Application of Machine Learning Methods for Recognition of Daily Patterns in Power Quality Time Series

Power electronic devices cause harmonic distortion that can have a negative effect on both the grid and the consumer side. Therefore, network operators conduct extensive measurement campaigns to monitor power quality and to identify problems at early stage. Due to the tremendous amount of measurement data, manual inspection is usually limited to simple analysis. As a result, the majority of useful information in the data stays unused and automatic methods are needed to process the big amount of measurement data effectively and to analyze them in depth. In this paper, five machine learning methods are used for automatic classification of daily patterns in current and voltage harmonics. The methods are described and applied to four data sets. Methods performance is evaluated and compared using four indices. This paper shows that both supervised and unsupervised methods can be successfully applied to harmonic measurement data, however with some limitations.