Term of Award

Spring 2020

Degree Name

Master of Science, Electrical Engineering

Document Type and Release Option

Thesis (open access)

Copyright Statement / License for Reuse

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Department

Department of Electrical and Computer Engineering

Committee Chair

Rami J. Haddad

Committee Member 1

Adel El-Shahat

Committee Member 2

Mohammad Ahad

Abstract

High penetration of distributed generation (DG) sources into a decentralized power system causes several disturbances, making the monitoring and operation control of the system complicated. Moreover, because of being passive, modern DG systems are unable to detect and inform about these disturbances related to power quality in an intelligent approach. This paper proposed an intelligent and novel technique, capable of making real-time decisions on the occurrence of different DG events such as islanding, capacitor switching, unsymmetrical faults, load switching, and loss of parallel feeder and distinguishing these events from the normal mode of operation. This event classification technique was designed to diagnose the distinctive pattern of the time-domain signal representing a measured electrical parameter, like the voltage, at DG point of common coupling (PCC) during such events. Then different power system events were classified into their root causes using long short-term memory (LSTM), which is a deep learning algorithm for time sequence to label classification. A total of 1100 events showcasing islanding, faults, and other DG events were generated based on the model of a smart distributed generation system using a MATLAB/Simulink environment. Classifier performance was calculated using 5-fold cross-validation. The genetic algorithm (GA) was used to determine the optimum value of classification hyper-parameters and the best combination of features. The simulation results indicated that the events were classified with high precision and specificity with ten cycles of occurrences while achieving a 99.17% validation accuracy. The performance of the proposed classification technique does not degrade with the presence of noise in test data, multiple DG sources in the model, and inclusion of motor starting event in training samples.

OCLC Number

1252206011

Research Data and Supplementary Material

No

Share

COinS