Training an LSTM Voltage Sags Classificator on a Synthetic Dataset

In this paper, an end-to-end deep learning method for voltage sag classification using a Long Short-Term Memory (LSTM) neural network is proposed. The network receives an unmodified three-phase voltage signal sequence as an input and outputs a label to indicate if a signal is not faulted or there is one of the seven standard voltage sag types present, labeled from A to G respectively. A database of over 450,000 synthetically generated signals is generated for network training purposes. When tested on synthetic signals, the network achieves an accuracy of over 92%, while it reports lower accuracy on real-world signals. However, the network has demonstrated the ability to generalize on measured signals despite being trained on synthetic signals, and mainly correctly classifies signal segments of serious severities which present the opportunity to improve the classification accuracy of arbitrary measured signals by training the network on an even larger synthetic dataset. The source code following the paper is available at https://bitbucket.org/sara_e21/voltage-sag-classifier.