Since the most important issue in the production of digital currencies is energy consumption, the use of illegal electricity in mining farms has become very popular. Illegal mining is particularly important in countries such as Iran where the price of electrical energy is extremely low. This issue has caused numerous problems such as frequent blackouts, large losses for industries and even daily power cuts in several large cities. Previous machine learning approaches for miner detection are mostly supervised methods which rely on labeled data. Due to the fact that the number of labeled data is very limited in reality, we propose unsupervised methods in this paper. A real data set from Markazi Province Distribution Company in Iran has been employed to produce the results. The classification process consists of two stages: in the first stage, Dynamic Mode Decomposition (DMD) has been used to extract new features which compose the set of features along with certain factors from the Advanced Metering Infrastructure (AMI). These features are selected for 58 subscribers with positive and negative labels. In the second stage, a number of unsupervised models are built from the results of the first stage. The highest accuracy of classification obtained is 74% from unsupervised algorithms and 85% for supervised algorithms, which is very significant considering the fact that unsupervised algorithms do not need labeled data.
