Cost Sensitive Class-Weighting Approach for Transient Instability Prediction Using Convolutional Neural Networks

Abstract

Transient instabilities triggered by critical faults can lead to rapidly developing blackouts in power systems. With the data driven methods using synchrophasor measurements collected from PMUs, it is possible to predict the instabilities just after the clearance of a critical fault. In this study, the performance of a classifier to be used for an early prediction of transient instability is enhanced by modifying the loss function used during the offline training phase. This study proposes to assign weights to the terms in the binary crossentropy loss function associated with each class, as misclassification of different classes generally causes different costs to the utility and its customers. The proposed method is able to determine the optimum values of the weights according to a pre-selected tolerance value, which represents how far the accuracy is away from being acceptable. The efficacy of the proposed method is examined on the 127-bus WSCC test system.