Development of empirical mode decomposition based neural network for power quality disturbances classification

Abstract

The complexity of the electric power network causes a lot of distortion, such as a decrease in power quality (PQ) in the form of voltage variations, harmonics, and frequency fluctuations. Monitoring the distortion source is important to ensure the availability of clean and quality electric power. Therefore, this study aims to classify power quality using a neural network with empirical mode decomposition-based feature extraction. The proposed method consists of 2 main steps, namely feature extraction, and classification. Empirical Mode Decomposition (EMD) was also applied to categorize the PQ disturbances into several intrinsic mode functions (IMF) components, which were extracted using statistical parameters and the Hilbert transformation. The statistical parameters consist of mean, root mean squared, range, standard deviation, kurtosis, crest factor, energy, and skewness, while the Hilbert transformation consists of instantaneous frequency and amplitude. The feature extraction results from both parameters were combined into a set of PQ disturbances and classified using Multi-Layer Feedforward Neural Networks (MLFNN). Training and testing were carried out on 3 feature datasets, namely statistical parameters, Hilbert transforms, and a combination of both as inputs from 3 different MLFNN architectures. The best results were obtained from the combined feature input on the network architecture with 2 layers of ten neurons, by 98.4 %, 97.75, and 97.4 % for precision, recall, and overall accuracy, respectively. The implemented method is used to classify PQ signals reliably for pure sinusoids, harmonics with sag and swell, as well as flicker with 100 % precision