WO2016136323A1

WO2016136323A1 - Energy predict system, energy predict method, computer program for causing execution thereof, recording medium whereupon said program is recorded, and operation assistance system

Info

Publication number: WO2016136323A1
Application number: PCT/JP2016/051186
Authority: WO
Inventors: 彰彦小川; 郷志清水; 聰子藏田
Original assignee: 株式会社Ｅ．Ｉ．エンジニアリング
Priority date: 2015-02-27
Filing date: 2016-01-15
Publication date: 2016-09-01
Also published as: JP6118975B2; JPWO2016136323A1

Abstract

The purpose of the present invention is to provide an energy predict system, energy predict method, computer program for causing execution thereof, recording medium whereupon said program is recorded, and operating assistance system, which are capable of computing high-precision prediction values in line with actual measured values on a given day. Provided is an energy predict system, comprising: an analysis condition setting unit 40 which sets a comparison time period and a number of comparison days whereupon a clustering analysis is carried out, and selects a comparison subject from load patterns which have been classified into categories; a time series data setting unit 50 which sets time series data of a given day and past time series data by actual measured values which have been classified into the selected load pattern; and a prediction value computation unit 60 which selects a plurality of instances of the past time series data which has a high degree of similarity, performs weighted summation according to the degree of similarity on the actual measured values from a prescribed time of the selected past time series data to a time which is a prescribed predicted amount of time in advance of the prescribed time, and computes a prediction value of an energy load from the prescribed time to a prediction time which is the predicted amount of time in advance on the given day.

Description

ENERGY PREDICTION SYSTEM, ENERGY PREDICTION METHOD, COMPUTER PROGRAM FOR EXECUTING THE SAME, RECORDING MEDIUM CONTAINING THE PROGRAM, AND OPERATION SUPPORT SYSTEM

The present invention relates to an energy prediction system, an energy prediction method, a computer program for executing the energy prediction method, a recording medium storing the program, and an operation support system. More specifically, an energy prediction system that predicts an energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water, an energy prediction method, a computer program that executes this, and a computer program The present invention relates to a recording medium on which a program is recorded and a driving support system.

Various energy loads such as heat and electric power for the predicted day are intricately intertwined with various factors such as the weather conditions of the day, the usage status of facilities and buildings (number of people, presence of events, etc.), and the operating status of installed equipment. to be influenced. In the conventional method of predicting a load using a multiple regression equation, a Kalman filter, or the like, it is theoretically possible to improve the accuracy of a predicted value by setting various explanatory variables. However, it is practically difficult to set an explanatory variable that covers (considerates) every situation, and there is a limit to improving the prediction accuracy.

As other energy load prediction methods, for example, those described in

Patent Documents

1 and 2 are known. In the method of Patent Document 1, the predicted load calculated using the ARMA model formula is corrected based on the external environment prediction information (the predicted maximum outside temperature on the next day). However, other factors than the outside temperature are not taken into consideration, and further improvement in prediction accuracy has been desired.

Moreover, in the method of patent document 2, the similarity between the past weather actual data and the weather forecast data on the prediction target day is calculated, and the power demand is predicted from the power demand data corresponding to the past weather actual data. . However, since it depends on weather data, it may be affected by sudden weather fluctuations, and further improvement in prediction accuracy has been desired.

JP-A-5-18565 JP 2013-66318 A

In view of such a conventional situation, the present invention provides an energy prediction system, an energy prediction method, a computer program for executing the same, and a computer program for executing the energy prediction system capable of calculating a highly accurate predicted value corresponding to the actual measurement value of the day. An object of the present invention is to provide a recording medium that records the above and a driving support system.

In order to achieve the above object, the energy prediction system according to the present invention is characterized in that the energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water is predicted. Performs cluster analysis with an actual value storage unit that stores the actual value of the energy load up to the present time for each measurement time, a load pattern classification unit that classifies the actual value into a load pattern based on the characteristics of the energy load every day An analysis condition setting unit that sets the comparison time and the number of comparison days and selects a comparison target for performing the cluster analysis from the classified load pattern, and the current date and time series data from the predetermined time on the predicted day to the set comparison time before, , The past time series for each day of the comparison days set in the past in the same time zone as the current time series data A time-series data setting unit that sets data based on measured values classified into the selected load pattern, and calculates a similarity to the current date / time series data for each of the past time-series data by the cluster analysis. A plurality of past time-series data having a high value are selected, and measured values of the selected past time-series data at a time ahead of a predetermined prediction time from the predetermined time are weighted and added according to the degree of similarity, and the predetermined time on the prediction day To a predicted value calculation unit that calculates a predicted value of the energy load at the predicted time ahead of the predicted time.
The cluster analysis used in the present invention is an evaluation of the similarity between clusters of different days, with the data extracted from the predetermined time on the prediction day to the comparison time as one cluster from the time series data of energy load. It is a technique to do.

According to the above characteristics, the time series data to be subjected to cluster analysis is set based on the actual measured value of energy load up to the present time. Since the actual measurement values are influenced by the environment and situation at the time of measurement, the time-series data set by the actual measurement values reflect all conditions (elements), and can be realized without complicated settings. It is possible to make predictions according to. In addition, since the measured values are classified into load patterns based on the characteristics of the selected energy load, the set time series data is based on the measurement results in a common or similar environment, They will be similar to each other. Therefore, the accuracy of cluster analysis is further improved. Since the predicted value is calculated by weighting and adding according to the calculated similarity, the accuracy is further improved.

The predicted value is calculated by adding a weighted average of the difference between the measured value at the predetermined time and the measured value at the predicted time of the selected past time series data to the measured value at the predetermined time on the prediction day. Good. Since the weighted average of the difference between the actual measurement value at the predetermined time on the prediction day and the actual measurement value at the prediction time is added to the actual measurement value at the predetermined time on the prediction day, a more accurate predicted value that is more realistic.

The predicted value calculation unit may select a plurality of past time series data in descending order of the similarity. Since the past time series data is selected in descending order of similarity, the accuracy is further improved.

The load pattern may be a combination of at least one selected from a group including at least all days, holidays, weekdays, special days, production plans, and weather information. Thereby, prediction according to the actual situation can be performed, and a highly accurate prediction value according to reality can be obtained.

The predicted value calculation unit weights and adds an actual measurement value of the selected past time-series data at a time that is a predetermined second predicted time ahead from the predetermined time according to the similarity, and from the predetermined time on the prediction day A temporary actual measurement value setting unit that sets a predicted value of the energy load at the next second prediction time ahead of the second prediction time as a temporary actual measurement value of the time, and the time-series data setting unit includes the prediction day New current date and time series data from the next second predicted time to the comparison time before, and past time series data for each day of the set comparison days in the same time zone as the new current date and time series data. A time-series data updating unit that updates the measured value classified into the selected load pattern and the temporary measured value, and the temporary measured value setting unit sets the temporary measured value for each second predicted time. The time series data update unit updates the time series data for each second prediction time, and the prediction value calculation unit repeatedly performs the cluster analysis for each second prediction time, The energy load from the predetermined time to the long-term predicted time ahead may be predicted. Since the predicted value calculated for each second prediction time is set as a temporary actual measurement value at that time and the time series data is updated, the cluster analysis is repeated every second prediction time until the long-term prediction time set The energy load can be predicted. For example, it is possible to predict an energy load until an arbitrary time elapses, such as 24 hours ahead or 48 hours ahead on the prediction day.

In this case, it is desirable that the temporary actual measurement value is replaced with the actual measurement value every time the actual measurement value at the next second prediction time on the prediction day is stored. As a result, the temporary actual measurement value is always replaced with the latest actual measurement value, so that the prediction accuracy is improved.

The load pattern classification unit may further include a long-term load pattern set in units of days of at least two consecutive days. As a result, changes that occur over a long period of time can be taken into consideration, and the prediction is more realistic, and the prediction accuracy can be further improved.

The energy facility is a complex facility or region including a plurality of facilities having a plurality of power generation devices and / or thermoelectric devices, and the prediction calculation unit calculates the prediction value for each facility and adds the prediction values. Then, the predicted value of the energy load of the complex facility or the entire area may be calculated. This not only predicts the energy load at a single facility with multiple generators and / or thermoelectric devices, but also predicts the overall energy load of a complex facility, apartment house, or area where multiple such facilities are assembled. It becomes possible and contributes to energy saving.

The cluster analysis may be, for example, an intergroup average distance method. In this case, the similarity may be a reciprocal of a distance of the past time series data with respect to the current date / time series data. The similarity may be a cosine of an angle between vectors of the past time series data and the current date and time series data.

In order to achieve the above object, the energy prediction method according to the present invention is characterized by an energy prediction method for predicting an energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water. The actual measured value of the energy load up to the present time is stored for each measurement time, the measured value is classified into load patterns based on the characteristics of the energy load for each day, and the comparison time and the number of comparison days for performing the cluster analysis are set. The comparison target for performing the cluster analysis is selected from the classified load patterns, and the current date / time series data from the predetermined time on the predicted day to the set comparison time and the past in the same time zone as the current date / time series data. The past time-series data of the set comparison days for each day was classified into the selected load pattern Set by measurement, calculate a similarity to the current time series data for each of the past time series data by the cluster analysis, and select a plurality of past time series data having a high similarity, and the selected past time series data Calculating a predicted value of the energy load at the predicted time ahead of the predicted time from the predetermined time on the prediction day by weighting and adding measured values at a predetermined time ahead of the predetermined time according to the similarity. It is in.

The energy prediction system described in any of the above is realized by a computer program for executing the system, and this computer program is recorded on a recording medium.

In order to achieve the above object, the operation support system according to the present invention is characterized in that, in a configuration that supports the operation of energy equipment that uses or manufactures energy such as electric power, cooling, heating, steam, hot water supply, etc. A measured value storage unit that stores measured values of energy load at each measurement time, a load pattern classification unit that classifies the measured values into load patterns based on the characteristics of the energy load every day, comparison time and comparison for performing cluster analysis An analysis condition setting unit that sets the number of days and selects a comparison target to be subjected to the cluster analysis from the classified load patterns, current date / time series data from a predetermined time on the predicted date to the set comparison time, and the current date / time Load pattern that selects past time series data for each day of the set comparison days in the same time zone as the series data A time-series data setting unit that is set based on classified actual measurement values, and a plurality of past time-series data having a high degree of similarity are calculated by calculating a similarity to the current time-series data for each of the past time-series data by the cluster analysis. The measured value of the selected past time-series data at the predetermined time ahead from the predetermined time is weighted according to the degree of similarity, and the predicted time at the prediction time ahead from the predetermined time on the prediction day A predicted value calculation unit for calculating a predicted value of the energy load, and supporting operation of the energy equipment based on the predicted value.

According to the features of the energy prediction system, the energy prediction method, the computer program for executing the program, the recording medium storing the program, and the driving support system according to the present invention, the prediction with high accuracy according to the actual measurement value of the day The value can be calculated.

Other objects, configurations, and effects of the present invention will be apparent from the following embodiments of the present invention.

It is a figure which shows the structure of the hardware of the energy prediction system which concerns on this invention. It is a figure which shows the structure of the software of the energy prediction system which concerns on this invention. It is a schematic diagram explaining the definition of the distance in the group average distance method, (a) is the current date and time series data, (b) is the past time series data of m days ago, (c) is the current date and time series data and m days ago. The distance dm with the past time series data is typically shown. It is a flowchart which shows a short-time prediction procedure. It is a flowchart which shows a long-term prediction procedure. It is a figure which shows an example of a comparison with long-term prediction and measured value (experience), (a) is a time of 5:30, (b) is a time of 7 o'clock, (c) is a time of 13:00, (d) is a time At 16:00, (e) shows an example of a change in the prediction diagram of the long-term prediction. It is a graph explaining the driving | operation assistance of an energy installation, (a) is the electric power demand prediction of the previous day, (b) is the electric power supply-and-demand balance prediction of the previous day, (c) is an example of adjustment of the electric power generation amount by the short-time prediction of the day Indicates. It is a figure which shows typically the angle between vectors showing a similarity instead of distance, (a) is present date series data, (b) is past time series data of (a) and similar shape, (c) is ( An angle between vectors between time series data of a) and (b) is shown typically. FIG. 8 is a diagram corresponding to FIG. 7 illustrating another example of an inter-vector angle representing a similarity instead of a distance.

Next, the present invention will be described in more detail with reference to the accompanying drawings as appropriate.
The energy prediction system 1 according to the present invention predicts an energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water. Examples of energy facilities include power generation equipment such as cogeneration, solar power generation, boiler equipment such as high pressure boilers, cold water equipment such as absorption refrigerators, hot water equipment such as hot water boilers, and hot water supply equipment such as hot water boilers. These are various power generation devices and thermoelectric devices such as buildings, factories, various stores and facilities including these devices, and complex facilities and regions including a plurality of these. Examples of the energy load include demand (consumption, usage) and production (supply) such as electric power, cold water, hot water, and steam.

As shown in FIG. 1, the hardware of the energy prediction system 1 according to the present invention generally includes a user interface 2 and a processing unit 3 that processes the software 10 of the energy prediction system 1. The user interface 2 includes a monitor 2a, a keyboard 2b, and a mouse 2c, and is used by the user to operate buttons and input fields on the screen displayed on the monitor 2a. The processing unit 3 is connected to the CPU 3a, the temporary storage memory 3b, the HDD 3c, and the like through a bus 3d such as a data bus and an address bus. The CPU 3a, the temporary storage memory 3b, the HDD 3c, and the like operate the software 10 in cooperation.

As shown in FIG. 2, the software 10 of the energy prediction system 1 generally includes an actual measurement value storage unit 20, a load pattern classification unit 30, an analysis condition setting unit 40, a time series data setting unit 50, a predicted value calculation unit 60, An output unit 70 and a recording unit 80 are included.

Measured value storage unit 20 stores measured values of energy load up to the present time for each measurement time. The measured value storage unit 20 receives measured values from various sensors of energy equipment connected to the energy prediction system 1. The measurement time is a cycle for measuring the energy load, and any unit of second (second), minute (minute), hour (hour) such as 30 seconds, 10 minutes, 1 hour or the like may be used. Although the prediction accuracy is improved by shortening the measurement time (cycle), the time (number of times) required for the calculation is increased, so it may be set as appropriate according to the required accuracy. It is also possible to capture meteorological information described later together with the actual measurement values.

The load pattern classification unit 30 classifies the actual measurement values stored in the actual measurement value storage unit 20 into load patterns based on the characteristics of the energy load every day. The load pattern is at least one combination selected from the group including at least all days, holidays, weekdays, singular days, production plans, and weather information. The special day indicates a day (environment) that is different from a normal business day such as holding an event in a commercial facility or a special sale day. The production plan indicates the number of machines operating in a factory and the production schedule of products to be produced. Meteorological information refers to, for example, various weathers published by the Japan Meteorological Agency, such as air temperature for a given time, daily average temperature, maximum temperature, minimum temperature, average relative humidity, maximum humidity, weather, atmospheric pressure, wind speed, precipitation, etc. Data and seasons. In this way, by classifying by characteristics that affect the pattern of energy load, unique differences between time series data to be compared are reduced, and similarity calculation described later is performed on time series data in similar environments. And the prediction accuracy can be improved.

The analysis condition setting unit 40 includes a comparison time setting unit 41 that sets a comparison time for performing cluster analysis, a comparison day setting unit 42 for setting comparison days for performing cluster analysis, and a load pattern classification unit for comparing comparison targets for performing cluster analysis. The load pattern selection unit 43 selects the load pattern classified by 30. The comparison time indicates a time zone (a period from the start time to the end time) of the time series data to be subjected to cluster analysis, and can be appropriately set regardless of the unit of time as described above. The comparison days indicate the number of later-described past time series data to be compared (similarity calculation) with the current date / time series data, and preferably two or more. Although the prediction accuracy is improved by increasing the number of comparison days, the calculation time (number of times) is increased, so that it may be set appropriately according to the required accuracy. In the present embodiment, there is an upper selection day setting unit 44 for setting the number of days (number) for selecting past time-series data having a high degree of similarity by cluster analysis described later.

The analysis condition setting unit 40 also sets a calculation time and a prediction time for calculating the similarity of the past time series data with respect to the current date and time series data. These times can be set as appropriate regardless of the unit of time as described above. The calculation time is a cycle for calculating the degree of similarity, and is matched with the above-described measurement time, for example. The prediction time is a predetermined time from a predetermined time on the prediction day to a predetermined time (prediction time) to be predicted. Furthermore, in this embodiment, the analysis condition setting unit 40 also sets the second predicted time. This second prediction time is a predetermined time from a predetermined time on the prediction day in the long-term prediction described later to a predetermined time to be predicted (next second predicted time). The predicted time and the second predicted time can be arbitrarily set and may be the same, for example, the same as the above-described measurement time.

The time series data setting unit 50 is set by the current date / time series data from the predetermined time on the predicted day to the comparison time set by the comparison time setting unit 41 and the comparison day setting unit 42 in the same time zone as the current date / time series data. The past time-series data for each past day of the comparison days thus set is set by the actually measured values classified into the load patterns selected by the load pattern selection unit 43.

Here, the current date / time series data from the time point t (predetermined time) to the time point N before the prediction date is expressed by the following formula 1. N is a number obtained by dividing the comparison time by the measurement time.

In addition, the past time series data m days before the forecast date in the same time zone as the current date and time series data is expressed by the following formula 2.

Further, the time series data setting unit 50 includes new current date / time series data from the next second prediction time that is the second prediction time ahead of the predetermined time on the prediction day to the comparison time set by the comparison time setting unit 41, The current time-series data and the past time-series data for each day of the comparison days set by the past comparison days setting unit 42 in the same time zone are classified into the load patterns selected by the load pattern selection unit 43. A time-series data update unit 51 that updates the measured values and the temporary measured values described later is included.

The predicted value calculation unit 60 includes a similarity calculation unit 61 that calculates the similarity, a past time series data selection unit 62 that selects past time series data having a high similarity, and a weight that calculates a weighting factor according to the similarity. A coefficient calculation unit 63 is provided. The similarity calculation unit 61 calculates the similarity to the current date / time series data for each past time series data created by the time series data setting unit 50 by cluster analysis. The past time series data selection unit 62 selects a plurality of past time series data according to the set selection conditions. For example, in the present embodiment, the past time series data for the number of days (number) set by the higher selection day setting unit 44 is selected from the set past time series data in descending order of similarity. The weighting factor calculation unit 63 calculates a weighting factor according to the similarity of the past time series data selected by the past time series data selection unit 62. Then, the predicted value calculation unit 60 weights the actually measured value at the time ahead of the predetermined prediction time from the predetermined time of the past time series data selected by the past time series data selection unit 62 with the weight coefficient calculated by the weight coefficient calculation unit 63. It is added and calculated as the predicted value of the energy load at the prediction time ahead of the predetermined prediction time from the predetermined time on the prediction day (short time prediction).

Here, in this embodiment, an intergroup average distance method is used for cluster analysis. This average distance method between groups calculates distances for all individual pairs (measured values for each calculation time) of the current time series data (cluster) and past time series data (cluster) and calculates the average as the similarity. To do. The greater the similarity (the reciprocal of the distance), the more similar the current date / time series data and the past time series data.
In the intergroup average distance method, the current date / time series data, the time series data m (m ≦ M) days ago, and the distance are defined by the following equation (3). M is the number of comparison days.

The current date / time series data (first cluster) P shown in FIG. 3A and the past time series data (second cluster) Q m days before the forecast date shown in FIG. ), It can be understood as points p and q arranged in a space formed by t, t−1, and t−2. In the group average distance method, the distance dm between the points p and q is calculated, and the distance dm that is short (small) has high similarity. The similarity calculation unit 61 calculates the distance dm for each past time series data, and the past time series data selection unit 62 selects a plurality of past time series data in descending order of the reciprocal of the distance dm (high similarity).

The weight coefficient calculation unit 63 calculates the weight coefficient wm from the reciprocal (similarity) of the distance dm of the past time series data selected by the past time series data selection unit 62. The weighting coefficient wm is the distance dm obtained from the selected past time-series data for the number of days (number, S) set by the upper selection days setting unit 44 as the reciprocal (similarity) of the distance dm, as shown in the following equation 4. It is the number (relative weight) divided by the sum of the reciprocals.

Then, the predicted value xt + 1 of the predicted time t + 1 that is a predetermined time (predicted time 1) from the time point t (predetermined time) is calculated by the following equation 5. In the present embodiment, the weighted average of the differences between the time points t + 1 and t in the selected past time series data for the number of days (number, S) set by the upper selection day setting unit 44 is the actual measurement value at the time point t on the prediction day. Add to xt to get the predicted value xt + 1.

Further, the predicted value calculation unit 60 weights and adds the actual measured values at the predetermined second predicted time ahead from the predetermined time of the selected past time-series data according to the similarity, and performs the second prediction from the predetermined time on the predicted day. A temporary actual measurement value setting unit 64 is provided that sets the predicted value of the energy load at the next second predicted time ahead as the temporary actual measurement value at that time. By repeating the same process as the short-time prediction described above every second prediction time of the actual measurement value, it is possible to perform a long-term prediction time (for example, prediction for 24 hours or 48 hours ahead (long-term prediction)). . The temporary actual measurement value is replaced with the actual measurement value every time the actual measurement value at the next second prediction time on the prediction day is stored.

The output unit 70 outputs, for example, a prediction value or a prediction graph obtained by the prediction value calculation unit 60 to a monitor or paper. For example, the measured value and the predicted value thereafter are distinguished from each other and displayed on the graph. In addition to changing the color of the long-term prediction graph, the average square error ratio (EEP) between the predicted value and the actually measured value is displayed. It is also possible to output these in the form of a table as a form. The recording unit 80 records various data such as the calculated predicted value and the output form.

Next, a procedure for short-term prediction of the energy load on the prediction day will be described with reference to FIG. 4a.
As shown in FIG. 4a, the short-term prediction procedure sets conditions for the day prediction by cluster analysis (S1), and calculates the similarity of past time series data by cluster analysis based on the set conditions (S2). A plurality of past time series data are selected in descending order of similarity (S3), a weighting factor is calculated, a predicted value ahead of a predetermined time is calculated using the weighting factor (S4), and the result is output ( S5). The actual measurement values are stored for each measurement time by the actual measurement value storage unit 20 and are classified into load patterns based on energy load characteristics by the load pattern classification unit 30.

In the condition setting step (S1), the comparison time is set by the comparison time setting unit 41, and the comparison days are set by the comparison day setting unit. The analysis condition setting unit 40 also sets calculation time and prediction time. For example, the current time (predetermined time) is set as 5:30, the comparison time is 4 hours, the comparison days are 30 days, the calculation time is 10 minutes, and the prediction time is 30 minutes. Further, the upper selection days setting unit 44 also sets the upper selection days, and is, for example, the top six (pieces).

Furthermore, the load pattern classified by the load pattern classification unit 30 is selected by the load pattern selection unit 43. For the selection of the load pattern, the same load pattern as that of the predicted day is selected. As a result, past actual measurement values in an environment similar to the prediction day can be set as a cluster analysis target, and the accuracy is further improved. Then, according to the setting conditions of the analysis condition setting unit 40, the time-series data setting unit 50 sets a plurality of current time-series data and past time-series data in the same time zone as the current date-and-time series data for each day based on actually measured values. For example, weekday data is selected as the same load pattern as the current date and time series data on the forecasted day of the weekday, and past time series data is created for each day in the same time zone (predetermined time from the specified time) for the latest 30 weekdays. Is done.

In the similarity calculation step (S2), the distance (similarity) for the current date / time series data is calculated for each past time series data set by the above-mentioned average distance method between groups by the similarity calculation unit 61. Then, in the past time series data selection step (S3), the past time series data selection unit 62 selects the top six places with the short distance (high similarity). For example, the distance to the current date / time series data of the predicted current day is calculated for each past time series data of 30 minutes on the most recent weekday, and the top six shortest distances dm (similarity 1 / dm is large) are selected as similar days. The

Next, in the predicted value calculation step (S4), the weighting factor of the similar day selected by the weighting factor calculation unit 63 is calculated, and the predicted time ahead of the prediction time in the past time-series data selected by the predicted value calculation unit 60 is set to a predetermined value. The predicted value is calculated by adding the weighted average of the difference between the actual measurement value and the time to the actual measurement value at the predetermined time on the prediction day. For example, in the past time series data of the selected similar date, a value obtained by multiplying the difference between the measured values of the predicted time ahead of the predicted time and the predetermined time (current time) by the weighting factor wm of the day is obtained. These are calculated in the top six places to obtain the total, and the total value is added to the actual measurement value at the predetermined time on the prediction day to obtain the predicted value of the predicted time.

Here, the long-term prediction shown in FIG. 4b can be performed simultaneously with the short-time prediction described above or independently.
As shown in the figure, the all-day prediction (until 24:00 (PM 12:00) on the prediction day) as the long-term prediction of the energy load on the prediction day, first, sets the conditions for the day prediction by cluster analysis (S11), A temporary actual measurement value at the next second predicted time is set (S12), the time series data is updated based on the set condition (S13), and the similarity of the past time series data updated by the cluster analysis is calculated ( S14), a plurality of past time-series data are selected in descending order of similarity (S15), a weighting factor is calculated, and the predicted value of the measurement time ahead is calculated using the weighting factor (S16). The above steps (S12 to S16) are repeated until the set long-term predicted time ahead (S17). If the prediction of the long-term prediction time (until 24:00 (PM 12:00) on the prediction day) is completed, the result is output (S18). The condition setting step (S11) is the same as the previous condition setting step (S1).

In the provisional actual measurement value setting step (S12), the provisional actual measurement value setting unit 64 calculates the predicted value of the next second prediction time in the second prediction time ahead from the predetermined time as the provisional actual measurement value. Calculation of the predicted value of the next second predicted time is the same as in the short-time prediction step (S2 to S4). For example, the predicted value of the next second predicted time 10 minutes (second predicted time, calculation time) ahead of a predetermined time (current time) is calculated as a temporary measured value. The calculation of the temporary actual measurement value is the same process as the short-time prediction described above.

Next, in the time series data update step (S13), the time series data update unit 51 changes the time zone of the previous date and time series data and the past time series data. For example, assuming that the next second predicted time at which the temporary actual measurement value was calculated in the previous step (S12) is a new predetermined time, the actual measurement value and the temporary actual measurement value in the time zone from that time to the comparison time before Update series data. Since there is no actual actual measurement value at the next second predicted time in the current date / time series data, the temporary date / time data described above is used as the current date / time series data. Then, similarly to the short-time prediction, a predicted value of a new next second predicted time ahead of the second predicted time (calculation time) is calculated. That is, in the short-term prediction, a prediction of a predetermined prediction time ahead (30 minutes in the above example) is performed, but in the long-term prediction, the prediction of the second prediction time ahead is performed every second prediction time (in this example, 10 minutes). The process is repeated until the long-term prediction time has elapsed (in this example, 24:00 (PM 12:00) on the prediction day). In this way, it is possible to predict the transition (fluctuation) of the energy load over time.

Here, when the second predicted time has elapsed, the actual measurement value of the energy load is stored in the actual measurement value storage unit 20. Therefore, the temporary actual measurement value corresponding to each time the actual measurement value is stored is replaced with the actual actual measurement value. And by performing the said process, prediction will be corrected (updated) for every 2nd prediction time progress, and it will become highly accurate prediction according to reality. For example, as shown in FIG. 5, the prediction line of the energy load is corrected by replacing the temporary actual measurement value with the actual actual measurement value as time elapses, so that highly accurate prediction is possible.

Furthermore, it is also possible to support the operation of the energy equipment using the energy load prediction (short-term prediction and / or long-term prediction) on the prediction day calculated as described above.
For example, as shown in FIG. 6 (a), for example, the power load for the previous day of a complex facility consisting of a store, a factory, a bank, and an office is predicted for each facility for its demand amount A1 to 4, and the next day of the entire complex facility. Power demand A is predicted. For example, a prediction method using a conventionally known Kalman filter can be applied to the previous day prediction.

Then, as shown in FIG. 6B, the power supply / demand balance for this power demand prediction A is predicted. Regarding the power generation amount Eb5 of renewable energy, for example, the solar power generation amount is predicted every 30 minutes, for example, by predicting the outside air temperature and the sunshine duration according to the weather forecast on the next day. Renewable energy is not limited to photovoltaic power generation but may be wind power generation. In this way, the power generation plan is determined in consideration of the power supply capacity and unit price of the power plant that is scheduled to supply power based on the predicted power demand A, including the power generation amount Eb5 from the above-mentioned renewable energy. The In the example of the figure, the power generation amounts Eb1, 2 and the first and second adjustment power amounts Eb3, 4 of the first and second power plants are determined.

The power usage of each facility on the day is different from the above-mentioned predicted power consumption A on the previous day, and the solar power generation amount E5 is also different. Therefore, for example, as shown in FIG. 6C, the energy prediction system 1 uses the measured power values E1 to E5 of each facility to calculate a predicted value for a predetermined time ahead (for example, 30 minutes ahead at 12:00) by the cluster analysis method. calculate. And based on the predicted value by the short-time prediction, for example, the power generation amount of the adjusted power 2 at the predicted time can be controlled. Furthermore, as shown in the figure, since power load prediction (long-term prediction) is performed until 24:00 on the prediction day, it is possible to predict power load predicted values Ef1 to Ef1 to 5 hours ahead of 1 hour and 2 hours. Therefore, when the adjustment power 2 cannot be adjusted, preparation for adjusting the power generation amount including the adjustment power 1 can be made. Furthermore, it is possible to respond in advance to power procurement from the power market as required. In this way, the operation of the energy facility can be controlled (supported) based on the predicted value calculated by the energy prediction system 1. In the figure, reference numerals E1 and E2 indicate actual measured power values of the first and second power plants, and reference numerals E3 and E4 indicate first and second adjusted power actual measured values. Moreover, the code | symbol Ef1,2 shows the predicted electric power generation value of a 1st, 2nd power plant, code | symbol Ef3,4 shows the 1st, 2nd adjustment electric power prediction value, and the code | symbol E5 shows a photovoltaic power generation prediction value. It is possible to predict not only the energy load in a single facility with multiple generators and / or thermoelectric devices, but also the overall energy load of a complex facility, apartment house, or area where multiple such facilities are assembled. .

Finally, mention is made of the possibilities of other embodiments of the invention. In addition, the same code | symbol is attached | subjected to the member similar to the above-mentioned embodiment.

As an example of the long-term prediction in the above-described embodiment, the all-day prediction is described with the long-term prediction time being 24:00 (PM 12:00) on the day of the short-term prediction. However, the long-term prediction time is not limited to the prediction until 24:00 on the prediction day, and for example, 24 hours, 48 hours (two consecutive days), or any other arbitrary period (time) can be set. However, depending on the long-term forecast time to be set, the forecast may extend to the next day as well as the forecast day. Therefore, in such a case, the load pattern classification unit 30 may further include a long-term load pattern set in units of days of at least two consecutive days. Here, at least two consecutive days are: 2 weekdays, all weekdays (Monday to Friday), weekends (Friday and Saturdays and Saturdays and Sundays, Friday to Sunday), holidays (Sundays and holidays and the following day) Weekdays), several days such as one week, one month, long vacation, event period, season, etc. In this way, by classifying the characteristics that affect the energy load pattern (long-term load pattern) for each period to which any long-term predicted time belongs, similar periods that take into account common changes that occur throughout the period Similarity calculation, which will be described later, can be performed on the time-series data in the above, and the prediction accuracy can be improved.

Especially, when the long-term load pattern is composed of a combination of days of the week, for example, the next day can be predicted without using weather conditions. In other words, it is possible to predict the energy of buildings and facilities that are geographically distant from each other, and even a distant point can be predicted by addition (aggregation) as shown in FIG. In the above embodiment, the prediction method using the Kalman filter is applied to the previous day prediction, but it is also possible to continuously predict from the previous day prediction to the current day prediction (for example, 48 hours) instead of the Kalman filter. Of course, in order to further improve the accuracy, it is possible to set a long-term load pattern in consideration of weather conditions in combination with the day of the week. In addition, in the long-term prediction that is predicted over the next day, similarly to the short-term prediction, prediction may be performed by applying each load pattern for each prediction day.

In the above embodiment, the distance between the past time series data and the current date / time series data is obtained as the similarity using the inter-group average distance method, but it is also possible to use the cosine of the angle between vectors instead.

The examples of the current date / time series data P1 shown in FIG. 7 (a) and the past time series data Q1 shown in FIG. 7 (b) have the same shape with only different ratios. Therefore, the angle between the vectors p1 and q1 of the time series data is 0 as shown in FIG. On the other hand, in the example of the current date / time series data P2 shown in FIG. 8A and the past time series data Q2 shown in FIG. 8B, as shown in FIG. 8C, between the time series data vectors p1 and q1. The angle is θ, and this vector angle is used.

Here, when the time series data is represented by a vector in an N + 1 dimensional space with the size at each time as a coordinate value, the cosines of the two vector angles θ of the current date and time series data and the past time series data m days ago are: It becomes as follows.

In the present embodiment, this cos θm is used as a similar index and corresponds to the reciprocal of the distance in the above embodiment. For example, it is preferable to extract the top S in the descending order of cos θm and use the weighting coefficient of Equation 7 below. As shown in FIG. 8, the similarity can be determined even if the time series data has different sizes.

In the short-time prediction, the size of the past time series data can be adjusted to the size of the current date series data by applying the ratio of the following formula (8).

In the similar determination based on the angle between vectors, if the shape of the time-series data is exactly the same, it becomes 1, and even if the actual value is very small, there is a possibility that it is most similar, and there is a possibility that the predicted value will be shifted. Therefore, the deviation of the predicted value is suppressed by the correction of the following formula 9.

In addition, cluster analysis is not limited to the average distance between groups method, for example, the distance between centroids (distance between centroids of two clusters), the nearest distance method (minimum value of the distance between two individuals belonging to different clusters) ), The farthest distance method (maximum value of the distance between two individuals belonging to different clusters), and the like can also be employed.

In the above embodiment, the distance dm is calculated for each past time series data of the extracted comparison days, and the past time series data having the shortest distance is selected from the top to the sixth. However, the number to be selected is merely an example and can be set as appropriate. Further, the selection is not limited to the order from the top, but, for example, a selection satisfying the following relationship 10 may be selected. Note that R can be set arbitrarily, for example, R = 1.62.

Further, in the short-time prediction in the above embodiment, the weighted average of the difference between the actual measurement value at the predetermined time and the actual measurement value at the prediction time of the past time series data for which the prediction value is selected is added to the actual measurement value at the predetermined time on the prediction day. Calculated. However, the present invention is not limited to this, and a weighted average of actually measured values of predicted times of past time-series data may be used as shown in the following formula 11.

In the above embodiment, the number of comparison days is 30 days, but it is only an example and can be set as appropriate. Further, instead of setting a specific number of days, it is also possible to target all or a specific period of past time series data having a common load pattern with the predicted day. Therefore, the accuracy of the predicted value is improved as the accumulation of actually measured values increases. In selecting a load pattern, for example, various known methods such as a Kalman filter are used to load data by using various data of weather forecasts (average temperature, maximum temperature, minimum temperature, average relative humidity, maximum humidity, hourly temperature). The patterns may be further classified, and the past time series data (similar date) to be subjected to cluster analysis may be set based on the classification result.

In the above embodiment, the measurement time (measurement time) is not limited to the time (time) at which the energy load is actually measured, and the actual measurement value is also limited to a numerical value that is actually measured at the measurement time (measurement time). Is not something For example, it is also possible to divide the actual measurement time (for example, 60 minutes) into a plurality of parts, obtain actual measurement values at the divided time (for example, 10 minutes), for example, by linear interpolation, and use these values as actual measurement values. It is.

The program for executing the energy prediction system according to the present invention is configured as a support system for operators by incorporating software into a personal computer as an operation support system for district heating and cooling and building heat sources, capturing actual load data and storing actual data, for example. be able to. It can also be used for optimal operation control by incorporating it into BEMS and FEMS control devices and control panels for electric power and centralized heat source equipment, and it can also be used by incorporating it into panels for household power visualization (HEMS) and the like. .

The present invention can be used as an energy prediction system. Moreover, it can utilize as a driving | operation assistance system which assists the driving | operation of an energy installation using the predicted value calculated by an energy prediction system. For example, the optimum operation pattern of the power generation facility and the thermoelectric facility can be assumed by predicting the load of the power, cold heat, heat, steam, and hot water supply on the next day of the power generation facility and the heat source facility. It is also possible to predict a load ahead of a predetermined time from actual measurement data at the beginning of operation of a building, factory, etc., and further use it for optimum operation management from a heat load prediction curve from the data at that time to the long-term prediction time ahead of the day. . In addition, the power load is similarly predicted for a predetermined time ahead, and the power load curve from the current power data to the long-term prediction time ahead of the day is used for demand control, and the production equipment is operated as necessary. It can be used for planning, and if it has power generation facilities, it can also be used for optimal operation management of power generation facilities. Furthermore, demand control can be performed based on the result of long-term prediction from data on the amount of power used every arbitrary time, not only in buildings and factories, but also at home. In addition, according to the present invention, the predicted values of all power consumptions that use power in a certain region can be aggregated every hour, and combined with a smart meter to predict the daily power usage in the region. By using the predicted load value, the number of distributed generators including solar power generation and wind power generation can be controlled and the load can be adjusted.

1: Energy prediction system, 2: User interface, 2a: Monitor, 2b: Keyboard, 2c: Mouse, 3: Processing unit, 3a: CPU, 3b: Temporary storage memory, 3c: HDD, 3d: Bus, 10: Software 20: Measured value storage unit, 30: Load pattern classification unit, 40: Analysis condition setting unit, 41: Comparison time setting unit, 42: Comparison day setting unit, 43: Load pattern selection unit, 44: Upper selection day setting unit , 50: time series data setting unit, 51: time series data update unit, 60: predicted value calculation unit, 61: similarity calculation unit, 62: past time series data selection unit, 63: weight coefficient calculation unit, 64: temporary Actual value setting unit, 70: output unit, 80: recording unit, 100: energy equipment, P: current date / time series data, Q: past time series data

Claims

An energy prediction system that predicts the energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water,
An actual value storage unit that stores the actual value of the energy load up to the present time for each measurement time;
A load pattern classification unit for classifying the actual measurement values into load patterns based on the characteristics of the energy load every day;
An analysis condition setting unit for setting a comparison time and a comparison day for performing cluster analysis and selecting a comparison target for performing the cluster analysis from classified load patterns;
The load in which the current date / time series data from the predetermined time on the prediction day to the set comparison time and the past time series data of the past set comparison days in the same time zone as the current date / time series data are selected. A time-series data setting unit that is set based on actual measurement values classified into patterns,
A plurality of past time series data having a high degree of similarity is selected by calculating a similarity to the current date / time series data for each of the past time series data by the cluster analysis, and a predetermined prediction is made from the predetermined time of the selected past time series data. A prediction value calculation unit that calculates the predicted value of the energy load at the prediction time ahead of the prediction time from the predetermined time on the prediction day by weighting and adding the actual measurement values at the time ahead in accordance with the similarity. Energy prediction system.
The predicted value is calculated by adding a weighted average of the difference between the measured value at the predetermined time and the measured value at the predicted time of the selected past time series data to the measured value at the predetermined time on the prediction day. The energy prediction system according to claim 1.
The energy prediction system according to claim 1, wherein the predicted value calculation unit selects a plurality of past time-series data in descending order of the similarity.
The energy prediction system according to any one of claims 1 to 3, wherein the load pattern is at least one combination selected from the group including at least all days, holidays, weekdays, singular days, production plans, and weather information.
The predicted value calculation unit weights and adds an actual measurement value of the selected past time-series data at a time that is a predetermined second predicted time ahead from the predetermined time according to the similarity, and from the predetermined time on the prediction day A temporary actual measurement value setting unit that sets a predicted value of the energy load at the next second prediction time ahead of the second prediction time as a temporary actual measurement value of the time, and the time-series data setting unit includes the prediction day New current date and time series data from the next second predicted time to the comparison time before, and past time series data for each day of the set comparison days in the same time zone as the new current date and time series data. A time-series data updating unit that updates the measured value classified into the selected load pattern and the temporary measured value, and the temporary measured value setting unit sets the temporary measured value for each second predicted time. And the time-series data update unit updates the time-series data for each second prediction time, and the prediction value calculation unit repeatedly performs the cluster analysis for each second prediction time, so that the prediction day The energy prediction system according to any one of claims 1 to 4, wherein the energy load is predicted from a predetermined time to a long-term predicted time ahead.
6. The energy prediction system according to claim 5, wherein the temporary actual measurement value is replaced with the actual measurement value each time the actual measurement value at the next second prediction time on the prediction day is stored.
The energy prediction system according to claim 5 or 6, wherein the load pattern classification unit further includes a long-term load pattern set in units of at least two consecutive days.
The energy facility is a complex facility or region including a plurality of facilities having a plurality of power generation devices and / or thermoelectric devices, and the prediction calculation unit calculates the prediction value for each facility and adds the prediction values. The energy prediction system according to any one of claims 1 to 7, wherein a predicted value of the energy load of the complex facility or the entire region is calculated.
The energy prediction system according to any one of claims 1 to 8, wherein the cluster analysis is an intergroup average distance method.
The energy prediction system according to claim 9, wherein the similarity is a reciprocal of a distance of the past time series data with respect to the current date and time series data.
The energy prediction system according to claim 9, wherein the similarity is a cosine of an angle between vectors of the past time series data and the current date and time series data.
An energy prediction method for predicting an energy load after the present time in an energy facility that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water,
Storing the actual measured value of the energy load up to the present time for each measurement time;
Classify the measured values into load patterns based on the characteristics of the energy load every day;
Set comparison time and comparison days for cluster analysis and select comparison target for cluster analysis from classified load patterns,
The load in which the current date / time series data from the predetermined time on the prediction day to the set comparison time and the past time series data of the past set comparison days in the same time zone as the current date / time series data are selected. Set according to the measured values classified into patterns,
A plurality of past time series data having a high degree of similarity is selected by calculating a similarity to the current date / time series data for each of the past time series data by the cluster analysis, and a predetermined prediction is made from the predetermined time of the selected past time series data. An energy prediction method for calculating a predicted value of the energy load at a prediction time ahead of the prediction time from the predetermined time on the prediction day by weighting and adding measured values at a time ahead in accordance with the similarity.
A computer program for executing the energy prediction system according to any one of claims 1 to 11.
A recording medium recording the computer program according to claim 13.
An operation support system that supports the operation of energy equipment that uses or manufactures energy such as electric power, cooling, heating, steam, and hot water,
An actual value storage unit that stores the actual value of the energy load up to the present time for each measurement time;
A load pattern classification unit for classifying the actual measurement values into load patterns based on the characteristics of the energy load every day;
An analysis condition setting unit for setting a comparison time and a comparison day for performing cluster analysis and selecting a comparison target for performing the cluster analysis from classified load patterns;
The load in which the current date / time series data from the predetermined time on the prediction day to the set comparison time and the past time series data of the past set comparison days in the same time zone as the current date / time series data are selected. A time-series data setting unit that is set based on actual measurement values classified into patterns,
A plurality of past time series data having a high degree of similarity is selected by calculating a similarity to the current date / time series data for each of the past time series data by the cluster analysis, and a predetermined prediction is made from the predetermined time of the selected past time series data. A prediction value calculation unit that calculates the predicted value of the energy load at the prediction time ahead of the prediction time from the predetermined time on the prediction day by weighting and adding the actual measurement values at the time ahead in accordance with the similarity. ,
A driving support system that supports driving of the energy facility based on the predicted value.