Applying Machine Learning Algorithms for Classifying Time-Frequency Failures in Power Grid Systems

In power grid systems, Power Quality (PQ) disturbances affect manufacturing process, cause malfunction of equipment and induce economic losses. This paper presents ten new features to identify PQ disturbances such as voltage sag, swell, interruption, harmonics and combined defaults. At first, Hilbert Transform (HT) and Phase Locked Loops (PLL) techniques are applied to estimate the frequency and phase of harmonic components of voltage signals in real time. Then new descriptors, i.e., features, based on Time-Frequency (TF) representations are used. These TF features are obtained from the R´enyi and Shannon entropy obtained with the Short-Time Fourier Transform (STFT), the Stockwell Transform (ST) and the Optimized Stockwell Transform (OST). In order to evaluate the proposed TF descriptors, machine learning algorithms are applied to effectively discriminate the different types of disturbances. Classification results show an accuracy of more than 99.4% even in 5 dB SNR high-noise condition.