Automatic classification of power quality disturbances using optimal feature selection based algorithm

The development of renewable energy sources and power electronic converters in conventional power systems leads to Power Quality (PQ) disturbances. This research aims at automatic detection and classification of single and multiple PQ disturbances using a novel optimal feature selection based on Discrete Wavelet Transform (DWT) and Artificial Neural Network (ANN). DWT is used for the extraction of useful features, which are used to distinguish among different PQ disturbances by an ANN classifier. The performance of the classifier solely depends on the feature vector used for the training. Therefore, this research is required for the constructive feature selection based classification system. In this study, an Artificial Bee Colony based Probabilistic Neural Network (ABCPNN) algorithm has been proposed for optimal feature selection. The most common types of single PQ disturbances include sag, swell, interruption, harmonics, oscillatory and impulsive transients, flicker, notch and spikes. Moreover, multiple disturbances consisting of combination of two disturbances are also considered. The DWT with multi-resolution analysis has been applied to decompose the PQ disturbance waveforms into detail and approximation coefficients at level eight using Daubechies wavelet family. Various types of statistical parameters of all the detail and approximation coefficients have been analysed for feature extraction, out of which the optimal features have been selected using ABC algorithm. The performance of the proposed algorithm has been analysed with different architectures of ANN such as multilayer perceptron and radial basis function neural network. The PNN has been found to be the most suitable classifier. The proposed algorithm is tested for both PQ disturbances obtained from the parametric equations and typical power distribution system models using MATLAB/Simulink and PSCAD/EMTDC. The PQ disturbances with uniformly distributed noise ranging from 20 to 50 dB have also been analysed. The experimental results show that the proposed ABC-PNN based approach is capable of efficiently eliminating unnecessary features to improve the accuracy and performance of the classifier