Detection and extraction of optimal features from power quality disturbances based on wavelet coefficient reconstruction and noise-robust methods

Abstract

The Discrete Wavelet Transform (DWT) is a well-established technique for detecting power quality disturbances, particularly transients. However, its practical implementation is often limited by its high sensitivity to noise. The downsampling process in Multiresolution Analysis (MRA) introduces aliasing distortion, which can be mitigated by carefully selecting decomposition and reconstruction filters. As a result, the accurate reconstruction of approximation and detail components is achieved, facilitating the extraction of key features essential for disturbance classification. This study proposes a noise-robust methodology that effectively filters noise while preserving interference characteristics for detecting disturbances with magnitude variations, such as sags, swells, and interruptions. These disturbances exhibit similar spectral and duration characteristics, making them difficult to distinguish from the pure signal, which has a fundamental frequency of 60 Hz and a magnitude variation of $+-10\text{\%}$. The approach employs DWT with MRA, followed by the reconstruction of approximation coefficients using the Inverse Discrete Wavelet Transform (IDWT). Feature extraction methods, including peak detection via local maxima identification, are combined with noise-robust techniques like Shannon entropy and energy to detect oscillatory transients, harmonics, flicker, notches, and complex disturbances, such as sag and swell with harmonics and sag and swell with oscillatory transients. The extracted feature vectors, consisting of eight elements, were used as input for six machine learning classifiers. Experimental results demonstrate that the proposed detection and feature extraction techniques achieve classification percentages exceeding 99%, even under varying noise conditions, allowing for the differentiation of each disturbance type. Most importantly, they enable clear distinction from the pure signal.