Thesis
PMUNET: Anomaly detection over concept drifting synchrophasor data streams
Washington State University
Master of Science (MS), Washington State University
12/2019
DOI:
https://doi.org/10.7273/000004185
Handle:
https://hdl.handle.net/2376/124691
Abstract
With push towards automation and digitization of the electric grid, number of Phasor Measurement Units (PMUs) have been installed for possible wide-area real-time monitoring and control of the power grid. PMUs data can have multiple anomalies and a fundamental challenge is to enable flexible, online and scalable anomaly detection over PMU data streams. While deep learning is often desirable to cope with data streams with rich features, training a deep model for anomaly detection over streaming PMU data is nontrivial due to concept drift and high learning cost. This paper proposes PMUNET, a novel deep learning framework to enable flexible, online and scalable anomaly detection over PMU data streams with concept drift. PMUNET is enabled by an active learning strategy, which incorporates a non-parametric multivariate concept drift detection component and serves as a periodic condition, and retrains the deep model over data streams only when necessary. We also develop parallel and distributed learning techniques for the active learning strategy, and show that these algorithms are scalable with provable performance guarantees in learning cost and data communication cost. Our experimental results verifies that the proposed active learning strategy outperforms several state-of-the-art shallow models and a passive learning strategy that continuously retrain the model without specific detection of concept drift.
Metrics
5 File views/ downloads
18 Record Views
Details
- Title
- PMUNET
- Creators
- Arman Ahmed
- Contributors
- Yinghui Wu (Advisor)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Identifiers
- 99900896440301842
- Language
- English
- Resource Type
- Thesis