A TECHNIQUE FOR ANALYSIS OF ANN BASED FAULT DIRECTION DISCRIMINATOR
Date
2003
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Degree Level
Masters
Abstract
Artificial Neural Networks (ANNs) have demonstrated their potential in modeling nonlinear processes and complex processes that consist of interactions of time varying voltages and currents experienced during power system faults. Generalization and prediction capabilities give them the ability to correctly process the data that broadly resemble the data they were trained on. ANNs trained to recognize non-noisy patterns are able to correctly identify patterns with either some segments missing or with some segments added to the input data. This property gives them a fault tolerant characteristic.
In spite of the numerous advantages of ANNs, there are some acceptability issues. One of the issues is that most training data are obtained from simulations. It is necessary that the data should correctly represent all the conditions the network is likely to encounter. Another issue is that the performance of an ANN can not be explained in the manner in which the performance of most electromechanical and electronic devices can be. The acceptability for their use in critical applications, such as power system
protection, would increase if a technique for analyzing their performance could be developed and demonstrated. This would allow the users to validate the performance of an ANN when it becomes necessary to investigate the correctness of its performance.
Analytical methods for explaining the weights and biases of a trained network are not readily available. The only means of verifying the performance of a trained network is to perform extensive testing. It is hard to verify that the neural network has properly trained for all possible scenarios of actual operating conditions.
Because the output of a neuron can be quantized to specified logic levels, the output was treated as a multi-valued logic in this project. At this time, no technique is available for minimizing multi-valued logic that could handle a large number of inputs when there are several neurons in the input, hidden and output layers.
This thesis presents a new technique for the multi-valued logic minimization for large number of inputs. The possibility of making the `explanation' capability an integral part of a trained ANN and the verification of the quality of the training it received is also
examined. The technique predicts the behavior of ANNs for the training data as well as for all possible inputs. The technique also makes it possible to implement ANNs on Programmable Logic Arrays. The proposed technique was applied to the ANN based fault directional discriminator. Some results are presented as well in this thesis.
Description
Keywords
Citation
Degree
Master of Science (M.Sc.)
Department
Electrical Engineering