Thesis
Anomaly detection in power distribution system measurements using machine learning
Washington State University
Master of Science (MS), Washington State University
12/2019
DOI:
https://doi.org/10.7273/000003996
Handle:
https://hdl.handle.net/2376/124676
Abstract
Sensor measurements of distribution system are uncertain due to sensor malfunctions, communication failure and cyber attacks. This thesis aims to perform anomaly detection on measurements utilizing data-driven approaches. The measurements considered are individual smart meter real power measurements and network-wide primary voltage magnitudes. Anomaly detection in individual smart meter measurements using gaussian probabilistic thresholds is explored. It flags non-anomalous data as verified by the comparison of smart meter real power and individual appliance consumption. To perform a real-time comparison for detection, Non-Intrusive Load Monitoring (NILM) is needed, which is difficult due to the associated consumer privacy issues. Alternatively, forecasting can be used for anomaly detection. So, single layer neural network models such as Multi-Layer Perceptron (MLP), and Long Short Term Memory (LSTM) with different features are tried. Even in training data, a poor performance is seen in these models, due to the smart meter profile variability. Hence, aggregated smart meter forecasting using neural networks can be used to detect anomaly in such aggregated measurements with a reasonable accuracy. Network-wide primary voltage measurements are correlated for a phase of feeder for different buses at a given time-step; this is extensively validated empirically. To leverage this, Principal Component Analysis (PCA) is used to reduce the dimensionality of this input data. Further, residual and subspace based methods are explored for network-level anomaly detection and identification. The results for the residual approach on missing and bad data cases are detailed for IEEE 13 bus and IEEE 8500 node test feeders. It is validated through simulations that residual-based approach on subspace projection matrix for the measurement data successfully performs anomaly detection and identification for primary network voltage measurements for the selected test cases. Further research is needed to validate the applicability and accuracy of the proposed framework during changes in the system operating conditions (topology changes, capacitor bank switching, etc.), and on real-world measurements form sensors deployed in the field.
Metrics
81 File views/ downloads
248 Record Views
Details
- Title
- Anomaly detection in power distribution system measurements using machine learning
- Creators
- Arun Abhishek Imayakumar
- Contributors
- Anamika Dubey (Advisor) - Washington State University, School of Electrical Engineering and Computer Science
- Awarding Institution
- Washington State University
- Academic Unit
- School of Electrical Engineering and Computer Science
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Identifiers
- 99900890797601842
- Language
- English
- Resource Type
- Thesis