Discovery of Signatures, Anomalies, and Precursors in Synchrophasor Data with Matrix Profile and Deep Recurrent Neural Networks (Final Project Report)

Abstract

The widespread deployment of phasor measurement unit (PMU) across the U.S. together with the burgeoning machine learning technology made it possible to develop data-driven PMU data analytics to improve grid security and reliability in a more insightful and effective manner. Although PMU applications have been explored for over a decade, the representative PMU usage is limited to the bulk power system monitoring mainly due to the data integrity issues associated with PMUs (typically missing, fragmented, and wrongly amplified data). To forge a breakthrough on this stalemate and embrace PMUs for power system control and protection as well, we applied various advanced machine learning and big data analysis technology to the power system event detection and classification as the first step toward the power system control and protection pertaining to grid security enhancement. In contrast to other related work, we capitalized on active power, reactive power, voltage magnitude, and frequency inspired by the system operator's generally monitored measurements and the input of the in-use disturbance recorders, which drastically raised the event detection and classification performance, highlighted in the following five contributions. First, the root of the PMU data integrity issue was addressed, developing an event-participation decomposition model with a novel online SPIKE-P (Stochastic Proximal Implicit Krasulina Event-Participation Decomposition) algorithm, which enables the prompt replacement of missing data with realistic data with the significantly low error between the two data. The same technical challenge was tackled, developing short-term forecasting of PMU data by an advanced attentional sequence to sequence (seq2seq) long term short memory forecast model, which enables us to replace missing data with realistic data even if all PMUs are out-of-service (i.e., all PMU data are missing). Next, three event detection algorithms were individually developed, which enables us to perform 1) the unrivaled search refinement with the spatiotemporal correlation encoding technique, 2) superb (voltage) event signature extraction with the newly invented separation extraction technique from the noise matrix, and 3) event-type-free non-event/event discrimination with no event label, based on the cutting-edge bidirectional generative adversarial network model. All three event detection algorithms can quickly detect the targeted event, which enables us to turn them into an online control and protection. Third, a groundbreaking classifier was produced with an innovative graph signal processing-based PMU sorting algorithm and information loading-based regularization technique, which enables us to 1) differentiate four event types (voltage event, frequency event, oscillation event, and non-event) with a comprehensive visualization, and 2) achieve sufficiently fast event classification (no longer than 2 seconds following events). Additionally, three representative data injection methods (fast gradient sign method, basic iterative method, DeepFool method) are employed to scrutinize the resilience of the aforementioned classifier against a typical cyberattack (i.e., false data injection attack or adversarial attack), which clarifies that subtle noisy signal injection (5-7% amplitude of the attack signal with relative to the original signal) would potentially deteriorate the classifier's performance (i.e., event type classification accuracy) by half. Furthermore, the classifier model trained with the Eastern Interconnection (EI) grid event data was reused to train another classifier model that classifies western electricity coordinating council (WECC) grid events. Another classifier successfully showed enough classification performance for the power system events in the WECC with no training in WECC grid event data, which enables the decrease in the model development task for other bulk power grids. Fourth, a dynamic behavior-based event signature library was designed using a deep neural network-based classifier with an empirical clustering and Shannon entropy, which enables us to 1) facilitate a reliable event signature database with more granular categorization than just voltage, frequency, and oscillation events, 2) effectively showcase how a generic event signature looks like for each categorization in education, and 3) efficiently discover a new type of event signature. Finally, the simulation-free PMU-oriented synthetic data for voltage and frequency events were generated with advanced probabilistic programming methods and deep cascaded convolutional generative models, differentiating between everpresent event signatures and remaining signatures, which enabled us to create highly realistic but artificially fabricated event data that may be used to augment training and testing datasets of supervised learning work such as an event classification. Other than the above, power system dynamic parameters were estimated using a physics-informed neural network, called neural ordinary differential equation, assuming that 1) the grid topology is disclosed, and 2) PMUs are deployed at all substations, which enables us to embed power system dynamic models (e.g., power swing equation) directly into the aforementioned neural network, and to identify unobservable/uncertain parameters, e.g., the grid inertia.