Misclassification Prediction for Transient Stability Assessment

Decisions of power system operators are dependent on indicators of power system stability that are generated near real-time. In this context identification of instability has to be fast and accurate in order to allow for timely response. Pretrained deep learning networks can provide such ultra-fast and accurate identification. However, even though the training is performed on a vast amount of data, a small number of borderline cases are usually classified erroneously. This issue can be resolved if the probability of misclassification is used as an additional metric. In this research on performing transient stability assessment, we propose to employ the classification deep learning methodology of per-sample misclassification prediction that is supported with theoretical guarantees originated from true class probability properties. We also show that the failure of stability prediction highly depends on the data window that is used to define the input data length for classification. The methodology is validated on the 39-bus IEEE and 179-bus WECC models.