| dc.description.abstract | Traditional power grids are a single-layered physical system, while smart grids are an
extension of traditional power grids that are cyber-physical networks, and the main difference is
smart grids include an information layer. There is a huge amount of information being managed
within recent smart grids, and the decentralized power generation adds an extra level of uncertainty
to smart grids. The standard methods of monitoring and security available cannot work as expected
when collecting and analyzing the large amount of data presented from different parameters in the
power network.
Compressive sensing is a signal processing tool that is used to monitor single and
simultaneous fault locations in smart distribution and transmission networks, to detect harmonic
distortions, and to recognize patterns of partial discharge. Compressive sensing reduces the
measurement cost and the management cost since it can detect or rebuild a signal from very few
samples. In this thesis, we propose to design and implement the fault detection via a feedforward
neural network using similar regularizations as in compressive sensing. We shall use the adaptivity
of neural networks to tackle with state changes in the smart grid, proving the scalability and the
decentralized capability of a neural network for fault detection in the grid.
Two codes have been created against two different databases, and it was found that indeed,
a feedforward autoencoder would be great at fault detection, however, many things should be
considered prior to implementing it on a large scale. The most important part of any autoencoder
generation is a good dataset. | en |
