Decision Fusion-Based Non-Intrusive Load Identification Involving Adaptive Threshold Event Detection

Abstract

Non-intrusive load monitoring (NILM) is an important measure to improve the intelligence level of the power demand side. Existing NILM methods have poor performance in identifying low-power devices with similar power, with the increasing diversity of household loads and the wide range of load fluctuations. This paper proposes a fusion-based load identification method for residential loads, considering the electrical characteristics of different load types. In the first stage, the adaptive threshold Cumulative Sum (CUSUM) algorithm is innovatively adopted to reduce the misjudgment of local high-power device switching fluctuations and the missed events of local low-power load operation in the global threshold. In the second stage, the minimum Bayesian decision fusion loss function is used to calculate the cost function of Voltage Current (UI) trajectory, power, and total harmonic distortion, which are input into the Softmax multi-classification regression model in parallel. The category corresponding to the prediction made by the minimum loss function is considered as the final output. Finally, the effectiveness of the proposed method in identifying multiple types of household loads was verified through experiments on the Plug-Level Appliance Identification Dataset (PLAID) dataset.