A new dynamic security assessment framework based on semi-supervised learning and data editing

In this paper, we propose a new online dynamic security assessment (DSA) framework based on semi-supervised learning and data editing. To reduce the number of labeled samples used by supervised learning in conventional DSA, which is required to ensure a high generalization performance of a classifier, we augment the training set with a large number of unlabeled samples that are easily computed. As an alternative to computationally expensive time-domain simulations, the unlabeled samples are labeled by an algorithm called tri-training. To reduce the noise that comes with incorrectly labeled samples, we use data editing, which significantly improves the classification performance. We demonstrate the performance of the proposed framework in a case study using the IEEE 39-bus New England test system with different levels of wind penetration. The results show that the proposed DSA framework reduces the number of labeled samples required to train the neural network used as an online transient stability classifier, which significantly reduces the computational burden associated with the training of the classifier.