An unsupervised learning schema for seeking patterns in rms voltage variations at the sub-10-minute time scale

This paper proposes an unsupervised learning schema for seeking the patterns in rms voltage variations at the time scale between 1 s and 10 min, a rarely considered time scale in studies but could be relevant for incorrect operation of end-user equipment. The proposed framework employs a Kernel Principal Component Analysis (KPCA) followed by a k-means clustering. The schema is applied on 10-min time series with a 1-s time resolution obtained from 44 different periods of a location south of Sweden. Then, ten patterns are obtained by reconstructing the 10-min time series from each cluster center. The results of the proposed schema show a good separation of cluster centers. Moreover, some statistical power-quality indices are applied to the whole dataset, showing voltage variation between (0.5–3) V over a 10-min window. Obtaining the most suitable indices and applying them to the ten obtained cluster centers and their belonging time series shows that the existing statistical indices may not be enough to show a complete picture of the sub-10 min actual variations. This outcome shows the necessity of extracting 10-min patterns through our proposed schema besides the existing statistics to quantify the voltage variations, levels, and patterns together. Findings of this paper are: Not forgetting the sub-10-min time scale; The necessity of employing both statistics and the proposed schema; Extraction of ten typical patterns; The need for the statistics and patterns that are justified as changes in equipment connected to the grid; and compressing a huge amount of data from power-quality monitoring. The proposed schema is applied to a much less understood phenomena/disturbance type so that this work will result in general knowledge beyond the specific case study.