first and second a function table in which the first and obtained comprising storing the second function in the memory means by the function deriving unit, the first that corresponds to the predicted impact factors of the prediction target time input from the input means of the computer Alternatively, a power demand forecasting apparatus that uses a computer to form a second function that reads out the second function from the function table and uses the read-out function to forecast the power demand at the time of the prediction target. 2. A method according to claim 1, wherein the first and second function <br/> number deriving means, prediction performance data of the predetermined period of new performance data of the power demand corresponding to the power demand power demand prediction apparatus and updates the function in addition to. 3. The method according to claim 2 , wherein the first and second function derivation means initialize the function when a period of new record data related to the update reaches a predetermined additional period. The power demand forecasting apparatus is characterized in that the function is newly obtained by using the actual result data including the new actual result data for the latest fixed period. 4. The method according to claim 1, wherein
The first and second function deriving means are used for deriving the function.
As the actual performance data,
Power demand forecast characterized by using the removed data
Measuring device. 5. The method according to any one of claims 1 to 3, wherein
The first and second function deriving means derives the function.
Then, reject the actual data that deviates from the derived function by a certain amount.
And re-derive the function using the remaining actual data.
A characteristic power demand forecasting device. 6. The method according to claim 1, wherein
The first and second function deriving means derives the function.
Then, compare the derived function with the actual data,
The function is corrected according to the degree of deviation of the actual product data of
A power demand forecasting device, characterized in that