Neural load disturbance analyzer

• US 5,469,528 A
• Filed: 09/16/1992
• Issued: 11/21/1995
• Est. Priority Date: 09/16/1992
• Status: Expired due to Term

• Current Assignee: ARINC Incorporated (Rtx Corporation)
• Original Assignee: Syseca, Inc. (GE Aerospace)
• Inventors: Douglas, Larry D., Hutchinson, Clarence W. Jr.
• Primary Examiner(s): Moore, David K.
• Assistant Examiner(s): HAFIZ, TARIQ R
• Application Number:
  US07/945,528
• Time in Patent Office: 1,161 Days
• Field of Search: 395/10, 395/20, 395/21, 395/22, 395/23, 395/24, 395/63, 395/64, 395/27, 395/25, 395/, 382/12-15
• US Class Current: 706/20
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: H02J 3/00 Circuit arrangements for ac mains or ac distribution networks
• CPC Class Codes: H02J 3/06 Controlling transfer of power between connected networks; Controlling sharing of load between connected networks

Abstract

A system and method for controlling the generation and distribution of power is disclosed. The system includes monitoring devices which acquire load data indicative of the system load. The load data are converted into patterns which are fed into a neural load analyzer configured to classify the patterns. The generators of the power system are controlled responsive to the resulting classifications.

Claims

1. A neural load disturbance analyzer for classifying a utility, power load disturbance into one of a plurality of disturbance modes, the neural load disturbance analyzer comprising:

a pre-processor which samples a signal indicative of a utility, power load at a plurality of times, the pre-processor generating pattern data representative of a pattern indicative of the samples;

a trained neural network disposed to receive the pattern and determine a winning output representative of the utility, power load disturbance;

a post-processor disposed to monitor the winning output and select a disturbance mode representative of the utility, power load disturbance based on the winning output; and

means for determining the probability of a utility, power load disturbance based on the number of occurrences of the winning output in the pattern data during a predetermined period relative to the total number of patterns presented to the neural network during the predetermined period.

2. The neural load disturbance analyzer of claim 1 wherein the pre-processor comprises:

means for generating a zero centered pattern data representative of the sample;

means for normalizing the zero centered pattern data; and

means for activating the trained neural network.

3. The neural load disturbance analyzer of claim 2 wherein the sample is representative of the load during a current time window, and wherein the means for generating a zero centered pattern representative of the sample comprises:

means for combining the sample with a sample representative of the load during a previous time window; and

means for ordering the combined samples according to a predetermined criteria.

4. The neural load disturbance analyzer of claim 1 wherein the neural network is self-organizing and comprises:

a plurality of neurodes connected by a set of connection weights and biases;

means for determining excitations of the neurodes based on the pattern data which is applied to the neural network and the set of connection weights and biases;

means for identifying a winning output based on the neurode excitations; and

means for activating the post-processor.

5. A method of training a self-organizing neural network having connections with weights and biases to distinguish between patterns representative of utility, power load disturbances, the method comprising the steps of:

retrieving data indicative of utility, power load disturbances at a plurality of times;

generating a set of pattern data groups representative of the utility, power load disturbances;

successively applying each pattern data group to an input slab of the self-organizing neural network, determining winning outputs based on the pattern data group and the weights and biases of the connections of the self-organizing neural network, and adjusting the weights and biases of the connections to segregate the utility, power load and determine responsive actions;

determining the number of times the winning output of a pattern wins during a training epoch; and

determining the probability of the occurrence of a disturbance corresponding to the winning output of the pattern by dividing the number of times the output wins during a training epoch by the number of pattern data groups.

6. The method of training a self-organizing neural network of claim 5 further comprising, prior to performing the step of successively applying the pattern data groups to the input slab of the neural network, the steps of:

normalizing the pattern data groups upon initialization; and

randomizing the connection weights and biases of the neural network.

7. The method of training a self-organizing neural network of claim 6 further comprising the steps of:

repeating the step of successively applying the pattern data groups to the input slab; and

storing the connection weights and biases of the neural network after the pattern data groups have been applied to the input slab.