JP2005328673A - Power demand prediction device, power demand prediction system, power demand prediction program, recording medium and power demand prediction method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a power demand prediction method for predicting a power demand that obtain not only a total power demand prediction value but also a demand prediction value for a short-period variation demand, and a power demand prediction device. <P>SOLUTION: This power demand prediction device for predicting a power demand on the basis of power demand data in the past chronological sequence is provided with a demand data separating means that reads out total power demand data in the past chronological sequence from a prescribed memory and separates the total power demand data into long-period variation demand data and short-period variation demand data; a short-period future demand prediction value computing means that computes a demand prediction value of a short period future for a long period variation demand by a transition vector method on the basis of the long-period variation demand, and that computes a demand prediction value of a short-period future for a short period variation demand by a local fuzzy re-constitution method on the basis of the short period variation demand data; and a demand prediction value calculating means that calculates the total power demand prediction value by adding the demand prediction value of the computed long-period variation demand and that of the short-period variation demand. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  TECHNICAL FIELD The present invention relates to a power demand prediction method and a power demand prediction device for use in power supply and demand adjustment, and more particularly to a short-time power demand prediction method and prediction device using chaos theory.

An apparatus for predicting power demand using deterministic chaos theory has been developed (see, for example, Patent Document 1). That is, it predicts near-future data until it loses deterministic causality within a range in which the behavior of time series data related to past power demand is chaotic. On the other hand, it is described that a chaos time-series short-term prediction device using a local fuzzy reconstruction method has also been developed and can be used for power prediction (for example, see Patent Document 2).
JP 7-123590 A JP-A-9-146915

  By the way, for example, an electric furnace that is a load that consumes a large amount of power and has a steep fluctuation does not occupy a small proportion of the total power demand. The ratio of the electric furnace load to the total power demand is, for example, about 6% to about 7%. For this reason, the problem which brings about a frequency fluctuation | variation to an electric power system will arise by the rapid fluctuation | variation of the electric furnace which consumes large electric power. Therefore, in obtaining the demand forecast value for the short time ahead of the electric power, not only the predicted value of the overall total power demand, but also the demand for the short time ahead may be predicted in real time, particularly with respect to the electric furnace and the like. is necessary.

  Accordingly, an object of the present invention is to provide a power demand prediction method and a power demand prediction apparatus capable of obtaining not only the predicted value of the overall total power demand but also the demand predicted value related to the short cycle fluctuation demand. It is in.

In order to solve the above-mentioned problem, the present invention is a power demand prediction device that predicts power demand based on past time-series power demand data,
Demand data separation means for reading past time-series total power demand data from a predetermined storage unit and separating the total power demand data into long-period fluctuation demand data and short-period fluctuation demand data;
For each of the long cycle fluctuation demand data and the short cycle fluctuation demand data, smoothing means for smoothing time-series fluctuations;
Based on the smoothed long-period fluctuation demand data, a short-term ahead demand forecast value is calculated for the long-period fluctuation demand by the transition vector method, and the local fuzzy reconstruction is performed based on the smoothed short-period fluctuation demand data. A short-term ahead demand forecast value calculating means for calculating a short-term ahead demand forecast value for short cycle fluctuation demand by the method,
A demand forecast value calculating means for calculating a demand forecast value of total power by adding the demand forecast value of the calculated long cycle fluctuation demand and the demand forecast value of the short period fluctuation demand. And

  In order to obtain a demand forecast value for a short time ahead of power in real time, the calculation is performed after separating into long-period fluctuation demand data and short-period fluctuation demand data in advance. For this reason, the overall forecast value of total power demand can provide an accurate forecast value that takes into account both long-period fluctuation demand and short-period fluctuation demand fluctuations. A value can be obtained. As a result, both the total power demand and short cycle fluctuation demand can be predicted in real time in real time, enabling high accuracy in manual control by humans and automatic control such as ELD control (Economic Load Dispatching Control). It is possible to adjust supply and demand accurately. For this reason, it is possible to improve the power quality by suppressing the frequency fluctuation, and thus the generator can be adjusted economically.

=== Hardware configuration ===
An embodiment of the present invention will be described with reference to the drawings. First, FIG. 1 shows a configuration diagram of a power demand prediction system including a demand prediction server 100 as a power demand prediction apparatus that predicts power demand based on past time-series power demand data. As shown in FIG. 1, a prediction result display personal computer (prediction result display computer) 200 is connected to the demand prediction server 100 via a LAN, and the server 100 and the personal computer 200 constitute a demand prediction system. The demand prediction server 100 is connected with a central power supply command station recording computer 300 via a LAN. The computer 300 outputs data related to the current value of the actual power demand to the demand prediction server 100. This server 100 is the actual value of total power demand, short cycle fluctuation demand, and current value of long cycle fluctuation demand, and the demand of each of long cycle fluctuation demand data and short cycle fluctuation demand (electric furnace load) data. Various data such as the predicted value and the predicted value of the total power demand are output to the personal computer 200. The personal computer 200 displays the received various data on the display screen as shown in the prediction image screen of FIG.

=== Electricity demand forecast method ===
A power demand prediction method executed by the demand prediction server 100 of this embodiment will be described with reference mainly to the flowchart of FIG. First, sampling data related to the total power demand (total demand in the figure) including the electric furnace load data is sequentially transmitted from the central power supply command station recording computer 300 to the demand prediction server 100, and the past time-series power demand. The data is accumulated in a database (storage unit) (S100). The server 100 reads the sampling data from the database, subtracts the electric furnace load from the total power demand at each past time point, calculates sampling data related to the long-period variable load, and stores it in the database. As a result, time-series sampling data obtained by separating the total power demand at each past time point into a long-period fluctuation load (long-period fluctuation demand) and an electric furnace load (short-period fluctuation demand) is obtained (S200 → S250 → S300). , S200 → S400). Prior to step S300 for obtaining sampling data related to the long-period fluctuation load, as will be described later, a process for smoothing the time-series fluctuation of the sampling data is executed (S250).

  The separation of the sampling data relating to the total power demand will be described. An electric furnace load (hereinafter referred to as a short cycle variable load) accompanied by a rapid fluctuation repeats a rapid fluctuation in units of several minutes to several tens of minutes. On the other hand, the remaining load (hereinafter referred to as long-period variable load) excluding this load from the total demand is a gradual fluctuation per day as seen in the daily load curve, and the behavior as time series data is different. . The actual power demand data used in this example has a large proportion of short-period fluctuating loads in the total demand, reaching a maximum of several percent, so prediction accuracy can be improved by separately predicting this. I can expect. Therefore, a method is adopted in which the separated short-cycle variable load and long-cycle variable load are predicted separately, and the sum of these is used as the total demand forecast value.

  Next, as shown in S250 and S600 of FIG. 3, time-series fluctuations are smoothed for the data to be obtained as the sampling data of the long-period fluctuation load in S300 and the sampling data of the electric furnace load obtained in S400. Execute the process to convert. In the present embodiment, the process of step S300 is executed after the process of step S250. However, the present invention is not limited to this, and the process of step S250 may be executed after the process of step S300. . In this smoothing process, a generally known smoothing method such as an exponential smoothing method or a moving average method may be used. In addition, as a method of smoothing a short-period fluctuation load accompanied by a sudden fluctuation without phase shift, a technique of connecting data represented by a graph with an envelope can be used. That is, as shown in the graph showing the concept of smoothing in FIG. 4A, a maximum value is obtained for each predetermined data section for sampling data (“Original Data” in the figure), and this is calculated at a predetermined interval. It is performed while shifting each time (numbers 1 to 7 circled in the figure), and each maximum value obtained is interpolated with a straight line ("Line by Maximums" in the figure). Further, the same processing is performed for the minimum value (“Line by Minimums” in the figure). Finally, these average values are defined as smoothed data (“Smoothed Data” in the figure).

This algorithm is briefly described as Step 1 to Step 4 below.
・ Step1: Find the maximum and minimum values from the time range set in the time series data.
・ Step 2: Perform the same operation as Step 1 while shifting the position one by one.
・ Step3: Connect the maximum and minimum values with a straight line next to each other.
Step 4: Calculate the average value of two lines at all time steps.

  In addition, the graph of FIG.4 (b) is the result of the smoothing by the moving average method which is the conventional typical smoothing method with respect to the data of actual short period fluctuation load (in the figure, "Actual Furnance Load"). And the result of the smoothing by a present Example is shown. When these are compared, a time lag occurs in the moving average method (refer to “Data with Moving Averages” in the figure), while a time lag hardly occurs in the smoothing method according to the present embodiment (“Data with Proposed Smooting” in the figure). reference). In addition, the fluctuation of about several minutes to be removed is smoothed well, but the characteristics of long-period fluctuation of about several tens of minutes are often left.

  Next, as shown in S700 and S800 of FIG. 3, a process of calculating a demand forecast value for a short time ahead by the transition vector method for the long-period fluctuation demand data smoothed in S250, and the electricity smoothed in S600 The furnace load data will be described with respect to a process for calculating a demand forecast value ahead for a short time by the local fuzzy reconstruction method.

<<< Local Fuzzy Reconstruction Method >>>
First, the local reconstruction method is well known and is as follows. First, let y (t) be the time-series data at time t and τ be the time delay. Next, using the delayed coordinate system, the vector X (t) = (y (t), y (t-τ), ..., y (t- (n-1) τ)) T (embedded vector) to make. This represents one point in the n-dimensional reconstruction state space (embedding dimension). Note that the current time is t = i. Further, by setting “h” as the number of past data and reducing the value of t by “1” from i, the vector X (t) is shifted one by one to the past, and an embedded vector is created. That is, “t = i, i−1, i−2,..., I − ((h−1) − (n−1) τ)”. As a result, a trajectory can be drawn in the n-dimensional reconstruction state space. In this n-dimensional reconstruction state space, a number of vectors (neighboring vectors) near the vector X (i) (latest vector) composed of the latest data are past data. In other words, the state at the predicted time is known. Therefore, this property is used to estimate and predict the near-future trajectory of the latest vector. Here, as the neighborhood vector, the number (1, 2,..., M) set in order from the closest Euclidean distance to the latest vector X (i) is selected.

In addition, for the neighborhood vector obtained by this local reconstruction method, the neighborhood vector search process is executed within a limited range. That is, in order to improve the calculation speed and enable real-time prediction, the neighborhood vector search range is limited as shown in FIG. Since power demand is more dependent on time, the neighborhood vector search range is limited to the same time in the past and the vicinity thereof. Assuming that the number of data points per day is day and the number of past data embedding days (the number of past data days used for forecasting) is 1, 2, ..., emday, i-em * day (where em = 1,2,3, ..., emday). Further, in order to set a search range around the same time, with respect to the thinning interval of the search in the past direction and the future direction, the time point from the present time pointed back to the past to the past direction τ _bf based on the same time point and the future direction τ The range up to _af . Also, with respect to data from the current time point in the past to the past direction bf * τ _bf and data to the future direction af * τ _af , the past direction thinning interval τ _bf and the future direction thinning interval τ Search for each _af (bf and af are natural numbers). The base point uses the first component for both the latest vector and the past embedded vector.

Further, in this embodiment, when calculating the demand forecast value ahead for a short time by the local fuzzy reconstruction method, by searching the neighborhood vector based on the attribute of the day of the week, the long-period fluctuation demand data by day of the week and the day of week Generate total electricity demand data. Among these, the neighborhood vector search method based on the day of week attribute will be described. First, if the day of the week attribute is inserted after the last component of the n-dimensional vector and the day of week attribute is week (for example, weekday = 1, Saturday = 2, Sunday = 3), the following vector is obtained. X (i): the latest vector, X (i) = (x 1, x 2, x 3, ···, x n, week), X (k): past of embedded vector, X (k) = (k ₁ , k ₂ , k ₃ , ..., k _n , week)

However, the day of the week to which the predicted time ahead (s step) belongs from the latest vector first component x ₁ is the day of week attribute of the latest vector X (i). Also, the day attributes of X (k) of the first predicted time destination component k ₁ (s steps) embedded belongs day vector past embedding vectors X (k). If the day of week attribute of X (i) matches the day of the week attribute of the same time, the Euclidean distance from the data up to the past direction bf * τ _bf and the future direction af * τ _af Calculate On the other hand, if the day of week attributes do not match, the Euclidean distance is not calculated. Then, in the calculated Euclidean distance, the number (1, 2,..., M) set in ascending order is selected as a neighborhood vector. FIG. 9 shows an explanatory diagram of an example of the neighborhood vector search using the day of week attribute, and FIG. 10 shows a search procedure flowchart.

  That is, as shown in FIG. 9, if the current day is i, the day of week attribute (for example, weekday = 1, Saturday = 2, holiday (Sunday or holiday) = 3), for example, at the same time one day before, If it is the same as the day of the week attribute, the search is performed. For example, the search is not performed if they are different at the same time two days ago. Also, as shown in FIG. 10, when a day of the week attribute is added to the latest vector (S10), a past embedded vector having the same day of the week attribute as the latest vector is read from the database. Then, the Euclidean distance is calculated for the latest vector and the past embedded vector having the same day attributes (S20). Next, extraction processing is executed with the set number (m) in the closest (smallest) order among the calculated Euclidean distances as neighborhood vectors (S30). FIG. 11 is a graph showing the effectiveness of searching for a neighborhood vector based on the day of week attribute. In the figure, it can be seen that the graph of “with data day classification” generally has a smaller average prediction error than the graph of “no data day classification”.

  Next, calculation of a demand forecast value ahead for a short time by the local fuzzy reconstruction method will be described. In FIG. 5, when predicting using these neighborhood vectors, for a plurality of vectors selected as neighborhood vectors, the closer to the latest vector, the greater the effect on the predicted trajectory, and the farther the effect, the smaller the effect. Express with fuzzy rules and make predictions based on this rule. (This part is a feature of the local fuzzy reconstruction method. Except for this part, the local reconstruction method is a prediction method that uses the Takens embedding theorem. This is a common method.)

The local fuzzy reconstruction method will be specifically described with reference to FIGS. First, as described above, after reconstructing time-series data into a multidimensional state space using a delayed coordinate system, X (k ₁ ), X (k ₂ ) are used as neighborhood vectors in order of Euclidean distance closer to the latest vector neighborhood. , X (k ₃ ) (when the neighborhood vector selection number m = 3) is selected. Subsequently, a membership function is determined for each component of these vectors as shown in FIG. In FIG. 6, among the selected plurality of neighboring vector l-dimensional components (1 ≦ l ≦ n, n: embedded dimension), the component having the furthest Euclidean distance from the latest vector l-dimensional component is graded 0.5. . The grades of the remaining neighborhood vector l-dimensional components are both determined to be 0.5 or more according to the Euclidean distance from the latest vector l-dimensional component (see “Equation 1” below). Finally, using these grades, a prediction vector is determined by “Expression 2” described later. FIG. 7 shows a flowchart of these prediction procedures. First, time-series data at equal sampling intervals is accumulated (S1). Then, processing for embedding time-series data in the multidimensional state space is executed (S2). Next, a search for neighborhood vectors is performed (S3), and a grade is calculated based on “Equation 1” described later (S4), and then a predicted value is determined (S5).

Where X (i): latest vector, X (i) = (x ₁ , x ₂ , x ₃ ,..., X _n ), X (k _j ): neighborhood vector, X (k _j ) = ( k _{1, j} , k _{2, j} , k _{3, j} ,..., k _{n, j} ), X (k _j + s): vector s step ahead of neighborhood vector, X (k _j + s) = (k _{1, j} ^s , k _{2, j} ^s , k _{3, j} ^s ,..., k _{n, j} ^s ), X (i + s): prediction vector, X (i + s) = (x ₁ ^s , x ₂ ^s , x ₃ ^s ,..., x _n ^s ), μ _l (j): membership function grade, l = 1,..., n, n: embedding dimension, j = 1,. .., m, m: Number of neighbors

Here, the calculation example of the actual demand forecast value is shown as follows. It can be calculated only by calculating x ₁ ^s which is the first component of the prediction vector. In this calculation example, the embedding dimension is 4 and the neighborhood vectors are 3. Then, the latest vector and the neighborhood vector are set as follows.
Latest vector X (i) = (790, 690,590, 490), neighborhood vector X (k ₁ ) = (800, 700, 600, 500), first component grade = 1 / (1+ | 800-790 | / | 820-790 |) = 0.75, neighborhood vector X (k ₂ ) = (810, 710, 610, 510), similarly, first component grade 0.6, neighborhood vector X (k ₃ ) = (820,720,620, 520)
Further, the grade of the first component is set to 0.5 as in the case of determining the membership function.

Then, the vector of the predicted time ahead (s step ahead) of the neighborhood vector is as follows.
X (k ₁ + s) = (850, 750, 650, 550), X (k ₂ + s) = (860, 760, 660, 560), X (k ₃ + s) = (870, 770, 670 , 570)
Then, the predicted value x ₁ ^s of the first component is obtained as follows. Predicted value x ₁ ^s = (0.75 x 850 + 0.6 x 860 + 0.5 x 870) / (0.75 + 0.6 + 0.5) = 858.6

<<< Transition vector method >>>
For details of the transition vector method of this embodiment, see Kiyoji Kawauchi and Hiroshi Sasaki: IEEJ Transactions B Vol. 124 No. 1 (2004) pp. 77-83 and the like.

The transition vector method of the present embodiment, first, the most recent vector X (i) for the m neighborhood vector X (k ₁₎ in the local fuzzy reconstruction method described above, ..., and X (k _m), the neighborhood vector S step ahead vector X (k ₁ + s), ..., X (k _m + s) (prediction time ahead vector) vector Z (k ₁ + s, k ₁ ) (= X (k ₁ + s) -X (k ₁ )), ..., Z (k _m + s, k _m ) (= X (k _m + s) -X (k _m )) is the s step ahead of each neighboring vector Defined as the transition vector of.

Further, in the transition vector method of the present embodiment, a function F (Z (k ₁ + s, k ₁ ),... For weighting the transition vector using the transition vector of each neighborhood vector to the s step destination as a variable. _{, Z (k m + s,} k m)) is defined as the transition vector W used in the prediction (i + s).

  Furthermore, in the transition vector method of the present embodiment, the transition vector W (i + s) used for prediction is combined with the latest vector X (i), thereby corresponding to the prediction vector in the local fuzzy reconstruction method described above. A prediction vector X (i + s) (= X (i) + W (i + s)) is obtained.

By the way, the function F (Z (k ₁ + s, k ₁ ),..., Z (k _m + s, k _m )) for weighting the transition vector is the selected m neighborhood vectors X (k _{1), ..., X (k} m) is generated from the membership functions defined in great influence so by the expected value as is in the vicinity to the latest vector X (i).

The membership function μ _l (j) (1 ≦ l ≦ n, 1 ≦ j ≦ m) corresponding to, for example, the l-th element w _l of W (i + s) This is the same function as the function form shown on the right side of “Equation 1” used in the construction method. However, in this function, “Eu _l (j)” is defined as the Euclidean distance of the 1st dimension between the latest vector and neighboring vectors, and “Eu _l ^Max ” is Eu _l (j) (1 ≦ j ≦ m) It is defined as the maximum value of.

Also, the relationship between, for example, the l-dimensional element w _l of w (i + s) and the above-mentioned membership function μ _l (j) is expressed by the following equation (2) used in the local fuzzy reconstruction method described above. It is defined by a relational expression in which “X _l ^s ” shown on the left side is replaced with w _l and “k _{l, j} ^s ” shown on the right side of the “Equation 2” is replaced with z _{l, j} . However, z _{l, j} is the l-th element of the transition vector Z (k _j + s, k _j ) corresponding to the j-th neighborhood vector.

  In the transition vector method of the present embodiment, the day of the week information is assigned to the embedded vector in the n-dimensional state reconstruction space as in the case of the local fuzzy reconstruction method described above, and is the same as the prediction date. A neighborhood vector is selected from embedded vectors to which day information is assigned. The day of the week information is classified into three types, weekdays, Saturdays, and Sundays, based on the above-mentioned trend of time series data. Holidays occurring on weekdays and Saturdays are attributed to Saturdays, holidays occurring on Sundays, Golden Weeks, and The year-end and New Year holidays are attributed to Sunday.

That is, as illustrated in FIG. 12, for the latest vector X (i) corresponding to “Sunday” (black circle in FIG. 12), neighboring vectors X (k ₁ ), X corresponding to “Sunday”. (k ₂ ), X (k ₃ ) (black circle in FIG. 12) are selected, and even if the Euclidean distance is short, the neighborhood vector corresponding to “Saturday” (white circle in FIG. 12) and the neighborhood corresponding to “weekday” A vector (double circle in FIG. 12) is not selected. As a result, as in the case of the local fuzzy reconstruction method described above, it is possible to obtain a predicted value even for power demand that varies depending on the day of the week.

Further, in the transition vector method of the present embodiment, as described above, first, the latest vector X is derived from the vector (embedded vector) embedded in the n-dimensional state reconstruction space using the delayed coordinate system for the time series data. For (i), m neighborhood vectors X (k ₁ ),..., X (k _m ) are selected in ascending order of the Euclidean distance. Next, from the m selected neighborhood vectors, further neighborhood vectors may be selected in ascending order of the Euclidean distance between the time differentiated vector and the latest vector time differentiated vector. That is, in the transition vector method of the present embodiment, neighborhood vectors may be further narrowed down in time-series data obtained by time differentiation. Here, since the time series data is discrete, it is assumed that time differentiation is performed by, for example, (y (t) -y (tu)) / u. However, y (t) is an observed value, t is time, and u is a natural number.

  Furthermore, in the transition vector method of the present embodiment, one or more of the embedding dimension (n), time delay (τ), and number of neighbors (m) used in the local fuzzy reconstruction method described above are used as parameters. The parameter may be determined for each prediction time so as to minimize the error between the demand prediction value and the actual measurement value. The error is, for example, an absolute value average error or a square average error. That is, if prediction is performed while the parameters n, τ, and m that are variable at each prediction time are searched (optimized) so as to minimize the above-described error, the power demands that greatly change with time More accurate prediction is possible. Here, the search method described above may be a rigorous solution that hits a predetermined range of n, τ, and m, or a metaheuristic method for obtaining a high-precision approximate solution such as a tab search. . Note that such parameter determination may be performed in a local reconstruction method including a local fuzzy reconstruction method.

According to the transition vector method of the present embodiment, for example, prediction of long-period fluctuation demand of power when there is no example in the past can be performed with high accuracy. This unprecedented prediction means that, as illustrated in FIG. 13, the Euclidean distance between the latest vector X (i) and the neighboring vectors X (k ₁ ), X (k ₂ ), X (k ₃ ) is It is a prediction of a state in which the neighborhood vector is not only separated but also biased in one direction. In this case, if the local fuzzy reconstruction method used for the short-period fluctuation demand forecast described above is applied, vectors X (k ₁ + s), X (k ₂ + s), X Since the prediction is performed using only (k ₃ + s), there is a possibility that the prediction is performed to X (i + s) _old (FIG. 13). On the other hand, according to the transition vector method, since the vector W (i + s) obtained by the transition vector is synthesized with the latest vector X (i), the prediction result is X (i + s) _new ( FIG. 13).

  Therefore, particularly when the maximum demand is continuously updated, i.e., when the distance between the neighborhood vector and the latest vector increases, the prediction of long-period fluctuation demand of power by the transition vector method of the present embodiment is as follows: Higher accuracy.

<<< Addition of long-period fluctuation demand forecast value and short-period fluctuation demand forecast value >>>>
As shown in S1000 of FIG. 3, the long-period fluctuation demand forecast value obtained in S700 and the short-period fluctuation demand forecast value obtained in S800 are added, and the total demand forecast value (demand forecast value of total power). Is calculated and stored in the database.

  The demand prediction server 100 executes the processes from S100 to S1100 described so far in real time. At this time, the demand prediction server 100 executes this real-time processing while sequentially reflecting the current value of the power demand received from the central power supply command station recording computer 300.

  Then, the demand prediction server 100 includes a total demand prediction value (a prediction value of total power demand), a long-period variable load (demand) prediction value, a day-of-week attribute search total demand prediction value (a prediction value of total power demand data by day of the week), The day-of-week attribute search long-period variable load (demand) predicted value and electric furnace load (short-cycle variable demand) predicted value are read from the database in real time as needed and output to the prediction result display personal computer 200. The personal computer 200 draws and displays the latest data and history data of each received predicted value in real time as shown in the display screen of FIG. 2 (S1200).

In this embodiment, the following operational effects are obtained.
In order to obtain a demand forecast value for a short time ahead of power in real time, the calculation is performed after separating into long-period fluctuation demand data and short-period fluctuation demand data in advance. For this reason, the overall forecast value of total power demand can provide an accurate forecast value that takes into account both long-period fluctuation demand and short-period fluctuation demand fluctuations. A value can be obtained. As a result, both the total power demand and short cycle fluctuation demand can be predicted in real time in real time, enabling high accuracy in manual control by humans and automatic control such as ELD control (Economic Load Dispatching Control). It is possible to adjust supply and demand accurately. For this reason, it is possible to improve the power quality by suppressing the frequency fluctuation, and thus the generator can be adjusted economically.

  Moreover, sample-like data can be obtained by smoothing each data which uses a demand predicted value for a calculation by a local fuzzy reconstruction method. Therefore, it is possible to obtain a highly accurate demand forecast value by calculating the demand forecast value based on sample data obtained by smoothing.

  Furthermore, an appropriate predicted value of power demand can be obtained even with respect to fluctuations in power consumption of an electric furnace, which is a load with excessive power consumption.

  Furthermore, by limiting the search range of the neighborhood vector, the processing load required for this search is reduced. As a result, the calculation speed is improved and each demand forecast value can be calculated in real time.

  In addition, it is possible to obtain a predicted value for power demand that varies depending on the day of the week.

  In addition, particularly when the maximum demand is continuously updated, the long-period fluctuation demand of power is predicted with higher accuracy by the transition vector method.

1 is a configuration diagram of a demand prediction system including a demand prediction server 100 as a power demand prediction apparatus according to an embodiment of the present invention. It is a schematic diagram which shows a mode that the prediction result of the electric power demand which concerns on one Example of this invention was displayed on the display screen of a personal computer. It is a flowchart which shows the prediction method of the electric power demand which concerns on one Example of this invention. It is a graph regarding the process of smoothing in the electric power demand prediction which concerns on one Example of this invention, (a) shows the concept of the process which smoothes a time-sequential fluctuation | variation about the sampling data of an electric furnace load, (b ) Shows the smoothed data by the conventional moving average method and the smoothed data by the present embodiment with respect to the data of the actual short cycle variable load. It is a conceptual diagram which shows the demand prediction method of the short time ahead by the local fuzzy reconstruction method in the prediction of the short period fluctuation demand of the electric power which concerns on one Example of this invention. It is a graph which shows the membership function used by the local fuzzy reconstruction method in the prediction of the short period fluctuation demand of the electric power which concerns on one Example of this invention. It is a flowchart which shows the process of the calculation of the demand forecast value of short time ahead by the local fuzzy reconstruction method in the prediction of the short period fluctuation demand of the electric power which concerns on one Example of this invention. It is a schematic diagram showing a mode that the neighborhood vector search range by a local fuzzy reconstruction method is limited in the prediction of the short period fluctuation demand of the electric power which concerns on one Example of this invention. It is explanatory drawing which shows an example of the neighborhood vector search using a day-of-week attribute in the electric power demand prediction which concerns on one Example of this invention. It is a flowchart which shows the search procedure of the neighborhood vector using a day-of-week attribute in the electric power demand prediction which concerns on one Example of this invention. It is a graph which shows the effectiveness in the case of searching a neighborhood vector based on the attribute of a day of the week in the electric power demand prediction which concerns on one Example of this invention. It is a conceptual diagram which shows an example of the neighborhood vector search corresponding to Sunday in the prediction of the long-period fluctuation demand of the electric power which concerns on one Example of this invention. It is a conceptual diagram which shows the demand prediction of the short time ahead by each of the local fuzzy reconstruction method and the transition vector method in prediction of the long-period fluctuation demand of the electric power which concerns on one Example of this invention.

Explanation of symbols

100 Demand forecast server 200 PC for forecast result display (computer for forecast result display)
300 Central power supply command center recording computer

Claims (13)

A power demand prediction device that predicts power demand based on past time-series power demand data,
Demand data separation means for reading past time-series total power demand data from a predetermined storage unit and separating the total power demand data into long-period fluctuation demand data and short-period fluctuation demand data;
Based on the long cycle fluctuation demand data, a demand forecast value of a short time ahead is calculated for the long cycle fluctuation demand by the transition vector method,
Based on the short-period fluctuation demand data, a short-term fluctuation demand is calculated for a short-period fluctuation demand by a local fuzzy reconstruction method.
Short-term future demand forecast value calculation means,
A demand forecast value calculation means for calculating a demand forecast value of total power by adding the demand forecast value of the calculated long cycle fluctuation demand and the demand forecast value of the short period fluctuation demand;
A power demand prediction apparatus comprising:   The transition vector method uses, for the local fuzzy reconstruction method, a vector generated based on a transition vector that is a difference vector between a neighborhood vector used in the local fuzzy reconstruction method and a vector ahead of the prediction time of the neighborhood vector. The power demand prediction apparatus according to claim 1, wherein the prediction vector used for the local fuzzy reconstruction method is generated by combining with the latest vector to be generated.   When calculating the demand forecast value by the transition vector method, the short-term ahead demand forecast value calculation means is configured to convert the neighborhood vector into a time-differentiated vector of the neighborhood vector and a local fuzzy reconstruction method. The power demand prediction apparatus according to claim 2, wherein a process of selecting a new neighborhood vector is executed in ascending order of the Euclidean distance from the time-differentiated vector of the latest vector used.   The short time ahead demand forecast value calculation means is an embedded dimension used in the local fuzzy reconstruction method when calculating the demand forecast value by the local reconstruction method including the transition vector method or the local fuzzy reconstruction method, The parameter consisting of one or more of the time delay and the number of neighbors is determined for each prediction time so as to minimize an error between the demand forecast value and the actual measurement value. Power demand forecasting device.   The smoothing means for smoothing time-series fluctuations for each of the long-period fluctuation demand data and the short-period fluctuation demand data used in the short-term prior demand forecast value calculation means is further provided. The electric power demand prediction apparatus in any one of 1-4.   2. The short time ahead demand forecast value calculating means, when calculating the demand forecast value by the local fuzzy reconstruction method, executes a neighborhood vector search process within a limited range. The electric power demand prediction apparatus in any one of -5.   When calculating the demand forecast value by the transition vector method or the local fuzzy reconstruction method, the short time ahead demand forecast value calculation means searches for the neighborhood vector based on the attribute of the day of the week as one component of the neighborhood vector. Thus, the long-term fluctuation demand data for each day of the week and / or the total power demand data for each day of the week is generated. 7. The power demand prediction apparatus according to claim 1, wherein: The power demand prediction device is configured with a server, and a prediction result display computer is connected to the power demand prediction device,
The power demand prediction device is at least one of the long-period fluctuation demand data, the short-period fluctuation demand data, the total power demand, the day-long long-period fluctuation demand data, and the day-by-day total power demand data. The power demand prediction apparatus according to claim 7, wherein data relating to the prediction value is output to the prediction result display computer.   The power demand prediction apparatus according to any one of claims 1 to 8, wherein the short cycle fluctuation demand data is load data of an electric furnace that consumes power. A power demand prediction system including a power demand prediction device that predicts power demand based on past time-series power demand data, and a prediction result display computer connected to the power demand prediction device,
The power demand prediction device
Demand data separation means for reading past time-series total power demand data from a predetermined storage unit and separating the total power demand data into long-period fluctuation demand data and short-period fluctuation demand data;
Based on the long cycle fluctuation demand data, a demand forecast value of a short time ahead is calculated for the long cycle fluctuation demand by the transition vector method,
Based on the short-period fluctuation demand data, a short-term fluctuation demand is calculated for a short-period fluctuation demand by a local fuzzy reconstruction method.
Short-term future demand forecast value calculation means,
A demand forecast value calculation means for calculating a demand forecast value of total power by adding the demand forecast value of the calculated long cycle fluctuation demand and the demand forecast value of the short period fluctuation demand;
While comprising
The prediction result display computer includes:
Data on at least one of the forecast values among the long-period fluctuation demand data, the short-period fluctuation demand data, the total power demand, the long-period fluctuation demand data by day of the week, and the total power demand data by day of the week. Receiving from the power demand prediction device, and displaying the predicted value on a display screen,
A power demand forecasting system characterized by this. On the computer,
In order to forecast power demand based on past time series power demand data,
Demand data separation means for reading past time-series total power demand data from a predetermined storage unit and separating the total power demand data into long-period fluctuation demand data and short-period fluctuation demand data;
Based on the long-period fluctuation demand data, a short-term fluctuation demand is calculated for the long-period fluctuation demand by the transition vector method, and the short-period fluctuation demand is short by the local fuzzy reconstruction method based on the short-period fluctuation demand data. By calculating the short-term demand forecast value calculation means for calculating the demand forecast value ahead of time, and the calculated demand forecast value of the long-period fluctuation demand and the demand forecast value of the short-period fluctuation demand, Demand forecast value calculation means for calculating the demand forecast value of total power,
Demand forecasting program for running On the computer,
In order to forecast power demand based on past time series power demand data,
Demand data separation means for reading past time-series total power demand data from a predetermined storage unit and separating the total power demand data into long-period fluctuation demand data and short-period fluctuation demand data;
Based on the long-period fluctuation demand data, a short-term fluctuation demand is calculated for the long-period fluctuation demand by the transition vector method, and the short-period fluctuation demand is short by the local fuzzy reconstruction method based on the short-period fluctuation demand data. By calculating the short-term demand forecast value calculation means for calculating the demand forecast value ahead of time, and the calculated demand forecast value of the long-period fluctuation demand and the demand forecast value of the short-period fluctuation demand, Demand forecast value calculation means for calculating the demand forecast value of total power,
The computer-readable recording medium which recorded the electric power demand prediction program for performing. A power demand prediction method for predicting power demand based on past time series power demand data,
The computer separates the past time-series total power demand data into long-period fluctuation demand data and short-period fluctuation demand data,
For each of the long-period fluctuation demand data and the short-period fluctuation demand data, the computer smoothes time-series fluctuations,
Based on the smoothed long-period fluctuation demand data, the computer calculates a short-term demand forecast value for the long-period fluctuation demand by the transition vector method, and based on the smoothed short-period fluctuation demand data, Calculate the demand forecast value ahead of short period fluctuation demand by fuzzy reconstruction method,
The computer calculates the predicted value of the total power demand by adding the calculated demand forecast value of the long-period fluctuation demand and the demand forecast value of the short-period fluctuation demand.
A power demand prediction method characterized by that.

Images