Classification of Power Quality Disturbances Using Convolutional Network and Long Short-Term Memory Network

The Electrical Power Quality (PQ) studies are commonly related to disturbances that alter the sinusoidal voltage features and/or current wave shapes. The classification approaches of electrical power quality disturbances found in the literature mainly consist of three steps: 1) signal analysis and feature extraction, 2) feature selection and 3) disturbances classification. However, there are some problems inherent in disturbances classification. The manual extraction of features is an imprecise and complex process, which can influence the resuits and, therefore, does not deal well with noisy signals. This paper proposes an approach based on Deep Learning using the raw data, without pre-processing, manual extraction or manual feature selection of the PQ disturbances signals for the classification of fifteen electrical power quality disturbances. A deep network is used, which consists of a hybrid architecture, composed by convolutional layers, a pooling layer, an LSTM layer, and batch normalization to extract features automatically. We adopted a 1-D convolution to adapt the input. The extracted features are used as input to fully connected layers, the last one being a SoftMax layer. The results are compared with state of the art methods based on the three steps, showing that the proposed approach had satisfactory performance even with noisy data.