JP5729282B2 - Prediction device, prediction program, and prediction method - Google Patents

Description

  The present invention relates to a prediction device, a prediction program, and a prediction method.

  In power consumption prediction, predict the trend of power consumption from a certain point in the forecasted day, such as when the power consumption reaches its peak and how much power consumption will be at the peak of power consumption. There is technology to do. In the following description, a prediction target day is referred to as a prediction target day.

  As an example of the above technique, there is a technique for predicting power consumption from past performance values of power consumption using the k-neighbor method. In the technology for predicting power consumption, the actual value pattern of the past day of the prediction target date is compared to the pattern of the actual value in the first time zone from a certain time point of the prediction target date to the start point of the prediction. From the patterns, a pattern having similar performance values in the first time zone is selected. And in said technique, the power consumption after the time of starting prediction is derived from the selected pattern. For example, in the technique described above, a plurality of patterns having similar performance values in the first time zone are selected, and the average value of the performance values after the time point when the prediction is started is calculated for the selected patterns. The power consumption after the starting point is derived. In this way, in the conventional technique, power consumption after the point in time when the prediction of the prediction target day is started is predicted.

  In addition, with regard to past time-series data, similar time-series data is clustered in the time zone from the first time point to the second time point for a plurality of time-series data for each day, and prediction is performed using the clustering result. There is a technique for predicting data after the second time point on the target date. In this technology, one cluster with the most similar time series data in the above time zone of the forecast date is selected, and a predetermined number with high similarity among the past time series data included in the selected cluster. The synthesized time series data is output as predicted time series data.

International Publication No. 97/25676

"Data science learned with R. 5. Pattern recognition", 2009, Kyoritsu Shuppan, Chapter 8, k-neighbor method, pages 86-99.

  However, the above technique has a problem that the accuracy of the prediction result may be lowered.

  An example will be described. 18 and 19 are diagrams for explaining an example of a conventional technique for predicting power consumption. The example of FIG. 18 shows a pattern 90 of actual values of power consumption in the time zone from 6:00 to 9:00 on October 1, 2011, which is the prediction target date. In the example of FIG. 18, a plurality of patterns other than the pattern 90 are in the time zone from 6 o'clock to 20 o'clock for 30 days from September 1, 2011 to September 30, 2011 before the prediction target date. It is a pattern of the actual value of power consumption.

  In the example of FIG. 18, a plurality of 30 patterns other than the pattern 90 can be classified into four groups. For example, as shown in the example of FIG. 18, it is possible to classify 6 patterns having a small variation with the peak of power consumption around 11 o'clock. In addition, 10 patterns having a large variation at the peak of power consumption at around 14:00 can be classified into group B. Further, eight patterns with large variations in power consumption peak from around 12:00 to around 13:00 and around 17:00 can be classified into group C. In addition, the peak is reached from 12:00 to 13:00, the power consumption is lower than the patterns classified into other groups, the variation is larger than the patterns classified into group A, and the patterns classified into groups B and C. Six small patterns can be classified into group D.

  In the example of FIG. 18, from 10:00 on October 1, 2011 to 20:00 using the pattern of the actual value in the time zone from 6:00 to 9:00 on October 1, 2011, which is the prediction target date. A case will be described in which power consumption in a time zone is predicted using a conventional technique for predicting power consumption. In the example of FIG. 18, in the conventional technology for predicting power consumption, first, for the time zone from 6 o'clock to 9 o'clock, there are a plurality of patterns 90 of actual values of prediction target days and power consumption of past actual values. Each of the patterns is compared to calculate the similarity. In the conventional technology for predicting power consumption, a power consumption pattern of predetermined past values, for example, six past performance values is selected from the one with the highest similarity. Here, a case where six patterns are selected as shown in the example of FIG. 19 will be described. In the example of FIG. 19, three patterns of September 5, September 14, and September 22 classified into the group B, September 6, September 11, and September classified into the group C The case where three patterns of 29 days are selected is shown. In the example of FIG. 19, with the conventional technology for predicting power consumption, an average value of actual values from 10 o'clock to 20 o'clock for performing power consumption prediction is calculated for the selected six patterns, and the calculated result is Output as prediction time series.

  However, in the conventional technology for predicting power consumption, as shown in the example of FIG. 19, there is a case where the prediction time series is not similar to any of the past patterns. Therefore, in the conventional technology for predicting power consumption, the prediction result may be significantly different from the actual result. Therefore, the conventional technology for predicting power consumption has a problem that the accuracy of the prediction result may be lowered.

  As shown in the example of FIG. 18, the pattern 90 of the prediction target day is most similar to the average value of the pattern of the group A among the average values of the patterns for the groups A to D. However, as shown in the example of FIG. 19, since the variation is large in the groups B and C, six patterns having high similarity to the pattern 90 are selected from the patterns of the groups B and C. There is a case. For this reason, the conventional technique for predicting power consumption may not select the group A pattern that should be selected. Therefore, the conventional technology for predicting power consumption has a problem that the accuracy of the prediction result may be lowered.

  Such a problem does not occur only when power consumption is predicted, but also occurs when data is predicted based on past performance values in the same manner. For example, in a production system in a factory that produces a product, the demand amount of the product is predicted from the time-series demand amount (time-series data) of the past product, and the production amount of the product is controlled according to the prediction result. The above problem also occurs in some cases.

  The disclosed technology has been made in view of the above, and an object thereof is to provide a prediction device, a prediction program, and a prediction method capable of suppressing a decrease in accuracy of a prediction result.

  The prediction device disclosed in the present application includes a classification unit, a first calculation unit, a selection unit, and a synthesis unit. The classification unit includes data having similar performance values in the first time zone and the second time zone among the plurality of first data including the actual value in the first time zone and the actual value in the second time zone. Classify them into the same group. The first calculation unit calculates a degree of similarity between the value corresponding to the actual value in the first time zone of the group and the actual value in the first time zone included in the second data. The second calculation unit selects a plurality of groups based on the degree calculated by the first calculation unit. The synthesizer synthesizes the first data portion in the second time zone included in the group for each group selected by the selector.

  According to one aspect of the prediction device disclosed in the present application, it is possible to suppress a decrease in accuracy of a prediction result.

FIG. 1 is a diagram illustrating an example of a functional configuration of the prediction apparatus according to the embodiment. FIG. 2 is a diagram illustrating an example of a pattern. FIG. 3 is a diagram illustrating an example of time-series data from September 1, 2011 to September 30, 2011. FIG. 4 is a diagram illustrating an example of time-series data from 6 o'clock to 20 o'clock on each day from September 1, 2011 to September 30, 2011. FIG. 5 is a diagram illustrating an example of prediction target data. FIG. 6 is a diagram illustrating an example of a method of calculating the dissimilarity by the classification unit. FIG. 7 is a diagram illustrating another example of a method of calculating the dissimilarity by the classification unit. FIG. 8 is a diagram illustrating an example of the result of clustering. FIG. 9 is a diagram for explaining an example of processing executed by the first calculation unit. FIG. 10 is a diagram for explaining an example of processing executed by the selection unit. FIG. 11 is a diagram for explaining an example of processing executed by the synthesis unit. FIG. 12A is a diagram for describing an example of processing executed by the synthesis unit. FIG. 12B is a diagram for describing an example of processing executed by the synthesis unit. FIG. 12C is a diagram for describing an example of processing executed by the synthesis unit. FIG. 13 is a diagram for explaining an example of processing executed by the second calculation unit. FIG. 14 is a diagram for explaining an example of processing executed by the second calculation unit. FIG. 15 is a diagram for explaining an example of processing executed by the prediction apparatus according to the embodiment. FIG. 16 is a flowchart illustrating the procedure of the prediction process according to the embodiment. FIG. 17 is a diagram illustrating a computer that executes a prediction program. FIG. 18 is a diagram for explaining an example of a conventional technique for predicting power consumption. FIG. 19 is a diagram for explaining an example of a conventional technique for predicting power consumption.

  Hereinafter, embodiments of a prediction apparatus, a prediction program, and a prediction method disclosed in the present application will be described in detail with reference to the drawings. The embodiments do not limit the disclosed technology.

  A prediction apparatus according to an embodiment will be described. FIG. 1 is a diagram illustrating an example of a functional configuration of the prediction apparatus according to the embodiment.

  In the present embodiment, a case will be described in which the prediction device 10 performs prediction of power consumption by calculating a predicted value of power consumption at each time from 10:00 to 20:00 on October 1, 2011. The prediction apparatus 10 includes time-series data including actual values of power consumption from 0:00 to 23:00 on each day from September 1, 2011 to September 30, 2011, and October 1, 2011 on the prediction target date. Prediction is performed using time-series data including actual values of power consumption from 0:00 to 9:00 of the day. Here, in the following description, time-series data including the actual power consumption value may be referred to as a pattern.

  The forecasting device 10 is the actual value of the power consumption in the first time zone from 6:00 to 9:00 on each day from September 1 to September 30, and the second from 10:00 to 20:00. A plurality of patterns including actual values of power consumption in the time zone are classified as follows. That is, the prediction device 10 classifies patterns having similar performance values in the first time zone and the second time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the 1st time slot | zone of the pattern of each group, and the actual value of the 1st time slot | zone contained in the pattern of prediction object day. Here, the degree of dissimilarity indicates that the greater the value, the lower the degree of similarity, and the smaller the value, the higher the degree of similarity. That is, it can be said that the dissimilarity indicates the degree of similarity. In the following description, the first time zone may be referred to as a comparison target time zone, and the second time zone may be referred to as a prediction target time zone. Subsequently, the prediction device 10 selects a predetermined number, for example, three groups, from the smaller calculated dissimilarity. Then, the prediction apparatus 10 synthesize | combines the part in the 2nd time slot | zone of the pattern contained in a group for every selected group. As described above, the prediction device 10 synthesizes a portion of the pattern included in the group in the second time zone for each predetermined number of groups similar to the actual value of the first time zone included in the pattern of the prediction target day. To do. Therefore, according to the prediction device 10, the synthesized part is similar to the part in the second time zone of any pattern from September 1st to September 30th. Therefore, according to the prediction device 10, if the synthesis result is the prediction result, it is possible to suppress the occurrence of an event in which the prediction result is significantly different from the actual result.

  Moreover, the prediction apparatus 10 classifies patterns having similar performance values in the first time zone and the second time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the 1st time slot | zone of the pattern of each group, and the actual value of the 1st time slot | zone contained in the pattern of prediction object day. That is, according to the prediction device 10, the average value of the actual values in the group is used as data to be compared when calculating the dissimilarity with respect to the pattern of the prediction target day. Therefore, according to the prediction device 10, regarding a group that is not similar to the prediction target day pattern, the selection of the pattern of the group is suppressed.

  Therefore, according to the above, according to the prediction apparatus 10, the fall of the precision of a prediction result can be suppressed.

[Functional configuration of prediction device]
As illustrated in FIG. 1, the prediction device 10 includes an input unit 11, an output unit 12, a storage unit 13, and a control unit 14.

  The input unit 11 inputs various information to the control unit 14. For example, the input unit 11 receives an instruction for executing a prediction process described later from the user, and inputs the received instruction to the control unit 14. As an example of the device of the input unit 11, there is a device that accepts a user operation such as a mouse or a keyboard.

  The output unit 12 outputs various information. For example, the output unit 12 is controlled by an output control unit 14f described later, and includes a prediction pattern including predicted values of power consumption at each time from 10:00 to 20:00 on October 1, 2011, and an implementation described later. Display the degree. An example of the device of the output unit 12 is a liquid crystal display.

  The storage unit 13 stores various information. For example, the storage unit 13 stores a pattern 13a and prediction information 13b.

  The pattern 13a is time-series data including actual power consumption values. FIG. 2 is a diagram illustrating an example of a pattern. In the example of FIG. 2, the pattern 13 a includes time-series data including actual values of power consumption from 0:00 to 23:00 on each day from September 1 to September 30, 2011, and the prediction target date. It includes time-series data including actual values of power consumption from 0:00 to 9:00 on October 1, 2011. The pattern 13 a is acquired from various devices such as an external device and stored in the storage unit 13.

  The prediction information 13b includes a prediction target time zone pattern synthesized by a later-described synthesis unit 14d and a realization degree calculated by a later-described second calculation unit 14e.

  The storage unit 13 is, for example, a semiconductor memory device such as a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 13 is not limited to the type of storage device described above, and may be a RAM (Random Access Memory) or a ROM (Read Only Memory).

  The control unit 14 has an internal memory for storing programs defining various processing procedures and control data, and executes various processes using these. As shown in FIG. 1, the control unit 14 includes a classification unit 14a, a first calculation unit 14b, a selection unit 14c, a synthesis unit 14d, a second calculation unit 14e, and an output control unit 14f. .

  The classification unit 14a classifies various information. For example, the classification unit 14 a compares the time-series data from September 1 to September 30 including the actual power consumption value in the comparison target time zone and the actual power consumption value in the prediction target time zone. Time series data having similar target time zones and prediction target time zones are classified into one group.

  To explain with a specific example, the classification unit 14 a first acquires the pattern 13 a from the storage unit 13. Thereby, the classification | category part 14a can acquire the performance value of the power consumption in each time from 0:00 on September 1, 2011 to 9:00 on October 1.

  Next, the classification unit 14a generates learning data by extracting time series data from 6 o'clock to 20 o'clock on each day from September 1, 2011 to September 30, 2011 from the pattern 13a.

  FIG. 3 shows an example of time-series data from September 1, 2011 to September 30, 2011. In the example of FIG. 3, the time-series data 20 indicates the actual power consumption values from 0:00 to 23:00 on each day from September 1, 2011 to September 30, 2011. For example, the time-series data 20 indicates that the actual power consumption value at 0:00 on September 1, 2011 is 300 kW. The same applies to the actual value of power consumption at each time of each day.

  FIG. 4 shows an example of time-series data from 6 o'clock to 20 o'clock on each day from September 1, 2011 to September 30, 2011. For example, the classification unit 14a calculates the time series from 6 o'clock to 20 o'clock on each day from September 1 to September 30, 2011 shown in the example of FIG. 4 from the time series data 20 shown in the example of FIG. Data is extracted as learning data 21. In the example of FIG. 4, the learning data 21 includes the actual power consumption value 21a in the first time zone from 6 o'clock to 9 o'clock and the power consumption in the second time zone from 10 o'clock to 20 o'clock. The actual value 21b is included.

  Moreover, the classification unit 14a generates prediction target data by extracting time-series data from 6:00 to 9:00 on October 1, 2011, which is the prediction target date, from the pattern 13a. FIG. 5 is a diagram illustrating an example of prediction target data. In the example of FIG. 5, the prediction target data 22 indicates the actual power consumption value in the first time zone from 6:00 to 9:00.

  And the classification | category part 14a is a combination of two days about 30 days from September 1, 2011 to September 30, 2011, for example, September 1 and September 2, September 1 and September 3 All combinations such as days are calculated. The total number of combinations is 435.

  Subsequently, the classification unit 14a performs the following process for all combinations. That is, the classification unit 14a calculates the absolute value of the difference between the actual values for each time from 6:00 to 20:00 using the learning data 21. Then, the classification unit 14a calculates a total of 15 absolute values calculated for each time from 6:00 to 20:00. Subsequently, the classification unit 14a calculates a value obtained by dividing the calculated sum by 15 as the dissimilarity of the time series data from 6 o'clock to 20 o'clock on the two days indicated by the corresponding combination. That is, the classification unit 14a uses the learning data 21 to calculate the degree of similarity between the time series data of the comparison target time zone and the prediction target time zone of the two days indicated by the combination. In this way, the classification unit 14a calculates the dissimilarity for all combinations.

  FIG. 6 is a diagram illustrating an example of a method of calculating the dissimilarity by the classification unit. In the example of FIG. 6, d1 (DAY1, DAY2) represents a function for calculating the dissimilarity from 6:00 to 20:00 in the combination of the day indicated by DAY1 and the day indicated by DAY2. In the example of FIG. 6, for example, the classification unit 14a uses this function to calculate the dissimilarity 243.33 in the combination of September 1, 2011 and September 2, 2011. Yes. In the example of FIG. 6, for example, the classification unit 14 a calculates the dissimilarity 288.33 in the combination of September 1, 2011 and September 3, 2011. In the example of FIG. 6, for example, the classification unit 14 a calculates the dissimilarity 249.67 in the combination of September 29, 2011 and September 30, 2011.

  The classification unit 14a has a larger value as the difference in the actual power consumption value at the time when the power consumption peaks between the time series data from 6 o'clock to 20 o'clock on the two days indicated by the combination. Thus, the dissimilarity can be calculated. Accordingly, a method for calculating such dissimilarity will be described. FIG. 7 is a diagram illustrating another example of a method of calculating the dissimilarity by the classification unit. Here, a combination of September 1, 2011 and September 2, 2011, and a combination of September 1, 2011 and September 3, 2011 will be described as examples. In the example of FIG. 7, the column (a) indicates the actual value of power consumption at each time on September 1st. In the example of FIG. 7, the column (b) indicates the actual power consumption value at each time on September 2nd. In the example of FIG. 7, the column (c) indicates the actual power consumption value at each time on September 3rd.

  In the example of FIG. 7, the column | (a) − (b) | includes the actual power consumption value at each time on September 1 and the actual power consumption value at each time on September 2. Indicates the absolute value of the difference. In the example of FIG. 7, the column | (a) − (c) | includes the actual power consumption value at each time on September 1 and the actual power consumption value at each time on September 3. Indicates the absolute value of the difference.

  In the example of FIG. 7, P (a) is a predetermined value, for example, 0.8, in the peak actual value 1100 among the actual power consumption values from 6 o'clock to 20 o'clock on September 1st. The actual value which becomes more than the value 880 (1100 × 0.8) multiplied by. In the example of FIG. 7, P (a) indicates the actual values 1020, 1100, 1050, and 900 at the respective times from 10:00 to 13:00.

  In the example of FIG. 7, P (b) is a predetermined value, for example, 0.8 in the actual value 1200 that is the peak among the actual values of power consumption from 6 o'clock to 20 o'clock on September 2. The actual value which becomes more than the value 960 (1200 × 0.8) multiplied by. In the example of FIG. 7, P (b) indicates the actual value 1050, 1150, 1200, 1150 at each time from 12:00 to 15:00.

  In the example of FIG. 7, P (c) is a predetermined value, for example, 0.8, to the peak actual value 1000 among the actual power values from 6 o'clock to 20 o'clock on September 3. An actual value that is 1000 (1000 × 0.8) or more multiplied by. In the example of FIG. 7, P (c) indicates the actual value 830, 900, 900, 900, 980, 1000, 900 at each time from 11:00 to 13:00 and from 15:00 to 18:00.

  In the example of FIG. 7, the column | (a) − (b) | of P (a) indicates the actual value of the power consumption on September 1 at the time corresponding to the actual value indicated by P (a). The absolute value of the difference from the actual power consumption value on September 2 is shown. Further, in the example of FIG. 7, the column of | (a) − (b) | of P (b) indicates the actual value of the power consumption on September 1 at the time corresponding to the actual value indicated by P (b). The absolute value of the difference from the actual power consumption value on September 2 is shown. In the example of FIG. 7, the column of | (a) − (c) | of P (a) indicates the actual value of the power consumption on September 1 at the time corresponding to the actual value indicated by P (a). The absolute value of the difference from the actual power consumption value on September 3 is shown. In the example of FIG. 7, the column of | (a) − (c) | of P (c) indicates the actual value of the power consumption on September 1 at the time corresponding to the actual value indicated by P (c). The absolute value of the difference from the actual power consumption value on September 3 is shown.

  As shown in the example of FIG. 7, the classification unit 14a calculates a value 243.33 obtained by dividing the sum by 15 in the same manner as described above for the combination of September 1 and September 2. And the classification | category part 14a is more than the value 880 which multiplied the predetermined value, for example, 0.8, to the actual value 1100 which becomes a peak among the actual values of the power consumption from 6:00 to 20:00 on September 1st. Specify the actual value. In the example of FIG. 7, the classification unit 14 a specifies the actual values 1020, 1100, 1050, and 900 at each time from 10:00 to 13:00.

  Further, the classification unit 14a sets the peak actual value 1200 of the actual power consumption values from 6 o'clock to 20 o'clock on September 2 to a predetermined value, for example, a value 960 or more multiplied by 0.8. Specify the actual value. In the example of FIG. 7, the classification unit 14a specifies the actual values 1050, 1150, 1200, and 1150 at the respective times from 12:00 to 15:00.

  And the classification | category part 14a carries out September about each time of 10 o'clock, 11 o'clock, 12 o'clock, and 13 o'clock corresponding to the specified actual value among the times from 6 o'clock to 20 o'clock on September 1 The absolute value of the difference between the actual power consumption value for 1 day and the actual power consumption value for September 2 is calculated. In the example of FIG. 7, the classification unit 14a calculates absolute values 270, 200, 0, and 250. Then, the classification unit 14a calculates the average value 180.00 ((270 + 200 + 0 + 250) / 4) of the calculated absolute value.

  Subsequently, the classifying unit 14a performs 9 times for each of the times 12:00, 13:00, 14:00, and 15:00 corresponding to the identified actual value among the times from 6:00 to 20:00 on September 2. The absolute value of the difference between the actual power consumption value on month 2 and the actual power consumption value on September 2 is calculated. In the example of FIG. 7, the classification unit 14a calculates absolute values 0, 250, 500, and 650. Then, the classification unit 14a calculates an average value 350.00 ((0 + 250 + 500 + 650) / 4) of the calculated absolute values.

  Thereafter, the classification unit 14a calculates the dissimilarity d1 (2011/2011) for the combination of September 1 and September 2 with the sum 773.33 of the calculated 243.33, 180.00, and 350.00. 09/01, 2011/09/02). Similarly, the classification unit 14a also calculates the dissimilarity d1 (2011/09/01, 2011/09/03) in the combination of September 1 and September 3. Further, the classification unit 14a can calculate the similarity in the same manner for all combinations. In this way, the classification unit 14a calculates the dissimilarity in consideration of the difference in the peak time zone of the actual value of the power consumption between the two time series data.

  Then, the classification unit 14a uses the dissimilarities calculated for all the combinations to calculate a predetermined number of time series data from 6 o'clock to 20 o'clock every day, for example, four clusters A by a known group average method. , B, C, and D for clustering. That is, the classification unit 14a classifies the time series data into a predetermined number, for example, four groups. In this manner, the classification unit 14a collects similar time series data into one cluster. As described above, when clustering is performed using the dissimilarity in consideration of the peak time zone of the actual power consumption between the two time-series data, the classification unit 14a displays the peak time zone. Clustering can be performed in consideration of differences.

  FIG. 8 is a diagram illustrating an example of the result of clustering. In the example of FIG. 8, in cluster A, September 1, 2011, September 26, 2011, September 24, 2011, September 27, 2011, September 21, 2011, September 8, 2011 The case where each time series data of a day is classified is shown. Further, the example of FIG. 8 also shows that the time series data indicated by each date / time is classified for the clusters B, C, and D as well.

  The first calculation unit 14b calculates various information. For example, the first calculation unit 14b calculates the average value of the actual value in the comparison target time zone of the time series data (pattern) classified into each clustering and the actual value in the comparison target time zone of the time series data of the prediction target date. The dissimilarity with is calculated.

  A specific example will be described. FIG. 9 is a diagram for explaining an example of processing executed by the first calculation unit. Each time series data shown in the example of FIG. 8 is classified in each of the clusters A, B, C, and D shown in the example of FIG. In the example of FIG. 9, d0 (DAY1, DAY2) is the absolute difference between the actual value of the time-series data indicated by DAY1 and the actual value of the time-series data indicated by DAY2 for each time in the comparison target time zone. It is a function for calculating an average value of sums of values as dissimilarity. For example, in the example of FIG. 9, the first calculation unit 14b calculates the dissimilarity using this function.

  For example, the first calculator 14b calculates the absolute value of the difference between the actual values of the pattern classified as clustering and the pattern of the prediction target day for each time from 6:00 to 9:00. And the 1st calculation part 14b calculates the sum total of four absolute values calculated about each time from 6:00 to 9:00. Subsequently, the first calculation unit 14b calculates an average value obtained by dividing the calculated sum by 4 as the dissimilarity of the pattern from 6 o'clock to 9 o'clock of the corresponding combination. In other words, the first calculation unit 14b, in this way, the classification unit 14a determines the dissimilarity for all combinations of all patterns classified into all clustering and the prediction target day patterns. calculate.

  For example, the example of FIG. 9 shows a case where the first calculation unit 14b calculates the dissimilarity 65.00 in the combination of September 1, 2011 and October 1, 2011.

  And the 1st calculation part 14b calculates the average value of the calculated dissimilarity for every cluster. For example, in the example of FIG. 9, d2 (CLUSTER1, DAY1) is the average of the dissimilarities of the comparison target time zones calculated for the pattern classified into the cluster indicated by CLUSTER1 and the prediction target day pattern indicated by DAY1. This is a function for calculating a value. For example, in the example of FIG. 9, the first calculation unit 14b uses this function to calculate the average value of dissimilarities.

  In the example of FIG. 9, the first calculation unit 14 b calculates the average dissimilarity value 65.00 for the cluster A. Further, in the example of FIG. 9, the first calculation unit 14 b calculates the average dissimilarity value 81.00 for the cluster B. In the example of FIG. 9, the first calculation unit 14 b calculates the average value 90.62 of the dissimilarity for the cluster C. In the example of FIG. 9, the first calculation unit 14 b calculates the average dissimilarity value 138.75 for the cluster D.

  The selection unit 14c selects a cluster based on the dissimilarity calculated by the first calculation unit 14b.

  For example, the selecting unit 14c first specifies the dissimilarity having the smallest value among the dissimilarities calculated by the first calculating unit 14b. For example, as shown in the example of FIG. 9, when four dissimilarities 65.00, 81.00, 90.62, and 138.75 are calculated for the clusters A to D by the first calculation unit 14b. The selecting unit 14c specifies the dissimilarity 65.00 having the smallest value. Then, the selection unit 14c newly specifies a dissimilarity equal to or less than a value obtained by multiplying the dissimilarity having the smallest value by a predetermined value, for example, 1.5, and selects a cluster corresponding to the specified dissimilarity. To do. Note that the selection unit 14c can newly specify a dissimilarity equal to or smaller than a value obtained by adding a predetermined value to the dissimilarity having the smallest value, and can select a cluster corresponding to the specified dissimilarity.

  FIG. 10 is a diagram for explaining an example of processing executed by the selection unit. In the example of FIG. 10, the case where the selection unit 14c specifies the dissimilarity 65.00 having the smallest value is shown. In addition, in the example of FIG. 10, the selection unit 14c newly adds dissimilarities 81.00 and 90.62 that are equal to or less than 97.5 by multiplying the dissimilarity 65.00 by a predetermined value, for example, 1.5. The case specified in is shown. Then, the selection unit 14c selects clusters A, B, and C corresponding to the specified dissimilarities 65.00, 81.00, and 90.62, respectively.

  For each cluster selected by the selection unit 14c, the synthesis unit 14d synthesizes a prediction target time zone portion of the pattern classified into clusters.

  A specific example will be described. 11 and FIGS. 12A, 12B, and 12C are diagrams for explaining an example of processing executed by the synthesis unit. First, the combining unit 14d calculates a value W obtained by dividing a predetermined number of patterns (number of time-series data) N by the number of clusters selected by the selection unit 14c. For example, when the number of patterns N is 6 and three clusters A, B, and C are selected by the selection unit 14c as illustrated in the example of FIG. 10, the synthesis unit 14d determines that the value W = 2 (6 / 3) is calculated.

  Then, the combining unit 14d selects, for each cluster, the top W patterns having high similarity (low dissimilarity) with the pattern of the comparison target time zone of the prediction target date. For example, when the value W = 2, as shown in the example of FIG. 11, for each of the clusters A, B, and C, the top two items having high similarity to the pattern of the comparison target time zone of the prediction target day Select the pattern. In the example of FIG. 11, among the patterns classified into the cluster A, the top two patterns having high similarity to the pattern of the comparison target time zone of the prediction target day are the patterns of September 26, 2011, This indicates that the pattern is September 24, 2011. In the example of FIG. 11, among the patterns classified into the cluster B, the top two patterns having high similarity to the pattern of the comparison target time zone of the prediction target day are the patterns of September 5, 2011, This indicates that the pattern is September 22, 2011. In the example of FIG. 11, among the patterns classified into the cluster C, the top two patterns having high similarity to the pattern of the comparison target time zone of the prediction target day are the patterns of September 6, 2011, This indicates that the pattern is on September 29, 2011.

  The synthesizing unit 14d may select, from each cluster, a pattern in which the dissimilarity with the pattern of the comparison target time zone of the prediction target day is equal to or less than α + β for the patterns classified into the selected clusters. it can. Here, α is the dissimilarity of the pattern having the smallest dissimilarity with the pattern of the comparison target time zone of the prediction target date. Β is a predetermined value.

  Subsequently, the synthesizing unit 14d synthesizes the selected top W patterns having high similarity for each cluster. For example, as illustrated in the example of FIG. 11, when two patterns are selected for each of the clusters A, B, and C, the combining unit 14d performs the following process. That is, the synthesizing unit 14d calculates an average value of actually measured values of the prediction target time zones of the two patterns for each of the clusters A, B, and C. Then, the combining unit 14d combines the two patterns in the prediction target time zone and the average value of the actual measurement values. In this way, the synthesizing unit 14d generates a pattern in the prediction target time zone, and generates predicted time series data (predicted pattern).

  As illustrated in the example of FIG. 12A, the combining unit 14d calculates an average value of the pattern of September 24 and the pattern of September 26 in the prediction target time zone from 10:00 to 20:00 for the pattern A. Thus, a predicted value (predA) of power consumption at each time in the prediction target time zone on October 1 is calculated. Then, the combining unit 14d combines the calculated predicted value, the September 24 pattern, and the September 26 pattern.

  Further, as illustrated in the example of FIG. 12B, the combining unit 14d calculates an average value of the pattern of September 6 and the pattern of September 29 of the prediction target time zone for the pattern C, A predicted value (predC) of power consumption at each time in the day prediction target time zone is calculated. Then, the combining unit 14d combines the calculated predicted value, the September 6 pattern, and the September 29 pattern.

  Further, as illustrated in the example of FIG. 12C, the combining unit 14d calculates an average value of the pattern of September 5 and the pattern of September 22 of the prediction target time zone for the pattern B, A predicted power consumption value (predB) at each time in the day prediction target time zone is calculated. Then, the combining unit 14d combines the calculated predicted value, the September 5 pattern, and the September 22 pattern.

  The second calculation unit 14e calculates, for each prediction value calculated by the synthesis unit 14d, a realization degree that is a degree that the prediction value can occur.

  For example, the second calculation unit 14e calculates the number of patterns classified into clusters for each cluster corresponding to the predicted value. FIG. 13 is a diagram for explaining an example of processing executed by the second calculation unit. For example, when the pattern is classified into each cluster by the classification unit 14a as shown in FIG. 8 and three clusters are selected by the selection unit 14c as shown in FIG. 11, the second calculation unit 14e Perform the following process. That is, as shown in the example of FIG. 13, the second calculation unit 14e has six patterns classified into cluster A, ten patterns classified into cluster B, and patterns classified into cluster C. The number is calculated as 8.

  Then, the second calculation unit 14e calculates the number of patterns of all clusters. For example, in the example of FIG. 13, the second calculation unit 14e calculates the number 24 (6 + 10 + 8) of patterns of all clusters.

  Subsequently, the second calculation unit 14e calculates, for each cluster, a value T obtained by dividing the smallest dissimilarity specified by the selection unit 14c by the dissimilarity corresponding to the cluster. Further, the second calculation unit 14e calculates, for each cluster, a value G obtained by dividing the number of cluster patterns by the number of patterns of all clusters. And the 2nd calculation part 14e calculates the multiplication value of the value T and the value G as a realization degree for every cluster.

  For example, in the example of FIG. 13, the second calculation unit 14 e calculates the degree of realization 0.25, which is the degree that a predicted value corresponding to the cluster A can occur. In the example of FIG. 13, the second calculation unit 14 e calculates a degree of realization of 0.334, which is a degree that a predicted value corresponding to the cluster B can occur. In the example of FIG. 13, the second calculation unit 14 e calculates a realization degree 0.239, which is a degree that a predicted value corresponding to the cluster C can be generated.

  Then, for each cluster, the second calculation unit 14e generates prediction information 13b including the degree of realization corresponding to the cluster and the synthesis result by the synthesis unit 14d, and stores the generated prediction information 13b in the storage unit 13. To do.

  FIG. 14 is a diagram for explaining an example of processing executed by the second calculation unit. In the example of FIG. 14, an example of the prediction information 13b calculated by the second calculation unit 14e is shown. In the example of FIG. 14, the degree of realization is added to the synthesis result.

  The output control unit 14f acquires the prediction information 13b stored in the storage unit 13, and controls the display of the output unit 12 so as to display the acquired prediction information. For example, the output control unit 14f performs control so as to display the prediction information 13b illustrated in the example of FIG.

  FIG. 15 is a diagram for explaining an example of processing executed by the prediction apparatus according to the embodiment. The left diagram of the example of FIG. 15 is a pattern indicated by the pattern 13a, and the right diagram illustrates a case where the prediction information 13b obtained as a result of the processing of the prediction device 10 is displayed. The prediction device 10 according to the example classifies patterns having similar performance values in the comparison target time zone and the prediction target time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the comparison target time zone of the pattern of each group and the actual value of the comparison target time zone included in the pattern of the prediction target day. Subsequently, the prediction device 10 selects a predetermined number, for example, three groups, from the smaller calculated dissimilarity. Then, the prediction apparatus 10 synthesize | combines the part in the prediction object time slot | zone of the pattern contained in a group for every selected group. That is, as illustrated in the example of FIG. 15, the prediction device 10 predicts the target time of the pattern included in the group for each predetermined number of groups similar to the actual value of the comparison target time zone included in the pattern of the target date of prediction. Combine the parts in the band. Therefore, according to the prediction device 10, the synthesized part is similar to the part in the prediction target time zone of any pattern from September 1st to September 30th. Therefore, according to the prediction device 10, by using the synthesis result as the prediction result, it is possible to suppress occurrence of an event in which the prediction result is significantly different from the actual result.

  Moreover, the prediction apparatus 10 classifies patterns having similar performance values in the comparison target time zone and the prediction target time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the comparison target time zone of the pattern of each group and the actual value of the comparison target time zone included in the pattern of the prediction target day. That is, according to the prediction device 10, the average value of the actual values in the group is used as data to be compared when calculating the dissimilarity with respect to the pattern of the prediction target day. Therefore, according to the prediction device 10, regarding a group that is not similar to the prediction target day pattern, the selection of the pattern of the group is suppressed.

  Therefore, according to the above, according to the prediction apparatus 10, the fall of the precision of a prediction result can be suppressed.

  The control unit 14 is an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA) or an electronic circuit such as a central processing unit (CPU) or a micro processing unit (MPU).

[Process flow]
Next, the process flow of the prediction apparatus 10 according to the present embodiment will be described. FIG. 16 is a flowchart illustrating the procedure of the prediction process according to the embodiment. The prediction process is executed, for example, at the timing when the control unit 14 receives an instruction for executing the prediction process from the input unit 11.

  As shown in FIG. 16, the classification unit 14a acquires the pattern 13a from the storage unit 13 (S101). Subsequently, the classification unit 14a generates learning data and prediction target data (S102). And the classification | category part 14a calculates dissimilarity about all the combinations (S103).

  Then, the classification unit 14a classifies the time series data from 6 o'clock to 20 o'clock every day into a predetermined number of clusters (groups) by a known group average method using the dissimilarities calculated for all combinations. Clustering is performed (S104). The first calculation unit 14b calculates the average value of the actual value in the comparison target time zone of the time series data (pattern) classified into each clustering and the actual value in the comparison target time zone of the time series data of the prediction target date. The dissimilarity is calculated (S105).

  Subsequently, the selecting unit 14c selects a cluster having a small dissimilarity based on the dissimilarity calculated by the first calculating unit 14b (S106). Subsequently, the synthesizing unit 14d selects, for each cluster, the top W patterns having high similarity to the pattern of the comparison target time zone of the prediction target date (S107). Then, the synthesizing unit 14d synthesizes the two patterns in the prediction target time zone and the average value of the actually measured values (S108).

  The second calculation unit 14e calculates, for each prediction value calculated by the synthesis unit 14d, a realization degree that is a degree at which the prediction value can occur (S109). Then, for each cluster, the second calculation unit 14e generates prediction information 13b including the degree of realization corresponding to the cluster and the synthesis result by the synthesis unit 14d, and stores the generated prediction information 13b in the storage unit 13. (S110).

  And the output control part 14f acquires the prediction information 13b memorize | stored in the memory | storage part 13, controls the display of the output part 12 so that the acquired prediction information may be displayed (S111), and complete | finishes a process.

  As described above, the prediction device 10 according to the embodiment classifies patterns having similar performance values in the comparison target time zone and the prediction target time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the comparison target time zone of the pattern of each group and the actual value of the comparison target time zone included in the pattern of the prediction target day. Subsequently, the prediction device 10 selects a predetermined number, for example, three groups, from the smaller calculated dissimilarity. Then, the prediction apparatus 10 synthesize | combines the part in the prediction object time slot | zone of the pattern contained in a group for every selected group. That is, as illustrated in the example of FIG. 15, the prediction device 10 predicts the target time of the pattern included in the group for each predetermined number of groups similar to the actual value of the comparison target time zone included in the pattern of the target date of prediction. Combine the parts in the band. Therefore, according to the prediction device 10, the synthesized part is similar to the part in the prediction target time zone of any pattern from September 1st to September 30th. Therefore, according to the prediction device 10, by using the synthesis result as the prediction result, it is possible to suppress occurrence of an event in which the prediction result is significantly different from the actual result.

  Moreover, the prediction apparatus 10 classifies patterns having similar performance values in the comparison target time zone and the prediction target time zone into the same group. And the prediction apparatus 10 calculates the dissimilarity of the average value of the actual value in the comparison target time zone of the pattern of each group and the actual value of the comparison target time zone included in the pattern of the prediction target day. That is, according to the prediction device 10, the average value of the actual values in the group is used as data to be compared when calculating the dissimilarity with respect to the pattern of the prediction target day. Therefore, according to the prediction device 10, regarding a group that is not similar to the prediction target day pattern, the selection of the pattern of the group is suppressed.

  Therefore, according to the above, according to the prediction apparatus 10, the fall of the precision of a prediction result can be suppressed.

  Although the embodiments related to the disclosed apparatus have been described above, the present invention may be implemented in various different forms other than the above-described embodiments. Therefore, another embodiment included in the present invention will be described below.

  For example, in the above-described embodiment, the case where the actual values at the same time are compared when calculating the dissimilarity has been described, but the disclosed apparatus is not limited thereto. For example, when the dissimilarity is calculated, the disclosed apparatus uses a well-known DTW (Dynamic Time Warping) distance (Y. Sakurai, et al .: FTW: fast similarity search under the time warping distance. PODS 2005). , pp 326-337). Thereby, it becomes possible to compare the actual values having the closest values within a predetermined time width.

  In addition, among the processes described in the embodiments, all or a part of the processes described as being automatically performed can be manually performed. In addition, among the processes described in this embodiment, all or a part of the processes described as being performed manually can be automatically performed by a known method.

  In addition, the processing at each step of each processing described in each embodiment can be arbitrarily finely divided or combined according to various loads and usage conditions. Also, the steps can be omitted.

  Further, the order of processing at each step of each processing described in each embodiment can be changed according to various loads and usage conditions.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific state of distribution / integration of each device is not limited to the one shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

[Prediction program]
Various processes of the prediction apparatus 10 described in the above embodiment can also be realized by executing a prepared program on a computer system such as a personal computer or a workstation. In the following, an example of a computer that executes a prediction program having the same function as that of the prediction apparatus 10 described in the above embodiment will be described with reference to FIG. FIG. 17 is a diagram illustrating a computer that executes a prediction program.

  As shown in FIG. 17, the computer 300 includes a central processing unit (CPU) 310, a read only memory (ROM) 320, a hard disk drive (HDD) 330, and a random access memory (RAM) 340. These units 310 to 340 are connected via a bus 350.

  The ROM 320 stores basic programs such as an OS. The HDD 330 is similar to the classification unit 14a, the first calculation unit 14b, the selection unit 14c, the synthesis unit 14d, the second calculation unit 14e, and the output control unit 14f described in the above embodiment. A prediction program 330a that performs the function is stored in advance. Note that the prediction program 330a may be separated as appropriate. The HDD 330 is provided with a pattern and prediction information. These patterns and prediction information correspond to the pattern 13a and the prediction information 13b described above.

  Then, the CPU 310 reads the prediction program 330a from the HDD 330 and executes it.

  Then, the CPU 310 reads the pattern and prediction information and stores them in the RAM 340. Further, the CPU 310 executes the prediction program 330 a using the pattern and prediction information stored in the RAM 340. Note that all data stored in the RAM 340 may not always be stored in the RAM 340. Data used for processing may be stored in the RAM 340.

  Note that the above-described prediction program 330a is not necessarily stored in the HDD 330 from the beginning.

  For example, the program is stored in a “portable physical medium” such as a flexible disk (FD), a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card inserted into the computer 300. Then, the computer 300 may read and execute the program from these.

  Furthermore, the program is stored in “another computer (or server)” connected to the computer 300 via a public line, the Internet, a LAN, a WAN, or the like. Then, the computer 300 may read and execute the program from these.

DESCRIPTION OF SYMBOLS 10 Prediction apparatus 12 Output part 13 Storage part 13a Pattern 13b Prediction information 14 Control part 14a Classification part 14b First calculation part 14c Selection part 14d Composition part 14e Second calculation part 14f Output control part

Claims (7)

Among a plurality of first data including the actual value of the first time zone and the actual value of the second time zone, data having similar performance values of the first time zone and the second time zone A classification unit for classifying into the same group;
A first calculation unit that calculates a degree of similarity between the actual value in the first time zone of the group and the actual value in the first time zone included in second data;
A selection unit that selects a plurality of groups based on the degree calculated by the first calculation unit;
A prediction apparatus comprising: a synthesis unit that synthesizes a portion of the first data in a second time zone included in the group for each group selected by the selection unit. Based on the degree calculated by the first calculation unit, the degree to which an event represented by the synthesized first data portion can occur for each portion of the first data synthesized by the synthesis unit The prediction apparatus according to claim 1, further comprising: a second calculation unit that calculates For each group selected by the selection unit, the combining unit includes one or a plurality of first items in descending order of similarity based on the degree of similarity calculated by the first calculation unit. The prediction apparatus according to claim 1 or 2, wherein data is selected and the second time zone portion of the selected first data is synthesized. The prediction apparatus according to claim 1, further comprising: an output unit that outputs a synthesis result by the synthesis unit. The prediction apparatus according to claim 2, further comprising: an output unit that outputs a synthesis result by the synthesis unit and the degree calculated by the second calculation unit. On the computer,
Among a plurality of first data including the actual value of the first time zone and the actual value of the second time zone, data having similar performance values of the first time zone and the second time zone Group them into the same group,
Calculating a degree of similarity between the actual value in the first time zone of the group and the actual value in the first time zone included in the second data;
Based on the calculated degree, select a plurality of groups,
The prediction program for performing each process which synthesize | combines the part of said 1st data in the 2nd time slot | zone contained in this group for every selected group. A prediction method performed by a computer,
Among a plurality of first data including the actual value of the first time zone and the actual value of the second time zone, data having similar performance values of the first time zone and the second time zone Group them into the same group,
Calculating a degree of similarity between the actual value in the first time zone of the group and the actual value in the first time zone included in the second data;
Based on the calculated degree, select a plurality of groups,
A prediction method, comprising: causing each selected group to execute each process of combining the first data portion in a second time zone included in the group.

Images