JP2018110016A - Wholesale power price prediction system and wholesale power price prediction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of predicting a wholesale power price.SOLUTION: A wholesale power price prediction system selects, on the basis of an attribute prediction value being a prediction value of attribute on the prediction object day and a decision tree showing a relation between the attribute and a plurality of clusters, an object cluster corresponding to the attribute prediction value from among the plurality of clusters, selects one specific prediction system from among a plurality of prediction systems on the basis of the object cluster and prediction system switching information, and calculates a time-series wholesale power price on the prediction object day as a time series prediction value using the specific prediction system according to a plurality of time-series actual values on the observation days.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a wholesale power price prediction system.

  In recent years, in the wholesale power trading market, the expansion of power trading in which prices are determined by a large number of participants is expected. It is desirable for each participant to be able to predict the price transition of the wholesale power trading market for smooth trading.

  Technologies for predicting future price trends in the wholesale power trading market are known.

  According to Patent Literature 1, time series information on power demand observed in the past is decomposed into fluctuations that occur periodically and fluctuations that occur randomly, and power is obtained by synthesizing predicted values of the fluctuations. It is described that it is possible to calculate the predicted value of the wholesale power price by calculating the predicted value of demand, multiplying the predicted value of power demand by a specific coefficient and adding a constant.

  Further, according to Patent Document 2, a multiple regression model is identified based on history information on wholesale power prices and history information on preset explanatory variables (day of week, temperature, crude oil price), and the identified multiple regression model is used. It is stated that it is possible to calculate the forecast value of wholesale electricity price.

JP 2004-145396 A JP 2008-084095 A

  However, in the conventional technology, when the explanatory variable set in advance is not the dominant factor that fluctuates the wholesale power price, or when there is a dominant factor other than the preset explanatory variable, etc. Accuracy is reduced. For example, when the price suddenly changes due to a factor that cannot be explained, the prediction accuracy is greatly reduced.

  The present invention has been made in consideration of the above points, and an object thereof is to improve the prediction accuracy of wholesale power prices.

  In order to solve the above-described problem, a wholesale power price prediction system according to an aspect of the present invention is a wholesale power price prediction system that predicts a wholesale power price on a prediction target date, and a plurality of past power price predictions from the prediction target date. About time-series wholesale power price observed on each observation date, time series actual value is stored, classification information indicating clusters corresponding to the plurality of observation days is stored, and attributes affecting wholesale power price Storing an attribute prediction value which is a prediction value of the attribute on the prediction target date, storing a decision tree indicating a relationship between the attribute and the plurality of clusters, and a plurality of predetermined prediction methods with the plurality of clusters Based on the attribute prediction value and the decision tree, a target cluster corresponding to the attribute prediction value is selected from the plurality of clusters based on the attribute prediction value and the decision tree. Based on the target cluster and the prediction method switching information, one specific prediction method is selected from the plurality of prediction methods, and using the specific prediction method, the time series actual values of the plurality of observation days, A calculation unit that calculates a time-series wholesale power price of the prediction target date as a time-series predicted value.

The structure of the power transaction system 4 of embodiment of this invention is shown. A functional configuration of the wholesale power price prediction system 6 is shown. The structure of the wholesale power price prediction system 6 is shown. Price prediction information 124A is shown. Price information 103A is shown. The price classification information 104A is shown. The prediction method switching information 107A is shown. External information 111A is shown. A conceptual diagram of decision tree information 105A is shown. Another form of the decision tree information 105A is shown. The operation of the price classification unit 101 is shown. The feature amount extracted by the price classification unit 101 is shown. The clusters classified by the price classification unit 101 are shown. The operation of the prediction method switching unit 122 is shown. The price prediction process of the prediction method A is shown. The price prediction process of the prediction method B is shown. The price prediction process of the prediction method C is shown. The price prediction process of the prediction method D is shown.

  Embodiments of the present invention will be described with reference to the drawings.

(1) Overall configuration

  FIG. 1 shows a configuration of a power trading system 4 according to an embodiment of the present invention.

  The power trading system 4 is a system for trading participants to conduct power trading with other trading participants in the power trading market. The power trading system 4 includes a trading participant system 1 used by a trading participant, a trading market manager system 2 used by a trading market operator, and an external information manager system 3 used by an external information manager.

  Trading participants are businesses that conduct power transactions in the power trading market, such as power retailers and power aggregators. The transaction participant system 1 includes a wholesale power price prediction system 6 and an information input / output terminal 13. The wholesale power price prediction system 6 includes a price information management device 10, an external information management device 11, and a price information prediction device 12.

  A trade market operator is a business operator that manages information necessary for trading participants to conduct power trading. The transaction market operator system 2 includes a price information distribution terminal 20 for distributing price information in the electric power transaction market.

  The external information manager is a company that manages external information such as weather information and economic information that can cause fluctuations in the wholesale power price. The external information manager system 3 includes an external information distribution terminal 30 for distributing external information.

  FIG. 2 shows a functional configuration of the wholesale power price prediction system 6.

  The price information management apparatus 10 includes an information input / output unit 100, a price classification unit 101, a decision tree creation unit 102, a prediction method switching information creation unit 106, a price information storage unit 103, a price classification information storage unit 104, and a decision tree information storage unit. 105. A prediction method switching information storage unit 107 is included. The information input / output unit 100 acquires price information 103 </ b> A from the price information distribution terminal 20 and stores it in the price information storage unit 103. The price information 103A includes time series information of wholesale power prices on each observation date. The price classification unit 101 creates price classification information 104A based on the price information 103A and stores it in the price classification information storage unit 104. The price classification information 104A includes a cluster number for time series information of wholesale power prices on each observation date. The decision tree creation unit 102 creates decision tree information 105A based on the price classification information 104A and the external information 111A, and stores it in the decision tree information storage unit 105. The decision tree information 105A indicates a decision tree for determining a cluster number from the external information 111A. The prediction method switching information creation unit 106 creates prediction method switching information 107A based on the price information 103A, price classification information 104A, and external information 111A, and stores the prediction method switching information 107A in the prediction method switching information storage unit 107. The prediction method switching information 107A indicates switching information for determining a prediction method from the cluster number.

  When creating the decision tree, the decision tree creating unit 102 first associates the time series information of the wholesale power price with the leaf of the decision tree, and then replaces each time series information with the cluster number to thereby determine the leaf of the decision tree. Is associated with the cluster number. FIG. 9 of the present embodiment shows a form in which one cluster number is associated with the end node (leaf) of the decision tree. In addition to the present embodiment, a plurality of cluster numbers may be associated with the end node of the decision tree as shown in FIG. In this case, a decision tree composed of nodes (corresponding to attributes, especially corresponding to cluster numbers in the leaves) and branches (corresponding to attribute values) cannot be composed of a sufficient number of attributes, and the leafs are uniquely cluster numbers. Appropriate processing can be carried out even if it does not correspond.

  The external information management device 11 includes an information input / output unit 110 and an external information storage unit 111. The information input / output unit 110 acquires the external information 111 A from the external information distribution terminal 30 and stores it in the external information storage unit 111. The external information 111A includes weather information and economic information on each observation date.

  The price information prediction apparatus 12 includes an information input / output unit 120, a cluster determination unit 121, a prediction method switching unit 122, a price prediction unit 123, and a price prediction information storage unit 124. The cluster determination unit 121 determines the cluster number for the time series information of the wholesale power price on the prediction target date based on the decision tree information 105A and the external information 111A. Based on the determined cluster number and prediction method switching information 107A, the prediction method switching unit 122 determines a prediction method for predicting the time series information of the wholesale power price on the prediction target date. The price prediction unit 123 creates price prediction information 124A based on the determined prediction method, price information 103A, and external information 111A, and stores it in the price prediction information storage unit 124. The price prediction information 124A includes time-series information on wholesale power prices on the prediction target date. The information input / output unit 120 displays a screen showing the price prediction information 124A, the waveform of the time series information of the wholesale power price of the forecast target date included in the price prediction information 124A, etc., as shown in FIG. The image is displayed on a display device such as the output terminal 13.

  The present invention is not limited to this embodiment. In the case where a plurality of cluster numbers are associated with a leaf of a decision tree, a prediction method corresponding to each cluster number is temporarily selected using FIG. 7, and then one of the prediction methods is finally selected. good. At this time, the prediction method corresponding to the cluster number of the cluster having the largest number of price time series classified into the cluster corresponding to the cluster number is finally selected from the plurality of prediction methods. Furthermore, in another embodiment, as a simpler method, the prediction method switching unit 122 uses the night and day, month number (1, 2,..., 12), time zone (for example, instead of the prediction method switching information 107A). The prediction method may be switched according to time information such as a time zone near sunset. The prediction method may be switched according to climate information such as season and weather. Further, the prediction method with the highest estimation accuracy may be selected in the immediately preceding time zone.

(2) Internal configuration

  FIG. 3 shows the configuration of the wholesale power price prediction system 6.

  The price information management device 10 includes a central processing unit (CPU) 1001, an input device 1002, an output device 1003, a communication device 1004, and a storage device 1000. The price information management device 10 is an information processing device such as a server device or a PC (personal computer), for example.

  The input device 1002 includes a keyboard, a mouse, and the like. The output device 1003 includes a display, a printer, and the like. The communication device 1004 is an interface device connected to a wired or wireless network 5 such as a local area network (LAN) or a universal serial bus (USB). The storage device 1000 is a semiconductor memory such as a random access memory (RAM) or a read only memory (ROM), a hard disk drive (HDD), or the like.

  The storage device 1000 stores computer programs for the information input / output unit 100, the price classification unit 101, the decision tree creation unit 102, the prediction method switching information creation unit 106, and the like. The CPU 1001 operates according to these computer programs, thereby realizing the information input / output unit 100, the price classification unit 101, the decision tree creation unit 102, and the prediction method switching information creation unit 106.

  The information input / output unit 100 is a program for registering or deleting price information 103 </ b> A in the price information storage unit 103.

  The price classification unit 101 acquires time series information of wholesale power prices observed in the past from the price information 103A acquired from the price information distribution terminal 20, and according to the outline of the acquired time series information of wholesale power prices. This is a program that classifies time series information of wholesale electricity prices and creates multiple clusters.

  The decision tree creation unit 102 obtains external information observed in the past from the external information 111A obtained from the external information distribution terminal 30, and determines a cluster from the external information based on the cluster created by the price classification unit 101. This program creates a decision tree.

  The prediction method switching information creation unit 106 acquires time series information of wholesale power prices observed in the past from the price information 103A acquired from the price information distribution terminal 20, and from the external information 111A acquired from the external information distribution terminal 30. This is a program that acquires external information observed in the past and creates prediction method switching information 107A for determining a prediction method from a cluster number based on cluster classification information created by price classification information 104A.

  Further, the storage device 1000 stores databases such as a price information storage unit 103, a price classification information storage unit 104, a decision tree information storage unit 105, and a prediction method switching information storage unit 107. The price information storage unit 103 holds price information 103A. The price classification information storage unit 104 holds price classification information 104A. The decision tree information storage unit 105 holds decision tree information 105A. The prediction method switching information storage unit 107 holds prediction method switching information 107A.

  The price information 103A is information including time series information of wholesale power prices on each observation date.

  The price classification information 104A is information including the cluster number for the time series information of the wholesale power price on each observation date.

  The decision tree information 105A is information including the relationship between external information and cluster numbers.

  The prediction method switching information 107A is information including the relationship between the cluster number and the prediction method.

  The external information management device 11 includes a CPU 1101, an input device 1102, an output device 1103, a communication device 1104, and a storage device 1100. The external information management device 11 is an information processing device such as a server or a PC, for example.

  The input device 1102 includes a keyboard, a mouse, and the like. The output device 1103 includes a display, a printer, and the like. The communication device 1104 is a wired or wireless network such as a LAN or USB. The storage device 1100 is a memory such as a RAM or a ROM.

  The storage device 1100 stores computer programs for the information input / output unit 110 and the like. The information input / output unit 110 is realized by the CPU 1101 operating according to the computer program.

  The information input / output unit 110 is a program for registering or deleting the external information 111A in the external information storage unit 111.

  The storage device 1100 stores a database such as the external information storage unit 111. The external information storage unit 111 holds external information 111A.

  The external information 111A is information including weather information and economic information on each observation date.

  The price information prediction device 12 includes a CPU 1201, an input device 1202, an output device 1203, a communication device 1204, a storage device 1200, and the like. The price information prediction device 12 is an information processing device such as a server or a PC, for example.

  The input device 1202 includes a keyboard, a mouse, and the like. The output device 1203 includes a display, a printer, and the like. The communication device 1204 is a wired or wireless network such as a LAN or USB. The storage device 1200 is a memory such as a RAM or a ROM.

  The storage device 1200 stores computer programs for the cluster determination unit 121, the prediction method switching unit 122, the price prediction unit 123, and the like. The CPU 126 operates according to these computer programs, thereby realizing the cluster determination unit 121, the prediction method switching unit 122, and the price prediction unit 123.

  The cluster determination unit 121 acquires external information on the prediction target date from the external information 111A acquired from the external information distribution terminal 30, searches the determination tree created by the decision tree creation unit 102 using the acquired external information, This is a program for determining a cluster number for time-series information of wholesale power prices on the forecast date.

  The prediction method switching unit 122 uses the cluster number determined by the cluster determination unit 121 to calculate time series information of the wholesale power price on the prediction target date based on the prediction method switching information 107A created by the prediction method switching information creation unit 106. This is a program for determining a prediction method for prediction.

  The price prediction unit 123 is a program that predicts time-series information on wholesale power prices on the prediction target date using the prediction method determined by the prediction method switching unit 122.

  The storage device 1200 stores a database such as the price prediction information storage unit 124. The price prediction information storage unit 124 stores price prediction information 124A.

  Note that the functions of any one of the price information management device 10, the external information management device 11, and the price information prediction device 12 may be realized by a single device. For example, the information input / output unit 100, the price classification unit 101, the decision tree creation unit 102, the prediction method switching information creation unit 106, the information input / output unit 120, the cluster determination unit 121, the prediction method switching unit 122, and the price prediction unit 123 It may be realized by one computer. Further, any one of the storage devices 1000, 1100, and 1200 may be realized by one device. Each of the information input / output unit 100 and the information input / output unit 120 may access a storage device in a local device, or may access a storage device in an external device.

(3) Details of each component

  FIG. 4 shows price prediction information 124A.

  The price prediction information 124 </ b> A is information created by the price prediction unit 123, and stores a price prediction value that is a prediction value of the wholesale power price calculated for the future prediction date.

  Specifically, the price prediction information 124A has an entry for each prediction target day. An entry corresponding to one prediction target date includes a prediction target date column 124A1 and a price prediction value column 124A2.

  The prediction target date column 124A1 stores the date of the prediction target date.

  The price forecast value column 124A2 stores a time series price forecast value (time series forecast value) that is time series information of a price forecast value on the forecast target date. The time-series price forecast value includes price forecast values in each of a plurality of predetermined time periods (periods). The price prediction value indicates a price per unit electric energy.

  Therefore, in the example of this figure, the time-series price prediction values on the prediction target date “August 1, 2014” are “35.25 £ / MWh” for period 1 and “37.94 £ / MWh” for period 2. ,..., A period 48 indicates “38.56 £ / MWh”. Here, the period is an index representing the delivery time of the electric power product. When the electric power product is delivered in units of 30 minutes, 0: 00-0: 30 is period 1, 0: 30-1: 00 is period 2, ..., 24: 00-24: 00 is a period 48. The time length of the period is determined by the wholesale power market, for example, and may be another time length.

  FIG. 5 shows price information 103A.

  The price information 103A is information acquired from the price information distribution terminal 20, and stores a price actual value that is an actual value of the wholesale power price observed on the past observation date.

  Specifically, the price information 103A has an entry for each observation date. The entry corresponding to one observation date includes an observation date column 103A1 and a price actual value column 103A2.

  The observation date column 103A1 stores the date of the observation date.

  In the price actual value column 103A2, a time series price actual value (time series actual value) which is time series information of the price actual value on the observation date is stored. The time-series price actual value includes the price actual value in each of a plurality of predetermined time zones.

  Therefore, in the example of this figure, the time-series price actual value of the observation date “January 1, 2012” is “45.46 £ / MWh” in period 1 and “47.19 £ / MWh” in period 2. ..., it is shown that the period 48 is "35.69 £ / MWh".

  FIG. 6 shows the price classification information 104A.

  The price classification information 104A is information created by the price classification unit 101, and stores the cluster number corresponding to the time-series price actual value of the past observation date. The time-series price actual value on each observation day is classified into one of a plurality of clusters depending on the feature. The cluster number is an identifier indicating any one of a plurality of clusters.

  Specifically, the price classification information 104A has an entry for each observation date. An entry corresponding to one observation date includes an observation date column 104A1 and a cluster number column 104A2.

  The observation date column 104A1 stores the date of the observation date.

  In the cluster number column 104A2, the cluster number corresponding to the time-series price actual value on the observation date is stored.

  Therefore, in the example of this figure, it is shown that the cluster number of the observation date “January 1, 2012” is “4”.

  FIG. 7 shows the prediction method switching information 107A.

  The prediction method switching information 107A is information created by the prediction method switching information creation unit 106, and stores the prediction method corresponding to the cluster number.

  Specifically, the prediction method switching information 107A has an entry for each cluster number. An entry corresponding to one cluster number includes a cluster number column 107A1 and a prediction method column 107A2.

  A cluster number is stored in the cluster number column 107A1.

  The prediction method column 107A2 stores the prediction method corresponding to the cluster number.

  Therefore, in the example of this figure, it is shown that the prediction method of the cluster number “1” is “B”.

  FIG. 8 shows the external information 111A.

  The external information 111A is information acquired from the external information distribution terminal 30, and includes an external information actual value that is a past actual value of external information observed and an external information predicted value that is a predicted value of future external information. Is stored.

  Specifically, the external information 111A has an entry for each observation date. The observation date here includes the past observation date and the future observation date. The entry corresponding to one observation date includes an observation date column 111A1 and an external information value column 111A2.

  The observation date column 111A1 stores the date of the observation date.

  The external information value column 111A2 stores the actual value and the predicted value of the external information corresponding to the observation date. When the external information includes a plurality of attributes, the actual value and the predicted value corresponding to the observation date are stored in the external information value column 111A2 for each of the plurality of attributes. Attributes include season, average temperature, sunset time, and the like. Note that the season separation method may be different for each country. For example, in the case of Japan, March to May is “Spring”, June to August is “Summer”, September to November is “Autumn”, and December to February is “Winter”. In this case, April to June may be “spring”, July to September may be “summer”, October to December may be “autumn”, and January to March may be “winter”. The seasons may be divided based on the calendar and climatic conditions (average temperature, average amount of solar radiation, frequency of occurrence of fine weather, rain, etc.), etc., instead of the month.

  Therefore, in the example of this figure, it is shown that the season of the observation date “January 1, 2012” is “winter”, the average temperature is “10 ° C.”, and the sunset time is “16:03”. .

  FIG. 9 shows a conceptual diagram of the decision tree information 105A.

  The decision tree information 105A is information created by the decision tree creation unit 102, and indicates a decision tree for determining a cluster number from external information.

  Specifically, the decision tree information 105A includes a node label column 105A1, a path label column 105A2, and a cluster number column 105A3.

  The node label column 105A1 stores attribute names.

  The value of the attribute is stored in the path label column 105A2.

  A cluster number is stored in the cluster number column 105A3.

  Therefore, in the example of this figure, it is shown that the paths corresponding to the node label “season” are “spring / autumn”, “summer”, and “winter”.

(4) Processing flow

  The processing procedure in the wholesale power price prediction system 6 is shown.

  The price information prediction device 12 executes the operations of the cluster determination unit 121, the prediction method switching unit 122, and the price prediction unit 123 for each prediction target day.

<Price classification unit 101>

  The price classification unit 101 classifies time series information of wholesale power prices observed in the past according to the outline, and calculates a cluster number of time series information of wholesale power prices on each observation date.

  FIG. 11 shows the operation of the price classification unit 101.

  First, in step S1011, the price classification unit 101 acquires the time-series price actual value on the observation date within the past classification target period from the price information 103A. The classification target period is a period from the past observation date to the latest observation date by a predetermined reference time length. The reference time length is, for example, 2 years. Next, the price classification | category part 101 extracts the feature-value used for clustering from the time-sequential price actual value acquired by step S1011 in step S1012. For example, the price classification unit 101 may perform frequency analysis by Fourier transforming the time-series price actual value, and may use the obtained Fourier coefficient as the feature amount. Note that the price classification unit 101 may calculate the feature amount from the time-series price actual value using another method such as a filter corresponding to the feature amount.

  FIG. 12 shows the feature amount extracted by the price classification unit 101.

  In the example of this figure, the price classification | category part 101 is extracting the Fourier coefficient obtained by carrying out the Fourier transform of the example of the price information 103A of FIG. 5 as a feature-value. Here, the price classification unit 101 extracts, as feature quantities, predetermined Fourier coefficients of frequency 1 to frequency 3 among Fourier coefficients of frequency 1 to frequency 25 determined by the time length of the period.

  Next, price classification section 101 sequentially selects each of a plurality of cluster number candidates and repeats steps S1013 to S1016. The cluster number candidate is a preset number of clusters. Next, in step S1014, the price classification unit 101 classifies the feature amount extracted in step S1012 into the selected cluster number candidate cluster, and calculates the cluster number of each observation date. Here, the price classification unit 101 can use an existing clustering method such as the K-means method or the shortest distance method for classifying the feature values.

  FIG. 13 shows clusters classified by the price classification unit 101.

  The example of this figure shows the result of plotting points represented by feature quantities for each observation day in the space centered on the above-described feature quantities and classifying the observation days into clusters. Here, the number of clusters is 6, and the price classification unit 101 classifies the feature amounts (frequency 1 to 3) shown in FIG. 12 into 6 clusters, and the time series price for each observation day as shown in FIG. The actual value is assigned to one of the cluster numbers 1 to 6. Next, in step S1015, the price classification unit 101 calculates a validity evaluation value representing the validity of the classification based on the price classification information 104A obtained in step S1014. For example, the price classification unit 101 uses an evaluation index such as AIC (Akaike information criterion) or BIC (Bayes information criterion) as a validity evaluation value.

  After step S1013 to step S1016 are repeated for all cluster number candidates, in step S1017, the price classification unit 101 compares the validity evaluation values calculated in step S1015 for each cluster number candidate, From the number candidates, the cluster number candidate corresponding to the best validity evaluation value is selected as the optimum cluster number. For example, when the validity evaluation value is AIC and the cluster number candidate is 2, the AIC is “12000”, when the cluster number candidate is 4, the AIC is “13000”, and when the cluster number candidate is 6, the AIC is “ When the AIC when the cluster number candidate is 8 is 12000, the AIC when the cluster number candidate is 10 and the AIC is 13000, the AIC when the cluster number candidate is 6 is the minimum. The classification unit 101 selects 6 as the optimal number of clusters. As a result, the price classification unit 101 stores the cluster number for each observation date calculated using the optimal number of clusters in S1014 in the price classification information storage unit 104 as price classification information 104A. Price classification unit 101 may create price classification information 104A using a preset optimum number of clusters without using a plurality of cluster number candidates.

  As described above, for example, the price classification unit 101 classifies a plurality of observation days into a plurality of clusters using the feature amount by calculating the frequency component of the time-series price actual value for each observation date as the feature amount. can do. Further, the price classification unit 101 can select the optimum cluster number from the plurality of cluster number candidates by calculating the validity evaluation value for each of the plurality of cluster number candidates.

<Decision tree creation unit 102>

  The decision tree creation unit 102 creates a decision tree for specifying the cluster number of the observation date based on the observation date external information 111A. First, the decision tree creation unit 102 acquires the external information actual value within the classification target period from the external information 111A. Next, the decision tree creation unit 102 uses the external information actual value and the price classification information 104A corresponding to each observation date in the classification target period to calculate the cluster number of the prediction target date from the external information prediction value of the prediction target date. Is determined and stored in the decision tree information storage unit 105 as decision tree information 105A. Here, the decision tree creating unit 102 can use an existing machine learning method such as ID3 or C4.5 for creating a decision tree. FIG. 9 is an example of a decision tree created based on the attributes “season”, “average temperature”, “sunset time” and price classification information 104A in the external information 111A.

<Prediction method switching information creation unit 106>

  The prediction method switching information creation unit 106 creates switching information for determining a prediction method from the cluster number.

  FIG. 14 shows the operation of the prediction method switching information creation unit 106.

  First, the prediction method switching information creation unit 106 sequentially selects each of a plurality of cluster numbers, and repeats steps S1221 to S1228.

  Next, the prediction method switching information creation unit 106 sequentially selects each of a plurality of prediction methods, and repeats step S1222 to step S1226. The plurality of prediction schemes are preset prediction scheme candidates. In the present embodiment, the prediction method switching information creating unit 106 includes, as a plurality of prediction methods, a prediction method A based on time series analysis, a prediction method B based on similar data, a prediction method C using multiple regression models, and a correlation analysis. Prediction method D is used.

  First, in step S1223, the prediction method switching information creation unit 106 refers to the price classification information 104A, and estimates a plurality of observation days corresponding to the selected cluster number from the observation days in the past classification target period. Specify the target date. Furthermore, the time series price actual value of the estimation target date is acquired from the price information 103A. For example, when the selected cluster number is “3”, the prediction method switching information creation unit 106 sets a plurality of observation dates in which the cluster number is “3” within the estimated reference period in the price classification information 104A. Identified as the estimated date. Furthermore, the prediction method switching information creation unit 106 acquires time-series price actual values for a plurality of estimation target dates from the price information 103A.

  Next, in step S1224, the prediction method switching information creation unit 106 calculates the estimated value of the time series information of the wholesale power price for each of the plurality of estimation target dates using the selected prediction method, and obtains a plurality of obtained values. The time-series information of the estimation target date is averaged for each period, and the estimated value of the obtained time-series information for one day is defined as a time-series price estimated value (time-series estimated value).

  Here, when the selected prediction method is the prediction method A, the prediction method switching information creation unit 106 calculates a time-series price estimated value according to the price prediction process (details of the process will be described later) shown in FIG. When the selected prediction method is the prediction method B, the prediction method switching information creation unit 106 calculates a time-series price estimated value according to the price prediction process shown in FIG. 16 (details of the process will be described later). When the selected prediction method is the prediction method C, the prediction method switching information creation unit 106 calculates a time-series price estimated value according to the price prediction process (details of the process will be described later) shown in FIG. When the selected prediction method is the prediction method D, the prediction method switching information creation unit 106 calculates a time-series price estimated value according to the price prediction process (details of the process will be described later) shown in FIG.

  Next, in step S1225, the prediction method switching information creation unit 106 uses the time-series price actual value acquired in step S1223 and the time-series price estimated value calculated in step S1224 to predict the prediction accuracy of the selected prediction method. A prediction accuracy evaluation value representing is calculated. As the prediction accuracy evaluation value, for example, an error evaluation value indicating the magnitude of the error of the time series price estimated value of the estimation target date with respect to the time series price actual value of the estimation target date such as an average absolute error rate or a mean square error is used. .

  After step S1222 to step S1226 are repeated for all prediction methods, finally, in step S1227, the prediction method switching information creation unit 106 corresponds to the best prediction accuracy evaluation value from among a plurality of prediction methods. The prediction method to be selected is selected as the optimal prediction method. For example, when determining the optimum prediction method for the cluster number “3”, the prediction accuracy evaluation value is the average absolute error rate, the average absolute error rate obtained by the prediction method A is 15%, and the prediction method B is obtained. Assuming that the average absolute error rate is 12%, the average absolute error rate obtained by the prediction method C is 18%, and the average absolute error rate obtained by the prediction method D is 16%, the prediction method switching information creation unit 106 The prediction method B that minimizes the average absolute error rate is selected as the optimal prediction method for the cluster number “3”. Thereby, the prediction method switching information creation unit 106 can select a prediction method having the smallest error from a plurality of prediction methods as the optimum prediction method for each cluster.

  The above prediction method switching information creation unit 106 selects the prediction method with the highest prediction accuracy for the target cluster by calculating the magnitude of the error of the time series price estimated value with respect to the time series price actual value on the estimation target date. be able to.

<Cluster determination unit 121>

  For example, the price information prediction device 12 operates according to the operation of the price information management device 10. The cluster determination unit 121 calculates the cluster number of the prediction target date based on the external information 111A of the prediction target date. First, the cluster determination unit 121 acquires an external information prediction value for the prediction target date from the external information 111A. At this time, the cluster determination unit 121 acquires the predicted value of the attribute existing in the node label field of the decision tree information 105A in the external information 111A. Next, using the decision tree information 105A, the cluster determination unit 121 determines a cluster corresponding to the acquired external information prediction value on the prediction target date as a target cluster, and determines a cluster number indicating the target cluster as a target cluster number. To do.

  As an example, the cluster determination unit 121 sets August 1, 2014 as the prediction target date, and uses the example of external information shown in FIG. 8 and the example of the decision tree shown in FIG. An example of determination will be described. First, since the first node label of the decision tree shown in FIG. 9 is “season” and the “season” on August 1, 2014 is “summer”, the cluster determining unit 121 selects the “summer” path. select. Next, since the second node label of the decision tree after the “summer” path is selected is “average temperature”, and the “average temperature” on August 1, 2014 is “28 ° C.” The cluster determination unit 121 selects a path of “20 ° C. or higher”. As a result, the cluster determination unit 121 determines the target cluster number as “3”.

  As described above, the cluster determination unit 121 can select the target cluster from the external information prediction value of the prediction target date by using the decision tree information 105A.

<Prediction method switching unit 122>

  The prediction method switching unit 122 determines an optimal prediction method for calculating the time series price prediction value of the prediction target date. The prediction method switching unit 122 uses the prediction method switching information 107A to determine the optimal prediction method corresponding to the cluster number of the prediction target date.

  As an example, if the cluster determination unit 121 determines that the cluster number of the prediction target date is “3”, the optimal prediction method is determined as “D” by using the example of the prediction method switching information shown in FIG. The

<Price prediction unit 123>

  The price prediction unit 123 calculates a time-series price prediction value for the prediction target date by performing price prediction processing using the optimal prediction method determined by the prediction method switching unit 122.

  Hereinafter, price prediction processing of prediction method A to prediction method D will be described using FIGS. 15 to 18, respectively.

<Prediction method A based on time series analysis>

  FIG. 15 shows a price prediction process of the prediction method A.

  First, in step S1251, the price prediction unit 123 identifies a plurality of observation dates within the prediction reference period as a plurality of reference observation dates, and acquires time-series price actual values for the plurality of reference observation dates from the price information 103A. The prediction reference period is a period from the past observation date by the reference time length to the prediction target date to the observation date immediately before the prediction target date. Next, in step S1252, the time series information of the wholesale power price acquired in step S1251 is decomposed into a periodic component that is a periodically appearing variation and a random component that represents a randomly appearing variation. Here, the periodic component represents a change in wholesale power price that is repeated at a constant cycle such as a daily cycle or a one-year cycle. For example, a peak occurs in the evening in the periodic component of the daily cycle in the time-series information of the wholesale power price. The random component represents the remaining fluctuation obtained by subtracting the periodic component from the time-series information of the wholesale power price.

  For example, in step S1252, the price prediction unit 123 calculates a moving average value by performing a moving average of the length of one day on the time-series price actual values of a plurality of reference observation days, and a plurality of reference observations The difference value is calculated by subtracting the moving average value from the time-series price actual value of the day, the difference value of the obtained multiple reference observation days is averaged for each period, and the obtained time-series information for one day is a periodic component And Further, the price prediction unit 123 calculates a random component by subtracting the periodic component from the time-series price actual value of the plurality of reference observation dates.

  Next, in step S1253, the price prediction unit 123 identifies a periodic component model using the periodic component calculated in step S1252. For example, the price prediction unit 123 can identify a periodic component model by calculating a Fourier coefficient corresponding to each frequency component included in the periodic component calculated in step S1252. Next, in step S1254, the price prediction unit 123 calculates the predicted value of the time series information of the periodic component of the prediction target day using the model of the periodic component identified in step S1253.

  Next, in step S1255, the price prediction unit 123 identifies the model of the random component calculated in step S1252. The random component is represented by a regression model such as an AR (autoregressive) model or an ARMA (autoregressive moving average) model. For example, when the wholesale power price of a certain period is represented by the AR model described by the previous period, the previous period, the previous period, the wholesale power price of the previous period, By using a square method or the like, a random component model can be identified by calculating a weighting factor for the wholesale power price of the previous period, the previous period, the previous period, and the previous period. Next, in step S1256, the price prediction unit 123 calculates the predicted value of the time-series information of the random component of the prediction target date using the model of the random component identified in step S1255. Finally, in step S1257, the price prediction unit 123 adds the predicted value of the time-series information of the random component calculated in step S1256 for each period to the predicted value of the time-series information of the periodic component calculated in step S1254. Thus, the time series price prediction value of the prediction target date is calculated.

  According to the prediction method A described above, the price prediction unit 123 can improve the accuracy of prediction by dividing the time-series price actual value into a periodic component and a random component and modeling them with different methods. .

<Prediction method B based on similar data>

  FIG. 16 shows a price prediction process of the prediction method B.

  First, in step S1261, the price prediction unit 123 specifies a plurality of observation days corresponding to the target cluster number as a plurality of reference observation dates within the prediction reference period from the observation dates in the price classification information 104A. Next, in step S1262, the price prediction unit 123 refers to the external information 111A, and acquires the external information actual value within the prediction reference period and the external information prediction value for the prediction target date. Next, in step S1263, the price prediction unit 123 selects an observation date whose external information actual value is similar to the external information prediction value of the prediction target date from the plurality of reference observation dates, and selects the selected observation date. The time-series price actual value is acquired from the price information 103A and is used as similar data. For example, the price prediction unit 123 uses the temperature as an attribute, and the Euclidean distance between the temperature of the external information predicted value (forecast value) of the prediction target day and the temperature of the external information actual value of each of the plurality of reference observation days. A predetermined number of observation days are selected from the plurality of reference observation dates in ascending order of the Euclidean distance, and the time series price actual value of the selected predetermined number of observation days is selected as the predetermined number of similar data. . Here, the price prediction unit 123 may select one observation day corresponding to the minimum Euclidean distance from the plurality of reference observation days.

  Finally, in step S1264, the price prediction unit 123 calculates a time-series price prediction value for the prediction target date using the similar data extracted in step S1263. For example, the price prediction unit 123 averages a predetermined number of similar data corresponding to the selected predetermined number of observation days for each period, thereby obtaining the time series information for one day obtained as the time series of the prediction target date. It is good also as a price forecast value.

  According to the prediction method B described above, the price prediction unit 123 selects the external information similar to the prediction target date from the observation days belonging to the same cluster as the prediction target date within the prediction reference period, thereby making a prediction reference. Compared to the case where external information similar to the prediction target date is selected from all observation days in the period, the accuracy of similar data can be improved.

<Prediction method C using multiple regression model>

  FIG. 17 shows price prediction processing of the prediction method C.

  First, in step S1271, the price prediction unit 123 specifies a plurality of observation days corresponding to the target cluster number as a plurality of reference observation days in the prediction reference period from the price classification information 104A, and the plurality of reference observation days. A series price actual value is acquired from price information 103A. Next, in step S1272, the price prediction unit 123 acquires external information on a plurality of reference observation dates from the external information 111A. Next, in step S1273, the price prediction unit 123 estimates the coefficient of the multiple regression model using the time-series price actual value acquired in step S1271 and the external information actual value acquired in step S1272. The coefficient of the multiple regression model is a plurality of weighting coefficients respectively corresponding to a plurality of attributes that explain the wholesale power price. Here, the price prediction unit 123 can estimate the weighting coefficient by using a least square method or the like. Finally, in step S1274, the price prediction unit 123 calculates a time series price prediction value for the prediction target date using the coefficient of the multiple regression model calculated in step S1273 and the external information prediction value for the prediction target date. For example, the price prediction unit 123 multiplies a plurality of attributes in the external information prediction value on the prediction target date by a plurality of weighting factors, and adds the multiplication results, thereby making a time-series price prediction for the prediction target date. Calculate the value.

  According to the above prediction method C, the price prediction unit 123 generates a multiple regression model based on the time series price actual value of the observation date belonging to the same cluster as the prediction target date from the prediction reference period, thereby predicting The accuracy of the multiple regression model can be improved as compared to the case where the multiple regression model is generated based on the actual time-series price values of all observation dates in the reference period.

<Prediction method D based on correlation analysis>

  FIG. 18 shows a price prediction process of the prediction method D.

  First, in step S1281, the price prediction unit 123 identifies a plurality of observation dates in the prediction reference period as a plurality of reference observation dates, and acquires time-series price actual values for the plurality of reference observation dates from the price information 103A. Next, in step S1283, the price prediction unit 123 extracts, as the price peak time, the time when the wholesale power price is maximum for each observation day based on the actual time-series price value acquired in step S1281. To do.

  Next, the price prediction unit 123 sequentially selects each of the plurality of attributes in the external information 111A, and repeats Steps S1284 to S1287. Next, in step S1285, the price prediction unit 123 acquires the actual value of the selected attribute from the plurality of reference observation dates from the external information 111A. In step S1286, the price prediction unit 123 calculates a correlation coefficient between the actual value of the attribute acquired in step S1282 and the price peak time extracted in step S1283.

  After step S1284 to step S1287 are repeated for all the attributes, in step S1288, the price prediction unit 123 selects an influence attribute that causes the wholesale power price to rise from a plurality of attributes. For example, the price prediction unit 123 selects an attribute corresponding to a correlation coefficient that is equal to or greater than a correlation coefficient threshold value as an influence attribute. The attribute may be a maximum temperature observation time that is a time when the highest temperature of the day is recorded, a peak time of evening rush, an average time of a company lunch time, or the like. For example, the attributes are “maximum temperature observation time” and “sunset time”, the correlation coefficient between the maximum temperature observation time and the price peak time is 0.3, and the phase between the sunset time and the price peak time is When the number of relationships is 0.7 and the correlation coefficient threshold is 0.6, the price prediction unit 123 selects the sunset time as the influence attribute.

  Next, in step S1289, the price prediction unit 123 classifies the plurality of observation days in the prediction reference period into a plurality of groups using the actual value of the influence attribute, and determines the prediction target date using the prediction value of the influence attribute. Classify into any of several groups. For example, the price prediction unit 123 classifies the sunset time into a predetermined time zone, so that the observation date when the sunset time is 18: 00-18: 15 is group 1 and the sunset time is 18: 15-18. : Classify observation day 30 to group 2.

  Finally, in step S1290, the price prediction unit 123 selects the observation date classified into the same group as the prediction target date in step S1285, and uses the time series price actual value of the selected observation date to predict the prediction target date. Calculate the time series price forecast for. For example, the price prediction unit 123 averages a plurality of time-series price actual values respectively corresponding to a plurality of selected observation days for each period, thereby obtaining the obtained time-series information for one day. It may be a time series price forecast value.

  According to the above prediction method D, since the price peak time has a great influence on the time series price prediction value, the prediction accuracy of the time series price prediction value can be improved by using the price peak time. In addition, by selecting an influence attribute from among a plurality of attributes using a correlation coefficient, an event that affects the price peak time can be detected.

  The above prediction methods A to D are used by the price prediction unit 123, and also used by the prediction method switching unit 122 in step S1223. In this case, the estimation target date is used instead of the prediction target date. That is, the prediction reference period is a period from the observation date in the past by the reference time length to the estimation target date to the observation date immediately before the estimation target date.

  In the wholesale power price prediction system 6 according to the present embodiment, the price classification unit 104 creates the price classification information 104A and the decision tree creation unit 102 creates the decision tree information 105A. The price classification information and the decision tree information may be held in the price classification information storage unit 104 and the decision tree information storage unit 105, respectively, and the price information prediction device 12 may perform prediction using these pieces of information.

  Further, the price information prediction device 12 may set and predict the prediction target date at a day interval, and the price information management device 10 may generate the decision tree information 105A at a different interval. For example, the price information management apparatus 10 determines whether or not a specific event has occurred by monitoring an external information actual value, a time-series price actual value, etc., and if it is determined that a specific event has occurred The decision tree information 105A may be updated by executing the processes of the price classification unit 101 and the decision tree creation unit 102. The specific event may be an instruction from the administrator of the price information management apparatus 10. Further, when the attribute is a stock index or the like, the specific event may be that the amount of change in the attribute of the day exceeds a predetermined change amount threshold. Further, the price information management apparatus 10 may update the decision tree information 105 </ b> A by periodically executing the processes of the price classification unit 101 and the decision tree creation unit 102.

  According to the price information prediction apparatus 12 described above, the cluster determination unit 121 selects a target cluster from a plurality of clusters, and the prediction method switching unit 122 selects an optimal prediction method corresponding to the target cluster. it can. As a result, it is possible to select past observation dates that are close to fluctuations in the wholesale power price on the prediction target date. Moreover, it is possible to improve the prediction accuracy when the controlling factor of the fluctuation of the wholesale power price changes. In addition, at least one of the forecast methods can estimate observations other than the target cluster by estimating the time series price prediction value from the actual time series price value of the observation date belonging to the target cluster within the estimated reference period past the prediction target date. The prediction accuracy can be improved as compared with the case of using the daily time-series price actual value.

(5) Other embodiments

  The prediction method switching unit 122 determines whether or not the variance of the specific time-series price actual value is equal to or greater than a predetermined variance threshold, and when it is determined that the variance is equal to or greater than the variance threshold, the prediction method switching unit 122 and the price prediction unit 123 are not processed, and the fluctuation parameters indicating the trend of the fluctuation of the wholesale power price on the prediction target day, such as the volatility of the wholesale power price on the prediction target day and the trend direction of the wholesale power price on the prediction target day May be predicted. For example, the prediction method switching unit 122 may determine whether or not the variance of the time series price actual values belonging to the target cluster is greater than or equal to the variance threshold. A standard deviation or the like may be used instead of the variance. Thereby, when the fluctuation of the wholesale power price is unstable, it is expected that the prediction accuracy of the time-series price prediction value is lowered, so the prediction method switching unit 122 calculates a fluctuation parameter different from the price prediction value. Thus, it is possible to output information that serves as a reference for fluctuations in the wholesale power price on the prediction target date. The wholesale power price prediction system 6 may calculate a fluctuation parameter in addition to the time series price prediction value. The wholesale power price prediction system 6 may calculate the reliability (accuracy) of the time series price prediction value based on the variance of the specific time series price actual value. Further, the price prediction unit 123 may calculate a variation parameter when it is determined that the variance of the time-series price prediction value on the prediction target date is equal to or greater than the variance threshold.

  The price classification unit 101 may select an observation date corresponding to a feature amount closest to the center of the cluster of the prediction target date as a central observation date in the feature amount space as shown in FIG. In this case, the information input / output unit 120 may display the waveform of the time series price actual value on the central observation date on the display device as information indicating the cluster of the prediction target date. Further, the information input / output unit 120 may display on the display device the result of statistical processing such as the volatility of the time series price actual value on the central observation date and the transition direction. Thereby, the administrator of the wholesale power price prediction system 6 can know the characteristics of the target cluster. Further, the information input / output unit 120 may display the variation parameter, the reliability of the time series price prediction value, and the like on the display device. As a result, the administrator of the wholesale power price prediction system 6 can know information that serves as a reference for the wholesale power price. Further, the information input / output unit 120 may display the name of the optimum prediction method on the display device. Thereby, the administrator of the wholesale power price prediction system 6 can know the prediction method used for the prediction of the wholesale power price.

  Terms for the expression of the present invention will be described. As the storage unit, storage devices 1000, 1100, 1200, and the like may be used. The CPU 1201 or the like may be used as the calculation unit. A CPU 1001 or the like may be used as the generation unit.

DESCRIPTION OF SYMBOLS 1 ... Trading participant system 2 ... Trading market operator system 3 ... External information manager system 4 ... Electric power trading system 5 ... Network 6 ... Wholesale electric power price prediction system 10 ... Price information management apparatus 11 ... External information management apparatus 12 ... Price Information prediction device 13 ... Information input / output terminal 20 ... Price information distribution terminal 30 ... External information distribution terminal 100 ... Information input / output unit 101 ... Price classification unit 102 ... Decision tree creation unit 103 ... Price information storage unit 104 ... Price classification information storage Unit 105 ... Decision tree information storage unit 106 ... Prediction method switching information creation unit 107 ... Prediction method switching information storage unit 110 ... Information input / output unit 111 ... External information storage unit 120 ... Information input / output unit 121 ... Cluster determination unit 122 ... Prediction Method switching unit 123 ... price prediction unit 124 ... price prediction information storage unit

Claims (11)

A wholesale power price forecasting system for forecasting a wholesale power price on a forecast date,
Storing a time series actual power value which is a time series wholesale power price observed on each of a plurality of observation days in the past from the prediction target date, storing classification information indicating a cluster corresponding to the plurality of observation days, For attributes affecting wholesale power prices, store attribute prediction values that are prediction values of the attributes on the prediction target date, store a decision tree indicating the relationship between the attributes and the plurality of clusters, and store the plurality of clusters. And a storage unit that stores prediction method switching information indicating a relationship between a plurality of predetermined prediction methods,
Based on the attribute prediction value and the decision tree, a target cluster corresponding to the attribute prediction value is selected from the plurality of clusters, and based on the target cluster and the prediction method switching information, the plurality of predictions Select one specific prediction method from among the methods, and use the specific prediction method, from the time series actual value of the plurality of observation days, the time series wholesale power price of the prediction target date as a time series prediction value A calculation unit for calculating,
Wholesale electricity price forecasting system with. The calculation unit, when using a first prediction method that is at least one of the plurality of prediction methods, a plurality of reference observations corresponding to the target cluster in the past from the prediction target date among the plurality of observation days. Selecting a day, and calculating the time series predicted value based on the time series actual value of the plurality of reference observation days,
The wholesale power price prediction system according to claim 1. The storage unit stores an attribute actual value that is an actual value of the attribute observed on the plurality of observation days,
The classification information indicating clusters corresponding to each of the plurality of observation days is generated by classifying the time series actual values of the plurality of observation days into the plurality of clusters, and based on the attribute actual values and the classification information The decision tree is generated, and further includes a generation unit that generates the prediction method switching information based on the time series actual value of the plurality of observation days, the attribute actual value, and the classification information.
The wholesale power price prediction system according to claim 2. The storage unit stores a plurality of attribute actual values observed on the plurality of observation days, and the plurality of attribute prediction values on the prediction target day,
When the second prediction method, which is at least one of the plurality of prediction methods, is used, the calculation unit detects a peak time in each time series actual value on the plurality of observation days, and from among the plurality of attributes Selecting a specific attribute having a high correlation with the peak time, and selecting a specific observation date from the plurality of observation days based on the predicted value of the specific attribute of the prediction target date and the actual value of the specific attribute And calculating the time-series predicted value based on the time-series actual value on the specific observation date,
The wholesale power price prediction system according to claim 3. The generating unit classifies the time series actual values of the plurality of observation days into the plurality of clusters based on frequency components of the time series actual values of the plurality of observation days.
The wholesale power price prediction system according to claim 4. The generation unit generates a plurality of cluster candidates for each of the plurality of cluster number candidates, calculates an evaluation index for each of the plurality of cluster candidates, and selects the best evaluation index from the plurality of cluster number candidates. Selecting a cluster number candidate corresponding to the number of the plurality of clusters,
The wholesale power price prediction system according to claim 5. The generation unit selects an estimation target date corresponding to the target cluster from the plurality of observation dates, and uses each of a plurality of predetermined prediction methods to observe observation dates past the estimation target date. From the time-series actual value, the time-series wholesale power price of the estimation target date is calculated as a time-series estimated value, and the estimation target date is calculated based on the time-series estimated value and the time-series actual value of the estimation target date. Calculating an error evaluation value indicating the magnitude of the error of the time series estimated value with respect to the time series actual value, and creating prediction method switching information based on the error evaluation value;
The wholesale power price prediction system according to claim 6. The calculation unit shows a tendency of fluctuations in the wholesale power price on the forecast target day when the magnitude of fluctuation in the time series actual value on the specific observation day among the plurality of observation days satisfies a predetermined condition. Predict fluctuation parameters,
The wholesale power price prediction system according to claim 7. The calculation unit selects a time-series actual value closest to the center of the target cluster, and causes the display device to display the selected time-series actual value.
The wholesale electric power price prediction system as described in any one of Claims 1 thru | or 8. The calculation unit causes the display device to display at least one of the specific prediction method and the time series prediction value.
The wholesale electric power price prediction system as described in any one of Claims 1 thru | or 8. A wholesale power price prediction method for predicting a wholesale power price on a forecast date,
Storing a time series actual value which is a wholesale power price of a time series observed on each of a plurality of observation days in the past from the prediction target date;
Storing classification information indicating a plurality of clusters indicating the classification of time series actual values of the plurality of observation days;
For attributes that affect wholesale power prices, store attribute prediction values that are prediction values of the attributes on the prediction target date;
Storing a decision tree indicating the relationship between the attribute and the plurality of clusters;
Storing prediction method switching information indicating a relationship between the plurality of clusters and a plurality of predetermined prediction methods;
Based on the attribute prediction value and the decision tree, select a target cluster corresponding to the attribute prediction value from the plurality of clusters,
Based on the target cluster and the prediction method switching information, select one specific prediction method from the plurality of prediction methods,
A wholesale power price prediction method comprising: calculating a time-series wholesale power price of the prediction target date as a time-series predicted value from the time-series actual values of the plurality of observation days using the specific prediction method.

Images