A highly accurate NILM: With an electro-spectral space that best fits algorithm’s national deployment requirements

Netzah Calamaro, Moshe Donko, Doron Shmilovitz

Research output: Contribution to journalArticlepeer-review


The central problems of some of the existing Non-Intrusive Load Monitoring (NILM) algorithms are indicated as: (1) higher required electrical device identification accuracy; (2) the fact that they enable training over a larger device count; and (3) their ability to be trained faster, limiting them from usage in industrial premises and external grids due to their sensitivity to various device types found in residential premises. The algorithm accuracy is higher compared to previous work and is capable of training over at least thirteen electrical devices collaboratively, a number that could be much higher if such a dataset is generated. The algorithm trains the data around 1.8 × 108 faster due to a higher sampling rate. These improvements potentially enable the algorithm to be suitable for future “grids and industrial premises load identification” systems. The algorithm builds on new principles: an electro-spectral features preprocessor, a faster waveform sampling sensor, a shorter required duration for the recorded data set, and the use of current waveforms vs. energy load profile, as was the case in previous NILM algorithms. Since the algorithm is intended for operation in any industrial premises or grid location, fast training is required. Known classification algorithms are comparatively trained using the proposed preprocessor over residential datasets, and in addition, the algorithm is compared to five known low-sampling NILM rate algorithms. The proposed spectral algorithm achieved 98% accuracy in terms of device identification over two international datasets, which is higher than the usual success of NILM algorithms.

Original languageEnglish
Article number7410
Issue number21
StatePublished - 1 Nov 2021


  • DSO— distributed system operator
  • E-V—electric vehicle
  • GMM—Gaussian mixture model
  • HGL—harmonic generating load (inspired from current’s physical components theory)
  • KDE—kernel density estimation
  • KNN—K-nearest neighbor
  • NILM—nonintrusive load monitoring
  • NIS—network information system
  • P-V—photo-voltaic
  • PCA—principal component analysis
  • RNN—recurrent neural network
  • SGD—stochastic gradient descent


Dive into the research topics of 'A highly accurate NILM: With an electro-spectral space that best fits algorithm’s national deployment requirements'. Together they form a unique fingerprint.

Cite this