JP2017010111A - Device, method, and program for assisting with noise data rejection - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device, method and program for assisting with noise data rejection that sufficiently removes noise data included in training data and makes the training data approximated to data consisting only of data having a true correlation.SOLUTION: A noise data rejection assistance device 100 comprises: a data storage unit 31 for storing, for multiple points of time, a history record that includes the value of an input variable and the actual value of an output variable at some point of time; and a preliminary prediction unit 12 for extracting, while changing a noise determination threshold within a prescribed range, a history record in which a difference between the predicted value of an output variable that is predicted from the value of an input variable and the actual value of an output variable is smaller than or equal to the noise determination threshold, calculating, on the basis of the extracted history record, a prediction function for predicting the value of an output variable from the value of an input variable and a prediction error of the prediction function, and selecting a noise determination threshold the calculated prediction error of which is smaller than others and outputting the selected threshold.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a noise data removal support apparatus, method, and program, and more particularly, to an abnormal data (noise data) removal technique for data input to an apparatus that learns a prediction model of a variable from past data.

  There exists a device that learns a model describing behavior of a variable value from past data and predicts a future variable value based on the model. Such devices are used in various fields, but in recent years, for example, the use in the field of predicting the amount of power generated by solar power generation has progressed.

  With the implementation of a fixed purchase system for renewable energy generated by sunlight and wind power, solar power generation has been spreading. It is an urgent issue to grasp the future power generation as accurately as possible from a few hours to a few days ago in order to plan the operation of power supply facilities that power companies bear and to reduce the operating costs in power market transactions. It has become. However, the output of photovoltaic power generation varies greatly depending on weather conditions, and its prediction is not easy.

  A technique for solving this problem is described in Patent Document 1. The apparatus described in Patent Document 1 uses a statistical method based on the least square method in order to predict the amount of photovoltaic power generation. This device statistically learns data representing past input / output relationships, using the feature value representing the weather condition at the target time at the target point as an input variable and the amount of photovoltaic power generation at the target time at the target point as an output variable. To create a prediction function at the target time. And this apparatus estimates the electric power generation amount by inputting an input variable value into the prediction function.

  This device provides a section determined by an upper limit threshold and a lower threshold of the amount of photovoltaic power generation for a predetermined weather situation, and when new data of the amount of photovoltaic power generation for a predetermined weather situation is input, the new data It is determined whether it is between the upper threshold and the lower threshold. If the new data is outside these ranges, the device excludes the new data as noise data. This apparatus is thereby improving the prediction accuracy.

JP 2008-77561 A

  The device described in Patent Document 1 cannot sufficiently remove noise data included in data representing past input / output relationships. The data to be learned in this way is generally called teacher data. That is, the apparatus described in Patent Document 1 cannot sufficiently remove noise data included in teacher data. Since the teacher data includes data that does not have a true correlation, this apparatus cannot create a highly accurate prediction function, and the prediction accuracy is not sufficiently improved.

  The apparatus described in Patent Document 1 is a section that includes an upper threshold value and a lower threshold value for the amount of photovoltaic power generation for each weather, such as sunny, cloudy, and rainy, to determine whether data is noise. Is set. However, the value is determined by the operator or user of the apparatus based on experience or the like, and may not necessarily be appropriate.

  It is an object of the present invention to provide a noise data removal support apparatus, method, and program that solve the above-described problems related to noise data removal.

  A noise data removal support apparatus according to an embodiment of the present invention includes a data storage unit that stores a history record including a value of an input variable and an actual value of an output variable at a certain time point, and a noise determination threshold value. A history record in which the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is less than or equal to the noise determination threshold is extracted and changed. Based on the history record, a prediction function for predicting the value of the output variable and a prediction error thereof are calculated from the value of the input variable, and a noise determination threshold value with a smaller prediction error than the others is selected and output. Prior prediction means.

  The method according to an embodiment of the present invention stores a history record including a value of an input variable and an actual value of an output variable at a certain time point for a plurality of time points, while changing a noise determination threshold value within a predetermined range. A history record in which the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is equal to or less than the noise determination threshold, and the input is performed based on the extracted history record. A prediction function for predicting the value of the output variable and its prediction error are calculated from the value of the variable, and a noise determination threshold value with a smaller prediction error than the others is selected and output.

  A program according to an embodiment of the present invention includes a data storage process for storing a history record including a value of an input variable and an actual value of an output variable at a certain time, and a noise determination threshold within a predetermined range. The difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is extracted from the noise determination threshold value or less based on the extracted history record. A prediction function for predicting the value of the output variable from the value of the input variable and a prediction error thereof, and a pre-prediction process for selecting and outputting a noise determination threshold value with which the calculated prediction error is smaller than others , Execute on the computer.

  The noise removal support apparatus according to the present invention can sufficiently remove the noise data included in the teacher data, that is, the teacher data can be brought close to data having only true correlation. Therefore, an apparatus that uses the data as a learning target can create a highly accurate prediction model, for example, a prediction function, and perform highly accurate prediction.

FIG. 1 is a configuration diagram of a noise data removal support apparatus 100 according to the first embodiment. FIG. 2 is a configuration diagram of the computer apparatus 200. FIG. 3 is a flowchart showing the overall operation of the noise data removal support apparatus 100. FIG. 4 is a flowchart showing the operation of the prior prediction unit 12. FIG. 5 is an operation flowchart of processing additionally executed by the prior prediction unit 12 according to the second embodiment. FIG. 6 is a diagram illustrating the dependence of the solar radiation amount prediction accuracy in Tokyo on the noise determination threshold in the case of 60 learning days. FIG. 7 is a diagram illustrating the dependence of the solar radiation amount prediction accuracy in Tokyo on the noise determination threshold in the case of 365 learning days. FIG. 8 shows the annual noise data generation rate of the teacher data calculated for each noise determination threshold ε. FIG. 9 is a configuration diagram of the noise data removal support apparatus 100 according to the third embodiment.

<First Embodiment>
<Configuration>
The noise data removal support apparatus 100 according to the present embodiment determines a noise determination threshold value that is a parameter for removing noise data from teacher data. The noise data removal support apparatus 100 may further generate a prediction function of the output variable from the teacher data from which the noise data has been removed based on the determined noise determination threshold value, and predict a future output variable value.

  FIG. 1 is a configuration diagram of a noise data removal support apparatus 100 according to the present embodiment. The noise data removal support device 100 is connected to the input device 40 and the output device 50 via a communication network or the like. The noise data removal support apparatus 100 includes a data acquisition unit 11, a prior prediction unit 12, a prediction unit 13, a data output unit 14, a main storage unit 20, and a data storage unit 31. The data storage unit 31 exists outside the noise data removal support apparatus 100 and may be connected via a communication network or the like.

  The input device 40 is a keyboard, a mouse, or the like. The output device 50 is a display, a printer, or the like.

  The data acquisition unit 11, the pre-prediction unit 12, the prediction unit 13, and the data output unit 14 include semiconductor logic elements, semiconductor storage elements, counters, and other circuit elements, and execute a pre-built algorithm. It is a logic circuit. The main storage unit 20 is a readable / writable semiconductor storage device or the like, and temporarily stores data. The data storage unit 31 is a storage device such as a hard disk drive or a solid state drive, and can always store programs and data.

  The data storage unit 31 stores teacher data that may include noise data in advance. The teacher data is a set of history records including the date and time, the actual value of the output variable at that time, and the predicted value of the input variable at that time or the actual value at that time. For example, the noise data removal support apparatus 100 statistically analyzes the teacher data to generate a prediction function, inputs an input variable value to the prediction function, and predicts an output variable value. The value of the input variable is input from the input device 40, for example.

  The reason why the prediction function can be generated is that there is a certain correlation between the input variable value and the output variable value. Many teacher data show this correlation, but some do not. Teacher data that does not show this correlation is noise data. In order to create a highly accurate prediction function from the teacher data, it is preferable that the teacher data does not include noise data.

  The teacher data is data for predicting, for example, the amount of renewable energy from various geography and weather data. Renewable energy is electric power generated by sunlight or wind power. In this case, the teacher data is history data including a renewable energy amount or the like as an output variable actual value and various geographical / weather data as input variable values. Teacher data includes, for example, latitude, longitude, date, time, power generation, solar radiation, temperature, humidity, wind direction, wind speed, precipitation, weather, upper cloud cover, middle cloud cover, lower cloud cover, total cloud cover, ground pressure, and This is historical data including sea level pressure. Here, the power generation amount is, for example, the amount of power generated by solar power generation or wind power generation.

  The teacher data stored in the data storage unit 31 calculates the difference between the actual value of the output variable value at the date and time when the history was acquired and the predicted value of the output variable value predicted before the date and time. Difference calculation data may be included. The difference calculation data may be the difference itself, a prediction value, or a prediction function description and an input variable value used for prediction. For example, when a history collected by a device that has predicted output variables in the past is used as teacher data, the device can include difference calculation data in the teacher data.

  The data storage unit 31 may further store operation parameters of the noise data removal support apparatus 100 in advance. The operation parameters are, for example, a rated value, an initial value / upper limit value / lower limit value / step size Δε of the noise determination threshold value ε, and the number of learning days.

  Here, the rated value is a normally expected output value such as the amount of solar radiation and the amount of power generation. The noise determination threshold value ε is a value for determining whether or not the weather data is noise data having no true correlation. If the prediction error exceeds this threshold, the meteorological data at that time is determined as noise data. The initial value of the noise determination threshold is a value at the start of processing for searching for an optimum threshold within a range that the threshold can take. The upper limit value / lower limit value of the noise determination threshold is an upper limit / lower limit value of a range that the threshold can take. The step width Δε of the noise determination threshold is a width between thresholds to be verified when searching for an optimum threshold within a range that the threshold can take. The learning days is a period in which teacher data used for learning is acquired.

  The data acquisition unit 11 performs initial settings such as reading information necessary for operation from the data storage unit 31 into the main storage unit 20. The prior prediction unit 12 determines a noise determination threshold value that is a parameter for removing noise data from the teacher data. The prediction unit 13 predicts an output variable value at a designated future time point based on the teacher data from which the noise data has been removed based on the determined noise determination threshold. The data output unit 14 outputs the predicted value of the output variable value to the output device 50.

  The noise data removal support apparatus 100 does not necessarily include the prediction unit 13, and may determine a noise determination threshold value and output it to another existing apparatus. The other device may remove the noise data from the teacher data based on the determined noise determination threshold, generate a prediction function, and predict the output variable value at the designated future time point.

  Furthermore, the noise data removal support apparatus 100 does not necessarily have to include the data acquisition unit 11 and the data output unit 14. The prior prediction unit 12 or the prediction unit 13 may include these functions.

  The noise data removal support apparatus 100 may be realized by the computer apparatus 200. FIG. 2 is a configuration diagram of the computer apparatus 200. The computer device 200 is connected to the input device 40 and the output device 50 via a communication network or the like. The computer device 200 includes a processor 10, a main storage unit 20, and an external storage device 30. The external storage device 30 exists outside the computer device 200 and may be connected via a communication network or the like. Inside the computer device 200, the processor 10, the main storage unit 20, the external storage device 30, the input device 40, and the output device 50 are connected to each other by a bus 60.

  The processor 10 functions as the data acquisition unit 11, the prior prediction unit 12, the prediction unit 13, and the data output unit 14 by executing the program 21 stored in the main storage unit 20. The external storage device 30 is a storage device such as a hard disk drive or a solid state drive, and is used as the data storage unit 31. The program 21 is initially stored in, for example, the external storage device 30 and loaded into the main storage unit 20 when the computer device 200 is initially set.

<Operation>
FIG. 3 is a flowchart showing the overall operation of the noise data removal support apparatus 100.

  When a start command is input from the input device 40, the data acquisition unit 11 temporarily stores the teacher data stored in the data storage unit 31, the initial value of the noise determination threshold, and the like in the main storage unit 20 ( S21). The start instruction includes one or more prediction target dates and times. The prediction target date and time is the current or future date and time.

  Next, the prior prediction unit 12 determines whether the noise determination threshold has already been determined (S22). The prior prediction unit 12 determines that the noise determination threshold has not been determined if the noise determination threshold is an initial value (No in S22), and has been determined if the noise determination threshold is a value other than the initial input value. Determine (Yes in S22).

  When the noise determination threshold has not been determined (No in S22), the prior prediction unit 12 determines the noise determination threshold by performing prior prediction (S23). This process will be described later.

  After determining the noise determination threshold, the prediction unit 13 deletes the noise data from the teacher data using the noise determination threshold determined by the prior prediction unit 12, and generates a prediction function based on the teacher data from which the noise data has been deleted. The output variable value of the prediction target date and time is predicted (S24). The same applies when the prior prediction unit 12 determines that the noise determination threshold has been determined (Yes in S22). The process of S24 will be described later.

  The prediction unit 13 repeatedly executes prediction of all prediction target dates and times designated by the start command (No in S25, S24). When the prediction unit 13 finishes predicting all the prediction target dates and times (Yes in S25), the data output unit 14 graphs the predicted value and outputs it to the output device 50 (S26), and the process ends.

<Details of Advance Prediction Process S23>
FIG. 4 is a flowchart showing the operation of the prior prediction unit 12. This process is executed in S23 of FIG.

  The prior prediction unit 12 first sets an upper limit value for the noise determination threshold ε (S30). Next, the prior prediction unit 12 sequentially decreases this value by Δε (S3A) until the minimum value of the prediction error is detected (Yes in S39), or the noise determination threshold ε is equal to or lower than the lower limit value (S31). Steps S31 to S3A are repeated until No).

  In the prediction process, the pre-prediction unit 12 takes out all the history records of the teacher data one by one (S35, S36), predicts the predicted value of the output variable and the actual value of the output variable predicted from the predicted value or actual value of the input variable. The difference from the value is compared with the current noise determination threshold value ε (S32). The prior prediction unit 12 records that the history record having the difference smaller than the current noise determination threshold ε is left in the teacher data (S33), and uses the history record having the difference equal to or larger than the current noise determination threshold ε as the noise data. Is recorded to be deleted (S34). The prior prediction unit 12 creates and stores a temporary extraction history table in the main storage unit 20 for example.

  Here, the prior prediction unit 12 may obtain the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable from the difference calculation data included in the history record. When the difference calculation data is, for example, a predicted value of the output variable value predicted before the date and time when the history is acquired, the prior prediction unit 12 calculates the difference between the actually measured value of the output variable of the history record and the predicted value. do it. If the difference calculation data is, for example, the difference between the predicted value of the output variable value predicted before the date and time when the history is acquired and the actual value of the output variable at the time of history acquisition, the prior prediction unit 12 calculates the difference. What is necessary is just to use in S32 as it is.

  Next, the prior prediction unit 12 creates a prediction function based on the teacher data from which the noise data determined based on the current noise determination threshold ε is deleted (S37). The prior prediction unit 12 determines the teacher data from which the noise data determined based on the current noise determination threshold ε is deleted with reference to the extraction history table of the main storage unit 20.

The prediction function is, for example, a function having the format shown in Expression (1).

Here t _i represents the i-th day from the first day included in the history of (i is a natural number from 1 to p). t _{i + m} represents the i + mth day (m is an integer). p indicates the number of days of teacher data-m. Y (t _{i + m} ) represents the predicted value of the output variable at the prediction target time on the i + mth day. Xn (t _i ) indicates the actual value of the n-th (n is an integer) input variable at the same time on the i-th day or the predicted value of the n-th input variable at the same time on the i-th day.

  That is, the function F of the formula (1) is a function for predicting the output variable value of the prediction target date of m days after the prediction target date, for example, several hours to several days later, using the value of the input variable in the teacher data. It is. For example, when m is 0, Expression (1) is a function that predicts the value of the output variable on the same day from the predicted value issued for the i-th day of the input variable. For example, when m is 1, Expression (1) is a function that predicts the value of the output variable of the next day from the actual value for the i-th day of the input variable. Here, m is stored in the data storage unit 31 as an operation parameter of the noise data removal support apparatus 100, for example.

  For example, the prior prediction unit 12 scans the teacher data from which the noise data has been deleted, and outputs the actual output variable value and the input variable value m days ago for the prediction target time of each day when the history is acquired. The function F is created by matching and statistical learning. Therefore, the prior prediction unit 12 uses multiple regression, a neural network, a support vector machine, or the like as a machine learning engine. In addition, the prior prediction unit 12 captures the input variable in the form of increasing the dimension of the input variable, for example, as a method for capturing the input variable into the prediction function F.

Here, the input variable _Xn is a weather parameter, for example, temperature, humidity, wind direction, wind speed, temperature, humidity, wind direction, wind speed, precipitation, weather, upper cloud cover, middle cloud cover, lower cloud cover, total cloud cover, ground surface Atmospheric pressure or sea level pressure.

  The output variable Y is, for example, the amount of power generation, the amount of solar radiation, or the wind speed. When the output variable is the solar radiation amount, the solar power generation amount is calculated, for example, by multiplying the solar radiation amount predicted value by a conversion coefficient. When the output variable is wind speed, the amount of wind power generation is obtained, for example, by converting the predicted wind speed value using a wind power generation conversion formula.

  In the example of Expression (1), the prior prediction unit 12 uses a value m days before the uniform prediction target date as the input variable value. The prior prediction unit 12 may be configured to use variable values for different days depending on input variables, or variable values for multiple days.

  After creating the prediction function (S37), the pre-prediction unit 12 substitutes the value of the input variable for the prediction function to obtain a prediction value, and further calculates a prediction error from the actual value and the prediction value of the output variable (S38). ). The prior prediction unit 12 calculates the predicted value of the output variable from the (m + 1) th day to the last day of the teacher data using, for example, the input variable values from the first day to the pth day of the teacher data. To do.

For example, the prior prediction unit 12 calculates a prediction error by MAPE (Mean Absolute Percentage Error) represented by Expression (2).

Here, y _{i + m} is Y (t _{i + m} ) of Equation (1), that is, the predicted value of the output variable on the i + m day. r _{i + m} is an actually measured value of the output variable on the i + m day.

  Next, the prior prediction unit 12 checks whether or not there is a minimum value of the prediction error curve (S39). When there is a minimum value in the prediction error curve with the noise determination threshold as the x component and the prediction error as the y component (Yes in S39), the prior prediction unit 12 outputs the noise determination threshold ε that gives the minimum value ( S3B), the operation ends. Here, the minimum value is a starting point at which the slope of the curve changes from positive to negative.

  When there is no minimum value in the prediction error curve (No in S39), the prior prediction unit 12 sequentially reduces the current noise determination threshold ε by Δε, initializes the extraction history table in the main storage unit 20, and performs the prediction process from S31. repeat. When the noise determination threshold ε is equal to or lower than the lower limit value without finding the minimum value (No in S39, No in S31), the prior prediction unit 12 outputs the noise determination threshold ε that gives the minimum value (S3C). End the operation.

  The noise determination threshold output by the prior prediction unit 12 may change when the number of days when the teacher data is acquired changes. In addition, the optimum value of the noise determination threshold value may change with the global weather fluctuation. For this reason, it is desirable that the administrator of the noise data removal support apparatus 100 updates the noise determination threshold at an appropriate period and uses the latest one as much as possible.

  When the history record does not include the difference calculation data, the prior prediction unit 12 first (before S30) learns teacher data including noise data to create a prediction function, and uses the prediction function to perform prediction. An error may be calculated and added to each history record. This process is similar to S37 and S38. The difference is a difference between whether the learning target is the teacher data after the noise data is deleted or the teacher data including the noise data.

  The prior prediction unit 12 may first set a lower limit for the noise determination threshold ε (change S30). In this case, the prior prediction unit 12 detects the minimum value of the prediction error while sequentially increasing this value by Δε (changes S3A) (Yes in S39), or until the noise determination threshold value ε reaches the upper limit value or more. (Change S31), the prediction process of S31 to S3A may be repeated. In this case, the minimum value is the starting point at which the slope of the curve changes from negative to positive.

  The prior prediction unit 12 may detect only the minimum value without detecting the minimum value.

<Details of Prediction Process S24>
First, the prediction unit 13 deletes the noise data from the teacher data using the noise determination threshold determined by the prior prediction unit 12, and generates a prediction function for the prediction target date and time based on the teacher data from which the noise data has been deleted. . This process is the same as the process from S32 to S37 in FIG. Here, it is assumed that the created prediction function is a function of Expression (1), for example.

Next, the prediction unit 13 inputs the input variable value m days before the prediction target date to the prediction function of Expression (1), calculates the output variable prediction value Y (t _{i + m} ) of the prediction target day, and outputs it. . For example, when m = 0 and the prediction target date is tomorrow, the prediction unit 13 obtains a prediction value for tomorrow of the input variable (weather data) from the input device 40 and substitutes it in the prediction function tomorrow. Predict output variable values. For example, when m = 3 and the prediction target day is tomorrow, the prediction unit 13 obtains the actual measurement value of the input variable from the history data two days ago and substitutes it in the prediction function to obtain the output variable value of tomorrow. Predict.

  When the prediction function created by the prior prediction unit 12 using the noise determination threshold is left in the main storage unit 20, the prediction unit 13 may divert the prediction function and omit the generation of the prediction function. good.

<Effect>
The noise removal support apparatus 100 according to the present embodiment can sufficiently remove the noise data included in the teacher data, that is, can bring the teacher data closer to data having only true correlation. This is because the prior prediction unit 12 selects and outputs a noise determination threshold value that gives a minimum value of the prediction error. Therefore, the device or the prediction unit 13 that uses the data as a learning target can create a highly accurate prediction function and perform highly accurate prediction.

  The reason for selecting the noise judgment threshold value that gives the minimum value of the prediction error is that it is a critical point where the accuracy of the teacher data is most improved and the accuracy deteriorates if the data is reduced too much. Further, if the minimum point is detected, the prior prediction unit 12 can cancel the search for the noise determination threshold value, and thus has an advantage of high processing efficiency.

  When the minimum value of the prediction error cannot be detected, the prior prediction unit 12 selects and outputs a noise determination threshold value that gives the minimum value. The reason for selecting the noise determination threshold value that gives the minimum value is that the accuracy of the teacher data is most improved.

<Data example>
FIG. 6 is a diagram illustrating the dependence of the solar radiation amount prediction accuracy in Tokyo on the noise determination threshold in the case of 60 learning days. FIG. 7 is a diagram illustrating the dependence of the solar radiation amount prediction accuracy in Tokyo on the noise determination threshold in the case of 365 learning days.

  In this example, the initial value / upper limit value of the noise determination threshold is 600 Wh, the step size of the noise determination threshold is 100 Wh, and the lower limit value of the noise determination threshold is 200 Wh. The above values are 60%, 10%, and 20% of the rated output of 1000Wh of solar radiation, respectively.

  In this example, in the Otemachi, Chiyoda-ku, Tokyo, the estimation of the integrated amount of solar radiation for each hour from 5:00 to 19:00 on the next day (hereinafter abbreviated as 1 hour of solar radiation) is performed at 18:00 on the previous day. The output variable is a 1 hour solar radiation value. The input variables are the upper cloud cover, the upper cloud cover, the middle cloud cover, the lower cloud cover, the temperature, the humidity, and the predicted amount of solar radiation from the target time, and the upper layer from 2 hours to 12 hours before the target time. These are forecast values for cloud cover, middle cloud cover, lower cloud cover, temperature, and humidity.

  The support vector machine machine-learned data for the past 60 days or the past 365 days regarding the correlation between the actual value of the output variable and the value of the input variable, and created a prediction formula for each target time.

  FIG. 6 and FIG. 7 show that there is an optimum value that can improve the accuracy most depending on the number of learning days when prediction is performed by changing the noise determination threshold. FIG. 6 shows that when the number of learning days is 60 days, the threshold with the smallest prediction error and the highest accuracy improvement is 400 Wh (40% of the rated output) shown in (b). FIG. 7 shows that when the number of learning days is 365 days, the threshold with the smallest prediction error and the highest accuracy improvement is 300 Wh (30% of the rated output) shown in (b). The accuracy at the threshold with the highest accuracy improvement is improved by about 0.04% regardless of the number of learning days when the present invention is not applied, that is, compared with the case of 600 Wh shown in FIGS.

  At this time, when the learning days are 60 days, the prediction unit 13 of the noise removal assisting apparatus 100 predicts the target date by setting the noise determination threshold ε that can be expected to improve the error to 400 Wh. Further, when the number of learning days is 365 days, the prediction unit 13 predicts the target date by setting the noise determination threshold ε that can be expected to improve the error to 300 Wh.

  In FIG. 6, the search range of the noise determination threshold is 300 Wh to 600 Wh. Therefore, the prior prediction unit 12 detects the optimum value 400Wh as a minimum value. In FIG. 7, the search range of the noise determination threshold is 200 Wh to 600 Wh. Therefore, the prior prediction unit 12 detects the optimum value 300Wh as a minimum value.

  However, if the lower limit value of the search range is 400 Wh in the case of FIG. 6, the prior prediction unit 12 detects the optimum value 400 Wh as the minimum value. If the lower limit value of the search range is 400 Wh in the example of FIG. 7, the prior prediction unit 12 still detects the optimum value 400 Wh as the minimum value. In this case, the prediction unit 13 of the noise removal assisting apparatus 100 predicts the target date with the noise determination threshold ε that is most expected to improve the error as 400 Wh regardless of whether the number of learning days is 60 days or 365 days.

<Second Embodiment>
The noise removal assisting apparatus 100 according to the present embodiment sets the search range of the noise determination threshold ε to be shorter than the range given by the upper limit value and the lower limit value of the activation parameter. Specifically, the noise removal assisting apparatus 100 detects the lower limit value of the noise determination threshold value ε at which noise data is not detected, and omits the search for a value range larger than the lower limit value. This is because there is no noise data for a range larger than the lower limit value, and the prediction function and the difference between the predicted value and the measured value do not change.

  Thereby, the noise removal assistance apparatus 100 of this Embodiment can shorten a prior prediction process time.

  The prior prediction unit 12 of the noise removal assisting apparatus 100 of the present embodiment detects the lower limit value of the noise determination threshold ε at which noise data is no longer detected, in addition to the processing of the prior prediction unit 12 of the first embodiment, The lower limit value is set as the upper limit value of the search range of the noise determination threshold value ε.

  FIG. 5 is an operation flowchart of processing additionally executed by the prior prediction unit 12 of the present embodiment.

  The prior prediction unit 12 first sets a lower limit value for the noise determination threshold ε (S40). Next, the prior prediction unit 12 sequentially increases this value by Δε (S48), and the frequency of detecting noise data becomes zero (Yes in S47), or the noise determination threshold ε is the upper limit given by the activation parameter. The following processing is repeated until the value becomes equal to or greater than the value (No in S41).

  The prior prediction unit 12 takes out all the history records of the teacher data one by one (S45, S46), the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable, and the current The noise determination threshold value ε is compared (S43). When the prior prediction unit 12 detects a history record having the difference equal to or greater than the current noise determination threshold ε (Yes in S43), the prior prediction unit 12 counts the number (S44).

  The prior prediction unit 12 may obtain the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable from the difference calculation data included in the history record. When the history record does not include the difference calculation data, the advance prediction unit 12 first learns the input teacher data (before S40) to create a prediction function, and uses the prediction function to calculate the difference. It may be calculated and added to each history record.

  When the noise determination threshold value ε at which the frequency of detecting noise data is zero is detected (Yes in S47), the prior prediction unit 12 replaces the upper limit value of the search range given as the activation parameter with the value (S48). ), And the addition process ends. If the noise determination threshold value ε is not detected (No in S47) without detecting the noise data detection frequency (No in S47), and the noise determination threshold value ε is equal to or greater than the upper limit given by the activation parameter (No in S41), the prior prediction unit 12 Finishes the addition process.

  When the addition process ends, the prior prediction unit 12 subsequently executes the process of the flowchart of FIG.

<Data example>
FIG. 8 shows the annual noise data generation rate of the teacher data calculated for each noise determination threshold ε. In this example, the initial value / lower limit value of the noise determination threshold is 0 Wh, the step size of the noise determination threshold is 100 Wh, and the upper limit value of the noise determination threshold is 1000 Wh. The above values are 0%, 10%, and 100% of the rated output of 1000Wh of solar radiation, respectively. These values are given to the noise data removal support apparatus 100 as activation parameters.

  FIG. 8 shows that no noise occurs when the noise determination threshold is 600 Wh (60% of the rated output) or more in both cases of learning days 60 and 365. Thus, the prior prediction unit 12 updates the upper limit value of the threshold search range as 60% of the rating. As a result of the update, the searched noise determination threshold ε is as shown in FIG. 6 and FIG.

<Third Embodiment>
FIG. 9 is a configuration diagram of the noise data removal support apparatus 100 according to the present embodiment. The noise data removal support apparatus 100 includes a prior prediction unit 12 and a data storage unit 31.

  The data storage unit 31 stores a history record including a value of an input variable and an actual value of an output variable at a certain time for a plurality of times.

  The prior prediction unit 12 has a history record in which a difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is not more than the noise determination threshold while changing the noise determination threshold within a predetermined range. To extract. Furthermore, the pre-prediction unit 12 calculates a prediction function for predicting an output variable value from an input variable value and its prediction error based on the extracted history record, and a noise determination threshold value in which the calculated prediction error is smaller than others. Select to output.

  The noise data removal support apparatus 100 according to the present embodiment can sufficiently remove noise data included in the teacher data, that is, allows the teacher data to approach only data having true correlation. This is because the prior prediction unit 12 selects and outputs a noise determination threshold value that gives a prediction error smaller than other values. Therefore, the device or the prediction unit 13 that uses the data as a learning target can create a highly accurate prediction function and perform highly accurate prediction.

  While the present invention has been described with reference to the embodiments, the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

DESCRIPTION OF SYMBOLS 10 Processor 11 Data acquisition part 12 Prior prediction part 13 Prediction part 14 Output part 20 Main memory part 21 Program 30 External storage device 31 Data storage part 40 Input device 50 Output device 60 Bus 100 Noise data removal assistance apparatus 200 Computer apparatus

Claims (10)

Data storage means for storing a history record including the value of the input variable and the actual value of the output variable at a certain time point for a plurality of time points;
While changing the noise judgment threshold within a predetermined range, a history record in which the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is equal to or less than the noise judgment threshold is extracted. Based on the extracted history record, a prediction function for predicting the value of the output variable from the value of the input variable and its prediction error are calculated, and a noise determination threshold value with a smaller prediction error than the others is selected. A noise data removal support device comprising: a prior prediction unit that outputs the data in advance.   Calculate and output a predicted value of an output variable from a prediction function calculated based on a history record whose difference is equal to or less than the noise determination threshold output by the prior prediction means and a given input variable value The noise data removal support apparatus according to claim 1, further comprising prediction means for performing the operation.   The noise data removal support apparatus according to any one of claims 1 to 2, wherein the prior prediction unit selects and outputs a noise determination threshold value that gives a minimum value of a prediction error with respect to a change in the noise determination threshold value.   The noise data removal support apparatus according to any one of claims 1 to 2, wherein the prior prediction unit selects and outputs a noise determination threshold value that provides a minimum value of a prediction error with respect to a change in the noise determination threshold value.   The pre-prediction means calculates the prediction function and its prediction error for a range larger than a value when the difference of all history records is equal to or less than the noise determination threshold even within the predetermined range. The noise data removal support apparatus according to any one of claims 1 to 4, wherein the processing to be performed is omitted. A history record including the value of the input variable and the actual value of the output variable at a certain time point is stored for a plurality of time points.
While changing the noise judgment threshold within a predetermined range, a history record in which the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is equal to or less than the noise judgment threshold is extracted. Based on the extracted history record, a prediction function for predicting the value of the output variable from the value of the input variable and its prediction error are calculated, and a noise determination threshold value with a smaller prediction error than the others is selected. Output method.   Further, the predicted value of the output variable is calculated from the prediction function calculated based on the history record whose difference is equal to or less than the noise determination threshold output by the prior prediction means, and the value of the given input variable. The method of claim 6, wherein   The method according to any one of claims 6 to 7, wherein a noise determination threshold value that gives a minimum value of a prediction error with respect to a change in the noise determination threshold value is selected and output. A data storage process for storing a history record including a value of an input variable and an actual value of an output variable at a certain time point for a plurality of time points;
While changing the noise judgment threshold within a predetermined range, a history record in which the difference between the predicted value of the output variable predicted from the value of the input variable and the actual value of the output variable is equal to or less than the noise judgment threshold is extracted. Based on the extracted history record, a prediction function for predicting the value of the output variable from the value of the input variable and its prediction error are calculated, and a noise determination threshold value with a smaller prediction error than the others is selected. A program that causes a computer to execute a prior prediction process to be output in response.   Calculate and output a predicted value of an output variable from a prediction function calculated based on a history record whose difference is equal to or less than the noise determination threshold output by the prior prediction means and a given input variable value The program according to claim 9, further causing the computer to execute prediction processing to be performed.

Images