Power system online stability assessment using active learning and synchrophasor data

Analysis of synchrophasor measurements using data mining tools, in pursuit of precise stability assessment, requires a sufficiently large training data set. Traditionally the process of learning the underlying power system behavioral patterns introduces a significant computational burden such that exhaustive simulations of all possible system operating conditions are necessary. Advancements in machine learning make it possible, in some cases, to reduce the amount of operating conditions that need to be analyzed without impacting the accuracy of stability assessment. By using a probabilistic learning tool in the described active learning scheme to interactively query operating conditions based on their importance, we show that significantly fewer data needs to be processed for accurate voltage stability and oscillatory stability estimation. Results show that the advantage of active learning is greater on more complicated power networks, where larger training data sets are involved.