Online ensemble learning for security assessment in PMU based power system

This paper presents an online Adaboost algorithm for ensemble of support vector machines (SVM) for power system security assessment utilizing online measurement data obtained from Phasor Measurement Units (PMUs). Our proposed learning scheme consists of a strong learner and multiple weak learners. The weak learners are linear SVMs which are easy to implement and incrementally updated with low computation complexity. The strong learner compensates for inevitable classification errors of linear SVMs by using the boosting approach. Since the data is unbalanced, i.e., the number unsecured scenarios is much smaller than the number of secured scenarios, conventional online Adaboost may result in the high misdetection rate. Hence, we propose an online Adaboost algorithm that can adapt itself to the unbalanced online data. In addition, efficient tradeoff between misdetection (i.e., failing to detect unsecured samples) and false alarm (i.e., classifying secured samples wrongly) can be achieved by tuning a design parameter. Numerical results show that our proposed scheme can achieve high security assessment efficiency and accuracy, which is potential for the advanced security monitoring application in future smartgrid.