The power quality indices corresponding to voltage event are provided by IEC 61000-4-30 and IEEE 1564 standards. The popularization of machine learning techniques in modern and smarter power grids inherently characterized with randomized big data motivates their inclusion. We aimed to establish key voltage disturbance event features that a machine learning model can use to attain a high classification and prediction accuracy. The feature extraction was achieved using the singular value decomposition and wavelet transform of a three phase system represented as a space phasor model with the corresponding ellipse parameters and the shape index feature extracted. The 6 by 500 dataset with the six (6) features (rms, event duration, 3^rd harmonic, 5^th harmonic, 7^th harmonic, shape index) was used to classify voltage sag, swell, interruption, harmonics, and normal conditions. MATLAB R2021a classification learner application was used to create a supervised machine learning model by training twenty nine (29) classifiers and comparing their performance using the accuracy (%), confusion matrix, total cost validation, prediction speed and training time. The results indicated that without the shape index, the voltage sag and voltage interruption events were misclassified in some scenarios for all the classifiers, with the highest accuracy obtainable of 99.6% by four (4) classifiers. The inclusion of shape index feature improved the trained models to 100% classification accuracy for fourteen (14) classifiers. The training, computing and processing is required to be of high performance and accuracy to increase the online situational awareness of network operators when such systems are implemented.

Keywords

Decomposition, Shape Index, voltage event and supervised learning.