Power Quality Disturbance Classification Using Deep BiLSTM Architectures with Exponentially Decayed Number of Nodes in the Hidden Layers

In recent years, there is growing interest in automatic power quality disturbance (PQD) classification using deep learning algorithms. In this paper, the average of instantaneous frequency and the average of spectrum entropy were used as time-frequency based feature extraction due to its discriminatory nature. Bidirectional Long Short-Term Memory (BiLSTM) architectures with exponentially decayed number of nodes in deep multilayers were utilized as Deep Recurrent Neural Network (DRNN) classifier. We experimentally generated fifteen classes of synthetic PQD signals. Each class contains 1000 samples divided randomly into training, validation, and testing. Results showed that four hidden layers of BiLSTM with exponentially decayed nodes interleaved with dropout layers provided the best classification accuracy of 99.23%.