Non-intrusive load monitoring: Comparative analysis of transient state clustering methods

Abstract

Non-intrusive load monitoring is one of the key tools in demand-side management (DSM). Recent advancements in the computational power of processors have accentuated the role of machine learning algorithms e.g., clustering, as a key function in the NILM solutions applied on power grids. In event-based NILM methods, the algorithm detects the transient states (load events) and clusters them based on the similarity of different features of the transient state. In this study, the performances of eight clustering algorithms are comprehensively investigated and the impact of choosing different input signals, e.g., P, Q, and I, on transient states clustering is analyzed. Various input signals from the BLUED dataset are fed to the clustering algorithms. By comparing the evaluation metrics including shape-based and ground-truth-based metrics, it is observed that the OPTICS algorithm fed by dual-stream input streams outperformed the rest of the investigated clustering algorithms and input sets. OPTICS algorithm groups load events based on their density in multi-dimensional space, using a dynamic radius. The OPTICS algorithm, as the best-performing transient state clustering algorithm for the low-frequency NILM purpose, is then tested with the downsampled input data in a wide frequency range, to observe the impact of the data-sampling frequency on the results, which simplifies the use of clustering algorithms in future studies.