Online Predictor Design for Conductor Galloping

Conductor galloping, like many other natural disasters, poses significant challenge to the reliable operation of power system. Such events happen in extreme weather conditions with complex physical triggering mechanisms. Hence, it is challenging to predict such events due to the limited (galloping) samples as well as complex physical dynamics. In this paper, we seek to employ a data-driven model to design galloping predictor. To mitigate the obstacles led by limited samples, we first design a greedy unit disk covering algorithm to extract the structural knowledge in the available dataset. Viewing such knowledge as states, we further propose a continuous time Markov chain to generate the pseudo online data. This data generator further allows us to design the data-driven online predictors for conductor galloping. Numerical studies highlight that k-nearest neighbor predictor works remarkably and more robust to the different degrees of data imbalance, compared with online logistic regression and online random forest.