Data Reduction in Smart Grid

The advancement of the electric grid has led to tremendous growth in data generated from the installed sensors. With this huge amount of data, the electric grid faces a big data problem. Therefore, it is necessary to implement a data reduction technique that would not only curtail the exponential data growth, but also be able to make smart decisions with the reduced data.This thesis deals with two data reduction techniques; data compression and dimensionality reduction. Some of the state of the art compression algorithms exploit the characteristics of load pro le data, where consecutive data samples have very small di erences. However, performance of these algorithms deteriorate with large value di erences. In this thesis, we propose a modi cation that improves the compression ratio when there are large value di erences. The algorithm is evaluated on smart meter load pro le data at di erent data resolution. We show that our proposed changes improve performance by 2 - 20 % for di erent data resolution. Also, this algorithm can be extended to other domains. Furthermore, we propose a lossless compression algorithm that uses i deep learning network to learn signal characteristics and make better prediction of the future values. The di erence between the real and predicted value is encoded to produce the nal bit string. We tested di erent deep learning architectures in four di erent scenarios on smart meter load pro le data. We show that as the model learns patterns in the data, the compression performance improves. Recent studies using linear dimensionality reduction techniques indicate that high dimensional smart grid data may actually lie in a lower dimension. Taking into account the complexity of the smart grid, the process is inherently non-linear. Therefore, we critically investigate the performance of non-linear dimensionality reduction techniques and their applications in smart grid. The techniques studied are principal component analysis (PCA), isometric feature mapping (ISOMAP), kernel principal component analysis (KPCA), locally linear embedding (LLE), laplacian eigenmaps, t-Distributed stochastic neighbor embedding (t-SNE) and autoenoders. The techniques are analysed for event detections using dimensionality reduction - long short term memory (LSTM) based anomaly detection algorithm (DR-LADA). Also, the proposed approach uses a moving window technique to mitigate the dynamic measurement of the grid. Similarly, we consider their applications in load pro ling relevant to demand side management using k-means cluster algorithm. Some non-linear dimensionality reduction techniques perform better than linear dimensionality reduction technique.

Name thesis:degree_name

Master of Applied Science (MASc)

Level thesis:degree_level

Master's

Discipline thesis:degree_discipline

Engineering - Electronic Systems

Grantor dc:publisher

Faculty of Graduate Studies and Research, University of Regina

Year dc:date.issued

2018

Author dc:creator

• Aleshinloye, Ahmed

Advisors dc:contributor.advisor

• Bais, Abdul
• Al-Anbagi, Irfan

Committee members dc:contributor.committeemember

• Laforge, Paul
• Wang, Zhanle

Language dc:language.iso

en

DOI

https://doi.org/10.82465/4664

OAI identifier oai:identifier

oai:uregina.scholaris.ca:10294/8847