Cryptocurrency Mining Detection Based on Time Series Clustering of Electricity Consumption Volume

A method with community detection in electricity consumption volume was developed to recognize suspected cryptocurrency miners from power-intensive users in China in authors' previous work. However, the detection result is parameter-sensitive and highly relies on selection of distance threshold value. The sensitivity can be resulted from either the clustering method or the distance metric. It has been found that the mining activity in China is featured with strong seasonality and can be detected with clustering-based methods. On this basis, in this paper, different clustering methods as well as different distance metrics are investigated for less parameter-dependence. As for clustering method, community detection, K-means clustering and hierarchical clustering are investigated. As for distance metrics, Pearson correlation-based distance, Euclidean distance and dynamic time warping (DTW) distance are investigated. The candidate method combinations are tested in Sichuan and Xinjiang cases and the detection effect as well as parameter dependency is compared. It is found that hierarchical clustering with DTW distance metric can generate accurate and stable results without parameter selection. The method can be applied for cryptocurrency detection in more general situations.