WO2017212880A1

WO2017212880A1 - Data prediction system and data prediction method

Info

Publication number: WO2017212880A1
Application number: PCT/JP2017/018334
Authority: WO
Inventors: 将人内海; 渡辺　徹; 郁雄茂森; 羊子 ▲崎▼久保; 敏之澤; 洋飯村; 広晃小川; 岡本　佳久
Original assignee: 株式会社日立製作所
Priority date: 2016-06-09
Filing date: 2017-05-16
Publication date: 2017-12-14

Abstract

This data prediction system for predicting future prediction values comprises: a data management device for managing data; and a prediction calculation device for correcting the future prediction values by modelling trends of error amounts in prediction calculation results calculated on the basis of correlations with main explanatory variables. The data management device is provided with a storage unit for storing prediction object past measurement data and explanatory factor data explaining the prediction object past measurement data, the prediction object past measurement data and the explanatory factor data being observed in accordance with time transition. The prediction calculation device is provided with: a first prediction calculation unit for performing predictions on the basis of correlations between the prediction object past measurement data and the explanatory factor data; a second prediction calculation unit for modelling trends of errors in calculation results from the first prediction calculation unit to perform predictions on future error amounts in calculation results from the first prediction calculation unit; and a correction unit for correcting the calculation results from the first prediction calculation unit by means of the calculation results from the second prediction calculation unit.

Description

Data prediction system and data prediction method

The present invention relates to a data prediction system and a data prediction method.

In the energy business field such as the electric power business and gas business, the communication business field, and the transportation business field such as taxi and delivery business, in order to operate facilities and allocate resources according to consumer demand, future demand Predict the value of.

However, due to the occurrence of various events that cause fluctuations in demand, there is a discrepancy between the predicted value calculated at a certain point in time and the actually observed value. Therefore, it is important for the prediction with high accuracy to newly update the predicted value based on the newly observed value.

JP 2013-5456 A (Patent Document 1) is disclosed as a technique for performing the above prediction. According to Patent Document 1, previous performance data having the same pattern as the prediction target period is extracted from performance data stored for each pattern that can be classified according to the day of the week.

∙ Based on the extracted data, change the coefficient included in the regression equation that includes the factor representing the weather condition and the load amount as variables, and calculate the predicted value for the prediction target period using the regression equation with the changed coefficient. The prediction value is corrected using the difference between the actual measurement value and the prediction value. At the time of correction, the correction amount is decreased as the previous time comes.

JP 2013-5456 A

The technique disclosed in Patent Document 1 is based on the assumption that most of the transition of the future value of the prediction target can be explained by main explanatory factors such as preset outdoor enthalpy, wet bulb temperature, and air temperature. Yes. However, depending on the data, there are many cases in which fluctuations in values that cannot be explained by the main explanatory variables set are observed.

For example, the periodicity of the data itself such that a certain amount is observed at night regardless of the temperature, the inertia that air conditioning demand remains constant even if the temperature changes, and fluctuations due to sudden events such as typhoons and events It is data to observe. For such data, the prediction according to Patent Document 1 is difficult.

In addition, the technique disclosed in Patent Document 1 does not make assumptions based on a model for the prediction error itself that could not be explained by the main explanation factors. Therefore, it is difficult to correct a predicted value that is relatively far in the future.

The present invention has been made in consideration of the above points, and models the tendency of the error amount of the prediction calculation result calculated based on the correlation with the main explanatory variables to correct the prediction calculation result. Accordingly, an object of the present invention is to provide a data prediction system and a data prediction that enable prediction of future values of data having fluctuations that are difficult to explain with main explanatory variables.

In order to solve this problem, one of the representative data prediction systems of the present invention is a data prediction system for predicting future prediction values, and a correlation between a data management device for managing data and main explanatory variables. A prediction calculation device that corrects a predicted value in the future by modeling a tendency of an error amount of a prediction calculation result calculated based on the A storage unit for storing the target past measurement data and the explanatory factor data for explaining the prediction target past measurement data is provided, and the prediction calculation device performs the prediction based on the correlation between the prediction target past measurement data and the explanatory factor data. A prediction calculation unit, a second prediction calculation unit that models a tendency of an error in the calculation result of the first prediction calculation unit, and predicts a future error amount of the calculation result of the first prediction calculation unit; Second plan And so and a correcting unit for correcting the calculation result of the first prediction computation unit by the calculation result of the arithmetic unit.

In order to solve this problem, one of the representative data prediction methods of the present invention is a data prediction method executed in a data prediction system that predicts a predicted value in the future. The data prediction system manages data. A data management device, and a prediction calculation device that corrects a predicted value in the future by modeling a tendency of an error amount of a prediction calculation result calculated based on a correlation with main explanatory variables, and data management A first step in which the apparatus stores prediction target past measurement data and explanation factor data explained by the prediction target past measurement data, which are observed along with a time transition, and a prediction arithmetic unit explains the prediction target past measurement data and explanation The second step of performing the prediction based on the correlation with the factor data, and the prediction calculation device models the error tendency of the prediction calculation result of the prediction, and the future of the prediction calculation result A third step of performing error amount prediction, the prediction arithmetic unit, and so and a fourth step of correcting the prediction calculation result by the calculation result of the error amount prediction.

According to the invention, it is possible to make a prediction reflecting even fluctuation components that are difficult to explain with main explanatory variables. Problems, configurations, and effects other than those described above will become apparent from the following description of embodiments.

It is a figure which shows the apparatus structure by 1st embodiment of a data management system. It is a figure which shows the function structure by 1st embodiment of a data prediction system. It is a figure which shows the data flow by 1st embodiment of a data prediction system. It is a figure which shows the processing flow by 1st embodiment of a data prediction system. It is a figure which shows the data flow by 1st embodiment of a prediction calculation process. It is a figure which shows the data flow by 2nd embodiment of a prediction calculation process. It is a figure which shows the data flow by 3rd embodiment of a prediction calculation process. It is a conceptual diagram which shows the effect of a data prediction system. It is a figure which shows the data flow by 2nd embodiment of a data prediction system. It is a figure which shows the processing flow by 2nd embodiment of a data prediction system. It is a figure which shows the data flow of 1st embodiment of a predicted value correction process. It is a figure which shows the data flow of 2nd embodiment of a predicted value correction process. It is a figure which shows the data flow of 3rd embodiment of a predicted value correction process. It is a figure which shows the data flow of 4th embodiment of a predicted value correction process. It is a figure which shows the concept of 4th embodiment of a predicted value correction process. It is a figure which shows the data flow by the modification of 2nd embodiment of a data prediction system. It is a figure which shows the apparatus structure by the modification of a data management system. It is a figure which shows the apparatus structure by another modification of a data management system. It is a figure which shows the apparatus structure by the further modification of a data management system. It is the conceptual diagram which showed the effect of the data prediction system. It is the conceptual diagram which showed the effect of the data prediction system.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

(1) First Embodiment (1-1) Configuration of Data Management System According to this Embodiment FIG. 1 shows the overall configuration of a data management system 1 according to this embodiment. The data management system 1 calculates a predicted value of an arbitrary future date and time based on the prediction target past measurement data and explanatory factor data that can explain the prediction target past measurement data, and based on the calculated prediction value, the physical facility It is a system that generates and executes an operation and control plan.

The data management system 1 includes a prediction calculation device 2, a data management device 3, a plan creation / execution management device 5, an information input / output terminal 4, a data observation device 6, and a data distribution device 7. The communication path 8 is, for example, a LAN (Local Area Network) or a WAN (Wide Area Network), and is a communication path that connects various devices and terminals constituting the data management system 1 so that they can communicate with each other.

The data management device 3 uses the past date and time set in advance via the information input / output terminal 4 for the prediction target past measurement data observed with time transition and the explanation factor data that can explain the prediction target past measurement data. Memorize up to the latest observation date.

In addition, the data management device 3 performs search and transmission in response to a data acquisition request from another device. The past measurement data to be predicted is, for example, energy consumption data such as electric power, gas, and water as a total of a measuring instrument or a plurality of measuring instruments, communication data of a certain communication base station, and time of a moving object such as a taxi. Data on the number of units in operation.

Explanatory factor data includes weather data such as temperature, humidity, solar radiation, wind speed and pressure, data indicating the occurrence of sudden events such as typhoons and events, the number of energy consumers, and communications connected to communication base stations. Data such as prediction target past measurement data such as the number of terminals. Or explanatory factor data is data which can explain the value of said prediction object past measurement data, such as said prediction object past measurement data itself.

The prediction calculation device 2 performs prediction based on the correlation between the prediction target past measurement data stored in the data management device 3 and the prediction explanation factor data. Furthermore, the prediction calculation device 2 models the tendency of the first prediction error. In addition, the prediction calculation device 2 calculates the result data of the first prediction calculation unit from the second prediction calculation unit that predicts the future error amount of the first prediction and the result data of the second prediction calculation unit. Make corrections.

The result data of the first prediction calculation unit and the second prediction calculation unit, the first prediction calculation result and the corrected data based on the second prediction calculation result, and the like are the future prediction values of the prediction target past measurement data. The data of the section which shows the fluctuation range of a predicted value are included. Or, the result data of the first prediction calculation unit and the second prediction calculation unit, the first prediction calculation result and the corrected data based on the second prediction calculation result, etc. are model equations for calculating the prediction value and Contains coefficient value data.

The plan creation / execution management device 5 creates and executes a physical facility operation plan so that a predetermined target can be achieved based on the prediction calculation result data generated and output by the prediction calculation device 2. Here, the physical equipment and its operation plan are, in the energy field, for example, a plan that satisfies a predicted future energy demand value or an energy demand plan value created based on the predicted future energy demand value. It is.

Physical equipment and its operation plan are, for example, in the energy field, specifically, the number of generators to be started and the output distribution plan of these generators, and the flow rate of gas and water flowing through gas conduits and water pipes. And pressure distribution plan.

The physical facility and its operation plan are, in the communication field, a control plan for the number of communication terminals connected to each communication base station, for example, so as not to exceed the capacity of the communication base station. In the transportation field, for example, it is a taxi dispatch plan that can satisfy the predicted number of users.

The facility operation plan is not limited to the direct execution by the subject who uses the plan creation / execution management device 5, but may be realized indirectly. Indirect facility operation is, for example, in the electric power field, physical facility operation by another person based on a direct relative transaction contract or a transaction contract via an exchange. In this case, the execution plan of the transaction contract corresponds to the operation plan of the equipment.

The information input / output terminal 4 inputs data to the predictive calculation device 2, the data management device 3, and the plan creation / execution management device 5, and displays data stored or output by these devices. The data observation device 6 periodically measures the prediction target past measurement data and the prediction explanation factor data at predetermined time intervals, and transmits them to the data distribution device 7 or the data management device 3. The data distribution device 7 stores the data received from the data observation device 6, and transmits the data to the data management device 3, the prediction calculation device 2, or both.

(1-2) Internal Device Configuration FIG. 2 shows a functional configuration of each device constituting the data prediction system 12 (FIG. 3) in the data management system 1. The data prediction system 12 includes a prediction calculation device 2 and a data management device 3.

The data management device 3 includes a CPU (Central Processing Unit) 31 that controls the operation of the data management device 3 in an integrated manner, an input device 32, an output device 33, a communication device 34, and a storage device 35. The data management device 3 is an information processing device such as a personal computer, a server computer, or a handheld computer.

The input device 32 is composed of a keyboard or a mouse, and the output device 33 is composed of a display or a printer. The communication device 34 includes a NIC (Network Interface Card) for connecting to a wireless LAN or a wired LAN. The storage device 35 is a storage medium such as a RAM (Random Access Memory) or a ROM (Read Only Memory). You may output the output result of each process part, and an intermediate result suitably via the output device 33. FIG.

The storage device 35 stores databases such as prediction target past measurement data storage means 351 and explanatory variable past measurement data storage means 352. The prediction target past measurement data storage unit 351 holds prediction target past measurement data 351A.

The prediction target past measurement data 351A is a past value measured at a predetermined time interval such as an interval of 30 minutes of data to be predicted by the prediction calculation device 2. The prediction target past measurement data 351A includes energy consumption data such as electric power, gas, and water as a total of a measuring instrument or a plurality of measuring instruments, communication data of a certain communication base station, and each time of a moving object such as a taxi. This data includes data on the number of units in operation.

The explanatory variable past measurement data storage unit 352 holds the explanatory variable past measurement data 352A. The explanatory variable past measurement data 352A is a past value measured at a predetermined time interval such as an interval of 30 minutes of explanatory factor data that can explain the increase or decrease of the value of the prediction target past measurement data 351A.

The explanatory variable past measurement data 352A is data including meteorological data such as temperature, humidity, amount of solar radiation, wind speed, atmospheric pressure, and data indicating whether or not a sudden event such as a typhoon or an event has occurred. The explanatory variable past measurement data 352A is data including the prediction target past measurement data 351A such as the number of energy consumers and the number of communication terminals connected to the communication base station, the prediction target past measurement data 351A itself, and the like.

The prediction calculation device 2 includes a CPU (Central Processing Unit) 21 that controls the overall operation of the prediction calculation device 2, an input device 22, an output device 23, a communication device 24, and a storage device 25. The prediction calculation device 2 is an information processing device such as a personal computer, a server computer, or a handheld computer.

The storage device 25 stores various computer programs such as the prediction calculation unit 251 and the prediction value correction unit 252. The prediction calculation unit 251 calculates first prediction calculation result data 302 (FIG. 3) based on the correlation between the prediction target past measurement data 351A and the explanatory variable past measurement data 352A.

The prediction calculation unit 251 calculates a prediction error by calculating a difference or the like based on the prediction target past measurement data 351A including the latest measurement value, and models the occurrence tendency of the prediction error. Thus, the prediction calculation unit 251 calculates the first prediction error amount at an arbitrary future time, and corrects the first prediction calculation result with the calculated future error amount.

The storage device 25 stores a database such as a prediction calculation result data storage unit 253. The prediction calculation result data storage unit 253 holds prediction calculation result data 253A.

The prediction calculation result data 253A is a prediction calculation result calculated by the prediction calculation unit 251, and is a representative value such as an expected value of the prediction calculation result, interval data such as a prediction confidence interval or a prediction interval, and a model used for prediction. Data including formulas and their coefficients. The prediction calculation unit 251 includes a second prediction calculation unit 251B (FIG. 3) that performs error series prediction.

(1-3) Overall Processing and Data Flow of Data Prediction System According to this Embodiment With reference to FIGS. 3 and 4, the processing and data flow of the data prediction system 12 in this embodiment will be described.

With reference to FIG. 3, the data flow of the data prediction process of the data prediction system 12 in this Embodiment is demonstrated.

The data management device 3 receives the explanatory variable past measurement data 352A transmitted from the data observation device 6 or the data distribution device 7, and stores it in the explanatory variable past measurement data storage means 352.

Further, the data management device 3 stores the prediction target past measurement data 351A transmitted from the data observation device 6 or the data distribution device 7 in the prediction target past measurement data storage unit 351.

The explanatory variable past measurement data 352A includes, for example, the above-described prediction target past measurement data 351A such as weather data such as temperature, humidity, solar radiation amount, wind speed, and atmospheric pressure, and data indicating whether or not a sudden event such as a typhoon or an event has occurred. It is data that can explain the value of.

In addition, the explanatory variable past measurement data 352A includes the data of the number of generation sources of the prediction target past measurement data 351A such as the number of energy consumers and the number of communication terminals connected to the communication base station, and the prediction target past measurement data 351A itself. It is data that can explain the value of the prediction target past measurement data 351A.

The prediction target past measurement data 351A includes, for example, energy consumption data such as power, gas, and water as a total of a measuring instrument or a plurality of measuring instruments, communication data of a certain communication base station, and a moving object such as a taxi. This is the number of operating units for each hour.

The prediction calculation device 2 acquires the prediction target past measurement data 351A and the explanatory variable past measurement data 352A stored in the data management device 3. The prediction calculation device 2 predicts a future value at an arbitrary point in time by the first prediction calculation unit 251A, and additionally stores the prediction calculation result data 253A in the prediction calculation result data storage unit 253.

Further, the prediction calculation device 2 inputs the prediction target past measurement data 351A stored in the data management device 3 and the latest observation data 303 transmitted from the data observation device 6 to the second prediction calculation unit 251B. The prediction calculation device 2 calculates the error of the first prediction from a predetermined past date and time, and predicts the error amount at any future date and time of the first prediction by modeling the tendency of occurrence of the error. .

Then, the prediction calculation device 2 uses the second prediction calculation result data 304 output from the second prediction calculation unit 251B and the first prediction calculation result data 302 output from the first prediction calculation unit 251A. Input to the predicted value correction unit 252.

The prediction calculation device 2 corrects the first prediction calculation result data 302 with the second prediction calculation result data 304, outputs the third prediction calculation result data 305, and transmits it to the plan creation / execution management device 5.

The first, second, and third prediction

calculation result data

302, 304, and 305 calculate, in addition to the future prediction value of the prediction target past measurement data 351A, the data of the section indicating the fluctuation range of the prediction value, or the prediction value. Data of the model formula and the coefficient value thereof.

With reference to FIG. 4, the processing procedure of the data prediction process of the data prediction system 12 in this Embodiment is demonstrated. This process is a process that starts when the predictive computation device 2 accepts an input operation from the device user or when the execution time set in advance via the information input / output terminal 4 is reached. Thus, the processing from step S401 to step S404 is executed.

In practice, processing is executed based on various computer programs stored in the CPU 21 and the storage device 25 of the predictive calculation device 2 and various computer programs stored in the CPU 31 and the storage device 35 of the data management device 3. . For convenience of explanation, the processing subject will be described as the predictive arithmetic device 2 and various computer programs included in the predictive arithmetic device 2.

First, the first prediction calculation unit 251A of the prediction calculation unit 251 acquires and receives the prediction target past measurement data 351A and the explanatory variable past measurement data 352A from the data management device 3. Next, the first prediction calculation unit 251A uses the information input / output terminal 4 based on the correlation between the value of the prediction target past measurement data 351A and the value of the explanatory variable such as the calendar date and weather information of the explanatory variable past measurement data 352A. The first prediction calculation result data 302 at a plurality of future points preset in advance is calculated. Thereafter, the first prediction calculation unit 251A additionally records the prediction calculation result data 253A in the prediction calculation result data storage unit 253 (S401).

A publicly known method may be applied as a method used when the first prediction calculation unit 251A performs prediction. Known methods include, for example, a prediction method (prediction method) based on an arithmetic average value of a similar past period (similar days, etc.) set in advance via the information input / output terminal 4 based on the day of the week or the temperature. . Other known methods include a prediction method using a single regression model or a multiple regression model, a prediction method using a neural network, a prediction method using time series analysis such as an AR model or an ARIMA model, and the like.

Also, the method used for the prediction may be a method in which only the lag having a statistically significant correlation is set as the order in the model order identification in the time series analysis. An example of a specific embodiment of the first prediction calculation unit 251A will be described later.

Next, the second prediction calculation unit 251B of the prediction calculation unit 251 acquires a prediction value from the first prediction calculation result data 302 of a predetermined past period from the prediction calculation result data 253A. The second prediction calculation unit 251B acquires actual measurement values for the same period from the prediction target past measurement data 351A or the latest observation data 303 acquired from the data observation device 6. The second prediction calculation unit 251B calculates first prediction error data (error series 310) as a difference between the predicted value and the actual measured value (S402).

An error generation tendency model is created from the calculated first prediction error data, and the first prediction error amount for a predetermined future period is calculated as the second prediction calculation result data 304 from the created model. (S403). The method used when the second prediction calculation unit 251B performs the prediction is the same as the method used when the first prediction calculation unit 251A described above performs the prediction, and the description thereof is omitted here.

Finally, the predicted value correction unit 252 uses the first prediction calculation result data 302 calculated by the first prediction calculation unit 251A based on the second prediction calculation result data 304 calculated by the second prediction calculation unit 251B. It correct | amends and calculates the 3rd prediction calculation result data 305 (S404). Specifically, for example, the prediction value of the second prediction calculation result data 304 is corrected by adding it to the prediction value of the first prediction calculation result data 302.

With the above processing, the data prediction processing in this embodiment is completed.

(1-4) Details of Each Component A first embodiment of the first prediction calculation unit 251A of the data prediction system 12 in the present embodiment will be described with reference to FIGS.

(1-4-1) First Prediction Calculation Unit FIG. 5 shows a data flow of the first prediction calculation unit 251A according to this embodiment. The prediction calculation according to the present embodiment is characterized in that a series of prediction target past measurement data 351A in the prediction target period is calculated based on the periodic feature amount of the prediction target past measurement data 351A.

(1-4-1-1) Clustering Unit A data classification unit (hereinafter referred to as a clustering unit) 251A1 uses a past period preset from the prediction target past measurement data 351A via the information input / output terminal 4 such as the past year. Prediction target past measurement data 351A is acquired, and the data is classified from the viewpoint of the measurement point and time based on the periodic feature amount. The clustering unit 251A1 includes a metric point clustering unit 251A11 that classifies data from the viewpoint of metric points and a time clustering unit 251A12 that classifies data from the viewpoint of time (FIG. 6).

A measuring point refers to a measuring instrument that measures data, a person who owns the measuring instrument, an object in which a measuring instrument is installed (such as a taxi equipped with GPS), or a building in which a measuring instrument is installed. .

The process flow of the clustering unit 251A1 will be specifically described with reference to FIG. First, the metric point clustering unit 251A11 acquires, from the prediction target past measurement data 351A, the prediction target past measurement data 351A of a preset past period such as the past year, for example, via the information input / output terminal 4.

Next, the weighing point clustering unit 251A11 processes the acquired data into time-series data with the weighing point granularity set in advance via the information input / output terminal 4. The measurement point granularity means, for example, the granularity of each measuring instrument, the granularity as the total of all measuring instruments, or external information associated with each measuring instrument, such as the region and contract type, via the information input / output terminal 4 It is a unit of a plurality of measuring devices set in advance.

Then, using a frequency analysis such as Fourier transform or wavelet transform, a feature amount indicating a periodic feature of each measurement point granularity data is calculated, and clustering processing is performed on the calculated feature amount. Through the above processing, for example, measurement points having similar waveform shapes of time series data for the past year are classified as measurement point clusters.

The time clustering unit 251A12 first generates time-series data of the total value by calculating the total value for each measurement time of the prediction target data for each of the measurement point clusters generated by the measurement point clustering unit 251A11.

Next, the time clustering unit 251A12 divides the generated time series data of the total value for each weighing point cluster with the time granularity set in advance via the information input / output terminal 4. The time granularity is, for example, a granularity in units of 24 hours, a granularity of one year, or a granularity of an arbitrary time set in advance via the information input / output terminal 4. The time granularity may apply the same granularity in all weighing point clusters, or may differ for each weighing point cluster.

Then, the time clustering unit 251A12 calculates a feature quantity indicating a periodic feature from the data divided by the time granularity using frequency analysis such as Fourier transform and wavelet transform, and performs clustering processing on the calculated feature quantity. .

Through the above processing, for example, weighing points having similar waveform shapes of time series data for the past year are classified as weighing point clusters. Further, the clustering result data 251A13 is output, in which the classified metric point clusters are classified as time clusters, for example, in which the time-series waveform shapes of the total value are similar in 24-hour units.

A known method may be applied to the clustering processing performed by the metric point clustering unit 251A11 and the time clustering unit 251A12. Known methods include k-means, EM algorithm, and spectral clustering, which are unsupervised clustering algorithms for neighborhood optimization.

Also, known methods include unsupervised SVM (Support Vector Vector), VQ algorithm, and SOM (Self-Organizing Maps), which are unsupervised clustering algorithms for discriminating plane optimization.

In determining the number of clusters to be classified, an index such as data similarity or data cohesion within a cluster calculated based on the variance within each cluster or a cluster separation calculated based on a distance between clusters may be used.

Before calculating the feature quantity indicating the periodic feature of the time series data in the metric point clustering unit 251A11 and the time clustering unit 251A12, normalization processing is performed so as to remove the magnitude information of the value of the time series data. Also good.

The normalization processing may be normalization such that the average is 0 and the variance is 1, for example. By calculating a feature amount indicating a periodic feature from the normalized time series data, it is possible to perform clustering based only on the waveform shape similarity of the time series data. The above is the first embodiment of the clustering unit 251A1.

(1-4-1-2) Profiling Unit The data generation unit (hereinafter referred to as profiling unit) 251A2 uses the clustering result data 251A13 and the explanatory variable past measurement data 352A output from the clustering unit 251A1. Thus, the profiling unit 251A2 specifies an explanatory variable that exists in common in each time cluster of the clustering result data 251A13 and calculates a range of the values.

Specifically, the profiling unit 251A2 uses an identifier such as a number or name that identifies each time cluster as a teacher label. Then, the profiling unit 251A2 calculates, as profiling result data, explanatory variables and their value ranges that are commonly present in each time cluster of each metric point cluster using a decision tree learning algorithm such as CART, ID3, and random forest. .

Note that the profiling result data calculation process is, for example, a process that places importance on the time series data of the past period highly correlated with the prediction period. Specifically, a weight value that places importance on time-series data of the past period highly correlated with the prediction period is applied to a decision tree learning algorithm such as CART, ID3, and random forest.

Here, the past period having a high correlation with the prediction period is, for example, the latest past day of the prediction target day or the past day of the same season as the prediction target day when there is a seasonal periodicity in the fluctuation of the prediction target. That is, the weight value W _i for data of a certain past date i is given as a function of the prediction period and the correlation C _i as follows:

Fig. 20 shows a conceptual diagram of the effect. Here, as a result of processing in the profiling unit 251A2, it is assumed that the attribute of daily average temperature is the dominant explanatory variable among the explanatory variables that exist in common in each time cluster.

As shown in the graph 801 of the daily average temperature throughout the year, the time series data of the future period of the prediction target is generated from the past time series data of the prediction target in the same season and the same average temperature as the prediction target day. .

The graph 803 for the most recent 7 days in which the prediction target day and the daily average temperature are the same is different from the graph 802 in the same month in the previous year where the prediction target day and the daily average temperature are the same, Assume that it has changed over time. Here, if time series data of the future period to be predicted is generated without applying the weight value, it is generated as an average before and after the secular change as indicated by a thick line in the graph 804.

When the weight value is applied, as shown by the thick line in the graph 805, data closer to the state of the recent time transition after the secular change is generated. As described above, it is possible to generate time series data of the future period of the prediction target described later as data that more accurately reflects the secular change of the prediction target. Although the case where the profiling unit 251A2 generates data has been described here, the data may be acquired from the outside.

(1-4-1-3) Predicted Value Calculation Unit The predicted value calculation unit 251A4 uses the clustering result data 251A13, the profiling result data, and the explanatory variable data 301, which is the explanatory variable data, to predict the future of the prediction target. Generate predicted values of time-series data for the period. The explanatory variable data 301 includes a forecast value that is an expected value of the explanatory variable related to the future period to be predicted.

Specifically, the predicted value calculation unit 251A4 determines a time cluster to which the waveform shape of the time series data of the prediction target period belongs based on the profiling result data and the explanatory variable data 301. Then, representative time-series data is generated from the past data belonging to the determined time cluster, for example, by arithmetic mean for each time.

This typical time-series data generation process is performed for each weighing point cluster. For the representative time-series data of each metric point cluster, for example, an extreme value such as a local maximum value in the morning or a local minimum value in the afternoon is predicted by the adjustment reference value calculation unit 251A3.

By adjusting the representative time series data so that the difference between the predicted extreme value and the extreme value of the representative time series data is minimized, the predicted value in the time series of the prediction target period of each weighing point cluster Is calculated and output as first prediction calculation result data 503. The first prediction calculation result data 503 may be one series data obtained by adding the prediction values in the time series of the prediction target periods of the respective measurement point clusters with the values of the same date and time.

(1-4-1-4) Adjustment Reference Value Calculation Unit The adjustment reference value calculation unit 251A3 is, for example, a maximum value in the morning or a minimum value in the afternoon of representative time series data of each time cluster of each measurement point cluster. Predict extreme values such as This prediction is performed based on the prediction target past measurement data 351A, the explanatory variable past measurement data 352A, and the clustering result data 251A13 output from the clustering unit 251A1.

Specific case of one measuring point cluster will be explained. First, prediction target past measurement data of a measurement point belonging to this measurement point cluster, for example, for a period set in advance, such as the past year, is acquired from the prediction target past measurement data 351A.

Next, the total value time series of this weighing point cluster is generated by summing the acquired data for each time, and the generated total value time series is set to the time granularity set when generating the time cluster of this weighing point cluster. Divide by. Each divided data is classified based on, for example, a day type such as a month, a day of the week, a weekday, or a holiday, and all actual observed values of respective extreme values of the classified data are calculated. A model for predicting the extreme value using a single regression model or a multiple regression model is generated using the actual observation value of the calculated extreme value and the past actual observation value of an explanatory variable such as temperature.

極 Using the model generated last, the predicted value of the extreme value of the prediction target period is calculated and input to the predicted value calculation unit 251A4. With the above processing, the first prediction calculation processing in the present embodiment is completed.

(1-4-1-5) Modified Example of First Prediction Calculation Unit The value calculated by the adjustment reference value calculation unit 251A3 is an extreme value in a prediction target period such as a maximum value in the morning or a minimum value in the afternoon, for example. However, the present invention is not limited to this, and may be an integrated value of prediction target data in the prediction target period (embodiment for calculating the integrated value).

In the embodiment for calculating the integrated value, the data input to the adjustment reference value calculating unit 251A3 is the time-series data of the integrated value for each period of the prediction target past measurement data 351A and the integration for each period of the explanatory variable past measurement data 352A. It becomes time-series data of representative values such as values and average values. The processing of the embodiment in which the value calculated by the adjustment reference value calculation unit 251A3 is the integrated value of the prediction target data in the prediction target period is the embodiment in which the value calculated by the adjustment reference value calculation unit 251A3 is the extreme value in the prediction target period. This is the same as the process.

In the embodiment for calculating the integrated value, the predicted value calculation unit 251A4 outputs the first predicted calculation result data 503. The predicted value calculation unit 251A4 represents the representative time series data so that the residual sum of squares of the predicted value of the representative time series data in the prediction target period and the integrated value in the prediction target period is minimized. Adjust the predicted value of.

The integrated value of predicted values of typical time series data input to the predicted value calculation unit 251A4 is calculated by the clustering unit 251A1 and the profiling unit 251A2. The integrated value in the prediction target period input to the predicted value calculation unit 251A4 is calculated by the adjustment reference value calculation unit 251A3.

(1-4-1-6) Another Modification of First Prediction Calculation Unit In addition, a new measurement point may be added or an existing measurement point may be removed as time passes. . Especially for newly added weighing points, there are cases where there are no measurement values of past time series data, or there are fewer measurement data than other measurement points even if measurement values exist. . In these cases, the newly added weighing point and the other weighing points cannot be subjected to clustering processing at the same time by the weighing point clustering unit 251A11, and the weighing point cluster to which the newly added weighing point should belong is determined. It cannot be specified.

Therefore, the feature amount indicating the periodic feature of the time series data obtained at the present time of the newly added measurement point cluster is calculated. You may perform the process which makes a new measurement point belong to the measurement point cluster with the closest feature quantity with this feature quantity. The metric point cluster having the closest feature amount uses, for example, the feature amount of the cluster center (cluster center point) in the same period of each metric point cluster already calculated by the metric point clustering unit 251A11 according to a scale such as the Euclidean distance. To calculate. The above processing may be performed at an arbitrary time timing or may be performed at regular time intervals.

In addition, in the measurement point cluster to which a new measurement point is added, when the scale of the measurement value of the newly added measurement point is large or the number of measurement points to be added is large, for example, the past, such as one year ago. The scale of the data to be predicted may be greatly different between the time point and the prediction target date. This causes a problem that the extreme value predicted by the adjustment reference value calculation unit 251A3 and the predicted result of the integrated value cannot be calculated appropriately.

In this case, newly added weighing points are first calculated from the representative time series data of the weighing point cluster calculated from the average time series data of the weighing point cluster or the center of gravity of the features of the weighing point cluster. Estimate the data time series in the past unmeasured period. Then, these estimated past data time series are added to the past time series data of the measurement points that already belong to the measurement point cluster, and then the extreme value and the integrated value are predicted in the adjustment reference value calculation unit 251A3. . This makes it possible to appropriately calculate the prediction results of extreme values and integrated values.

(1-4-2) Second Prediction Calculation Unit Next, a first embodiment of the second prediction calculation unit 251B of the data prediction system 12 in the present embodiment will be described with reference to FIG.

The second prediction calculation unit 251B in the present embodiment generates a model of an error generation tendency, predicts a sequence of error amounts of the first prediction in the prediction target period using the generated model, and outputs a second prediction calculation result Output as data 304. The error generation tendency model is based on a series of prediction errors calculated up to the present by the first prediction calculation unit 251A.

The second prediction calculation unit 251B includes a model generation unit (hereinafter referred to as an error sequence generation unit) 251B1 that generates a sequence of error occurrence tendency, an error model identification unit 251B2 that predicts an error amount system, and an error prediction amount And an error prediction amount calculation unit 251B3.

Specifically, first, the error series generation unit 251B1 calculates a prediction error time series 310B which is a difference between the first prediction calculation result data 302 and the latest observation data 303. Further, the error series generation unit 251B1 has a prediction error time series 310A that is a difference between the prediction calculation result data 253A that is the first prediction calculation result data 302 in the predetermined past period and the prediction target past measurement data 351A in the same past period. Is calculated. The error series generation unit 251B1 connects the two prediction error series as an error series 310 which is one time series data.

Next, the error model identification unit 251B2 uses the time series analysis method to determine the order in the time series model such as the AR model or the ARIMA model and to estimate the coefficient. For determining the order, a known technique is used in which the Akaike information criterion (AIC) under several orders is calculated, and the order in which the value of the Akaike information criterion is the minimum is applied. Also good. In order to determine the order, a method may be used in which a lag in which the value of autocorrelation or partial autocorrelation of time series data is statistically significant is applied as the order. In the estimation of the coefficient, a known method such as estimation by the least square method under the applied order may be applied.

Then, the error prediction amount calculation unit 251B3 calculates the predicted value of the time series error amount of the first prediction calculation result in the prediction target period using the generated model, and outputs it as the second prediction calculation result data 304. To do.

With the above processing, the second prediction calculation processing in this embodiment is completed. When the first prediction calculation is performed every predetermined period such as every 24 hours, the prediction error sequence may be a discontinuous sequence at the boundary of the period. When the second prediction calculation is performed based on the discontinuous series, an appropriate second prediction calculation result cannot be obtained. Therefore, the second prediction calculation may be performed after removing discontinuous points (discontinuous points at the engine breaks) of the prediction error series, for example, by smoothing processing or the like.

The error model identification unit 251B2 and the error prediction amount calculation unit 251B3 have been described as applying a time series analysis method such as an AR model or an ARIMA model. However, the present invention is not limited to this, and the process of the first prediction calculation unit 251A is performed. May be substituted. In this case, the error model identification unit 251B2 is replaced by the processing of the clustering unit 251A1 and the profiling unit 251A2 shown in FIG. 5, and the error prediction amount calculation unit 251B3 is replaced by the predicted value calculation unit 251A4.

In addition, regardless of which method is applied by the error model identification unit 251B2, the processing is performed with emphasis on the time series 310B of the prediction error of the past period highly correlated with the prediction period. Specifically, weight values that place importance on the first prediction calculation result data of the past period highly correlated with the prediction period are used for error model identification processing using a time series analysis method such as an AR model or an ARIMA model. Apply.

Or, when the processing of the clustering unit 251A1 and the profiling unit 251A2 shown in FIG. 5 is applied to the error model identification unit 251B2, the weight value is applied to a decision tree learning algorithm such as CART, ID3, and random forest.

Here, the past period having a high correlation with the prediction period is, for example, the latest past day of the prediction target date, or the past date in the same season as the prediction target date when the prediction target has a seasonal periodicity. That is, the weight value W _i for data of a certain past date i is given as a function of the prediction period and the correlation C _i as in the equation (1).

Fig. 21 shows a conceptual diagram of the effect. First, the prediction error in the first prediction calculation result is as shown in the graph 851 for the whole year. Among them, the graph 853 of the data for the most recent 7 days of the prediction target date shows a change over time with respect to the prediction target date and the graph 852 of the data for the 7 days of the same month in the previous year as the error of the first prediction for each day Suppose you are.

Here, if the second prediction calculation result data is generated without applying the weight value, it is generated as an average error before and after the secular change as indicated by a thick line in the graph 854. On the other hand, when the weight value is applied, as shown by the thick line in the graph 855, data closer to the recent error state after the secular change is generated. As described above, it is possible to generate the second prediction calculation result data 304 that more accurately reflects the secular change of the prediction error.

(1-5) Description of Effects of the Present Invention Next, effects of the data prediction system 12 in the present embodiment will be described with reference to FIG.

FIG. 8 illustrates a prediction calculation result and an actual observation result (upper part of FIG. 8) in a certain period, and a prediction error amount (lower part of FIG. 8) in the same period. The current time when the data prediction system 12 in this embodiment is operating is indicated as “current 306” in the figure.

The upper graph in FIG. 8 shows “prediction target past measurement data 351A” and “latest observation” after “first prediction calculation result data 302 and prediction calculation result data 253A” calculated by the first prediction calculation unit 251A. The state where data 303 "is obtained is shown.

In FIG. 8, it is assumed that the latest observation data is collected two periods later than “Current 306”. Therefore, the second prediction calculation unit 251B calculates a series of errors indicated by “error series 310 of the first prediction calculation result” in the lower part of FIG. 8 as the prediction error confirmed at the “current 306” time point.

Further, the second prediction calculation unit 251B generates an error generation tendency model from the error series, thereby obtaining an error amount of the first prediction calculation until a future period set in advance via the information input / output terminal 4. , Calculated as “second prediction calculation result data 304” shown in FIG.

Here, for the generation of the error generation tendency model performed by the second prediction calculation unit 251B, a known method such as a method based on time series analysis using an AR model or an ARIMA model may be applied as described above. However, the method shown in FIGS. 5 and 6 may be applied as described above. Finally, the predicted value correction unit 252 outputs “third predicted calculation result data 305” by adding “second predicted calculation result data 304” to “first predicted calculation result data 302”. To do.

As described above, the second prediction calculation unit 251B models and predicts the fluctuation of the prediction target data that is difficult to explain by the result of the first prediction calculation unit 251A that performs prediction using the explanatory variable data 301. By correcting the first prediction calculation result data 302 of the first prediction calculation unit 251A with the second prediction calculation result data 304 that is a prediction result, fluctuation components that are difficult to explain with main explanatory variables can be obtained. Realize the reflected forecast. Third prediction calculation result data 305 is output as a prediction result.

The fluctuation of the prediction target data that is difficult to explain by the result of the first prediction calculation unit 251A includes, for example, fluctuation of data due to the periodicity of the data itself such that a constant amount is observed at night regardless of the temperature. Moreover, the fluctuation | variation of the prediction object data difficult to explain with the result of the 1st prediction calculating part 251A includes the fluctuation | variation of the data by inertia that air-conditioning demand maintains fixed even if temperature changes. In addition, fluctuations in the prediction target data that are difficult to explain with the results of the first prediction calculation unit 251A include data fluctuations due to sudden events such as typhoons and events.

As described above, the second prediction calculation unit 251B models and predicts the fluctuation of the prediction target data that is difficult to explain by the result of the first prediction calculation unit 251A that performs prediction using the explanatory variable data 301. By correcting the first prediction calculation result data 302 of the first prediction calculation unit 251A with the second prediction calculation result data 304 that is a prediction result, fluctuation components that are difficult to explain with main explanatory variables can be obtained. Realize the reflected forecast. The fluctuations in the forecasted data that are difficult to explain are the periodicity of the data itself, such that a certain amount is observed at night regardless of the temperature, the inertia that air conditioning demands remain constant even if the temperature changes, typhoons and events, etc. Fluctuations due to sudden events.

(2) Second Embodiment (2-1) Overall Processing and Data Flow of Data Prediction System According to this Embodiment This embodiment includes a plurality of prediction calculation units other than the first and second prediction calculation units. And providing a second prediction value correction unit that corrects the prediction value based on the results of a plurality of prediction calculation units, thereby enabling prediction that reflects data fluctuations that cannot be explained yet in the first embodiment. To do.

This embodiment will be described with reference to FIG. 9 and FIG. FIG. 9 shows a data flow between the functions of the data prediction system 12 in the present embodiment. Here, the second prediction value correction unit 252B and the fourth prediction calculation unit 251C, which are different from the first embodiment shown in FIG. 3, will be described.

First, the fourth prediction calculation unit 251C of the prediction calculation unit 251 includes a prediction value of a future period set in advance via the information input / output terminal 4 based on the prediction target past measurement data 351A and the latest observation data 303. The prediction calculation result data 701 is calculated.

The fourth prediction calculation result data 701 includes, in addition to the predicted value, data of a section indicating a fluctuation range of the predicted value or a model formula for calculating the predicted value and data of its coefficient value. This is the same as the first prediction calculation result data 302, the second prediction calculation result data 304, and the third prediction calculation result data 305.

Note that the target to be predicted by the fourth prediction calculation unit 251C is not the prediction target past measurement data 351A, but the total value for each measurement point cluster obtained from the result of the measurement point clustering unit 251A11 shown in FIG. May be. In this case, the result data of the metric point clustering unit 251A11 illustrated in FIG. 6 is input to the fourth prediction calculation unit 251C.

Then, the second predicted value correction unit 252B performs a correction process based on the first predicted calculation result data 302, the third predicted calculation result data 305, and the fourth predicted calculation result data 701, thereby obtaining the fifth predicted value. Calculation result data 703 is calculated.

FIG. 10 shows a processing procedure of the data prediction system 12 in the second embodiment. This process is a process that starts when the predictive calculation device 2 accepts an input operation from a device user or when an execution time set in advance via the information input / output terminal 4 is reached. In this process, the predictive computation device 2 executes the processes from step S401 to step S404, step S801, and step S802.

Actually, processing is executed based on various computer programs stored in the CPU 21 and the storage device 25 of the prediction arithmetic device 2 and various computer programs stored in the CPU 31 and the storage device 35 of the data management device 3. The For convenience of explanation, the processing subject will be described as the predictive arithmetic device 2 and various computer programs included in the predictive arithmetic device 2. Here, step S801 and step S802, which are different from the first embodiment shown in FIG. 4, will be described.

First, the fourth prediction calculation unit 251C of the prediction calculation unit 251 includes a prediction value of a future period set in advance via the information input / output terminal 4 based on the prediction target past measurement data 351A and the latest observation data 303. The prediction calculation result data 701 is calculated (S801). In the calculation, the fourth prediction calculation unit 251C uses a time series analysis method such as an AR model or an ARIMA model. Here, in model order identification in time series analysis, only a lag having a statistically significant correlation may be set as the order.

Next, the second prediction value correction unit 252B of the prediction value correction unit 252 includes the first prediction calculation result data 302 calculated by the first prediction calculation unit 251A, and the third calculation calculated by the first prediction correction unit 252A. Based on the prediction calculation result data 305 and the fourth prediction calculation result data 701 calculated by the fourth prediction calculation unit 251C, any one of the prediction calculation result data or all the prediction calculation result data is corrected, thereby obtaining the fifth. Is calculated (S802), and is transmitted to the plan creation / execution management device 5.

Here, there are several embodiments for the processing of the second predicted value correction unit 252B. Hereinafter, description will be made with reference to FIGS. 11, 12, 13 and 14.

(2-2) Processing of Second Prediction Value Correction Unit (2-2-1) First Embodiment of Second Prediction Value Correction Unit According to this Embodiment FIG. The data flow between the functions in 1st embodiment of the part 252B is shown. The second predicted value correction unit 252B in the present embodiment calculates the amount of fluctuation of the predicted value of the prediction target future period set in advance via the information input / output terminal 4 from the plurality of prediction calculation result data.

The second predicted value correction unit 252B causes the index value to be within the minimum or the range set in advance via the information input / output terminal 4. The index value is a variance or standard deviation of the final predicted value variation, a confidence interval or a prediction interval, or VaR (Value at Risk) calculated from the distribution. The second predicted value correction unit 252B calculates final predicted calculation result data by combining a plurality of predicted calculation result data. A specific processing procedure and data flow will be described with reference to FIG.

First, the prediction value assumption unit 252B2 receives a model formula and coefficient values for performing the prediction calculation included in the first, third, and fourth prediction

calculation result data

302, 305, and 701. Further, the explanatory variable future assumption unit 252B1 calculates an assumed value in a prediction target period of explanatory information such as weather information such as temperature and calendar date based on the input explanatory variable past measurement data 352A.

A known method may be applied to the assumed value calculation method. Known methods include, for example, a method using time series analysis, a method using a single regression or multiple regression model, and a prediction method using a neural network. As a known method, there is a method using a probability model such as modeling as a distribution of measurement trends of past values or modeling as a stochastic process. Moreover, the estimated value of the prediction target period of the explanatory variable to be calculated may be one for each explanatory variable, or a plurality of values may be calculated for each explanatory variable.

The prediction value assumption unit 252B2 is a prediction target calculated by the explanatory variable future assumption unit 252B1 for the model formulas and coefficient values of the input first, third, and fourth prediction

calculation result data

302, 305, and 701. Substituting the assumed value of the explanatory variable in the period. Thus, the predicted value assumption unit 252B2 calculates the expected value of the predicted value in the first, third, and fourth predicted

calculation result data

302, 305, and 701 in the prediction target period.

In the predicted value synthesis unit 252B4, the predicted values of the predicted values in the first, third, and fourth predicted

calculation result data

302, 305, and 701 calculated by the predicted value assumption unit 252B2 are converted into preset synthesis ratio values. It is calculated as candidate result data of the fifth prediction calculation result data 503A by performing proportional distribution based on this and performing weighted averaging. The candidate result data of the fifth prediction calculation result data 503A includes an index value.

The composition ratio calculation unit 252B3 determines whether the index value is the minimum value or less than a preset threshold value. If the index value is the minimum value or not less than the preset threshold value, the first, third and The composite ratio value of the fourth prediction calculation result is changed.

Note that a known optimization algorithm such as a genetic algorithm may be applied to the process of changing the composition ratio value. Further, as in the Monte Carlo method, a process of extracting a composite ratio value whose index value is a minimum value or less than a preset threshold value from among a plurality of randomly set composite ratio values may be used.

(2-2-2) Second Embodiment of Second Prediction Value Correction Unit According to this Embodiment Next, referring to FIG. 12, the second prediction value correction unit 252B in the second embodiment The data flow between functions and the flow of processing will be described. The second predicted value correction unit 252B in the present embodiment calculates the amount of fluctuation of the predicted value of the prediction target future period set in advance via the information input / output terminal 4 from the plurality of prediction calculation result data.

The second predicted value correction unit 252B calculates the utility value in the prediction target period from the calculated fluctuation amount of the predicted value and the input external information. The second predicted value correction unit 252B is configured so that the representative value such as the final expected value of the utility value or the index value is minimized or maximized or set in advance via the information input / output terminal 4. A plurality of prediction calculation result data are synthesized so as to fall within the range. Thus, the second predicted value correction unit 252B calculates final prediction calculation result data.

The utility value quantifies the utility for the user of the data management system 1 that increases or decreases as a result of the operation of the physical facility created and executed by the plan creation / execution management device 5 that has received the prediction calculation result data. Is. For example, it is a loss amount such as an imbalance cost caused by a difference between a predicted or planned power demand value existing in an electric power business field or the like and an actual demand value.

Specific data and processing flow will be described with reference to FIG. Only the utility value calculation unit 252B5 and the composition ratio calculation unit 252B3, which are the differences from FIG. 11, will be described. The utility value calculation unit 252B5 uses the result candidate data of the fifth prediction calculation result data 503B calculated by the prediction value synthesis unit 252B4 and the external information data 502B received from the outside, and uses the utility in the prediction target period. Calculate the assumed value of the value and the fluctuation amount of the assumed value. Here, the fluctuation amount of the assumed value includes an index value of the assumed value.

A specific processing flow of the utility value calculation unit 252B5 in this embodiment will be described. The utility value calculation unit 252B5 first calculates the assumed value of the basic data for calculating the utility value in the prediction target period, based on the external information data 502B and the explanatory variable past measurement data 352A received. The external information data 502B is necessary information for calculating the utility value received from the outside of the data prediction system 12. For example, in the electric power field, the external information data 502B is wholesale transaction history data including past contract prices and contract information on wholesale power exchanges, or past history data of imbalance settlement prices calculated by the system operator.

In addition, the process of calculating the assumed value of the basic data for calculating the utility value in the prediction target period based on the external information data 502B is performed using the method described when the first prediction calculation unit 251A described above performs the prediction. You may go.

Next, the composite ratio calculation unit 252B3 has a minimum or maximum representative value such as an expected value or a total value of expected values of utility values in the prediction target period calculated by the utility value calculation unit 252B5, or an index value. It is determined whether it is within the range set in advance via. When the composite ratio calculation unit 252B3 has a determination result that is not minimum, maximum, or not within a preset range via the information input / output terminal 4, the composite ratio calculation unit 252B3 determines the first, third, and fourth prediction calculation results. Change the composite ratio value. The process for changing the composition ratio value is the same as that in the embodiment shown in FIG.

With the above processing, the utility value calculation unit 252B5 in the present embodiment, based on the utility values predicted or assumed in the prediction target period, the first, third, and fourth prediction

calculation result data

302, 305 And 701 are corrected, and fifth prediction calculation result data 503B is calculated.

In the case of positive utility such as income and profit, the final value of the utility value will be maximized in the case of negative utility such as expenditure and loss so that the representative value such as the final expected value of utility value will be maximized. Processing may be performed so that a representative value such as an expected value is minimized.

Further, although the utility value here has been described as a value belonging to a monetary concept, it is not limited thereto, and may be a value belonging to a sensory concept such as comfort. Further, the utility value calculated by the utility value calculation unit 252B5 has been described as being calculated as a future value that may occur in the prediction target period, but is not limited thereto, and is calculated as a utility value in a predetermined past period. May be.

In this case, the composite ratio calculation unit 252B3 has a minimum value, a maximum value, or a representative value such as an average value or a total value of past utility values calculated in a predetermined past period, via the information input / output terminal 4. To determine whether it is within the preset range. When the determination result is not minimum, not maximum, or not within a range set in advance via the information input / output terminal 4, the synthesis ratio calculation unit 252B3 outputs the first, third, and fourth prediction

calculation result data

302, 305, and 701. Change the composite ratio value.

(2-2-3) Third Embodiment of Second Predicted Value Correction Unit According to This Embodiment Next, referring to FIG. 13, in the third embodiment of the second predicted value correction unit 252B The data flow between functions and the flow of processing will be described. In the second predicted value correction unit 252B in the present embodiment, the prediction target period is based on the predicted calculation result data 253A storing the predicted

calculation result data

302, 305, and 701 from the first, third, and fourth past. The fifth prediction calculation result data 503C is calculated by determining the prediction calculation result data to be applied.

Specifically, first, the state learning unit 252B6 obtains state information that is relational information between each prediction calculation result and attribute information of the prediction target period from the input prediction calculation result data 253A and explanatory variable past measurement data 352A. calculate. The attribute information of the prediction target period is, for example, a month type, a day of the week, a time zone, a day type indicating a weekday or a holiday, a weather information such as temperature or humidity, or a previous prediction target period. It is shown as information including these information.

The state of the forecast target period is, for example, “month = 1”, “day of the week = month”, “day type = weekday”, “average temperature = 7”, “average temperature difference from previous forecast target period = −3” . In this case, the state information that is the relationship information between the state of the prediction target period and the first prediction calculation result is “average error rate 4.32%”. In addition, the state information that is the relationship information between the state of the prediction target period and the second prediction calculation result is “average error rate 1.22%”. Further, the state information which is the relationship information between the state of the prediction target period and the third prediction calculation result is “average error rate 6.01%”.

As an example of the state information, the average error rate in each prediction calculation has been described as being related, but not limited to this, for example, label information such as “appropriate” and “inappropriate” may be associated. Further, information representing an error sequence of each prediction calculation result, for example, an error sequence array, result information obtained by performing frequency analysis such as Fourier transform on the error sequence, or the like may be related.

The processing for calculating the relationship information between each prediction calculation result and the state of the prediction target period may be set in advance via the information input / output terminal 4, for example. A known learning algorithm may be applied to this calculation process. Known algorithms include decision tree learning algorithms such as CART, ID3, and random forest, and discriminator learning algorithms such as SVM (Support Vector Vector) and Naive Bayes. A known algorithm uses, for example, information indicating an error sequence of each prediction calculation result, for example, an error sequence array or result information obtained by performing frequency analysis such as Fourier transform on the error sequence as a teacher label.

Next, the applied prediction calculation result determination unit 252B7 calculates the prediction calculation result to be applied to the prediction target period based on the state information calculated by the state learning unit 252B6 and the predicted value of the explanatory variable data related to the prediction target period. judge. The prediction calculation result switching unit 252B8 selects one of the first, third, or fourth prediction

calculation result data

302, 305, and 701 based on the information of the prediction calculation result to be applied calculated by the application prediction calculation result determination unit 252B7. Is selected and calculated as fifth prediction calculation result data 503C.

With the above processing, the utility value calculation unit 252B5 in the present embodiment is based on the prediction calculation result data 253A that stores the prediction calculation result data from the first, third, and fourth past. Prediction calculation result data to be applied in the step is determined. The utility value calculation unit 252B5 calculates fifth prediction calculation result data 503C based on the determination result.

In the third embodiment of the second predicted value correction unit 252B, the prediction calculation result to be applied to the prediction target period is determined based on the relationship information between the prediction calculation result and the state of the prediction target period. Although described, the present invention is not limited to this. For example, you may determine the prediction calculation result applied to a prediction object period based on the data which the predicted value assumption part 252B2 shown in FIG. 11 and FIG. 12 calculated.

In this case, the prediction calculation result switching unit 252B8 has a representative value such as an expected value of the predicted value in the first, third, or fourth prediction target period calculated by the predicted value assumption unit 252B2, an index value, or the like. A process of switching to the minimum prediction calculation result data is performed.

In addition, the prediction calculation result switching unit 252B8 is not limited to this, based on the information on the prediction calculation result of the utility value in each prediction calculation result data in the prediction target period calculated by the utility value calculation unit 252B5 illustrated in FIG. You may determine the prediction calculation result applied to a prediction object period. Further, the prediction calculation result switching unit 252B8 determines a prediction calculation result to be applied to the prediction target period based on information on the result of the utility value in each prediction calculation result data in the past period calculated by the utility value calculation unit 252B5. Also good.

In this case, the prediction calculation result switching unit 252B8 represents a representative value such as an expected value or total value of the utility value in each prediction calculation result data calculated by the utility value calculation unit 252B5, an index value of the utility value, or the like. Based on this, the process of switching the prediction calculation result data is performed. Alternatively, the prediction calculation result switching unit 252B8 switches the prediction calculation result data based on a representative value such as an average value or total value of utility values in each prediction calculation result data in a predetermined past period, an index value of the utility value, or the like. Process. The predicted calculation result data is set to a value that is maximum for a positive utility and minimum for a negative utility.

Further, the prediction calculation result switching unit 252B8 calculates a plurality of prediction calculation results (preliminary prediction calculation results) calculated and output at a certain past time point, and each prediction calculation result calculated and output again at a time point after the past time point. (The latest prediction calculation result) may be compared. By this comparison, a prediction calculation result applied to the prediction target period may be determined. More specifically, an error that is a difference between each of a plurality of prediction calculation results and the latest prediction calculation result is calculated, and the prediction calculation result with the smallest error is applied to the prediction target period. Calculated as the result.

(2-2-4) Fourth Embodiment of Second Prediction Value Correction Unit According to This Embodiment Next, with reference to FIG. 14 and FIG. 15, the fourth prediction value correction unit 252B has the fourth Processing in the embodiment will be described.

FIG. 14 shows a data flow showing an embodiment in which a prediction calculation result correction unit 252B9 described later is added to the third embodiment of the second prediction value correction unit 252B shown in FIG. FIG. 15 is a graph display of data input to the second predicted value correction unit 252B in the present embodiment and data output as a result of the processing of the second predicted value correction unit 252B in the present embodiment. It is. Here, for the sake of simplicity, the processing when the first and fourth prediction

calculation result data

302 and 701 are input will be described.

The first prediction calculation result data 302 and the fourth prediction calculation result data 701 as shown in the upper part of FIG. 15 are input to the prediction calculation result correction unit 252B9. The prediction calculation result correction unit 252B9 corrects other prediction calculation result data with reference to preset prediction calculation result data.

For example, when the prediction calculation result data used as the reference is the fourth prediction calculation result data 701, the first prediction calculation result data 302 has the smallest residual at the same time as the fourth prediction calculation result data 701. Correct so that

The specific correction process may be, for example, a process of correcting so that the least-square residual sum at the same time is minimized. 14 shows the first prediction calculation result data 302 after the correction processing, and the corrected prediction calculation result data is output as fifth prediction calculation result data 503D. Here, the prediction calculation result data used as a reference in the correction processing may be set, for example, via the information input / output terminal 4 or applied prediction calculation result data calculated by the application prediction calculation result determination unit 252B7 shown in FIG. It is good also as prediction calculation result data on the basis of.

The description has been given of the case where the first and fourth prediction

calculation result data

302 and 701 are input. Similarly, when the third prediction calculation result data 305 is input, the reference is also used as a reference. Correction processing based on the prediction calculation result data is performed. Further, the fifth predicted calculation result data 503D output from the second predicted value correction unit 252B in the present embodiment is the fourth predicted

calculation result data

302, 305, or 701 after the correction, the first, third, or reference. An arithmetic average value for each time may be used. Further, the fifth prediction calculation result data 503D output from the second prediction value correction unit 252B may be the corrected first, third, or fourth predicted

calculation result data

302, 305, or 701 after correction.

Alternatively, based on the synthesis ratio calculated by the synthesis ratio calculation unit 252B3 shown in FIGS. 11 and 12, the corrected fourth, fourth, or reference prediction calculation result data is prorated, and the fifth You may output as prediction calculation result data 503D. In addition, here, the correction based on the fourth prediction calculation result data 701 has been described. However, the present invention is not limited thereto, and correction processing based on actually measured prediction target data may be performed.

(2-3) Modified Example According to this Embodiment In the description of the second embodiment of the data prediction system 12, the prediction calculation unit 251 includes the first, second, and fourth three

prediction calculation units

251A, 251B, and 251C. However, the present invention is not limited to this, and may be composed of four or more prediction calculation units. In this case, the prediction calculation result data input to the second prediction value correction unit 252B is assumed to be four or more, and the processing of the second prediction value correction unit 252B is performed.

Particularly, a configuration in which a prediction calculation unit that replaces the first prediction calculation unit 251A shown in FIGS. 3 and 9 may be added. In this case, the series of prediction errors input to the second prediction calculation unit 251B performs a process of selecting one of the prediction calculation result data calculated from the first prediction calculation unit 251A and the added prediction calculation unit. . This modification will be described with reference to FIG.

FIG. 16 shows a data flow of the prediction calculation result data selection unit 254 in this modification. Prediction calculation result data or a prediction error sequence to be input to the second prediction calculation unit 251B is determined from the three prediction calculation result data. The three prediction calculation result data are the first prediction calculation result data 302, the sixth prediction calculation result data 1501, and the seventh prediction calculation result data 1502.

As shown in FIGS. 3 and 9, the first prediction calculation unit 251A calculates the first prediction calculation result data 302. Similarly, the sixth prediction calculation unit calculates sixth prediction calculation result data 1501 and the seventh prediction calculation unit calculates seventh prediction calculation result data 1502.

First, the continuity test unit (hereinafter referred to as an error series test unit) 254A calculates each prediction calculation result from the prediction target past measurement data 351A and the first, sixth and seventh prediction

calculation result data

302, 1501 and 1502. Generate a series of prediction errors. Next, the error series verification unit 254A performs a test of whether or not the prediction error series of each prediction calculation result is a stationary series, or calculates a numerical value indicating the degree of continuity of the prediction error series of each prediction calculation result. . In the test of whether or not it is a stationary series, a known test method such as an ADF (AugmentedmentDickey Fuller) test is used. As the numerical value indicating the degree of stationarity, for example, a numerical value indicating stationarity or non-stationarity such as a t value calculated as a result of the test process or a determination coefficient may be used.

Then, the data selection unit (hereinafter referred to as prediction calculation result data switching unit) 254B inputs to the second prediction calculation unit 251B based on the determination result calculated by the error series test unit 254A or a numerical value indicating the degree of continuity. Select the prediction calculation result data to be output. The prediction calculation result data is selected from the first, sixth or seventh prediction calculation result data and the prediction calculation result data 253A which is the first, sixth or seventh past prediction calculation result data.

With the above processing, the operation of the prediction calculation result data selection unit 254 is completed. The prediction error sequence of the prediction calculation result data selected or output by the prediction calculation result data selection unit 254 is a stationary sequence or a sequence closest to the stationary sequence among the prediction errors of each prediction calculation result data. . For this reason, it is possible to improve the prediction accuracy of the second prediction calculation result data calculated by the second prediction calculation unit 251B, and thus improve the prediction accuracy of the corrected third prediction calculation result data. it can.

Note that the prediction calculation processing of the sixth and seventh prediction calculation units that calculated the sixth and seventh prediction calculation result data is disclosed with reference to FIGS. 5 and 6 in the description of the first prediction calculation unit 251A. Alternatively, a known technique or a known technique may be applied. Known methods include a prediction method using a single regression model or multiple regression model, a prediction method using a neural network, and a prediction method using time series analysis such as an AR model or an ARIMA model. Also, as a known method, there is a prediction method based on an arithmetic average value in a similar past period set in advance via the information input / output terminal 4 based on a day of the week or an air temperature.

(3) Modification of system configuration (3-1) Application example to energy business field such as electric power, gas and water Next, referring to FIG. An embodiment when 1 is applied will be described.

The data management system 1A calculates a predicted value in time series of energy demand in a predetermined future period. The data management system 1A is a system for generating and controlling execution of an operation plan of an energy supply device that can be operated based on the calculated energy demand, for example, a generator, a gas or water delivery pump, and the like. Alternatively, the data management system 1A is a system for generating and executing a plan for an energy procurement transaction directly from another energy company or from an exchange.

Energy company 1000A, grid operator 7000A, transaction market operator 8000A, public information provider 9000A and customer 2000A, and various devices and terminals that each has. The energy company 1000A is a company composed of a supply and demand manager 1000A1, a facility manager 1000A2, and a transaction manager 1000A3.

Demand / supply manager 1000A1 is a department or person in charge who manages the amount of energy procurement. The department or person in charge who manages the amount of energy procurement explains the time series data of energy demand from the past measured for each customer, all customers, or each predetermined customer group, and fluctuations in energy demand. Based on the obtained factor data, the future energy demand is predicted. The factor data is, for example, weather, and the future unit of energy demand is, for example, 30 minutes. The department or person in charge who manages the energy procurement amount manages the energy procurement so that the predicted energy demand can be satisfied.

The supply and demand manager 1000A1 includes a data management device 3A that stores time-series data of measured energy demand from the past and factor data such as weather that can explain fluctuations in the energy demand. Further, the supply and demand manager 1000A1 includes a prediction calculation device 2A for calculating a predicted value of energy demand and an information input / output terminal 4A for exchanging data with these devices.

The facility manager 1000A2 is a department or a person in charge who plans and executes an operation plan of an energy supply facility owned by the company or an energy supply facility that is not owned by the company and can be incorporated into the energy procurement plan of the company.

The facility manager 1000A2 includes information management of energy supply facilities, planning of operation plans for energy supply facilities, and a facility management device 5A1 for transmitting control signals for execution. The facility manager 1000A2 also includes a control device 5A2 for receiving a control signal from the facility management device 5A1 and actually executing control of the energy supply facility.

The transaction manager 1000A3 is a department or a person in charge who plans and executes a transaction for procuring energy through a direct contract with another energy provider or through an exchange. The transaction manager 1000A3 includes a transaction management device 5A3 for managing information on energy procurement transaction plans and contracted energy procurement contracts, and exchanging messages related to transactions with other energy providers and exchanges.

System operator 7000A is a business operator that manages energy supply system facilities over a wide area, measures the actual energy demand of each customer in the area, and stores the measured values. The grid operator 7000A includes a grid information management apparatus 7A1 for distributing data of the measured energy demand of the consumer.

The transaction market operator 8000A is a business operator that comprehensively manages information and procedures necessary for conducting energy transactions with a plurality of energy business operators. The transaction market operator 8000A includes a market operation management device 7A2 for distributing information relating to energy transactions and performing an ordering process for orders received from each energy company.

The public information provider 9000A is a provider that provides past history information and future forecast information related to weather such as temperature, humidity, atmospheric pressure, wind speed, precipitation, and snowfall. And a public information distribution device 7A3 for distributing forecast information.

The customer 2000A is an individual or a corporation having energy consumption facilities and supply facilities. The customer 2000A transmits to the energy provider 1000A or the grid operator 7000A information that may affect the trend of energy demand and supply, such as owned equipment and facilities, type of business, number of people in the room, and location. Information input / output terminal 6A2. Further, the customer 2000A includes a measuring device 6A1 for measuring a demand amount and a supply amount of energy at a predetermined time interval and transmitting the energy demand amount and supply amount to the data management device 3A, the prediction arithmetic device 2A, or the system information management device 7A1.

The above is an example of the apparatus configuration in the embodiment when the data management system 1 is applied to an energy business field such as electric power, gas, or water. Note that the processing of each device and the processing unit included in each device in the present embodiment is the same as the processing in the already disclosed embodiment.

However, the prediction target past measurement data 351A is measurement data of energy consumption in time series. The time series is a time series for each consumer or energy meter or a time series as a total value for each time of all consumers or all energy meters. Alternatively, the time series is a time series as a total value for each time of a consumer group unit or energy meter group unit determined in advance by an energy provider. The measurement data may be not only energy consumption but also measurement data of an energy supply device such as a solar power generator.

In addition, the explanatory variable past measurement data 352A includes time-series data of calendar days, data of time-series of weather, time-series data of events that cause temporary fluctuations in energy demand or supply, Includes social trend data and attribute data. The calendar day is a day type such as a month, a day of the week, a weekday, or a holiday, and the weather is temperature, humidity, temperature, wind speed, precipitation, snowfall, or the like. Events that cause temporary fluctuations in energy demand or supply include the presence or absence of typhoons or events such as sports events.Social trends refer to fluctuations in energy demand or supply such as industry dynamic information. It is a trend to bring. The attribute is information related to the customer such as the type of energy supply contract of the customer, the type of business, the type of building, and the floor area.

(3-2) Application Example to Transportation Business Field such as Taxi Next, an embodiment when the data management system 1 is applied to a transportation business field such as a taxi will be described with reference to FIG.

The data management system 1B calculates a predicted value in time series of the transportation demand amount in a predetermined future period in each business office, a total of a plurality of business offices, or a specific place or region. The data management system 1B is a system for generating and executing an operation plan for an operable facility, such as a taxi, based on the calculated transportation demand. The data management system 1B is composed of a transportation company 1000B, a public information provider 9000B, a mobile body 2000B, and various devices and various terminals included in each.

The transportation company 1000B is a company composed of an operation status manager 1000B1 and an operation commander 1000B2. The operation status manager 1000B1 is a department or a person in charge who predicts a future transportation demand amount, for example, every 30 minutes. Future transportation demand is predicted based on time-series data of transportation demand from the past and factor data such as temporary events. The transportation demand is measured at a sales office unit, a total of a plurality of sales offices, or a specific place or region. The temporary event is, for example, a sports event that can explain the change in the transportation demand.

The operation status manager 1000B1 includes a data management device 3, a prediction calculation device 2B for calculating a predicted value of the transportation demand, and an information input / output terminal 4B for exchanging data with these devices. The data management device 3 stores time-series data of the measured transportation demand from the past and factor data that can explain fluctuations in the transportation demand.

The operation commander 1000B2 is a department or a person in charge for planning and executing an operation plan for facilities related to transportation such as taxis. The operation commander 1000B2 includes an operation management device 5B1 for transmitting facility information management, facility operation plan planning, and execution instructions, and a command execution device 5B2. The command execution device 5B2 receives an instruction from the operation management device 5B1 and actually executes or supports the execution.

The public information provider 9000B is a business entity that provides information on the presence / absence of sports events or the like, past history information on weather such as temperature, humidity, atmospheric pressure, wind speed, precipitation, and snowfall, and future forecast information. . The public information provider 9000B includes a public information distribution device 7B for distributing past history information and forecast information.

The mobile body 2000B is a facility for carrying information, and includes an information input / output terminal 6B2 and a measuring device 6B1 for measuring the amount of transportation demand at a predetermined time interval and transmitting it to the data management device 3B or the predictive computing device 2B. Prepare. The information input / output terminal 6B2 transmits information that may affect the transportation situation such as the location of the facility to the carrier 1000B.

The above is an example of the device configuration in the embodiment when the data management system 1 is applied to the transportation business field. Note that the processing in each device and the processing unit included in each device in the present embodiment is the same as the processing in the already disclosed embodiment.

However, in the present embodiment, the prediction target past measurement data 351A is the past measurement data of the transportation demand amount in time series measured for each sales office, the total of a plurality of sales offices, or each specific place or region.

The explanatory variable past measurement data 352A is a time series of factors that cause fluctuations in transportation demand, such as calendar time series data, weather time series data, presence / absence of typhoons, presence / absence of sports events, etc. It is data in.

(3-3) Application Example to the Telecommunications Business Field Next, an embodiment when the data management system 1 is applied to the telecommunications business field will be described with reference to FIG.

The data management system 1C calculates a predicted value in time series of the data communication amount in a predetermined future period in units of base stations or in total of a plurality of base stations. The data management system 1C is a system for generating and executing an operation plan for an operable facility, for example, a facility such as an exchange, based on the calculated data communication amount. The data management system 1C includes a telecommunications carrier 1000C, a public information provider 9000C, a base station 2000C, and various devices and terminals included in each. The communication carrier 1000C is a carrier composed of a communication status manager 1000C1 and a facility manager 1000C2.

The communication status manager 1000C1 is a department or person responsible for predicting the future data communication volume for every 30 minutes, for example, based on the time-series data of data communication volume from the past and factor data such as temporary events. It is a person. The amount of data communication from the past is measured for each base station, for a plurality of base stations, or for each predetermined base station group. The temporary event is, for example, a sporting event or the like that can account for fluctuations in data traffic.

The communication status manager 1000C1 includes a data management device 3C, a prediction calculation device 2C for calculating a predicted value of the data communication amount, and an information input / output terminal 4C for exchanging data with these devices. The data management device 3C stores time-series data of the measured data communication amount from the past and factor data that can explain the fluctuation of the data communication amount.

The facility manager 1000C2 is a department or a person in charge who creates and executes an operation plan for facilities related to data communication such as an exchange. The facility manager 1000C2 receives the control signal from the facility management device 5C1 for transmitting the control signal for managing the facility information, planning the operation plan of the facility, and executing it, and actually receiving the control signal from the facility management device 5C1. And a control device 5C2 for executing the control.

Public information provider 9000C provides information on the presence or absence of sports events, past history information on future weather information such as temperature, humidity, atmospheric pressure, wind speed, precipitation, and snowfall. Business operators. The public information provider 9000C includes a public information distribution device 7C for distributing past history information and forecast information.

The base station 2000C is a facility for controlling data communication, and is an information input / output terminal 6C2 and a measuring device for measuring the amount of data communication at predetermined time intervals and transmitting it to the data management device 3C or the prediction arithmetic device 2C. 6C1. The information input / output terminal 6C2 transmits information that can affect the tendency of data communication such as the location of equipment to the communication carrier 1000C.

The above is an example of the device configuration in the embodiment when the data management system 1 is applied to the telecommunications business field. Note that the processing in each device and the processing unit included in each device in the present embodiment is the same as the processing in the already disclosed embodiment.

However, in the present embodiment, the prediction target past measurement data 351A is the past measurement data of the data communication amount in time series measured for each base station, a plurality of base stations, or a predetermined base station group.

The explanatory variable past measurement data 352A is a time series of factors that cause fluctuations in the amount of data communication such as calendar time series data, meteorological time series data, presence / absence of typhoon landing, presence / absence of sports events, etc. The data at.

DESCRIPTION OF SYMBOLS 1 ... Data management system, 2 ... Prediction arithmetic device, 3 ... Data management device, 4 ... Information input / output terminal, 5 ... Plan creation and execution management device, 6 ... Data observation device, 7 ... Data Distribution device, 8 ... communication path.

Claims

In a data prediction system that calculates a predicted value,
A data management device for managing data;
A predictive arithmetic unit,
The prediction calculation device includes a first prediction calculation unit that calculates a predicted value;
A second prediction calculation unit for calculating a prediction value of an error of the calculation result of the first prediction calculation unit;
A data prediction system comprising: a prediction correction unit that corrects a calculation result of the first prediction calculation unit using a calculation result of the second prediction calculation unit.
The first prediction calculation unit includes:
A data generation unit that generates or acquires representative time-series data in the prediction period of the prediction target data;
A reference value calculation unit for calculating a reference value for correcting the representative time-series data;
The prediction value calculation part which correct | amends the calculated said representative time series data based on the said reference value, and calculates the predicted value in the prediction period of prediction object data is provided. Data prediction system.
The first prediction calculation unit includes:
A reference value calculation unit that calculates a maximum value, a minimum value, or both of the prediction values in a predetermined period of the prediction period of the prediction target data;
The maximum value, the minimum value, or both in the same period of representative time series data in the prediction period of the prediction target data match the calculated maximum value, the minimum value, or both, or the residual is the minimum The data prediction system according to claim 1, further comprising: a predicted value calculation unit that corrects the representative time series data.
The first prediction calculation unit includes:
A reference value calculation unit that calculates a predicted value of the integrated value during a predetermined period in the prediction period of the prediction target data;
Predicted value calculation for correcting the representative time series data so that the integrated value of the representative time series data in the prediction period of the prediction target data coincides with the calculated integrated value or the residual is minimized. The data prediction system according to claim 1, further comprising: a unit.
The first prediction calculation unit includes:
A data classification unit that classifies time series data for each predetermined period unit of the prediction target data based on an index value indicating a periodic characteristic;
The data prediction unit according to claim 1, further comprising: a data generation unit configured to generate representative time series data in a prediction period of a prediction target from the group of time series data classified by the data classification unit. system.
The data generator is
Generating a representative time-series data by assigning a predetermined weight to each of the time-series data classified so as to place importance on the time-series data of the past period highly correlated with the prediction period. The data prediction system according to claim 5.
The prediction target data is a total value of values measured by a plurality of data observation devices, and the time series data is data for each predetermined period measured by each of the data observation devices,
The predicted value calculation unit
The data prediction system according to claim 5, wherein a predicted value for each group is calculated from the group of the time series data classified by the data classification unit, and the combined predicted value of the prediction target data is calculated.
The second prediction calculation unit
A model identifying unit that generates a model generated by time series analysis from error time series data that is time series data of errors of a first predicted value from a predetermined past;
A prediction amount calculation unit that calculates a prediction amount of an error of the first prediction value in a predetermined period of the prediction period from the generated model;
The data prediction system according to claim 1, further comprising:
The second prediction calculation unit
The error time series data related to the time transition of the error of the first predicted value for each predetermined period is classified based on an index value indicating a periodic characteristic of the error time series data, and the group of the classified error time series data A model identifying unit that generates representative error time-series data of the prediction target in the prediction period,
A prediction amount calculation unit for calculating a predicted value of an error of the first prediction calculation result from the error time series data;
The data prediction system according to claim 1, further comprising:
The model identification unit generates a model by assigning a predetermined weight to each of the error time series data so as to place emphasis on the error time series data of the past period highly correlated with the prediction period. The data prediction system according to claim 8.
The prediction calculation device is
A continuity test processing unit that performs a continuity test on each of the error time series data of the prediction related to each first prediction calculation result calculated by the plurality of first prediction calculation units that calculate the predicted value When,
A data selection unit that selects one of the first prediction calculation results based on the result of the test, and
The prediction correction unit
The prediction correction is performed based on a first prediction calculation result selected by the data selection unit and a second prediction calculation result related to a prediction error related to the selected first prediction calculation result. Data prediction system.
The prediction calculation device is
A prediction target period related to the first prediction calculation result corrected using each first prediction calculation result or second prediction calculation result calculated by the plurality of first prediction calculation units that calculate the prediction value A predicted value assumption unit for calculating a fluctuation range at
Based on the fluctuation range of each prediction value calculated by the prediction value assumption unit, the apportioning ratio of each prediction value is set such that the final fluctuation range of the prediction value in the prediction target period is the minimum or a predetermined value or less. A composition ratio calculation unit to calculate,
The prediction value composition part which carries out a weighted average of a plurality of 1st prediction calculation results using the distribution ratio which the composition ratio calculation part computed, and outputs as a prediction calculation result. Data prediction system.
The prediction calculation device is
A prediction target period related to the first prediction calculation result corrected using each first prediction calculation result or second prediction calculation result calculated by the plurality of first prediction calculation units that calculate the prediction value A predicted value assumption unit for calculating a representative value of utility value or a fluctuation range of utility value in
A composite ratio calculation unit for calculating a proportional ratio for maximizing or minimizing a representative value of the utility value, or a proportional ratio for which the variation range of the utility value is minimum or less than a predetermined value;
The prediction value composition part which carries out a weighted average of a plurality of 1st prediction calculation results using the distribution ratio which the composition ratio calculation part computed, and outputs as a prediction calculation result. Data prediction system.
The prediction calculation device is
Attribute information related to the first prediction calculation result corrected using each first prediction calculation result or second prediction calculation result calculated by the plurality of first prediction calculation units that calculate the prediction value; A state learning unit that calculates state information that is a combination with information specifying a prediction error or a prediction calculation result to be applied;
Based on the state information and the prediction information of the attribute information in the prediction period, an applied prediction calculation result determination unit that selects a prediction calculation result used in the prediction period;
The data prediction system according to claim 1, further comprising: a prediction calculation result switching unit that outputs the selected prediction calculation result as a prediction value.
The prediction calculation device is
It is a difference between a plurality of prior prediction calculation results generated in advance from a plurality of the first prediction calculation units and a previous prediction calculation result that is an output of the first prediction calculation unit calculated again after the previous prediction calculation. An applied prediction calculation result determination unit that performs an error calculation, and selects a prediction calculation result based on the prior prediction calculation result that makes an index related to the error time series less than or equal to a predetermined value;
The data prediction system according to claim 1, further comprising: a prediction calculation result switching unit that outputs the selected prediction calculation result as a prediction value.
The prediction calculation device is
Among the prediction calculation results in the prediction period calculated by two or more prediction calculation units, so as to match at least one of the prediction calculation results, or so that the difference from the certain prediction calculation result is minimized. The data prediction system according to claim 14, further comprising: a prediction calculation result correcting unit that corrects and outputs the other prediction calculation result.
The data classification unit
The time series data measured by the data observation device is classified into a plurality of groups from index values indicating periodic characteristics of the data measured by the data observation device, and measurement in a second period that is less than the predetermined period is performed. The data measured by the second data observation device was used, the index value indicating the typical periodic characteristics of each group in the second period, and the periodic features of the data measured by the second data observation device A metric point clustering unit that classifies data into a group of data observation devices based on a distance from an index value indicating
The predicted value calculation unit
The data prediction system according to claim 5, wherein when adding the predicted values for each group of the time series data, correction is performed based on a classification result of the metric point clustering unit.
In a data prediction method to be executed in a data prediction system that calculates a predicted value in the future,
The data prediction system includes:
A data management device for managing data;
A prediction calculation device that corrects a predicted value in the future by modeling a tendency of an error amount of a prediction calculation result calculated based on a correlation with a main explanatory variable, and
A first step in which the data management device stores prediction target past measurement data and explanatory factor data described by the prediction target past measurement data, which are observed with time transition;
A second step in which the prediction arithmetic unit performs prediction based on a correlation between the prediction target past measurement data and the explanatory factor data;
A third step in which the prediction calculation device models a tendency of an error in the prediction calculation result of the prediction and predicts a future error amount of the prediction calculation result;
And a fourth step of correcting the prediction calculation result based on the calculation result of the error amount prediction.