A Hybrid Machine Learning Framework for Enhancing PMU-based Event Identification with Limited Labels

The energy industry is experiencing rapid and dramatic changes on both the generator side and the load side, necessitating faster, more accurate, and robust event detection methods for situational awareness. Growing installations of PMU devices that provide high resolution synchronized measurements combined with the advancement of artificial intelligence and big data analytics techniques have recently attracted the R&D community interest. Some supervised learning techniques have been proposed using PMU measurements, however, they are facing challenges in 1) limited interpretability, 2) biased learning models/results, and 3) insufficient labeled data for learning. To address these issues, we propose a machine learning-based framework for physically-meaningful interpretability, hybrid-learning method with indexes, and a flexible data-preparation approach. Specifically, a thoroughly designed feature selection method is proposed for discovering event signatures. Then, a hybrid machine learning process is constructed to reduce biases of different machine learners due to their diversified working mechanisms. Finally, we propose to utilize unlabeled data via semi-supervised learning and add strategical event data via active learning, e.g., simulations. The goal is to significantly improve the supervised learning results via computational efficient techniques. Extensive simulations are conducted using a commercial power system dynamics simulator and synthetic realistic transmission grid models. Significant improvements are observed via hybrid supervised learning methods, semi-supervised learning, and active learning.