A comprehensive harmonic monitoring program has been designed and implemented on a typical medium-voltage distribution system in Australia. The monitoring program involved measurements of the three-phase harmonic currents and voltages from the residential, commercial, and industrial load sectors. Data over a three year period have been downloaded and available for analysis. The large amount of acquired data makes it difficult to identify operational events that significantly impact the harmonics generated on the system. More sophisticated analysis methods are required to automatically determine which part of the measurement data are of importance. Based on this information, a closer inspection of smaller data sets can then be carried out to determine the reasons for its detection. In this paper, we classify the measurement data using data mining based on the minimum message length technique comprising unsupervised learning to assign cluster designation and supervised learning to describe the generated clusters and to predict the occurrences of unusual clusters in future measurement data. The clusters obtained from the unsupervised learning process provide the engineers with a rapid, visually oriented method of evaluating the underlying operational information contained within the clusters.