Non-intrusive load monitoring (NILM) is a low-cost alternative to appliance level sub-metering, that leverages signal processing and machine learning techniques to estimate the power consumption of individual appliances from whole-home measurements. However, the difficulty associated with obtaining training data sets for the commonly used supervised NILM classification algorithms is a major obstacle in wide commercial adoption of the technology. The diversity of electrical load signatures (patterns of appliances' power draw) demands in-situ training (labeling of the signatures), which often needs to be performed by ordinary users through usersystem interaction. To produce the example signatures required for training, continuous interaction with users might be required, which could reduce the success of the training process due to user fatigue. Pre-populating the training data set could help facilitate the process by reducing the number of user-system interactions needed for labeling. Taking into consideration all the issues described above, a study to test the feasibility of autonomous clustering of similar appliances' signatures based on hierarchical clustering was investigated. The information contained in the structure of the binary cluster tree was used for clustering without the need for a priori selection of the number of clusters. The assessment, carried out on data collected from a residential setting, showed promising results (with accuracy above 90%, calculated based on the ground truth labels) supporting the feasibility of the approach for unsupervised clustering.