JPH04372046A - Method and device for predicting demand amount - Google Patents

Abstract

PURPOSE:To improve the predicted accuracy of the demand. CONSTITUTION:The past recorded data of the demand amount with the influence factor is stored in a data base 2. On the basis of past record data extracted from the data base 2, a collation function I between the consecutive influence factor and the demand amount is obtained for each inconsecutive influence factor or for each combination. On the area other than the past record data, pseudo data is prepared by a pseudo data preparation means 11, and a function II with the pseudo data added is derived from a function derivation means 12 to be stored in function tables 4 and 13. A prediction execution means 7 reads out the function I or II corresponding to the predicted influence factor at the estimated objective point from a function table 4 by the prediction execution mean 7, predicting the amount of demand at the predicated objective point using the function. The addition of new past record data is desired for updating the functions I and II. In this case, when the updating is performed during the prescribed period, a function is desired to be newly obtained by initializing the functions I and II.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for predicting future demand for electricity, water, commercial products, etc.

0002

BACKGROUND OF THE INVENTION It is known that the amount of electricity and water demanded varies greatly depending on factors (hereinafter collectively referred to as influencing factors) such as differences in weather, types of days of the week, social events, and temperature. Similarly, it is known that the demand for certain products fluctuates depending on these influencing factors. In order to smoothly supply electricity, water, products, etc. in accordance with fluctuations in demand, we forecast future demand and adjust power operations, water supply operations, product production and distribution according to the forecast. It is important to plan and manage such matters.
For example, as a conventional power demand forecasting method,
The one described in Publication No. 0-102822 has been proposed. According to this, a day is divided into multiple time periods, the actual value of electricity demand in each time period is collected in association with the aspects of influencing factors that are fluctuation factors of demand, and the collected data is used to calculate demand The relationship between the relationship and the influencing factor is stored as the frequency of occurrence in the corresponding element of the actual data table.
Using the actual data table, find the demand occurrence frequency vector from the time period to be predicted and the elements that match the aspect of the prediction influence factor in that time period, and based on this, calculate the maximum frequency in the time period to be predicted. The power demand amount is determined as a predicted value.

[0003] Furthermore, according to what is described in Japanese Patent Application Laid-Open No. 106325/1983, based on the past relationship between daily maximum temperature and power demand, the power demand characteristics with respect to daily maximum temperature are calculated using the least squares method. By expressing the following equation and substituting the expected maximum temperature on the prediction target day into this relational expression, the power demand on the prediction target day is predicted.

0004

However, according to the above-mentioned prior art, there are the following problems to be solved. First, according to the technique disclosed in Japanese Patent Application Laid-open No. 60-102822, when creating and updating a performance data table, performance data is stored cumulatively, so the influence of past data patterns is comparatively reduced. There is a problem of receiving a lot of attention. In particular, prediction accuracy tends to deteriorate around sudden changes in weather or special days.

[0005] Furthermore, according to the technique disclosed in Japanese Patent Application Laid-Open No. 106325/1983, the predicted maximum temperature on the forecasted day is simply substituted into the quadratic equation expressing the relationship between the daily maximum temperature and the amount of demand. The problem is that there is a limit to the accuracy of predictions because it does not take into account fluctuations in power demand due to weather changes such as past temperatures.

[0006] Such problems are similar in the case of forecasting the demand for water and products.

[0007] An object of the present invention is to provide a demand forecasting method and apparatus that can improve forecast accuracy while using the latest demand record data as possible.

[0008]

[Means for Solving the Problems] In order to achieve the above object, the demand quantity forecasting method of the present invention collects actual demand quantity data with a plurality of influencing factors that influence increases and decreases in demand quantity, Separate into continuous influencing factors, where the relationship between changes in influencing factors and changes in demand is continuous, and discontinuous influencing factors, where the relationship is discontinuous, and use actual data for a recent certain period of the aforementioned actual data. , find a correlation function between the continuity influence factor and the demand quantity for each discontinuity influence factor or for each combination of discontinuity influence factors, and use the function corresponding to the prediction influence factor at the time of prediction from among the functions. The goal is to predict the amount of demand at the time of the forecast.

[0009] Furthermore, when determining the function, pseudo data at the time of prediction is created based on past performance data for areas that deviate from the performance data for a recent certain period of the performance data, and the pseudo data for the certain period are combined with the pseudo data for the certain period. It is desirable to use actual data.

[0010]Also, by combining them and finding the first function using actual data over a recent certain period,
A second function is obtained using the pseudo data and the actual data for a certain period of time, and a function corresponding to the prediction influence factor at the time of the prediction target is used from among the first and second functions to calculate the demand at the time of the prediction target. The amount can be predicted.

[0011] Furthermore, it is desirable to update the function by adding new performance data of the demand amount corresponding to the prediction result to the performance data for the certain period. In this case, when the period of new performance data related to the update reaches a predetermined additional period, the function is initialized, and the performance data of the recent certain period of performance data including the new performance data is used to It is desirable to find a new function.

[0012] A multiple regression model or a neural network model can be used as a means for determining the function. At that time, the convergence conditions of the model may be changed depending on the reliability of the actual data used, or the convergence conditions of the model may be changed when using the pseudo data when calculating the second function rather than when using the actual data. It is desirable to set it loosely.

[0013] Furthermore, if the performance data used for deriving the function is data obtained by removing the base portion that does not fluctuate from the actual performance data, the sensitivity of the function will be increased. Furthermore, after deriving the function, we can discard the actual data that deviates from the derived function by a certain amount and derive the function again using the remaining actual data, or we can compare the derived function and the actual data and In order to improve prediction accuracy, it is desirable to correct the function according to the degree of deviation in the actual data.

The demand forecasting device of the present invention can be realized by applying the above demand forecasting method and configuring it using an information processing device such as a computer.

[0015] In other words, there is a database in which actual data on demand is stored with a plurality of influencing factors that influence increases and decreases in demand, and a database in which the relationship between changes in influencing factors and changes in demand is discontinuous. Sequentially extracting performance data for a recent certain period corresponding to each predetermined discontinuity influencing factor or each combination of discontinuity influencing factors from the database,
a function deriving means for calculating a correlation function between the continuity influencing factor and the demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors based on the extracted performance data; A prediction that inputs a stored function table and a prediction influence factor at a prediction target time, reads a function corresponding to the prediction influence factor from the function table, and uses the read function to predict the demand amount at the prediction target time. and an execution means.

[0016]Furthermore, a pseudo data creation means is added to the above configuration for creating pseudo data at a prediction target point based on past performance data for a region that deviates from the extracted performance data, and the pseudo data is input to the function derivation means. Add it to the data.

[0017]

[Function] With this structure, according to the present invention, the object of the present invention can be achieved by the following function.

First, for each discontinuity influencing factor or for each combination thereof, a correlation function between the continuity influencing factor and the demand quantity is determined, and based on the function, the demand quantity at the time of the prediction target is predicted. Unique days on which differences or social events occur can also be predicted using functions based on past data, improving prediction accuracy. In addition, since predictions are made by calculating the correlation function only for continuity-affecting factors, excluding discontinuity-affecting factors, the reliability of the function can be increased even if the number of actual measurement data is small. Performance data can be used, and the influence of past performance can be minimized. Moreover, if a multiple regression model or a neural network is applied to the function model, the period covered by the performance data can be further shortened.

Furthermore, by creating pseudo data, the prediction accuracy is improved even when there is little actual data or when there is no actual data corresponding to the continuity influencing factor at the time of the prediction target.

[0020] Moreover, if new performance data is added and the function is updated, the prediction accuracy can be further improved. In particular, when the period of new performance data related to an update reaches a predetermined additional period, the influence of past old performance data can be reduced by initializing and re-deriving the function.

[0021] A multiple regression model or a neural network model is used as a means for determining the function, and the convergence conditions of the model are changed depending on the reliability of the actual data used, or the convergence conditions of the model are relaxed when pseudo data is used. By setting , reasonable predictions can be achieved.

[0022]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be explained below with reference to illustrated embodiments. FIG. 1 shows a configuration diagram of an electric power demand prediction device according to an embodiment in which the present invention is applied to electric power demand prediction. As shown in the figure, performance data on power demand is input into a database 2 from a performance data input means 1 . The input performance data includes new performance data obtained through operations in addition to past performance data accumulated. The function deriving means (I) 3 reads out actual performance data in the database 2 as appropriate, derives a function to be used for prediction according to a preset procedure, and stores it in a function table (I) 4. The function update means (I) 5 basically has the same configuration as the function derivation means (I) 3, but the difference is that it reads new performance data collected in accordance with the execution of prediction from the database 2, and It calculates a function that takes into account the data and replaces the function in the function table with a new one.

On the other hand, a prediction execution command is inputted from prediction command input means 6. This prediction command includes designation of a prediction target time such as a prediction target date, and prediction conditions such as influencing factors such as weather, day of the week, and temperature conditions related to the prediction target time. The prediction execution means 6 reads a function corresponding to the input prediction condition from the function table (I) 4, uses the function to determine the amount of power demand at the time of the prediction target, and outputs it as a predicted value to the prediction result output means 8. It looks like this.

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

Next, the detailed configuration of each part will be explained. The performance data input from the performance data input means 1 includes:
There are multiple influencing factors that affect the increase or decrease in electricity demand. Here, the influencing factors refer to various factors that influence increases and decreases in power demand, and an example of classifying these factors into homogeneous factor groups is shown below. In addition, the influencing factor group is further divided into continuous influencing factors, in which the relationship between changes in the influencing factor and changes in demand is continuous, and discontinuous influencing factors, in which the relationship is discontinuous.

(Example of discontinuity influencing factors) Type of weather (
- Sunny, rainy, cloudy, snowy, etc.) - Type of day of the week (weekdays, Sundays and holidays, Saturdays, mid-day holidays, factory holidays, etc.) - Time of day - Social events (sporting events, etc.) (continuous (Example of factors that influence discontinuity) ・Daily maximum temperature, daily minimum temperature, other sensible temperature, discomfort index, humidity, solar radiation. Cases are divided into cases for each combination, and within each case, the correlation between the influence factor of continuity and the amount of power demand is converted into a function and predicted.

Therefore, the database 2 is configured to store performance data in a tabular format as shown in the conceptual diagram shown in FIG. In the following, in this embodiment, a case will be described in which the amount of power demand is predicted in consideration of the type of weather, the type of day of the week, the time of day, the daily maximum temperature, and the daily minimum temperature.

The function deriving means (I) 3 is configured to obtain a correlation function between the continuity influencing factor and the demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors, and 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, the functions of power demand, daily maximum temperature, and daily minimum temperature for the time of the prediction target day are calculated. demand. Note that the performance data to be extracted is limited to data for a recent certain period from the point of derivation of the function.

FIG. 3 shows an example of a change in power demand over a certain day, and FIG. 4 shows an example of the relationship between power demand, daily maximum temperature, and daily minimum temperature at a certain time. In FIG. 3, the vertical axis indicates power demand and the horizontal axis indicates time. When predicting the power demand that changes in this way hourly, the function deriving means (I) 3 derives 24 functions for each time. It is also possible to calculate functions for each year, day, minute, second, etc. On the other hand, as is clear from the relationship curve H between daily maximum temperature and electricity demand at a certain time and the relationship curve L between daily minimum temperature and electricity demand at a certain time shown in Figure 4, there is a relationship between temperature and electricity demand. A kind of nonlinear relationship can be observed. Note that the vertical axis in the figure represents the power demand, and the horizontal axis represents the temperature. The existence of such a trend is considered to be due to the following reasons. That is,
In summer, when temperatures rise, electricity demand tends to increase rapidly due to the demand for air conditioners such as air conditioners. On the other hand, electricity demand remains almost constant during periods when temperature changes are small, such as from early spring to early summer, or from early autumn to early winter, when temperature changes are not large. In addition, demand for heating increases during the winter season when temperatures are low, so electricity demand tends to increase as the temperature decreases. The highest temperature on such a day,
A functional relationship is derived based on the relationship between daily minimum temperatures, and a function table (I) 4 is created for each discontinuity influencing factor or for each combination thereof.
I am trying to store it in .

As a calculation model for function derivation by the function derivation means (I) 3, a well-known multiple regression model or a model using a neural network can be applied. Here, a specific example using a neural network will be described with reference to FIG. This is because, in general, when the object is nonlinear, the utility of using a neural network in finding functional relationships is great. As shown in FIG. 5, a neural network connects elements 51 called neurons to connecting lines 52a, 52b, . ．． ．．
This is a model that simulates the behavior learned by human brain cells by connecting an input layer, one or more intermediate layers, and an output layer in a hierarchical manner. In this case, data for learning, such as influence factors that are input, such as daily maximum temperature and daily minimum temperature, are α (℃) and β (
℃), the data that the power demand, which is the output, is γ (MW) is called teaching data. Learning in a neural network means that when temperature data, which is teacher data, is input, connection lines 52a, 52b, . ．． ．． The purpose is to find the weighting coefficient for . There are various methods for determining the weighting coefficients, and in this embodiment, a backpropagation method will be described below, but the method is not limited to this.

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

[0032]

[Math. 1] Xi= f(ΣWij・Xj−θi)xj=
f(ΣWjk・Xk−θj) where Wpq: bond lines 52a, 52b, . ．． ．． Weighting coefficient p, q: i, j, k i, j, k: 1, 2,... θp: Threshold value f at which the neuron starts firing: Sigmoid function f(n)=(1/(1+exp( -n)) Now, suppose that when certain teacher data is input, the output Xi is different from the expected Xi0 and deviates. Completion of learning is equivalent to the following inequality being satisfied.

[0033]

[Equation 2] Σ(Xi-Xi0)2<ε Where Σ: Symbol representing the sum ε: Preset error threshold Therefore, in order to satisfy Equation 2, Wij is proportional to the error on the left side of Equation 2. and reversely correct it. The function deriving means (I) 3 is configured using such a neural network.

Now, regarding the embodiment configured as described above, the more detailed configuration and operation of the basic components will be explained using FIGS. 6 to 8. Additionally, the detailed configuration of the additional portion will be explained along with its operation. Function derivation means (I) 3
The processing procedure is shown in FIG. First, in step 61, in order to obtain all functions for combinations of discontinuity influencing factors among prediction target conditions, combinations are sequentially set and changed. In step 62, a function corresponding to the set/changed combination of discontinuity influencing factors is initialized. In other words, make it a meaningless function. Next, step 6
3, performance data corresponding to the set combination is extracted from the data table 2. For example, from among daily performance data for a recent certain period, performance data for which the weather is "sunny" and the day of the week is "weekdays" is read out. Based on this read data, in step 64, a correlation function between daily maximum temperature or daily minimum temperature and power demand is derived for each time. The functions derived in this way are stored in the function table (I) 4. Then, in step 65, it is determined whether or not the derivation of functions related to all combinations has been completed. If not, the process returns to step 61 to change the combination, and the above-mentioned procedure is continued until derivation of all functions is completed. Repeat the process.

[0035] Here, the processing in step 64 will be specifically explained. Now, the correlation function between the influencing factors and the power demand is expressed by Equation 3, and the weighting coefficient of the neural network is set so as to satisfy the condition of Equation 4 to derive the function.

[0036]

[Math 3] P(i,h)=g(X1,X2,...,Xn)

00
37]

[Equation 4] |P(i,h)-P'(i,h)|<ε However, the meanings of the symbols are as follows.

P(i,h): Power demand (MW) at time h on day i P'(i,h): Output of neural network at time h on day i g: Symbol indicating functional relationship Xj: Influencing factors representing climate, social events, etc. (j = 1, 2,...) ε: Output error threshold In order to increase the reliability of function derivation using a neural network, as is generally done, , it is desirable to repeatedly perform learning using teacher data. For example, learning is repeated 20 times for one combination of daily maximum temperature and demand amount, and then the combination is changed and repeated 100 times in total. this is,
Since the influence of previous learning results remains, this influence is removed through repetition. In this way, a demand quantity prediction function related to continuity influencing factors can be obtained with high accuracy. The reason why discontinuity influencing factors are not converted into functions is because, for example, even if the weather is the same, the trend in demand may be completely different depending on the type of day of the week, and the accuracy of function conversion cannot be improved. be.

FIG. 7 shows the processing procedure of the prediction execution means 7. As shown in the figure, in step 71, the prediction conditions of the prediction target are input from the prediction command means 6. As a prediction condition, for example, assume that the daily change in power demand is to be predicted on what month and day, and input the expected weather for that day as sunny, the day of the week as a weekday, the expected daily maximum temperature and daily minimum temperature for that day, etc. Next, in step 72, an initial value of the prediction target time is set. In the next step 73,
As will be explained later, it is determined whether the condition of the prediction target is within or outside the area in which actual performance data exists, that is, it is determined whether or not interpolation is required. In the case of interpolation, proceed to step 74, read out the neural network of the corresponding function shown in Equation 3 from function table 3, input the prediction condition to its input layer, and calculate the output of the power demand at the relevant time. Use as predicted value. Subsequently, in step 75, it is determined whether prediction has been completed for all times, and the process returns to step 72 to repeat the above prediction process until the prediction is completed. When all predictions are completed, the process proceeds to step 76 and the prediction results are output.

The function updating means (I) 5 updates the functions in the function table (I) 4 to reflect new performance data according to the procedure shown in FIG. Actual data at the time corresponding to the above prediction result is sent to the database 2 via the input means 1.
is input from time to time. The function updating means (I) 5 determines whether new performance data has been input in step 81 of FIG. , function table (
I) Update 4. In other words, the weighting coefficients of the neural network are changed. The processing content of this step 82 corresponds to steps 62, 63, 64, 66, and 67 in FIG. 6, so the details will be omitted. Thereafter, each time new performance data is input, the function is repeatedly updated, but it is determined whether or not the period of this repetition exceeds a preset period (additional learning period) (step 83). In this case, proceed to step 84 to initialize the relevant functional relationship, and use the neural network at each point in time to newly find the influencing factors of the initial learning period and the function of power demand, and create a new function. Reset.

An example of the results of power demand prediction performed using the above procedure will be explained next.・Continuity influencing factor conditions: daily maximum temperature, daily minimum temperature ・Discontinuity influencing factor conditions: Weather - sunny, day of the week - weekday ・Initial learning period: 1 month ・Additional learning period: 1 month Months in Figure 9 The results of the prediction are shown below. In the neural network, ○ is the actual amount one month before the prediction execution which becomes the training 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 within the range marked with a circle, the predicted amount almost matches the actual amount, indicating that highly accurate predictions can be made (interpolation case). However, predicted quantities in other ranges have a very low degree of agreement and poor accuracy (extrapolation case).

Based on these, the accuracy of demand forecasting is evaluated in two ways: only the interpolation case and the extrapolation case and the interpolation case together. Criteria for evaluating the accuracy of demand forecasting are broadly classified into two categories, and in the case of next-day forecasting, the following two points are used.

・Reducing the daily average error ・Reducing the variation in the daily average error Here, the daily average error is calculated by the following formula 5.
This is done using the accuracy evaluation formula.

[0044]

[Math 5]

[0045] However, abs: symbol representing absolute value Σ: symbol representing sum i=1, 2, . ．． ．． , 24 The above formula calculates the relative error by dividing the absolute value of the difference between the predicted demand value and the actual demand value at each time point i by the actual value, and then divides the sum of the relative errors at each time point by the number of time points. By doing this, the daily average error is determined.

Various modifications can be considered as a method for evaluating such an error. Moreover, although the evaluation was performed using relative errors in this embodiment, it is of course possible to use the absolute values of the predicted value and the demand value as evaluation criteria.

As an example of the error, a prediction result based on the past 1 year and 5 months' worth of actual data is shown in FIG. As shown in the figure, the average value of the daily average error is 2.18% when only the interpolation case is involved, and 2.18% when the extrapolation case is included.
．． It can be observed that the accuracy is 84%, which is worse when including the extrapolation case than when only the interpolation case is included.

Here, the additional parts 11 to 14 of the embodiment in FIG.
The details will be explained below. This additional part was added to improve prediction accuracy, and is particularly characterized by the provision of pseudo data creation means 11. The reason for providing this pseudo data creation means 11 will be explained. As mentioned above, in order to perform accurate demand forecasting, it is not desirable to perform demand forecasting based on the extrapolation case. Therefore, in order to reduce the number of cases of extrapolation, in order to change the influencing factors at the predicted time point, which originally corresponds to the case of extrapolation, to the case of interpolation, pseudo training data is used as function derivation means (II) 1.
2. An example of pseudo data will be explained using FIG. 11. FIG. 11 is an example in which the daily power demand for the next month is calculated based on actual data for a certain month in early spring. In the figure, the black ellipse is the range of actual data for a certain month, and the hatched ellipse is the range of actual data for the next month, which is the target of prediction. Also, the vertical axis shows the amount of electricity demand, and the horizontal axis shows the temperature. Since there is no actual data for the month in which the prediction is executed within the temperature range in which the actual data serving as learning data exists, this corresponds to a case of extrapolation when executing the prediction, and as described above, the accuracy of the prediction deteriorates. Generally, in order to resolve such cases of extrapolation, there is a method of accumulating training 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 prediction accuracy is reduced due to the large influence of past data. In order to solve these problems, in this example, a pseudo A method is introduced to reduce the number of cases of extrapolation by adding data represented by the range of the white ellipse. In principle, pseudo data is created automatically using a computer, etc. However, if necessary, it is also possible to generate pseudo data by manually inputting human know-how by an operator.

Various methods can be considered for creating pseudo data in the pseudo data creating means 11. In this embodiment, a typical example thereof is shown below. Now, Tmin(i): Daily minimum temperature of data for the same month of the previous year of the forecast month Tmax(i): Daily minimum temperature of data of the same month of the previous year of the forecast month P(i, h): Time of data of the same month of the previous year of the forecast month Let it be the power demand at h.

First, data at the same time one year ago (
Tmin(i), Tmax(i), P(i, h))), then (Tmin(i), Tmax(i), P"(i, h
)). however,

[0052]

[Math 6]

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

Furthermore, when the function deriving means (II) 12 derives a nonlinear function using the above-mentioned pseudo data,
Since the reliability of pseudo data is not 100%, in this embodiment, the method for deriving the function is changed depending on the reliability of the data. Specifically, when a neural network is used as a function model, the weighting coefficients of the neural network are determined under loose convergence conditions, while when actual data exists, the weighting coefficients of the neural network are determined under strict convergence conditions. shall be. In other words, 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 actual data is used. That is, the weighting coefficient of the neural network is updated so that the function type of Equation 7 satisfies the condition of Equation 8.

[0055]

[Math 7] P”(i,h)=g(X1,X2,...,Xn)

0
056]

[Equation 8] |P''(i, h)-P'(i, h)|<ε' However, the meanings of the symbols are as follows.

P''(i,h): Pseudo data of 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 : Influencing factors representing climate, social events, etc., where (j=1, 2,...) ε': Threshold of output error for pseudo data (ε<ε
') The pseudo data thus generated is added to the actual data, and the function deriving means (II) 12 calculates the function II. The structure of the function deriving means (II) 12 is as follows:
) Since it is the same as 3, the explanation will be omitted. Furthermore, the processing procedure for creating pseudo data and deriving functions is shown in step 66 in FIG.
, 67, and Function I and Function II are calculated in parallel.

Note that instead of the processing procedure shown in FIG. 6, the judgment result (step 73) of whether or not the condition of the prediction target in the prediction execution means (I) 5 falls outside the range of the actual performance data that is the basis of the function I is The pseudo data may be output to the pseudo data creation means 11, pseudo data may be created in response to this, and the function derivation means (II) 12 may be activated to derive the function II.

Furthermore, in the embodiments shown in FIGS. 1, 6, and 7, the processing of the function derivation means (I) 3, 12 and the prediction execution means 7 are not directly related, and the processing is executed independently. explained. However, the present invention is not limited to this, and the prediction commands and conditions inputted from the prediction command input means 6 are input to the function derivation means (I) 3, 12, and only the actual data matching the prediction conditions are input. In addition to extracting from the database 2, pseudo data can be created as needed, and only functions related to the prediction target can be derived in response to the input of the prediction command. According to this, derivation of unnecessary functions can be omitted.

Modifications of the above embodiment will be further explained below. With reference to FIG. 12, an example for further improving prediction accuracy will be described. In this example, as shown in FIG. 12(a), the daily change in power demand is a certain minimum power demand B
Considering that it will not be lower than the base value, the demand quantity after subtracting the base removal amount B (Figure 12(b))
By targeting the prediction target, the prediction sensitivity is increased and the prediction accuracy is thereby improved. That is, the actual prediction data (Tmin(i), T
max(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): Daily minimum temperature Tm of the i-th teacher data
ax(i): Daily maximum temperature P(i,
h): Electricity demand amount at time h of the i-th training data B: Base removal amount Also, with reference to FIGS. 13 and 14, the accuracy of prediction is improved by correcting the function derived in the above example. An example will be explained. FIG. 13 conveniently expresses the concept of function modification on a two-dimensional plane. As shown in FIGS. 13 and 14, the function derivation means (I) 3 and 12 compare the derived function with actual data based on the result of deriving the function, and identify inappropriate functions that do not match the demand trend from among the actual data. Data is detected based on preset criteria (steps 101 and 111), and a function is derived again using only proper data by rejecting the inappropriate data (steps 102 and 112). ). This method can be applied to the derivation of either function I or II. As a specific method for rejecting data, it is possible to substantially reject the data by gently adjusting the threshold value of the output error of the neural network when re-deriving the function.

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

[0062]

[Math. 9]

Here, g(X1,...Xn): derived functions X1,... ．． ．． , Xn: Influencing factor Yi: Power demand m: Number of actual data n: Number of influencing factors By the correction value E obtained by the above formula 9, as shown in FIG. 15,
The function g before correction is translated by E to obtain the function g' after correction.
seek. Thereby, prediction accuracy can be improved.

[0064] Although not taken into consideration in the above embodiment,
When performing demand forecasting, special days, such as the middle of the Topping Stones holiday,
Forecasts for the first weekday after the Obon holidays often have large errors between predicted values and actual values. In order to deal with this point, the error between the predicted value and the actual value of the past singular day is stored in the database 2 or inputted through the actual data input means 1, and this data is used to calculate the function derivation means (I) 3, By correcting the functions I and II obtained in step 12, it is possible to perform accurate prediction.

[0065] The present invention has been described above based on an embodiment in which the present invention is applied to predicting the amount of electricity demand. However, as mentioned above, the present invention is not limited to electricity, but can also be applied to predictions of water demand, bread demand, etc. It is clear that this method can be applied to objects whose demand increases or decreases depending on influencing factors such as temperature.

[0066]

[Effects of the Invention] As explained above, according to the present invention,
It has the following effects. For each discontinuity influencing factor or for each combination thereof, find the correlation function between the continuity influencing factor and the demand quantity,
Since the demand amount at the time of the prediction target is predicted based on the function, it is possible to predict differences in days of the week and special days when social events are held using the function based on actual data, thereby improving prediction accuracy.

[0067] Furthermore, since the correlation function for the continuity influencing factor is calculated and predicted, the reliability of the function is high even if the number of actual measurement data is small, so it is possible to use actual data for a recent certain period. The influence of past performance can be minimized to the necessary minimum. In particular, if a multiple regression model or a neural network is applied to the function model, the period covered by the performance data can be further shortened.

Furthermore, by creating pseudo data, the prediction accuracy is improved even when there is little actual data or when there is no actual data corresponding to the continuity influencing factor at the time of the prediction target.

Furthermore, if the function is updated by adding new performance data, the prediction accuracy is further improved. In particular, when the period of new performance data related to an update reaches a predetermined additional period, the influence of past old performance data can be reduced by initializing and re-deriving the function.

[0070] As a means of determining the function, a multiple regression model or a neural network model is used, and the convergence conditions of the model are changed depending on the reliability of actual data, etc., or the convergence conditions of the model are relaxed when pseudo data is used. Depending on the settings, reasonable predictions can be achieved.

[0071] In addition, it is possible to use actual values obtained by removing base removal from actual data, to reject actual data that deviates significantly from the derived function and to derive the function again, or to calculate the degree of deviation between the derived function and actual data. If the function is corrected in accordance with the above, the prediction accuracy can be further improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a demand forecasting device 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 showing an example of daily changes in power demand.

FIG. 4 is a diagram showing an example of the correlation between 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 the processing procedure of the function deriving means.

FIG. 7 is a flowchart of the processing procedure of the prediction execution means.

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

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

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

FIG. 11 is a diagram for explaining a method of creating pseudo data.

FIG. 12 is an explanatory diagram of an example in which a function is derived using the actual demand value obtained by subtracting the base removal amount from the actual demand amount, (a) is the actual demand value including the base removal amount, (b)
is a diagram of the 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 the main part of the processing procedure of the embodiment for rejecting performance data that deviates from the derived function.

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

FIG. 16 is a flowchart of the main part of the processing procedure of the embodiment for correcting a function according to the degree of performance 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 update means (II)

Claims (23)

[Claims] [Claim 1] Actual data on demand quantity is collected with a plurality of influencing factors that influence increases and decreases in demand quantity, and the relationship between changes in the influencing factors and changes in demand quantity is continuous influencing factors. and discontinuous influence factors whose relationship is discontinuous,
Using the performance data for a recent certain period of the performance data,
Determining the correlation function between the continuity influencing factor and the demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors,
A demand amount forecasting method comprising predicting a demand amount at a prediction target time using a function corresponding to a prediction influence factor at the prediction target time from among the functions. [Claim 2] Actual data on the amount of demand is collected with multiple influencing factors that influence increases and decreases in the amount of demand, and the relationship between changes in the influencing factors and changes in the amount of demand is a continuous influencing factor. and discontinuous influence factors whose relationship is discontinuous,
Create pseudo data at the time of prediction target based on past performance data for areas that deviate from the performance data for a recent certain period of the performance data, and use the pseudo data and the performance data for the certain period to calculate discontinuity effects. For each factor or for each combination of discontinuity influencing factors, a function of the continuity influencing factor and the demand quantity is determined, and from among the functions, the function corresponding to the predicted influencing factor at the forecasting point is used to calculate the function at the forecasting point. A demand forecasting method that involves forecasting demand. [Claim 3] Actual data on the amount of demand is collected with a plurality of influencing factors that influence increases and decreases in the amount of demand, and the relationship between changes in the influencing factors and changes in the amount of demand is a continuous influencing factor. and discontinuous influence factors whose relationship is discontinuous,
Using the performance data for a recent certain period of the performance data,
Calculate the first function of the continuity influencing factor and demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors, and based on past actual data for areas that deviate from the actual data for the certain period of time. Create pseudo data at the time of the forecast target, and use the pseudo data and the actual data for the certain period to calculate the continuity impact factors and demand for each discontinuity impact factor or for each combination of discontinuity impact factors. A demand quantity forecasting method comprising calculating a second function and predicting a demand quantity at a prediction target time using a function corresponding to a prediction influence factor at the prediction target time from among the first and second functions. [Claim 4] In any one of claims 1 to 3,
A demand forecasting method, comprising: updating the function by adding new performance data of the demand corresponding to the predicted demand to the performance data for the certain period. 5. In claim 4, the function is initialized when the period of new performance data related to the update reaches a predetermined additional period, and the function is initialized when the period of new performance data related to the update reaches a predetermined additional period, and A method for predicting demand quantity, characterized in that the function is newly determined using actual data of. [Claim 6] In any one of claims 1 to 3,
A demand forecasting method characterized in that the performance data used for deriving the function is data obtained by removing a base portion from actual performance data. [Claim 7] In any one of claims 1 to 3,
A demand forecasting method characterized in that, after deriving the function, performance data that deviates by a certain amount from the derived function is discarded, and the function is derived again using the remaining performance data. [Claim 8] In any one of claims 1 to 3,
After deriving the function, the derived function is compared with actual data, and the function is corrected according to the degree of deviation of the actual data from the function. 9. The demand forecasting method according to claim 1, wherein the function is determined using a multiple regression model or a neural network model. 10. Claims 1 to 3, characterized in that the function is determined using a multiple regression model or a neural network model, and the convergence conditions of the model are changed depending on the reliability of actual data used at that time. demand forecasting method. 11. In claim 3, the first and second
The function of is determined using a multiple regression model or a neural network model, and when determining the second function, the convergence conditions of the model are set more gently when using the pseudo data than when using the actual data. Characteristic demand forecasting method. [Claim 12] A database in which actual data on demand quantity is stored with a plurality of influencing factors that influence increases and decreases in demand quantity, and a database in which the relationship between changes in the influencing factors and changes in demand quantity is discontinuous. Actual data for a recent certain period corresponding to each predetermined discontinuity influencing factor or combination of discontinuity influencing factors is sequentially extracted from the database, and the discontinuity influencing factor is determined based on the extracted actual data. a function deriving means for determining a correlation function between the continuity influence factor and the demand quantity for each combination of discontinuity influence factors, a function table in which the determined function is stored, and a prediction influence factor at the time of prediction target A demand amount forecasting device comprising: a prediction execution unit that inputs a function that corresponds to the prediction influence factor, reads a function corresponding to the prediction influence factor from the function table, and predicts the demand amount at the prediction target time using the read function. [Claim 13] A database in which the actual data of the demand quantity is stored with a plurality of influencing factors that influence an increase or decrease in the demand quantity, and a database in which the relationship between changes in the influencing factors and changes in the demand quantity is discontinuous. Performance data for a recent certain period corresponding to each predetermined discontinuity influencing factor or combination of discontinuity influencing factors is sequentially extracted from the database, and past performance data is extracted for areas that deviate from the extracted performance data. pseudo data creation means for creating pseudo data at the prediction target point based on the above, and for each discontinuity influencing factor or for each combination of discontinuity influencing factors, based on the pseudo data and the extracted actual data; A function deriving means for obtaining a correlation function between a continuity influencing factor and a demand quantity, a function table in which the determined function is stored, and a function deriving means for inputting a predicted influencing factor at a prediction target time and a function corresponding to the predicted influencing factor. 1. A demand forecasting device comprising a prediction execution unit that reads out a function table and uses the read function to predict a demand at the time of the prediction target. [Claim 14] A database in which the actual data on the demand quantity is stored with a plurality of influencing factors that influence increases and decreases in the demand quantity, and a database in which the relationship between changes in the influencing factors and changes in the demand quantity is discontinuous. Actual data for a recent certain period corresponding to each predetermined discontinuity influencing factor or combination of discontinuity influencing factors is sequentially extracted from the database, and the discontinuity influencing factor is determined based on the extracted actual data. a first function deriving means for calculating a first correlation function between the continuity influencing factor and the demand quantity for each case or for each combination of discontinuity influencing factors; pseudo data creation means for creating pseudo data at the prediction target point based on the above, and for each discontinuity influencing factor or for each combination of discontinuity influencing factors, based on the pseudo data and the extracted actual data; a second function deriving means for calculating a second correlation function between the continuity influencing factor and the demand quantity; a function table storing the determined first and second functions; and a prediction influencing factor at the prediction target time point. A forecasting execution means for reading out the first or second function inputted and corresponding to the forecasting influence factor from the function table, and predicting the demand at the forecasting target time using the read function. Prediction device. 15. In any one of claims 12 to 14, the function deriving means updates the function by adding new actual data of the demand amount corresponding to the predicted demand amount to the actual data for the certain period. A demand forecasting device characterized by: 16. In claim 15, the function deriving means initializes the function when the period of the new performance data related to the update reaches a predetermined additional period, and the function derivation means initializes the function when the period of the new performance data related to the update reaches a predetermined additional period, and generates the performance data including the new performance data. A demand forecasting device characterized in that the function is newly calculated using actual performance data for the recent certain period. 17. In any one of claims 12 to 14, the function deriving means calculates a correlation function between the continuity influencing factor and the demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors. A demand forecasting device comprising a multiple regression model or a neural network model. 18. In any one of claims 12 to 14, the function deriving means calculates a correlation function between the continuity influencing factor and the demand quantity for each discontinuity influencing factor or for each combination of discontinuity influencing factors. 1. A demand forecasting device comprising a multiple regression model or a neural network model for deriving the following, and changing the convergence conditions of each model depending on the reliability of actual performance data, etc. used. 19. In claim 14, the second function deriving means is characterized in that the convergence conditions of the model are set more gently when using the pseudo data than when using the actual data. Demand forecasting device. 20. The demand forecasting method according to claim 1, wherein the demand amount is one of electricity, water, and merchandise. 21. In any one of claims 1 to 11, the demand amount is electricity, and the discontinuity influencing factor includes any one of weather type, day of the week type, social event, etc. . A demand forecasting method, wherein the continuity influencing factor includes any one of daily maximum temperature, daily minimum temperature, sensible temperature, discomfort index, humidity, amount of solar radiation, etc. 22. The demand forecasting device according to claim 12, wherein the demand amount is one of electricity, water, and merchandise. 23. In any one of claims 12 to 19, the demand amount is electricity, and the discontinuity influencing factor includes any one of a type of weather, a type of day of the week, a social event, etc. . A demand forecasting device, wherein the continuity influencing factor includes any one of daily maximum temperature, daily minimum temperature, sensible temperature, discomfort index, humidity, amount of solar radiation, etc.