Evaluation of non-intrusive load monitoring algorithms on open access datasets

Abstract

Non-Intrusive Load Monitoring (NILM), which aims to decompose the power usage behavior of individual devices from the total power signal, is an important technique for energy consumption analysis. However, in practice, NILM usually relies on low-frequency smart meter data. As the sampling frequency decreases, transient features and state transitions gradually become lost, which in turn affects decomposition accuracy and the evaluation of algorithm performance. This thesis investigates the effect of reduced sampling frequency on the evaluation results of representative NILM methods based on the UK-DALE dataset. The experiments compare optimization methods with deep learning methods and analyze the role of event-based optimization improvement in the combined optimization algorithm. All experiments are evaluated along three dimensions: event/state detection, point-wise error, and energy estimation. The results show that a decrease in sampling frequency changes both the relative performance between algorithms and the decomposability of different devices. Short-duration loads are more difficult to recognize events under low-frequency conditions, while cyclic and multi-state loads remain somewhat interpretable at the energy level. Therefore, algorithm performance should be evaluated using both error-based and energy-based metrics under very low-frequency conditions. The comparison of CO and CO+ further suggests that the event-based optimization method primarily improves error and energy metrics and weakens as sampling becomes coarser. Overall, this study reveals the role of low-frequency conditions in reshaping NILM assessment and informs method selection in constrained sampling environments.