Load prediction based on-line and off-line training of neural networks

• US 20070185823A1
• Filed: 07/28/2006
• Published: 08/09/2007
• Est. Priority Date: 07/28/2005
• Status: Active Grant

• Current Assignee: Siemens AG
• Original Assignee: Siemens Energy Incorporated (Siemens AG)
• Inventors: Chen, Dingguo
• Granted Patent: US 7,552,100 B2
• Time in Patent Office: Days
• US Class Current: 706/15
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: H02J 3/008 involving trading of energy or energy transmission rights
• CPC Class Codes: Y02E 40/70 Smart grids as climate change mitigation technology in the energy generation sector
• CPC Class Codes: Y04S 10/50 Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
• CPC Class Codes: Y04S 50/10 Energy trading, including energy flowing from end-user application to grid

Abstract

A method and system is provided for predicting loads within a power system through the training of on-line and an off-line neural networks. Load data and load increments are used with an on-line load prediction scheme to generate predicted load values to optimize power generation and minimize costs. This objective is achieved by employing a method and system which predicts short term load trends through the use of historical load data and short term load forecast data.

Claims

1. A method of energy management in a power system comprising the steps of:

Training an off-line neural network with load data;

Training the off-line neural network with a first set of on-line predicted load values;

Using the first set of on-line predicted values to predict a second set of predicted load values;

Training an on-line neural network with load data;

Training the on-line neural network with a third set of on-line predicted load values

Using the third set of on-line predicted load values to predict a fourth set of on-line predicted load values

Comparing the second set of predicted load values with actual load values and incrementing a first counter if the difference between the second set of predicted load values and the actual load values are within an acceptable range;

Comparing the fourth set of predicted load values with actual load values and incrementing a second counter if the difference between the fourth set of predicted load values and the actual load values are within an acceptable range; and

Calculating final predicted load values based on the values of the first and second counter.

2. The method of claim 1, further comprising the step of comparing values in the first and second counters, selecting the off-line load prediction scheme for use on the on-line neural network if the value in the first counter is greater than the value in the second counter;

and selecting the on-line load prediction scheme for use on the on-line neural network if the value in the second counter is greater than the value in the first counter.

3. The method of claim 1 wherein the step of training the off-line neural network comprises the step of calculating load increments between load data, and using the load increments to train the off-line neural network.

4. The method of claim 3, wherein the step of training the off-line neural network comprises the step of normalizing the load increments and using the normalized load increments to train the off-line neural network.

5. The method of claim 4, wherein the step of using the load increments to train the off-line neural network comprises the step of using the load increments to calculate the first set of on-line predicted load values.

6. The method of claim 4, wherein the step of normalizing the load increments comprises the step of using a standard deviation from the best matching date calculated in off-line neural network training.

7. The method of claim 1 wherein the step of training the on-line neural network comprises the step of calculating load increments between load data, and using the load increments to train the on-line neural network.

8. The method of claim 7, wherein the step of training the on-line neural network comprises the step of normalizing the load increments and using the normalized load increments to train the on-line neural network.

9. The method of claim 8, wherein the step of using the load increments to train the on-line neural network comprises the step of using the load increments to calculate the third set of on-line predicted load values.

10. The method of claim 8, wherein the step of normalizing the load increments comprises the step of using a standard deviation from the best matching date calculated in on-line neural network training.

11. The method of claim 10 further comprising the step of using gradient optimization to train the on-line neural network.

12. The method of claim 6 further comprising the step of using gradient optimization to train the off-line neural network.

13. The method of claim 2 wherein the step of calculating the final predicted load values comprises generating an accuracy probability for each of the on-line predicted load values based on on-line neural network training and off-line predicted load values based on off-line network training wherein corresponding probabilities are generated by calculating the ratio of accurate predictions to the total number of accurate predictions.

14. The method of claim 13 further comprising the step of comparing the generated probabilities and selection the load prediction scheme with the highest probability.

15. A computer-readable medium having stored thereon instructions which when executed by a processor, cause the processor to perform the steps of:

Training an off-line neural network with load data;

Training the off-line neural network with a first set of on-line predicted load values;

Using the first set of on-line predicted values to predict a second set of predicted load values;

Training an on-line neural network with load data;

Training the on-line neural network with a third set of on-line predicted load values

Using the third set of on-line predicted load values to predict a fourth set of on-line predicted load values

Comparing the second set of predicted load values with actual load values and incrementing a first counter if the difference between the second set of predicted load values and the actual load values are within an acceptable range;

Comparing the fourth set of predicted load values with actual load values and incrementing a second counter if the difference between the fourth set of predicted load values and the actual load values are within an acceptable range; and

Calculating final predicted load values based on the values of the first and second counter.

16. A system for predicting loads within an energy management system of a power system, comprising:

a processor for executing energy management applications;

an off-line neural network in communication with the processor;

an on-line neural network in communication with the processor and the off-line neural network;

a load database for storing current and historical load values and predicted load values and accessible by the processor, the off-line neural network;

an on-line load prediction module in communication with the processor, the off-line neural network and load database;

a neural network training module in communication with the on-line neural network and off-line neural network;

a decision algorithm module for selecting a load prediction scheme;

a memory in communication with the processor for storing energy related data; and

an energy management module for executing energy management functions comprising,

Training the off-line neural network with a first set of on-line predicted load values;

Using the first set of on-line predicted values to predict a second set of predicted load values;

Training an on-line neural network with load data;

Training the on-line neural network with a third set of on-line predicted load values

Using the third set of on-line predicted load values to predict a fourth set of on-line predicted load values

Comparing the second set of predicted load values with actual load values and incrementing a first counter if the difference between the second set of predicted load values and the actual load values are within an acceptable range;

Comparing the fourth set of predicted load values with actual load values and incrementing a second counter if the difference between the fourth set of predicted load values and the actual load values are within an acceptable range; and

Calculating final predicted load values based on the values of the first and second counter.

17. The system of claim 16, wherein the energy management functions further comprise the step of comparing values in the first and second counters, selecting the off-line load prediction scheme for use on the on-line neural network if the value in the first counter is greater than the value in the second counter;

and selecting the on-line load prediction scheme for use on the on-line neural network if the value in the second counter is greater than the value in the first counter.

18. The system of claim 17 wherein the step of calculating the final predicted load values comprises generating an accuracy probability for each of the on-line predicted load values based on on-line neural network training and off-line predicted load values based on off-line network training wherein corresponding probabilities are generated by calculating the ratio of accurate predictions to the total number of accurate predictions and using the predicted load values from the neural network having the higher probability.

19. The system of claim 16, wherein the step of training the off-line neural network and the on-line neural network comprises the step of calculating load increments between load data, and using the load increments to train the off-line neural network and on-line neural network

20. The system of claim 19, wherein the step of training the off-line neural network and on-line neural network comprises the step of normalizing the load increments and using the normalized load increments to train the off-line neural network and on-line neural network.

21. The system of claim 20, wherein the step of normalizing the load increments comprises the step of using a standard deviation from the best matching date calculated in off-line neural network training.