Nonintrusive load monitoring in residential households with low-resolution data Academic Article uri icon

abstract

  • © 2019 Elsevier Ltd Detailed information on individual appliance consumption is beneficial for improving energy efficiency and managing demand response. Nonintrusive load monitoring (NILM) aims to estimate the device-level energy consumption from the load data of an entire household. Because the majority of households can only provide load data at a normal smart-meter level, this paper introduces a novel similar time window (STW) algorithm to perform NILM with lower-resolution data. Derived from k-nearest neighbors (kNN), the proposed STW algorithm compares both the time and frequency domain similarities between windows of interest and historical data segments, and then selects the most similar time windows by instance-based learning to determine the device-level energy consumption. The desirable features of this algorithm include (1) reductions in the costs of and requirements for sensing equipment, (2) improvements in privacy preservation, and (3) a significant enhancement of the computational efficiency. To facilitate the selection of the data resolution and to satisfy the NILM application requirements in a cost-effective way, the paper also investigates the relationship among the input/output data resolution, time window length and prediction accuracy. To enable the generalizability of this algorithm, a cross-prediction approach is proposed to obtain the device-level consumption from a “library” of a group of households, without knowing each one's own historical data. Simulation results using four real-world public datasets demonstrate the competitive performance of the proposed STW algorithm with respect to traditionally used approaches for low-resolution NILM.

author list (cited authors)

  • Shi, X., Ming, H., Shakkottai, S., Xie, L. e., & Yao, J.

complete list of authors

  • Shi, Xin||Ming, Hao||Shakkottai, Srinivas||Xie, Le||Yao, Jianguo

publication date

  • October 2019