Data Anomaly Detection and Correction in PMU Measurements for Wide-Area Monitoring Applications

Open Access
- Author:
- Mahapatra, Kaveri
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 11, 2020
- Committee Members:
- Nilanjan Ray Chaudhuri, Dissertation Advisor/Co-Advisor
Nilanjan Ray Chaudhuri, Committee Chair/Co-Chair
Minghui Zhu, Committee Member
Mehdi Kiani, Committee Member
James Freihaut, Outside Member
Kultegin Aydin, Program Head/Chair - Keywords:
- Synchrophasor
PMU
Wide Area Monitoring
Cyberphysical Security
Oscillation Monitoring
Event Detection
Disturbance Detection
Disturbance Classification
Wide area control
Damping Control
Neuromorphic Computing
Spiking Neural Network
Anomaly Detection
Corruption Detection
Event Classification
Data Anomaly - Abstract:
- This dissertation presents new algorithms for detection and characterization of corruption as well as genuine power system disturbances and classification of the later for situational awareness, and proposes methods of data correction for applications in wide-area oscillation monitoring and damping control in power grid from Phasor Measurement Unit (PMU) signals. First, for the purpose of characterization of outliers in PMU data, a linearized framework is established that relates the system's intrinsic dynamical properties to synchrophasor measurements on a sliding window, which includes pre-disturbance data and that triggered by genuine faults or disturbances. Using singular value perturbation theory, a set of features based on principal component scores were derived to distinguish between fault-induced outliers versus measurement anomalies. An online classifier for characterization of outliers was also developed to demonstrate the performance of the features in discriminating against outliers in synchrophasor data streams. However, detection and characterization of outliers is not enough, since mode metering applications can deteriorate not only in presence of spiky anomalous outliers, but also with continuous injection of corrupted data from cyber-attack. To address this problem, a Principal Component Pursuit (PCP)-based data pre-processor is proposed to provide resilience against malicious data corruption. By solving PCP, a low rank matrix approximating the true system response can be recovered from incoming PMU data matrix in presence of gross sparse errors originating from cyber-attacks. It is shown that in spite of continuous injection of anomalous data in multiple signals, use of reconstructed data obtained through the pre-processor improves the estimation of poorly-damped modes. However, a high computational burden is imposed by solving PCP on a block of data that can lead to a delayed estimation of frequency and damping, which can be detrimental to oscillation monitoring applications. Consequently, a vector processing algorithm based on online robust Principal Component Analysis (PCA) is proposed, where the actual signals can be recovered in presence of attack by solving a l1-norm minimization problem. This algorithm is not only capable of identifying the corrupted signals at any instant, but also providing an estimate close to the true measurements that belong to the current signal subspace. A mechanism is provided for selecting a signal subspace to be used with optimization at an operating condition from a subspace library. Extensive testing has shown that simultaneous corruption in up to 20% of the signals can be successfully corrected in this method. To improve the detection when upto 30-40% of the signals are corrupted simultaneously, an lp-norm based nonconvex optimization algorithm as an improvement over l1 is proposed. This algorithm not only improves the wide-area oscillation monitoring, but also the damping control performance during cyberattacks. Improved damping of inter-area oscillations are observed by feeding the reconstructed signals from the proposed pre-processor to the controller during attacks on multiple signals. Finally, a novel methodology based on neuromorphic computing architecture for event-driven classification of different types of power system disturbances (events) has been proposed. In conjunction with a QR decomposition-based signal selection technique to select PMU measurements relevant to disturbance event classes through an offline analysis, a Spiking neural network-based classifier is proposed for identifying different types of events. Key benefits such as much less processing requirements and low energy consumption as compared to traditional CPU/GPU- based machine learning techniques make it a suitable candidate for real time situ- ational awareness application in control rooms.