JP6245030B2 - Power consumption prediction method, power consumption prediction program, and power consumption prediction apparatus - Google Patents

Description

  The present invention relates to a power consumption prediction method, a power consumption prediction program, and a power consumption prediction apparatus that perform power consumption prediction.

  Along with the current power situation, there is a demand for detailed understanding of how the daily power demand is likely to fluctuate depending on the time and time zone, as well as forecasting the peak value for each day's power demand at the office. . There are several fluctuation patterns in the power demand time series of the offices of the day, such as “If the morning demand is such a power demand, the afternoon will be this much power demand,” depending on the situation of the day. One of them tends to appear. Focusing on this point, the present inventor lists several daily fluctuation patterns that are likely to appear on each day for the power demand time series of the establishment, and the patterns that are most likely to appear are accurately listed. A prediction device for prediction is proposed (Patent Document 1).

JP2013-114629A

  However, the above-described conventional techniques have the following problems. A diurnal variation pattern different from the previous one may suddenly appear at the turn of the season (hereinafter also referred to as “state change”). At that time, in the conventional method, an appropriate daily fluctuation pattern cannot be selected, and an accurate predicted daily fluctuation pattern may not be created. And it may be repeated for a while after the state change.

  In one aspect, an object of the present invention is to provide a power consumption prediction method that suppresses deterioration in prediction accuracy.

  In one aspect, the power consumption prediction method uses 24 hours starting from a predetermined time as a day unit, and includes a predetermined number of groups of actual time series for each unit day associated with power consumption and time. Power consumption and time in the first time zone on the prediction target unit date by acquiring a representative time series calculated from the time series group classified using dissimilarity and the actual time series included in the time series group A reference time series including the above is acquired, and based on the results of past power consumption prediction, it is set to either a state change when the power consumption trend changes or a normal time when the change does not appear Appearance state information is acquired, and when the acquired appearance state information is a state change, the acquired reference time series and the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before With each of N actual time series The similarity is calculated, the actual time series having the smallest calculated dissimilarity is identified, the dissimilarity between the identified actual time series and each of the representative time series is calculated, and the calculated dissimilarity is The representative time series that is the minimum is selected, and the portion corresponding to the second time zone that follows the first time zone of the selected representative time series is set as the prediction time series that predicts the power consumption of the prediction target unit day. Acquire and output the acquired prediction time series.

    According to one aspect of the present invention, when a situation in which the appearance tendency of the daily fluctuation pattern is greatly changed is detected, the selection scale of the daily fluctuation pattern is switched, so that deterioration of prediction accuracy can be suppressed.

It is a block diagram which shows an example of the hardware constitutions of a power consumption prediction apparatus. It is a graph which shows an example of a power consumption time series. It is a figure which shows an example of the record layout of power consumption time series DB. It is a flowchart which shows an example of the process sequence of a power consumption prediction process. It is a flowchart which shows an example of the process sequence of the prediction time series production | generation process based on Pf. It is a flowchart which shows an example of the process sequence of the prediction time series production | generation process based on Nc. It is a flowchart which shows an example of the process sequence of the appearance state information update process of a time series pattern. It is the table | surface which showed the result of the cluster selection based on Pf. It is the graph which showed the reference time series and the partial time series for the past 3 days. It is a table | surface which shows the calculation result of the dissimilarity of a reference time series and each partial time series. It is the table | surface which showed the selection error of each cluster, and the prediction error by each cluster. It is the graph which showed the representative cluster and the performance time series. It is a graph which shows a result at the time of performing prediction by switching a cluster selection method. It is a list of evaluation results of prediction results. It is explanatory drawing which shows an example of a prediction result output. It is a block diagram which shows the function structural example of a power consumption prediction apparatus.

  FIG. 1 is a block diagram illustrating an example of a hardware configuration of the power consumption prediction apparatus 1. The power consumption prediction apparatus 1 includes a CPU (Central Processing Unit) 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, a mass storage device 14, a communication unit 15, and a reading unit 16. Each component is connected by a bus.

  The CPU 11 controls each part of the hardware according to the control program 1P stored in the ROM 13. The RAM 12 is, for example, SRAM (Static RAM), DRAM (Dynamic RAM), or flash memory. The RAM 12 temporarily stores data generated when the CPU 11 executes the program.

  The mass storage device 14 is, for example, a hard disk or an SSD (Solid State Drive). The large-capacity storage device 14 stores a power consumption time series DB 141, a predicted time series DB 142, and a result / state DB 143. Further, the control program 1P may be stored in the mass storage device 14. The power consumption time series DB 141, the prediction time series DB 142, and the results / state DB 143 may be stored in another computer connected to the power consumption prediction apparatus 1 via the network N.

  The communication unit 15 communicates with other computers via the network N. The reading unit 16 reads a portable storage medium 1a including a CD (Compact Disk) -ROM and a DVD (Digital Versatile Disc) -ROM. The CPU 11 may read the control program 1P from the portable storage medium 1a via the reading unit 16 and store it in the mass storage device 14. Further, the CPU 11 may download the control program 1P from another computer via the network N and store it in the mass storage device 14. Furthermore, the CPU 11 may read the control program 1P from the semiconductor memory 1b.

  In the following description, 24 hours starting from a predetermined time is set as a day unit. Each day is called a unit day. For example, when the starting time is 6:00 am, one unit day is from 6:00 am to 6:00 am the next day. Moreover, the day which performs prediction is called a prediction target unit date or a prediction target date. In the present embodiment, since 24 hours from midnight to midnight the next is one unit day, the unit day and the calendar day are the same.

  FIG. 2 is a graph showing an example of a power consumption time series. FIG. 3 is a diagram illustrating an example of a record layout of the power consumption time series DB 141. The horizontal axis in FIG. 2 is the date and time, and the vertical axis is the time series value of power consumption. The time series values shown in FIGS. 2 and 3 are ratioless when the actually observed power consumption value is set to 1 as the reference power value, and the unit is dimensionless. The value of power consumption is, for example, an average of two values when the maximum value of power consumption at 0 to 29 minutes per hour is a (Kw) and the maximum value of power consumption at 30 to 59 minutes is b (Kw). Value = (a + b) / 2 is recorded as the next hour (0 minute) value. For example, the maximum value of power consumption from 10:00 to 10:29 and the average value of the maximum power consumption from 10:30 to 10:59 are recorded as values of 11:00. . In the present embodiment, it is assumed that the target for measuring power consumption is a company office. Power consumption is not recorded or predicted for weekends and holidays. For this reason, the data shown in FIGS. 2 and 3 have missing dates.

  The time series value of power consumption handled in this embodiment is a value obtained by dividing an actually measured value by the maximum value measured on that day. As for the prediction target day, since the maximum value is not fixed at the time of prediction, the time series value is a value obtained by dividing by the average value of the maximum values of each business day until the past one week. The time series value is used as the ratio value so that even if there is a variation in the value from day to day, the value is normalized so that the fluctuation trend of the value of one day can be compared with the fluctuation trend of the value of another day. This is because of In the following description, unless otherwise specified, the power consumption value is a normalized ratio value.

  A rectangle d0 surrounds the time series for one day on August 24th. A rectangle d1 surrounds the time series from 6 o'clock to 20 o'clock on August 24. A rectangle d3 encloses a time series from 6 o'clock to 20 o'clock on September 27 when the prediction is performed.

  The time series is divided into three time zones. Reference time zone (first time zone), predicted time zone (second time zone), and other time zones. The reference time zone is from 6 o'clock to 9 o'clock, and this time zone is not predicted. The value of power consumption in the reference time zone is used to predict power consumption in the prediction time zone. The predicted time zone is from 10:00 to 20:00, and is a time zone in which power consumption is predicted. In the predicted time zone, the power consumption value is also stored in the power consumption time series DB 141. Other time zones are from 21:00 to 5:00 the next day. For other time zones, the power consumption value is stored in the power consumption time series DB 141.

  The settings for the reference time zone, the predicted time zone, and other time zones may be determined according to the purpose of the prediction. In the present embodiment, since the purpose is to predict power consumption every morning until the night of the day, the setting is as described above. Further, when making a prediction, it is necessary to decide how many days before the date of the prediction date to use the data. In this embodiment, it is up to 28 days ago. In addition, since there is no measurement data about a closed day, the number of past measurement data used for prediction may fluctuate depending on the prediction date, but this is allowed. That is, even if the number of past measurement data is small, the past data is not used more than 28 days ago. This is because, if going back to the past too much, there is a possibility that the fluctuation pattern of the power consumption is significantly different from the latest fluctuation pattern, which may lead to a decrease in accuracy. The above settings are merely examples. For example, the power consumption from 10:00 on that day to 20:00 on the next day may be predicted using a time series from 6:00 to 9:00 every morning. In determining the number of days before the measurement data set in the reference time zone, predicted time zone, and other time zones, perform preliminary experiments using past data and make predictions with the desired accuracy. It is desirable to determine whether or not this is possible.

  As shown in FIG. 3, the power consumption time series DB 141 includes a date / time column and a value column. The date and time column stores the date and the hour of every hour. The value column stores the actual value of power consumption.

  In this embodiment, it is assumed that the prediction for September 27 is performed. That is, the prediction from 10:00 to 20:00 is performed from the actual value from 6:00 to 9:00 on September 27.

  FIG. 4 is a flowchart illustrating an example of a processing procedure of power consumption prediction processing. The CPU 11 of the power consumption prediction apparatus 1 acquires the reference time series and the learning partial time series from the power consumption time series DB 141 (step S1). The reference time series is data from 6:00 to 9:00 on the predicted date. In FIG. 3, it is the data in the rectangle which attached | subjected the code | symbol of 141a. The partial time series for learning (hereinafter, referred to as “partial time series”) is time series data from the day before the prediction date to 28 days before, and is data of actual results in the reference time zone and the prediction time zone. . Here, it is time series data from August 30 to September 26, and is data from 10:00 to 20:00 on each day.

ここで、ｒｅｆｅｒは参照時間帯を、ｐｒｅｄは予測時間帯を示す。時間帯ｐｅｒｉｏｄの時点数をｎｕｍ（ｐｅｒｉｏｄ）で表す。ｓ_i［ｔ］は、部分時系列ｓ_i上の時刻ｔの値を示す。非類似度の算出はすべての組み合わせについて行う。 First, the CPU 11 calculates dissimilarity for any two partial time series (step S2). The calculation of the dissimilarity distRP (s _i , s _j ) is performed according to the following formula (1), for example.

Here, refer indicates a reference time zone, and pred indicates a predicted time zone. The number of time points in the time period period is represented by num (period). s _i [t] indicates a value at time t on the partial time series s _i . The dissimilarity is calculated for all combinations.

Next, the CPU 11 clusters the partial time series for learning using the obtained dissimilarity (step S3). In the present embodiment, it is classified into five clusters (time series group). For example, in each cluster C _k , clustering is performed by the k-means method of the number of clusters K, with the average value at each time point of the partial time series belonging to C _k as the center point. The first K center points are generated by the KKZ method. However, the present invention is not limited to this, and other clustering methods may be used.

As for the number of clusters, the number of clusters that matches the time-series pattern of the target office is determined based on the results of preliminary experiments. For example, it is assumed that a certain office is divided into four patterns: (1) the day after the holiday, (2) before the holiday, (3) the regular departure date, and (4) other days. In this case, considering the possibility of appearance of an irregular pattern other than 4 patterns, good prediction accuracy can be obtained by dividing the number of 5 clusters by adding 1 to 4. As another method for determining the number of clusters, it is also possible to use the Sturges formula (for the number of elements N, (1 + l _og2 N) is the number of clusters).

Subsequently, the CPU 11 generates a representative time series rep (C _k ) for each cluster (step S4). For example, the average value at each time point of the partial time series belonging to the cluster C _k is set as the representative time series rep (C _k ).

  Further, the CPU 11 determines whether or not the appearance state is normal (step S5). The normal time refers to a state in which there is no change in the power consumption trend and the prediction by the conventional method is assumed to be effective. When the state changes, there is a change in the power consumption trend due to the occurrence of a factor that greatly fluctuates the partial time series pattern. When the state changes, there is a change in the power consumption trend, so it is assumed that the prediction accuracy will deteriorate in the prediction by the conventional method.

  If the CPU 11 determines that the appearance state is normal (YES in step S5), the CPU 11 generates a prediction time series based on Pf, which is the conventional method (step S6). Subsequently, the CPU 11 displays the generated prediction time series (step S7), stores the prediction time series (normal prediction time series) in the prediction time series DB 142 (step S11), and ends the process. The record layout of the prediction time series DB 142 is the same as that of the power consumption time series DB 141.

  When the CPU 11 determines that the appearance state is a state change (NO in step S5), the CPU 11 generates a prediction time series based on Nc (step S8). Subsequently, the CPU 11 displays the generated prediction time series (step S9). Further, a prediction time series based on Pf is also generated (step S10). Then, the predicted time series based on the generated Pf is stored in the predicted time series DB 142 (step S11), and the process ends.

  Next, prediction time series generation processing based on Pf will be described. FIG. 5 is a flowchart showing an example of the processing procedure of the prediction time series generation processing based on Pf. First, the CPU 11 calculates the reference time series and the dissimilarity between clusters (step S21). The cluster is the cluster obtained in step S3 of FIG. Calculate dissimilarity with reference time series for all clusters.

ここで、Ｔ_pは予測対象日を表す。ｓ_p［ｔ］は予測対象日の部分時系列を示す。予測を実行する時点では、参照時間帯（ｒｅｆｅｒ）の消費電力は明らかになっている。各クラスタについて、クラスタに含まれるすべての時系列について、非類似度ｄｉｓｔＲ（ｓ_i，Ｔ_p）の算出を行う。 First, the dissimilarity distR (s _i , T _p ) between the time series s _i and the reference time series included in each cluster is calculated. The calculation of the dissimilarity distR (s _i , T _p ) is performed according to the following formula (2), for example.

Here, T _p represents the prediction target date. s _p [t] indicates a partial time series of the prediction target date. At the time when the prediction is executed, the power consumption in the reference time period (refer) is clear. For each cluster, the dissimilarity distR (s _i , T _p ) is calculated for all time series included in the cluster.

ここで、ｋ＝０から５の自然数である。Ｃ_kはクラスターを示す。ｓ_k（j）は、クラスタ−Ｃ_kに含まれる部分時系列を示している。 Using the obtained dissimilarity distR (s _i , T _p ), the dissimilarity DistR (C _k , T _p ) between the cluster and the reference time series is calculated. The calculation of the dissimilarity DistR (C _k , T _p ) is performed according to the following formula (3).

Here, k is a natural number from 0 to 5. C _k represents a cluster. s _{k (j)} indicates a partial time series included in the cluster -C _k .

ここで、ｎｕｍ（Ｓ）＝ｎｕｍ（Ｃ_０）＋ ｎｕｍ（Ｃ_１）…＋ ｎｕｍ（Ｃ_４）、総メンバー数である。 Next, the CPU 11 selects a cluster (step S23). First, for each cluster, based on the dissimilarity DistR (C _k , T _p ) obtained in step S21 and the number of members of the cluster, the cluster realization degree Pref (C _k , T _p ) for the prediction target date T _p . Is calculated (step S22). The realization degree Pref (C _k , T _p ) is calculated according to the following mathematical formula (4).

Here, num (S) = num (C ₀ ) + num (C ₁ )... + Num (C ₄ ), the total number of members.

  The CPU 11 selects the cluster with the highest degree of realization. The CPU 11 generates a portion corresponding to the predicted time series of the representative time series of the selected cluster as a predicted time series (step S24), and returns the process to the caller.

Next, prediction time series generation processing based on Nc will be described. FIG. 6 is a flowchart illustrating an example of a processing procedure of predicted time series generation processing based on Nc. First, the CPU 11 acquires a partial time series from the day before the prediction target day to c days before the prediction target day (step S31). In the present embodiment, c = 3. The partial time series time zones are a reference time zone and a predicted time zone. Subsequently, the CPU 11 calculates the dissimilarity between the reference time series and the acquired partial time series (step S32). The calculation of the dissimilarity distP (s _i , T _p ) is performed according to the above equation (2). Therefore, before reaching step S32, for those that have already been calculated and stored in the RAM 12 or the mass storage device 14, the stored values can be reused, so the calculation can be omitted. It is.

ここで、ｓ_iをクラスターの代表時系列、ｓ_jを選択した部分時系列として、算出を行う。 Subsequently, the CPU 11 selects the partial time series having the minimum value among the calculated dissimilarity distP (s _i , T _p ) (step S33). Next, the CPU 11 calculates the dissimilarity between the representative time series and the predicted partial time series for each cluster (step S34). The dissimilarity distP (rep (C _k ), s _ic ) is calculated according to the following formula (5).

Here, the calculation is performed with s _i as the representative time series of the cluster and s _j as the selected partial time series.

  Next, the CPU 11 selects a cluster having the smallest dissimilarity with the partial time series (step S35). Then, the CPU 11 generates a predicted time series from the selected cluster (step S36), and returns the process to the caller. The prediction time series is a portion of the prediction time zone of the representative time series of the selected cluster. As described above, by the processing shown in FIGS. 4 to 6, the power consumption prediction device 1 predicts the prediction time zone for the day (after 10 o'clock) after the elapse of the reference time zone for the prediction target day (after 9 o'clock). The power consumption can be predicted.

  Next, updating of the appearance state information of the time series pattern will be described. The update of the appearance state information of the time-series pattern is a process performed after the prediction time zone of the prediction target day has elapsed. This is because the update of the appearance state requires acquisition of the actual time series corresponding to the predicted time series, and the actual time series cannot be acquired unless the predicted time period has elapsed.

ここで、ｓは予測時系列、ｓ_pは実績時系列、Ｔ_pは予測対象日、ｐｒｅｄは予測時間帯、ｎｕｍ（ｐｒｅｄ）は時間帯ｐｒｅｄの時点数、ｓ［ｔ］は、予測時系列ｓ上の時刻ｔの値を示す。同様にｓ_p［ｔ］は、実績時系列ｓ_p上の時刻ｔの値を示す。 FIG. 7 is a flowchart illustrating an example of a processing procedure of time series pattern appearance state information update processing. The CPU 11 of the power consumption prediction apparatus 1 obtains a prediction time series and an actual time series based on Pf of the prediction target date to be evaluated (step S41). The CPU 11 calculates the obtained prediction time-series prediction results (step S42). The predicted performance value is, for example, a value of MAD (Mean Absolute Deviation average absolute deviation). In this case, the calculation of the predicted score value is performed according to the following formula (6).

Here, s is the prediction time series, s _p is the actual time-series, T _p is the prediction target day, pred prediction time zone, num (pred) is the number of time of time of day pred, s [t] is, prediction time series The value of time t on s is shown. Similarly, s _p [t] indicates the value of time t on the actual time series s _p .

  Next, the CPU 11 determines whether or not the calculated predicted result satisfies the standard (step S43). Here, if the predicted result value is smaller than 0.07, the CPU 11 determines that the predicted result satisfies the standard. If the predicted result value is 0.07 or more, the CPU 11 determines that the predicted result does not satisfy the standard. When the CPU 11 determines that the predicted result satisfies the criterion (YES in step S43), the CPU 11 determines whether or not pf_counter (counter) exceeds 0 (step S44). When pf_counter exceeds 0 (YES in step S44), the CPU 11 subtracts 1 from pf_counter (step S45). The CPU 11 determines whether or not pf_counter is 0 (step S46). If pf_counter is 0 (YES in step S46), the CPU 11 sets the appearance state to normal (step S47). If pf_counter is not 0 (NO in step S46), the CPU 11 ends the process. If it is determined in step S44 that pf_counter is 0 or less (NO in step S44), the CPU 11 ends the process.

  If the CPU 11 determines that the predicted result does not satisfy the standard (NO in step S43), the CPU 11 determines whether or not pf_counter is less than 2 (step S48). If pf_counter is less than 2 (YES in step S48), the CPU 11 adds 1 to pf_counter (step S49). The CPU 11 determines whether pf_counter is 2 (step S50). When pf_counter is 2 (YES in step S50), the CPU 11 sets the appearance state at the time of state change (step S51). If pf_counter is not 2 (NO in step 50), the CPU 11 ends the process. If pf_counter is 2 or more in step S48 (NO in step S48), the CPU 11 ends the process.

  By the process shown in FIG. 7, when the predicted result does not satisfy the standard for two consecutive days, the appearance state is set as the state change. When the predicted result satisfies the standard for two consecutive days, the appearance state is set as normal. The reference value 0.07 is an example. An appropriate value is set for the reference value after an experiment. As for the value of 2 which is a threshold for switching the appearance state, an appropriate value is set after an experiment.

  Next, calculation of a prediction time series when using the data shown in FIGS. 2 and 3 will be described. The calculation conditions here are as follows. The reference time zone is from 6 o'clock to 9 o'clock, and the predicted time zone is from 10 o'clock to 20 o'clock. The data used for the prediction is data on business days from the day before the prediction date to 28 days before, and the number of clusters for clustering is 5. The forecast date is September 27 and 19 business days before 28 days.

FIG. 8 is a table showing the results of cluster selection based on Pf. The number of cluster members is 7 for cluster C ₀ (August 30, August 31, September 3, September 11, September 12, September 13, September 14), and C ₁ 2 (September 24, September 25), C ₂ is 4 (September 7, September 19, September 20, September 21) C ₃ is 5 (September 4, September) 5 days, September 6, September 10, September 18), _{C 4} is 1 (September 26). A reference time series prediction target day (Tp), the dissimilarity between each cluster (Ck) has, in order from the cluster _{C 0,} in 0.2919,0.2367,0.2357,0.3155,0.1913 is there. It is cluster C ₄ that takes the minimum dissimilarity. However, the realization degree Pref (C _k , T _p ) is 0.2414, 0.085, 0.1708, 0.1595, 0.0526 in order from the cluster C ₀ , and the cluster C ₀ takes the largest value. Therefore, cluster _C0 is selected. Here, when the prediction error (MAD) is obtained for the representative time series of each cluster using the actual time series on September 27, 0.1486, 0.0866, 0.1020, _{0 in} order from the cluster C0. .1397,0.0646 next, cluster of the minimum value is found to be cluster _{C 4.} That is, the best cluster should be selected was C _4.

  Next, cluster selection based on Nc will be described. Here, the number of data (c) used to select a cluster is 3. That is, the past three days from the day before the prediction target date are used. FIG. 9 is a graph showing a reference time series and partial time series for the past three days. The horizontal axis is the time, and the vertical axis is the actual value. The reference time series of the prediction target date September 27, the partial time series for the past three days of September 24, September 25, and September 26 are shown. FIG. 10 is a table showing the calculation result of the dissimilarity between the reference time series and each partial time series. The dissimilarity values are 0.2317, 0.2417, and 0.1913 sequentially from September 24, and the minimum value is September 26. Of the partial time series, the partial time series of September 26 is selected.

FIG. 11 is a table showing cluster selection and prediction error by each cluster. The upper part of the table shows the dissimilarity value between the selected partial time series of September 26 and the representative time series of each cluster. In order from the cluster C _0, they are 0.1124, 0.069, 0.068, 0.1065, and 0.0. Therefore, the cluster C ₄ that takes the minimum value 0.0 is selected. The lower part of the table shows the prediction error (MAD) of the representative time series of each cluster using the actual time series of September 27. This is the same as FIG. In accordance with the cluster selection is Nc, it can be seen that the best cluster C ₄ is selected. FIG. 12 is a graph showing representative clusters and actual time series. FIG. 12A shows the representative time series of cluster 0 and the actual time series of September 27. FIG. 12B shows the actual time series of a representative time series of C _4, 9 May 27 days. In both cases, the horizontal axis represents time, and the vertical axis represents normalized power consumption. Be compared with the FIGS. 12A and 12B, the can be seen cluster C ₄ is the best cluster.

  FIG. 13 is a graph showing the results when prediction is performed by switching the cluster selection method. The horizontal axis is the date, and the vertical axis is the MAD value or pf_counter value. The graph 13a shows the prediction result value (MAD) of the prediction time series when the prediction by Pf and the prediction by Nc are switched. The graph 13b shows the predicted performance value (MAD) of the predicted time series by Pf. The graph 13c shows the value of pf_counter. The graph 13d shows the date when the prediction by Nc is executed. As can be seen from FIG. 13, for the day determined to be a state change, the performance is better when the prediction by Nc is used together than the prediction by the conventional Pf.

FIG. 14 is a list of evaluation results of predicted results. The average μ, standard deviation σ, and average + standard deviation (μ + σ) values of the predicted results are evaluated. The three columns on the left side of FIG. 14 are the evaluations for the predictions for all the days, and the three columns on the right side are the evaluations for the predictions of the days determined to be in the state change. The improvement rate is calculated according to the following formula (7).
Improvement rate = (value of only Pf) − (value of Pf−Nc switching) / (value of Pf) (7)

  The smaller the value is, the better the predicted result is, so it can be said that the average value is improved. Since the standard deviation is an index representing variation, a smaller value is better. Similarly, the smaller the average value + standard deviation, the better. From the list shown in FIG. 14, the average value μ, the standard deviation σ, and the average value + standard deviation all decrease. It shows that the prediction using Nc together is more accurate than the prediction using only Pf. In particular, when the appearance state is only when the state changes, as shown in FIG. 14, it is clear that the result is slightly improved as compared with the case of the entire data.

  FIG. 15 is an explanatory diagram illustrating an example of a prediction result output. FIG. 15 shows an example in which the predicted time series for September 27 is displayed. The horizontal axis is the date and time, and the vertical axis is the actual value or the predicted value. As described above, the value on the vertical axis is a ratio value when the maximum power consumption of the day is set to 1 every day, and is dimensionless. The predicted time series is a value obtained by dividing the predicted power consumption value by the average value of the maximum values of each business day until the past one week. A graph 151 is a predicted time series of September 27th. A graph 152 is an actual time series from 0:00 on August 30 to 9:00 on September 27. For the actual time series, it is determined whether the value is in the reference time zone (from 6 o'clock to 9 o'clock), predicted time zone (from 10 o'clock to 20 o'clock), or other time zone (from 21 o'clock to the next 5 o'clock) It can be done. The state change time mark 153 indicates that the appearance state is not the normal time but the state change time. In the example of FIG. 15, it is determined that September 25th, 26th, and 27th are the time of state change. The 25th and 26th days indicate that the appearance state is determined to be a state change time based on the prediction time series based on Pf and the implementation seat time series, and for the 27th day that is the prediction target day, the appearance state is Based on the determination that the state is changing, the prediction time series based on Nc is output. Since the power consumption prediction device 1 displays, as an output, a state change time mark 153 indicating whether or not the state has changed, in addition to the actual time series and the prediction time series, a power generation plan and a power saving plan based on that Can be executed.

  FIG. 16 is a block diagram illustrating a functional configuration example of the power consumption prediction apparatus 1. The power consumption prediction apparatus 1 includes an acquisition unit 11a, a reference acquisition unit 11b, a state acquisition unit 11c, a calculation unit 11d, a specifying unit 11e, a second calculation unit 11f, and a second acquisition unit 11g. When the CPU 11 executes the control program 1P and the like, the power consumption prediction apparatus 1 operates as follows.

  The acquisition unit 11a includes a time series group in which a plurality of days of actual time series for each unit day associated with power consumption and time are classified into a predetermined number of groups using dissimilarity and the time series group. A representative time series calculated from the included actual time series is acquired. The reference acquisition unit 11b acquires a reference time series including power consumption and time in the first time zone on the prediction target unit date. Based on the results of past power consumption prediction, the state acquisition unit 11c acquires appearance state information indicating a change in the power consumption tendency or a normal time in which the change does not appear. When the acquired appearance state is a state change, the calculation unit 11d acquires the acquired reference time series and N pieces of N from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before The dissimilarity with each actual time series is calculated. The specifying unit 11e specifies one actual time series having the smallest calculated dissimilarity. The second calculation unit 11f calculates the dissimilarity between the identified actual time series and each of the representative time series. The second acquisition unit 11g selects a representative time series having the smallest calculated dissimilarity, and selects a portion corresponding to a second time zone subsequent to the first time zone of the selected representative time series as the prediction Obtained as a predicted time series in which the power consumption of the target unit day is predicted.

The technical features (components) described in each embodiment can be combined with each other, and new technical features can be formed by combining them.
The embodiments disclosed herein are illustrative in all respects and should not be considered as restrictive. The scope of the present invention is defined not by the above-mentioned meaning but by the scope of the claims, and is intended to include all modifications within the meaning and scope equivalent to the scope of the claims.

  Regarding the above embodiment, the following additional notes are disclosed.

(Appendix 1)
A power consumption prediction method for predicting power consumption by a computer,
24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each unit day associated with power consumption and time are classified using a dissimilarity into a predetermined number of groups, and actual times included in the time series group Obtain the representative time series calculated from the series,
Obtain a reference time series including power consumption and time in the first time zone on the prediction target unit date,
Based on the results of past power consumption prediction, obtain the appearance state information set to either the state change when the trend of power consumption appears or the normal time when the change does not appear,
When the acquired appearance state information is when the state changes, the acquired reference time series and N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before And calculate dissimilarity with
Identify the actual time series of 1 with the smallest calculated dissimilarity,
Calculate the dissimilarity between the identified time series and each of the representative time series,
Select the representative time series with the smallest calculated dissimilarity,
Obtaining a portion corresponding to a second time zone following the first time zone of the selected representative time series as a predicted time series in which power consumption of the prediction target unit day is predicted;
A power consumption prediction method that outputs the obtained prediction time series.

(Appendix 2)
The power consumption prediction method according to claim 1, wherein the appearance state information is output together with the obtained prediction time series when the state changes.

(Appendix 3)
Calculating the dissimilarity between the reference time series and the actual time series included in each of the time series groups;
Based on the calculated dissimilarity, the realization degree indicating the high possibility of appearing as the actual time series of the prediction target unit date is calculated for each time series group,
Select the time series group with the highest degree of realization calculated,
Storing a portion corresponding to the second time zone of the representative time series of the selected time series group as a normal prediction time series;
The power consumption prediction method according to attachment 1, wherein when the appearance state is normal time, the obtained time series until the day before the prediction target unit date, the reference time series on the prediction target unit date, and the normal prediction time series are output. .

(Appendix 4)
Obtain the latest normal prediction time series, the actual time series corresponding to the normal prediction time series,
Based on the acquired actual time series, calculate an evaluation value that is inversely proportional to the forecast performance of the latest normal forecast time series,
Get the count value of the counter that shows the deterioration of the predicted performance,
When the calculated evaluation value is less than a predetermined reference value and the acquired count value exceeds 0, the count value is decremented by 1. When the count value is 0 after the subtraction, the appearance state information is set at normal time. And
When the calculated evaluation value is less than or equal to the reference value and the acquired count value is less than a predetermined upper limit value, the count value is incremented by 1, and when the added count value is equal to the upper limit value, the appearance state information The power consumption prediction method according to appendix 3, wherein: is set when the state changes.

(Appendix 5)
The N is 3. The power consumption prediction method according to any one of supplementary notes 1 to 4.

(Appendix 6)
The evaluation value of the prediction time series is an average absolute deviation between the prediction time series and the actual time series,
The power consumption prediction method according to appendix 4 or 5, wherein the reference value is 0.07.

(Appendix 7)
The upper limit value is 2. The power consumption prediction method according to any one of supplementary notes 4 to 6.

(Appendix 8)
24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each day with which power consumption and time are associated are classified into a predetermined number of groups using dissimilarity and the actual results included in the time series group Obtain the representative time series calculated from the time series,
Obtain a reference time series including power consumption and time in the first time zone on the prediction target unit date,
Based on the results of past power consumption prediction, obtain the appearance state information set in either the state change when the change in power consumption trend appears or the normal time when the change does not appear,
When the acquired appearance state information is when the state changes, the acquired reference time series and N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before And calculate dissimilarity with
Identify the actual time series of 1 with the smallest calculated dissimilarity,
Calculate the dissimilarity between the identified time series and each of the representative time series,
Select the representative time series with the smallest calculated dissimilarity,
Power consumption that causes a computer to execute a process of acquiring a portion corresponding to a second time zone following the first time zone of the selected representative time series as a predicted time series in which the power consumption of the prediction target unit day is predicted Prediction program.

(Appendix 9)
24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each unit day associated with power consumption and time are classified into a predetermined number of groups using dissimilarity, and an actual time series included in the time series group An acquisition unit for acquiring a representative time series calculated from
A reference acquisition unit that acquires a reference time series including power consumption and time in the first time zone on the prediction target unit date;
Based on the results of past power consumption prediction, a state acquisition unit that acquires appearance state information indicating a change in power consumption trend or a normal time when the change does not appear;
When the acquired appearance state is a state change time, the acquired reference time series, and each of N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before A calculation unit for calculating the dissimilarity of
A specifying unit for specifying the actual time series of 1 with the calculated dissimilarity being the minimum,
A second calculation unit for calculating a dissimilarity between the identified actual time series and each of the representative time series; and selecting a representative time series having the smallest calculated dissimilarity, and selecting the first of the selected representative time series A power consumption prediction apparatus comprising: a second acquisition unit that acquires a portion corresponding to a second time zone subsequent to the time zone as a prediction time series in which the power consumption of the prediction target unit day is predicted.

DESCRIPTION OF SYMBOLS 1 Power consumption prediction apparatus 1a Portable storage medium 1b Semiconductor memory 11 CPU
12 RAM
13 ROM
14 Mass storage device 141 Power consumption time series DB
142 Forecast Time Series DB
143 Results / Status DB
15 Communication unit 16 Reading unit

Claims (5)

A power consumption prediction method for predicting power consumption by a computer,
24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each unit day associated with power consumption and time are classified using a dissimilarity into a predetermined number of groups, and actual times included in the time series group Obtain the representative time series calculated from the series,
Obtain a reference time series including power consumption and time in the first time zone on the prediction target unit date,
Based on the results of past power consumption prediction, obtain the appearance state information set to either the state change when the trend of power consumption appears or the normal time when the change does not appear,
When the acquired appearance state information is when the state changes, the acquired reference time series and N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before And calculate dissimilarity with
Identify the actual time series of 1 with the smallest calculated dissimilarity,
Calculate the dissimilarity between the identified time series and each of the representative time series,
Select the representative time series with the smallest calculated dissimilarity,
Obtaining a portion corresponding to a second time zone following the first time zone of the selected representative time series as a predicted time series in which power consumption of the prediction target unit day is predicted;
A power consumption prediction method that outputs the obtained prediction time series. The power consumption prediction method according to claim 1, wherein the appearance state information is a state change time and is output together with the acquired prediction time series. Calculating the dissimilarity between the reference time series and the actual time series included in each of the time series groups;
Based on the calculated dissimilarity, the realization degree indicating the high possibility of appearing as the actual time series of the prediction target unit date is calculated for each time series group,
Select the time series group with the highest degree of realization calculated,
Storing a portion corresponding to the second time zone of the representative time series of the selected time series group as a normal prediction time series;
The power consumption prediction according to claim 1, wherein when the appearance state is normal time, an actual time series up to the day before the prediction target unit date acquired, a reference time series on the prediction target unit date, and a normal prediction time series are output. Method. 24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each day with which power consumption and time are associated are classified into a predetermined number of groups using dissimilarity and the actual results included in the time series group Obtain the representative time series calculated from the time series,
Obtain a reference time series including power consumption and time in the first time zone on the prediction target unit date,
Based on the results of past power consumption prediction, obtain the appearance state information set in either the state change when the change in power consumption trend appears or the normal time when the change does not appear,
When the acquired appearance state information is when the state changes, the acquired reference time series and N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before And calculate dissimilarity with
Identify the actual time series of 1 with the smallest calculated dissimilarity,
Calculate the dissimilarity between the identified time series and each of the representative time series,
Select the representative time series with the smallest calculated dissimilarity,
Power consumption that causes a computer to execute a process of acquiring a portion corresponding to a second time zone following the first time zone of the selected representative time series as a predicted time series in which the power consumption of the prediction target unit day is predicted Prediction program. 24 hours starting from a given time is a day unit,
A time series group in which a plurality of days of actual time series for each unit day associated with power consumption and time are classified into a predetermined number of groups using dissimilarity, and an actual time series included in the time series group An acquisition unit for acquiring a representative time series calculated from
A reference acquisition unit that acquires a reference time series including power consumption and time in the first time zone on the prediction target unit date;
Based on the results of past power consumption prediction, a state acquisition unit that acquires appearance state information indicating a change in power consumption trend or a normal time when the change does not appear;
When the acquired appearance state is a state change time, the acquired reference time series, and each of N actual time series from the prediction target unit day N days before (N is a natural number of N> 2) to the prediction target unit day before A calculation unit for calculating the dissimilarity of
A specifying unit for specifying the actual time series of 1 with the calculated dissimilarity being the minimum,
A second calculation unit for calculating a dissimilarity between the identified actual time series and each of the representative time series; and selecting a representative time series having the smallest calculated dissimilarity, and selecting the first of the selected representative time series A power consumption prediction apparatus comprising: a second acquisition unit that acquires a portion corresponding to a second time zone subsequent to the time zone as a prediction time series in which the power consumption of the prediction target unit day is predicted.

Images