Electricity theft is a great trouble for power companies. As the means of tampering with smart meters continue to increase, the electricity theft behaviours become more diversified and covert, which are difficult to be identified using the existing electricity theft detection method. In addition, the existing methods usually cannot estimate the economic losses caused by electricity theft. To address these issues, a combined unsupervised learning approach for electricity theft detection and loss estimation is proposed in this study. First, three anomaly measurement indexes including the mean index, fluctuation index, and trend index are proposed to capture different anomalies respectively. Then, based on historical electricity consumption data, we develop two unsupervised learning techniques including the sample-to-subsamples decomposition algorithm and clustering algorithm to obtain the typical ranges of index values, and the load samples whose index values are not in the typical ranges will be considered fraudulent. Furthermore, three anomaly measurement indexes are combined to judge whether the load sample is fraudulent, and the user whose most load samples are judged fraudulent will be considered as an electricity thief. Finally, an economic loss estimation method is proposed, which quantifies the losses of electricity theft. Numerical experiments are carried out based on the Irish smart meter dataset, and the results demonstrate the effectiveness and the superior performance of the proposed method compared with a series of electricity theft detection methods.