With widely deployed smart meters, non-intrusive energy measurements have become feasible, which may benefit people by furnishing a better understanding of appliance-level energy consumption. This work is a step forward in using graph signal processing for non-intrusive load monitoring (NILM) by proposing two novel techniques: the spectral cluster mean (SC-M) and spectral cluster eigenvector (SC-EV) methods. These methods use spectral clustering for extracting individual appliance energy usage from the aggregate energy profile of the building. After clustering the data, different strategies are employed to identify each cluster and thus the state of each device. The SC-M method identifies the cluster by comparing its mean with the devices’ pre-defined profiles. The SC-EV method employs an eigenvector resultant to locate the event and then recognize the device using its profile. An ideal dataset and a real-world REFIT dataset are used to test the performance of these two techniques. The f-measure score and disaggregation accuracy of the proposed techniques demonstrate that these two techniques are competitive and viable, with advantages of low complexity, high accuracy, no training data requirement, and fast processing time. Therefore, the proposed techniques are suitable candidates for NILM.
Widespread inculcation of smart meter data in modern grid is motivating stakeholders to utilize it for demand response management and achieving energy sustainability goals. One of the methods being used in regard is Non-Intrusive Load Monitoring (NILM), for disaggregating individual devices from a combined load signature. This study combines two spectral clustering strategies using voting based spectral clustering technique in such a way as to achieve the benefits of both parent strategies. The voters in the consensus are taken to be the accuracies achieved using Spectral Cluster-Mean (SC-M) and Spectral Cluster-Eigen Vector (SC-EV) algorithms with different window sizes to achieve diversity. Currently, Spectral Clustering for NILM has been used by few research works and so far, no one technique has achieved higher accuracy in detecting various kinds of devices. The proposed strategy was evaluated on real world data set (REFIT). The results have shown enhanced overall performance by up to 6%. An in-depth analysis of various tuning parameters of SC-M and SC-EV is also presented. These novel contributions increase the feasibility of spectral clustering and voting based consensus clustering for NILM and may open further avenues of research in this direction.