Identify data sources for neural network

• US 7,403,928 B2
• Filed: 02/01/2006
• Issued: 07/22/2008
• Est. Priority Date: 02/03/2005
• Status: Active Grant

• Current Assignee: Siemens AG
• Original Assignee: Siemens Power Transmission & Distribution Incorporated (Siemens AG)
• Inventors: Chen, Dingguo
• Primary Examiner(s): Holmes; Michael B
• Application Number:
  US11/345,438
• Publication Number:
  US 20060184472A1
• Time in Patent Office: 902 Days
• Field of Search: 706/15
• US Class Current: 706/15
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A system, method, and device for identifying data sources for a neural network are disclosed. The exemplary system may have a module for determining load curves for each selected data set. The system may also have a module for determining a global difference measure and a global similarity measure for each load curve of each selected data set. The system may have a module for determining a set of data sets with lowest value global difference measure. The system may also have a module for determining a set of data sets with largest value global similarity measure. The system may also have a module for determining a union of the sets of lowest value difference measure and the sets of largest value similarity measure. The system may also have a module for determining for each set in the union one of a local similarity measure and a local difference measure and a module for selecting a set of reduced data sets based on one of the local similarity measure and the local difference measure.

Claims

1. A computer implemented system for making efficient use of data sources for use in offline training of a neural network for predicting electric power loads comprising:

a processor for;

calculating load curves for each of selected data sets;

calculating a global difference measure and a global similarity measure for each load curve of each selected data sets;

calculating a set of data sets with lowest value global difference measure;

calculating a set of data sets with largest value global similarity measure;

calculating a union of the sets of lowest value difference measure and the sets of largest value similarity measure;

calculating for each set in the union one of a local similarity measure and a local difference measure;

generating a set of reduced data sets based on one of the local similarity measure and the local difference measure effective to provide more efficient use of the data sets; and

using the set of reduced data sets for performing offline training of a neural network for predicting electric power loads.

2. The system of claim 1, wherein the processor is further configured for generating the set of reduced data sets from the sets in the union based on data sets with the largest value of the local similarity measure.

3. The system of claim 1, wherein the processor is further configured for generating the set of reduced data sets from the sets in the union based on data sets with the smallest value of the local difference measure.

4. The system of claim 1, wherein the data sets are data for a one-day period and wherein the processor is further configured for modifying reference load curve based on Daylight Savings Time.

5. The system of claim 1, wherein the data sets are data for a one-day period and wherein the processor is further configured for generating the selected data set'"'"'s days based on a reference day data set.

6. The system of claim 5, wherein the processor is further configured for generating the selected data sets from data sets including any weekdays between Tuesdays and Thursdays of the same week when the next day after the reference day is between Tuesday and Thursday inclusively;

selecting the selected data sets from data sets including previous Mondays, Sundays and Tuesdays when the next day after the reference day is Monday; and

selecting the selected data sets from data sets including previous Thursdays, Fridays and Saturdays when the next day after the reference day is Friday.

7. The system of claim 5, wherein the processor is further configured for generating the selected data sets from data sets including previous Fridays, weekends and Mondays when the next day after the reference day is any weekend day;

and generating the selected data sets from data sets including the previous same holidays and days preceding or following these holidays by no more than two days when the next day after the reference day is a holiday.

8. A method for making efficient use of data sources for use in offline training of a neural network for predicting electric power loads comprising:

calculating load curves for each selected data sets;

calculating a global difference measure and a global similarity measure for each load curve of each selected data sets;

calculating a set of data sets with lowest value global difference measure;

calculating a set of data sets with largest value global similarity measure;

calculating a union of the sets of lowest value difference measure and the sets of largest value similarity measure;

calculating for each set in the union one of a local similarity measure and a local difference measure;

generating a set of reduced data sets based on one of the local similarity measure and the local difference measure effective to provide more efficient use of the data sets; and

using the set of reduced data sets for performing offline training of a neural network for predicting electric power loads.

9. The method of claim 8, further comprising the following actions:

generating the set of reduced data sets from the sets in the union based on data sets with the largest value of the local similarity measure.

10. The method of claim 8, further comprising the following actions:

generating the set of reduced data sets from the sets in the union based on data sets with the smallest value of the local difference measure.

11. The method of claim 8, wherein the data sets are data for a one-day period and further comprising the following actions:

modifying reference load curve based on Daylight Savings Time.

12. The method of claim 8, wherein the data sets are data for a one-day period and further comprising the following actions:

generating the selected data set'"'"'s days based on a reference day data set.

13. The method of claim 12, further comprising the following actions:

generating the selected data sets from data sets including any weekdays between Tuesdays and Thursdays of the same week when the next day after the reference day is between Tuesday and Thursday inclusively;

generating the selected data sets from data sets including previous Mondays, Sundays and Tuesdays when the next day after the reference day is Monday; and

generating the selected data sets from data sets including previous Thursdays, Fridays and Saturdays when the next day after the reference day is Friday.

14. The method of claim 12, further comprising the following actions:

generating the selected data sets from data sets including previous Fridays, weekends and Mondays when the next day after the reference day is any weekend day; and

generating the selected data sets from data sets including the previous same holidays and days preceding or following these holidays by no more than two days when the next day after the reference day is a holiday.

15. The method of claim 12, further comprising the following actions:

generating the selected data sets of any day not including the reference day.

16. A computer implemented system for identifying matching data sources for more efficient use in offline training of a neural network of a neural network-based very short term load prediction for operating a power generation unit comprising:

a processor for;

calculating load curves for each of a selected date data sets;

calculating a global difference measure and a global similarity measure for each load curve of each selected date data sets;

calculating a set of date data sets with lowest value global difference measure;

calculating a set of date data sets with largest value global similarity measure;

calculating a union of the date data sets of lowest value difference measure and the date data sets of largest value similarity measure;

calculating for each date data set in the union one of a local similarity measure and a local difference measure;

generating a data set of reduced date data sets based on one of the local similarity measure and the local difference measure effective to provide more efficient use of the date data sets; and

using the set of reduced date data sets for performing offline training of a neural network to predict electric power loads.

17. The system of claim 16, wherein the processor is further configured for generating the data set of reduced date data sets from the date data sets in the union based on date data sets with the largest value of the local similarity measure.

18. The system of claim 16, wherein the processor is further configured for generating the data set of reduced date data sets from the data sets in the union based on date data sets with the smallest value of the local difference measure.

19. The system of claim 16, wherein the processor is further configured for modifying reference load curves based on Daylight Savings Time.

20. The system of claim 16, wherein the processor is further configured for:

generating the selected date data sets from date data sets including any weekdays between Tuesdays and Thursdays of the same week when the next day after the reference day is between Tuesday and Thursday inclusively;

generating the selected date data sets from date data sets including previous Mondays, Sundays and Tuesdays when the next day after the reference day is Monday; and

generating the selected date data sets from date data sets including previous Thursdays, Fridays and Saturdays when the next day after the reference day is Friday.