MACHINE LEARNING BASED INTELLIGENT POWER QUALITY MONITORING AND DISTURBANCE CLASSIFICATION

Abstract

The quality of electrical power supplied to homes, industries, and commercial buildings has become a serious concern as the use of sensitive electronic equipment has grown. Voltage sags, swells, interruptions, harmonics, transients, flicker, and notches can cause equipment malfunction, data loss, overheating, and production stoppages. At the same time, the increasing use of power electronic converters, variable speed drives, and renewable energy sources has made such disturbances more frequent. This paper presents a machine learning based intelligent power quality monitoring and disturbance classification system that captures voltage waveforms, extracts signal features, and automatically identifies the type of disturbance present. The monitoring hardware consists of a step-down voltage transformer and a current sensor, followed by a signal conditioning circuit that scales and offsets the signal for a microcontroller analog-to-digital converter sampling at several kilohertz. Waveform windows of ten cycles are transmitted to a processing unit. To build a large and balanced training set, disturbance signals are also generated synthetically using parametric equations consistent with the IEEE 1159 recommended practice, covering normal supply, sag, swell, interruption, harmonics, oscillatory transient, impulsive transient, flicker, notch, and combined disturbances such as sag with harmonics, with white noise added at several signal-to-noise ratios. Signal processing forms the core of the feature extraction stage. The discrete wavelet transform decomposes each waveform into detail and approximation levels, and the energy, standard deviation, and entropy of the coefficients at each level are computed. The fast Fourier transform provides harmonic magnitudes and total harmonic distortion, and time domain features such as RMS variation, peak value, crest factor, and kurtosis are added. These features are used to train and compare k-nearest neighbours, a support vector machine, a decision tree, random forest, gradient boosting, and a multilayer perceptron. The models are evaluated using accuracy, precision, recall, F1 score, and confusion matrices across noise levels. Random forest and the support vector machine achieved the highest accuracy and retained good performance even at low signal-to-noise ratios, while wavelet energy features proved the most informative. A dashboard displays the live waveform, RMS voltage, frequency, total harmonic distortion, the classified disturbance type, and a log of events with time stamps and durations. The proposed system offers a low-cost alternative to commercial power quality analysers for use in academic laboratories, small industries, and distribution substations. It helps engineers identify the source and nature of disturbances and plan corrective measures such as filters and voltage regulators. Future work includes threephase monitoring, deployment of the classifier on the microcontroller itself, and the use of deep learning directly on raw waveforms.