JPH08308108A - Method and device for predicting electric power demand - Google Patents

Abstract

PURPOSE: To improve the power demand prediction accuracy of an electric power demand predicting device on a power system by accurately modeling the relation between the past demand of the system and factors affected the past demand. CONSTITUTION: An electric power demand predicting device is composed of a history data base recording the past electric power demand fetched from an objective power system and factors affected the demand, prediction executing device 105 which predicts a device which fetches affected factors for predicting the electric power demand of the system by using a model created by means of a predicting model creating and updating device which creates a data pre- processor predicting model for processing the data recorded in the data base to an easily usable form for predicting the demand, and actual demand fetching device which fetches an actual demand after prediction. Therefore, the future electric power demand of the system can be found with high accuracy based on the relation between the past electric power demand of the system and factors affected the demand.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a power demand forecasting method and apparatus for forecasting the power demand of a power system.
In particular, the present invention relates to a power demand prediction method and apparatus for predicting power demand using historical data of a power system.

[0002]

[Prior art]

(1) Japanese Unexamined Patent Application Publication No. 5-18995 Japanese Patent Laid-Open No. 5-18995 In order to predict the electric power demand, in many conventional methods, the relationship between the factor affecting the demand calculated by regression analysis and the demand , Forecasting was often done using the factors of the forecast day. When this method is used, the prediction error tends to increase when there is a tendency, such as the turning of the season, between the demand volume and the factors that affect it that cannot be represented by a mathematical model. . In the known example, in order to reduce this error, the neural network model is used to correct the prediction model to improve the prediction accuracy.

(2) Japanese Unexamined Patent Publication No. 6-105465 No. 6 Maximum Power Demand Forecasting Method and Apparatus In this known example, regarding the problem of obtaining the power demand, there is a relation between the weather condition A consisting of the maximum temperature, the minimum temperature and humidity and the power demand. Based on the correlation, we created the weather data of multiple points in the supply area, the weather conditions of the days closest to the forecast day, the weekday weekday gap, and the demand forecast model at the transition of seasons using linear regression analysis. , Maximum demand forecast.

(3) Maximum power prediction system using Tatsuyoshi Kato, Yoshiteru Ueki, Tetsuro Matsui, Hiroshi Endo, Neuro-Fuzzy. 1995 Annual Meeting of the Institute of Electrical Engineers, 6-298 (199).
5). In this known example, the factors that affect the power demand amount,
First, the relationship with the power demand is obtained using a neural network model. After that, the forecast model for each season is used to make predictions for periods when the demand trend does not show a certain pattern, such as the transition of seasons as described in the publicly known example (1). Model and summer model) The results are combined using fuzzy reasoning to obtain the demand forecast result.

(4) Junichi Murata, Takashi Miyake, Setsuo Sagara, Prediction of Weekly Electric Power Demand by Automatic Sample Selection. Transactions of the Institute of Electrical Engineers of Japan,
112 (1992), 1077-1084. In this publicly known example, an algorithm for automatically selecting a sample from past historical data has been developed in order to make a mechanical / objective forecast in the forecast of power demand. The abnormal value existing in the quantity value is effectively removed from the sample, and the predicted value of the power demand in the normal period is optimized.

[0006]

The problem common to all the above-mentioned known examples is how to accurately construct a prediction model. In either method, a large prediction error appears.

Particularly, there is a problem in a highly accurate p-prediction method in the case where the demand trend cannot be expressed in a fixed pattern as in the known example (1). In the known example (1), although the prediction result is corrected, a single prediction model is used as the base prediction model. Moreover, in the known example (2), prediction is performed using a plurality of models. Further, in the publicly known example (3), prediction is performed with a plurality of models, and the prediction results obtained from the plurality of models are weighted and combined to try to improve the prediction accuracy. However, although this method is effective for forecasting demand at the turn of the season, it is not so effective for forecasting when demand trends that appear during a particular season are not constant.

Further, in the case of selecting an appropriate past demand amount history data by the method of the known example (4) and creating a demand amount prediction model, it is necessary to hold a large amount of past history data regarding the power demand amount. . Usually, since there is a small amount of history data that is effective for power demand prediction, when the sample selection algorithm in this known example is used, as a result, it becomes necessary to execute prediction with a small amount of data, which is unreliable. There is.
As described above, in the conventional method, the prediction model, which is the basis of the prediction, is changed at most seasonally. Especially when the relationship between the demand and the factors affecting it does not show a certain tendency,
It is considered necessary to make a model that more closely matches the characteristics of the demand amount data than shown in the above-mentioned known example.

According to the present invention, even when demand forecasting is performed using past history data in which the demand trend does not show a constant pattern, a forecasting model is formed by combining a plurality of models in the conventional known example. While the present invention has been created, the present invention has an object to solve the creation of a demand quantity prediction model that improves prediction accuracy by using a single model.

[0010]

In order to achieve the above object, according to the present invention, the electric power demand amount of the electric power system is determined from the relationship between the electric power demand amount based on the historical data of the electric power system and the factors affecting the electric power demand amount. In the power demand forecasting method or device for predicting the power demand, the distribution density of the history data is obtained from the correlation between the power demand data selected from the history data and the factors affecting the power demand. Clustering historical data based on
The weighting coefficient of each history data is determined based on this clustering result, and the power demand of the power system is predicted using a prediction model using this weighting coefficient.

In order to solve the above problems, the present invention stores past history data in the prediction model / update device based on the correlation between the factors affecting the demand amount and the demand amount. Analyze past historical data by dividing it using criteria. Based on this analysis result, an appropriate demand forecast model is constructed in order to make the forecast. Further, in the present invention, in order to solve the above-mentioned problem, first, the relationship between the demand amount and the factors affecting it is shown in a correlation diagram. The correlation diagram is divided by a linear line or a plane showing an average tendency, and the existence density of past history data in the divided region is obtained. The order of the model is determined according to the magnitude of this existence density, and after determining the order of the model, the weighting coefficient for each past history data used for model creation is determined.
By using the weighted history data, the coefficient of the model is obtained by the method of least squares, and the optimum prediction model is created.

[0012]

In the known example of the related art, a prediction model is created by combining a plurality of models, whereas in the present invention, it is possible to improve prediction accuracy by using a single model.

Further, compared with the power demand forecast shown in the publicly known example, it becomes possible to perform modeling more meticulously in accordance with the demand data characteristics.

[0014]

FIG. 1 is a diagram for explaining the embodiment of the present invention with reference to the drawings.
FIG. 1 is a configuration diagram of a demand forecasting method and device according to the present invention.

The present invention includes a target power system 101, a history database 102 that stores history data of past demand and factors affecting it, and imports appropriate data for creating a prediction model from the history database. , A data preprocessing device 103 for converting into a suitable form for prediction execution, and based on data created by the data preprocessing device, creates and updates a model between demand and factors affecting it Prediction model creating / updating device 104,
An influence factor capturing device 106 that captures factors that affect the demand on the forecast day from the inside or outside of the power system,
A demand amount actual result fetching device 107 for fetching the actual demand amount from the power system, a prediction result output device 108 for outputting the forecast result on a screen or a form, information of a forecast object, for example, a total demand forecast or an individual unit (for example, a region Prediction condition input device 109 for inputting which of the demand forecasts in (1) to (3) is to be executed.

Next, a data flow in the demand forecasting method and apparatus of the present invention will be described. In the power system,
System information is obtained from the measuring instruments installed in each facility that composes the power system in a fixed time cycle. The past history data regarding this information is stored in the past history database through the communication line 151. In order to forecast demand, it is necessary to obtain the necessary data from this historical database. In order to obtain the data here, the history database information is transferred to the data preprocessing device 10 via the communication line 152.
Send to 3. The data preprocessor 103 includes a communication line 160.
The information of the prediction condition input device is also input at the same time.
The data preprocessing device 103 selects appropriate data for demand forecasting from the information stored in the history database and converts it into an appropriate form. The converted data is sent to the prediction model creating / updating device 104 via the communication line 153. In the prediction model creation / update / device, when the actual data is found after the prediction, in order to update the prediction model, the actual demand amount is fetched from the power system through the communication line 157. The captured data is sent to the prediction model creation / update device 104 through the communication line 158. The prediction model created by the prediction model creation / update device 104 is the communication line 1
It is sent to the prediction execution device 105 via 54. In the prediction executing device 105, the influencing factors affecting the demand amount at the time of executing the prediction are fetched from the power system by the device 106 through the communication line 155, and the fetched data are sent to the prediction executing device through the communication line 156. The prediction execution device makes a prediction using the data acquired from the device 104 and the device 106, sends the result to the prediction result output device 108 through the communication line 169, and is easy for a person who uses the system incorporating the contents of the present invention to see. To output the prediction result.

Next, detailed examples of each device of the present invention will be described. In the past history database 102, FIG.
The relationship between the past demand amount and the factors affecting it is stored in a data format represented by 201. The demand amount here refers to, for example, an example relating to the power demand amount, and is the sum of the demand amount related to the active power in all regions of the target power system or the demand related to the active power for each individual region. Volume or the sum of the reactive power demands of all target regions, or the reactive power demand for each individual region. Further, as the factors that affect the demand amount, when the total demand active power is targeted, temperature, humidity at the time of data collection, weather conditions represented by weather, and social factors represented by sports events. And so on. In the case of individual active power, for example, a known example (1)
Nakajima, Okubo, Matsumoto, Ishida, Tamura, Development of several hours ahead power flow state prediction (DPF) system for the next central power dispatching center.
IEEJ 1995 National Convention, 1396, 1995. As shown in Figure 2, the ratio of each substation to the total demand is a major factor. Influential factors related to reactive power are, for example, publicly known cases (2) edited by the Japan Electric Joint Research Group, stable power system operation technology. Electric Joint Research, 121, 1991. It depends on the size of the active power. Such factors are saved together with the year, month, day, time when the past data was collected, whether it is the total demand amount or the individual demand amount, and the equipment name of the data collection target. Further, FIG. 2202 is an example of a lower system switching database, which will be described later, which is necessary when executing the individual demand amount prediction. In this database, switching source information 251, switching destination information 252, and switching target amount information 253 are described.

Next, details of the data preprocessing device 103 will be described. As described above, the data preprocessing device 103 acquires appropriate data for model creation for executing demand forecasting from the database of the device 102 and processes it into a proper form for forecasting demand. Is a device for performing.

Details of the processing here will be described with reference to FIG. First, the means 301 selects data in the past range necessary for constructing the prediction model. Then the device
Based on the prediction target method input in 109, the means 302 determines whether or not the demand amount of the prediction target is the total demand amount prediction. When it is the total demand forecast, the means 307 selects necessary data while determining the classification of the item 211 in the database of FIG. When the means 302 determines that it is not the total demand, the means 308 selects data for individual load prediction.

The details are shown below. When the individual substation load prediction is performed, the connection state of the lower system connected to the lower order of the substation load to be predicted becomes a problem as shown in the known example (1). Therefore, there is no switching of load data in the database as in the known example (1),
It is necessary to convert to a so-called normal system state. With the information from the switching information data 202, the means 3
The data read in 01 is converted into a normal system state by means 303. Using the resulting data, the means 304 determines based on the data used by the means 109 whether it is active power prediction or reactive power prediction. When the determination result is active power prediction, the means 305 selects the active power data according to the total demand amount individual demand amount classification flag of FIG. When the determination result of the means 304 is reactive power, the means 306 selects reactive power load data. These results are sent to the prediction target data creation device 309. The prediction target data creation device 309 processes the data into a shape required when creating the prediction model. For example, a publicly known example
As shown in (1), in the case of individual active power load prediction,
Replace the active power demand value for each substation with a value that is divided by the total demand value, or in the case of individual reactive power load prediction, use the ratio with the active power load as in known example (2). To ask.

Next, details of the prediction model creation / update device 104 will be described. After obtaining the data for creating the demand amount prediction model in the device 103, an example of the relationship between the demand amount and the factors affecting the demand amount can be simply expressed in a schematic diagram as shown in FIG. . Graph 40 in FIG.
In the case of 1, there is little variation in the relationship between the influencing factors and the demand amount, and it is expected that the prediction can be executed accurately. However, in the graph 402 in FIG. 4, although there is some relationship between the influencing factors and the demand amount, there are large variations and it is difficult to create an accurate prediction model. In particular, the present invention provides a model that can be appropriately modeled in the latter case where the relationship between the influence factors and the demand amount varies.

The details will be described with reference to the graphs of FIGS. 5 and 6 and the flowchart of FIG.

In the present embodiment, for the sake of simplicity, a conceptual diagram will be described in the case where there is one influencing factor affecting the demand amount as shown in FIG. In addition, the target of model creation in this embodiment is the power demand data. The feature of the power demand data is that when the relationship between the influencing factors and the demand is graphed as a graph 501 in FIG. 5, it becomes a downward convex curve or a straight line expressed by a linear expression in many cases. It is a point. As the influencing factors, it is possible to make a graph by using the various factors described in FIG.

In the flowchart of FIG. 7, first, the means 701 reads model creation target data. Next means 70
In 2, create a diagram to find the correlation between demand and the influencing factors that affect it. Here, when modeling based on the data 502 having a large variation like the graph 501 in FIG. 5, the modeling is performed by a straight line represented by the following equation 1 or represented by the following equation 2. The problem is how to model with a curved line.

Y = ax + b (Equation 1) y = ax ² + bx + c (Equation 2) Therefore, in the present embodiment, first, the average tendency shown in the graph 503 for the data 502 is shown based on the least squares method. 7 means 703. The average tendency obtained by the means 703 in FIG. 7 makes it possible to divide the data into areas 504 and 505 in FIG. 5 (means 704 in FIG. 7). Next, the means 705 in FIG. 7 obtains the data existence densities D1 and D2 in the respective regions 504 and 505 in FIG. 5 by using the following equation 3.

(Presence Density) = (Number of Data) ÷ (Area Density) (Equation 3) The densities of the respective areas obtained here are equal to each other within a preset range of error ε, that is, data distribution If there is no deviation in the line, it is judged that it is appropriate to perform modeling with a straight line (FIG. 7 means 706). When the data distribution of the upper half region 504 and the lower half region 505 of the straight line 503 of FIG. 5 is deviated by the means 706 of FIG. 7, the data needs to be modeled by a curve. Here, as an example, an embodiment will be described in which a model is obtained using the model of the above-mentioned Expression 2. For that purpose, the area 505 in FIG. 5 is further divided into areas 604 and 605 as shown in a graph 601 in FIG. 6 as a clustering means (means 609 in FIG. 7). Here, the reason why the region 504 in FIG. 6 is not divided is that the power data, which is the target of the present embodiment, can be approximated to a downward convex curve or a straight line, as described above. The area 505 in FIG. 5 is replaced with the area 6 in FIG.
04 and the area 605 are divided into straight lines 503 in FIG.
A straight line parallel to is obtained as shown in FIG. Here, as an example, it is assumed that a straight line 606 having the same distance as the straight line 503 represented by the line segments 607 and 608 in FIG. 6 is obtained. In addition to this method, it is also possible to divide the area on the basis that the areas of the data area 604 in FIG. 6 and the data area 605 in FIG. 6 are equal. Next, the means 710 of FIG. 7 obtains the data existence density D3 of the area 604 and the data existence density D4 of the area 605. The above-mentioned formula 3 is used for the method of obtaining. Next, the means 711 shown in FIG. 7 checks whether or not the data distribution densities D3 and D4 obtained previously are equal to each other within a preset error range. 7 means 71
When it is determined that they are equal to each other, the area 505 is set to the area 60.
The data is modeled using the model of Expression 2 without being divided into four and the region 605. At that time, in order to obtain the model more accurately, the regions 504 and 505 in FIG.
A weighting coefficient is set for each of the data. As the weighting coefficient here, a larger weighting coefficient is set for the data included in the area 505 than the data included in the area 504. Means 711
If there is a large difference between the data existence densities D3 and D4 described above, the areas 504, 604, and 604 shown in FIG.
A weighting coefficient of data is set for each area 605. The weighting factor setting method for each area here is
As described above, in consideration of the property that the power data that is the target of the present embodiment can be approximated to a downward convex curve or a straight line, the weighting coefficient for the region 605 is the largest and the weighting coefficient for the region 504 is the largest. The weighting factor of the region 604 is set to be small, and an intermediate value of the weighting factors for the above-described two regions is given. Using the weighting factors thus obtained, the means 714 models the data using the quadratic equation shown in Equation 2. Based on the above result, the means 715 uses the weighted least squares method to obtain the coefficient of each equation. The model created by the device 104 needs to be updated as the actual result data is determined by the demand amount actual result intake device 107. The updating method here is shown in FIG.
This can be achieved by adding the newly entered performance data to the data read in.

Next, in the prediction executing device 105, the model obtained by the above-mentioned prediction model creating / updating device 104 and the factors affecting the demand amount at the time when the demand amount prediction taken in by the device 106 is actually performed. Perform the prediction using.

Finally, the result of the device 105 is output by the prediction result output device 108. An example of the output screen will be described with reference to FIG. The screen 801 is assumed to be a graphic display as an example of an output device. By pressing buttons 803, 804, and 805 on the screen 802, the graphs 901, 902, and 903 in FIG. 9 are output, and various information is provided to the user. Graph 901
Represents the change in total daily demand. In the graph, the upper and lower limits of the prediction error that may occur at the time of prediction are shown. A graph 902 represents a fluctuation graph of the total demand amount or the individual demand amount for the week. Further, reference numeral 903 indicates the relationship between the factor affecting the demand amount and the demand amount. The area 932 is the data used for model creation, and 932 is the prediction result. From this result, it is possible to determine whether the prediction result is a specific day. Buttons 806, 807, and 808 are operation buttons. The button 806 is a model creation instruction button, the button 807 is a prediction execution button, and the button 808 is a button for issuing an instruction to modify the model when the demand amount prediction result is far from the past actual data. Also,
In the above-described one embodiment of the present invention, the actual power system state is detected and used as history data as history data. However, in addition to this, for example, virtual data history data of the present invention is used. Needless to say, can be used as a simulator for power system operation.

[0029]

As described above, the present invention has the following effects.

In a demand quantity predicting apparatus for predicting future demand quantity from the relationship between the demand quantity and the factors affecting it from the past history data, the demand quantity and the factors affecting it at the time of creating the demand quantity model. A highly accurate demand forecast result can be obtained by dividing the data based on the correlation of (1) and appropriately selecting the formulas used for the determination of the weighting factor and the forecast model.

[Brief description of drawings]

FIG. 1 is a diagram showing the features of the present invention.

FIG. 2 is a diagram showing an example of storing a past history database according to the present invention.

FIG. 3 is a flowchart of the processing of the data preprocessing device of the present invention.

FIG. 4 is an example of past history data used for demand forecasting.

FIG. 5 is a conceptual diagram showing the relationship between the influencing factors and the demand amount, and the state of division.

FIG. 6 is a conceptual diagram showing the relationship between the influential factors and the demand amount, and the state of division.

FIG. 7 is a flowchart of the process of the prediction model creation / update device.

FIG. 8 shows an example of a prediction result output device.

FIG. 9 is an example of a screen pattern displayed on the prediction result output device.

[Explanation of symbols]

101 ... Electric power system, 102 ... Past history database,
103 ... Data preprocessing device, 104 ... Prediction model creation
Updating device, 105 ... Prediction executing device, 106 ... Influencing factor capturing device, 107 ... Demand amount actual capturing device, 108
... prediction result output device, 109 ... prediction condition input device, 15
1 ... Communication line connecting power system and past history database, 152 ... Communication line connecting past history database and data preprocessing device, 153 ... Communication line connecting data preprocessing device and prediction model creating / updating device 154 ... A communication line connecting the prediction model creating / updating device and the prediction executing device, 155 ... A communication line connecting the power system and the influence factor capturing device, 156 ... A communication line connecting the influence factor capturing device and the prediction executing device, 157 ... Communication line connecting power system and demand record capturing device, 158 ... Communication line connecting demand record capturing device and prediction model creating / updating device, 159 ... Communication line connecting prediction execution device and prediction result output device , 160 ... A communication line connecting the prediction target input means and the data preprocessing device, 201 ... An example of a storage method of a past history database, 20. ... An example of storage of lower system switching information, 211 ... Column for storing total demand amount / individual demand amount classification flag, 212 ... Column for storing equipment name, 213 ... Column for storing year of capturing data, 214 ... Data Column for storing the month in which the data was captured, 215 ... column for storing the date in which the data was captured, 216 ... column for storing the time when the data was captured, 217 ... column for storing the presence or absence of a social event on the date when the data was captured 218 ... Column for storing meteorological factors, 219 ... Column for storing actual amount of active power, 220 ... Column for storing actual amount of reactive power, 251 ... Column for storing system switching source, 252 ... Store system switching destination Columns to
253 ... Column for storing system switching amount, 301 ... Initial data reading means, 302 ... Total demand amount forecasting / individual demand amount forecasting judging means, 303 ... Regular system state converting means, 304
... Active power / reactive power prediction determination means, 305 ... Active power data selection means, 306 ... Reactive power data selection means, 3
07 ... Total demand amount data selecting means, 308 ... Individual demand amount data selecting means, 309 ... Prediction target data creating device, 40
1, 402 ... Correlation diagram between influencing factors and demand amount, 501 ... Correlation diagram between influencing factors and demand amount, 502 ... Collection area of data used for model creation, 503 ... Trend line showing average tendency of model creation data , 504, 505 ... Area after data division, 601 ... Correlation diagram of influence factors and demand, 602 ...
Data collection area used for model creation, 603 ... Trend lines showing average tendency of model creation data, 604, 605 ...
Area after data division, 606 ... Trend line indicating division of model creation data, 607 ... Distance from average trend line 503, 608 ... Trend line 50 indicating division of model creation data
Distance from 6, 701 ... Target data reading means, 70
2 ... Influencing factor / demand correlation diagram creating means, 703 ... Means for obtaining average tendency of data, 704 ... Data area dividing means, 705 ... Distribution density calculating means, 706 ... Distribution density uneven distribution judging means, 707 ... Linear model selection Means, 709 ... Region dividing means, 710 ... Distribution density calculating means, 711 ... Distribution density uneven distribution determining means, 712 ... Data weighting coefficient determining means, 7
13 ... Data weighting coefficient determining means, 714 ... Quadratic model selecting means, 715 ... Model coefficient determining means, 801 ... Screen display example, 802 ... Demand amount graph display area, 803, 8
04,805 ... Screen selection buttons, 806, 807, 80
8 ... Operation selection buttons, 901, 902, 903 ... One example of screen display.

Claims (10)

[Claims] 1. A power demand amount from historical data of a power system,
In a power demand amount prediction method for predicting the power demand amount of the power system from the relationship with factors affecting the power demand amount, the power demand amount data selected from the history data and the correlation between the affecting factors The distribution density of the history data is obtained more, the history data is clustered based on the distribution density, the weighting coefficient of each history data is determined based on the clustering result, and the prediction model using the weighting coefficient is used to A power demand forecasting method characterized by forecasting power demand of a power system. 2. The power demand forecasting method according to claim 1, wherein when clustering historical data based on the distribution density, the power demand and a factor affecting it are correlated. A method for predicting power demand, characterized by dividing the history data to determine the weighting factor. 3. The power demand forecasting method according to claim 2, wherein the history data is divided according to a linear equation or a quadratic equation when dividing the history data. Prediction method. 4. The power demand forecasting method according to claim 2, wherein when dividing the history data according to the linear equation or the quadratic equation, the history data is divided by a piece of a distribution of the history data after the division. A power demand forecasting method characterized by using a linear equation or a quadratic equation. 5. The power demand forecasting method according to claim 1 or 2, wherein a specific mathematical formula is selected from a plurality of mathematical models according to the distribution density of the history data. A power demand forecasting method comprising selecting a model and using a forecasting model based on this mathematical model. 6. The power demand forecasting method according to claim 1, wherein actual past history data of the power system is used as the history data. 7. The demand forecasting method according to claim 1, wherein virtual history data of the power system is used as the history data. 8. A past history database that holds the relationship between the demand amount in the past and factors that affect the demand amount, and at the same time selects appropriate data for predicting the demand amount from the database, and at the same time, demand amounts other than the influencing factors. A data pre-processing device that removes fluctuation factors, a device that creates and updates a mathematical model for forecasting the demand amount from the data selected by the device, a factor that influences the forecast relevant day is taken from the power system, and the actual demand amount is calculated. The past demand amount and the factors affecting it are composed of a data capturing device that captures values, a prediction execution device that executes demand amount prediction based on the prediction model and the influencing factors, and an output device that outputs the prediction result. A demand forecasting device that obtains future demand by finding relationships. 9. In the demand forecasting method and the forecast model creating / updating apparatus of claim 8, the demand forecast data of the period selected from the history database and the correlation between factors affecting the past The distribution density of the history data of is calculated, the data is clustered based on the distribution density, the weighting coefficient of each data is determined based on the clustering result, and then the prediction model is created by the weighted mathematical model. A demand forecasting method and device. 10. The demand forecasting method and the forecasting model creating / updating device of claim 9, wherein an appropriate forecasting model is selected from a plurality of mathematical models according to the distribution density of the history data. And a demand forecasting method and device.