Device Classification for NILM using FIT-PS compared with Standard Signal Forms

Dimensionality Expansion of Load Monitoring Time Series and Transfer Learning for EMS

“…HMM appears to perform well on sequential data thus are well suited method for solving NILM problems. HMM was soon outperformed by other methods of shallow machine learning like SVM [11], kNN [12], Naive Bayes, Logistic Regression Classifier and Decision Tree [13].…”

“…For solving NILM and ILM problems, several methods have been proposed some based on combinatorics [6], thresholding [7], shallow machine learning such as Hidden Markov Models (HMM) [8], [9], [10], SVM [11], kNN [12], Naive Bayes, Logistic Regression Classifier and Decision Tree [13] and, more recently, deep learning such as recurrent neural networks (RNN) [14], [15], [16], [17], convolutional neural networks (CNN) [18], [14] or autoencoders [14], [19]. While machine learning-based solutions often yield superior appliance recognition results, only few such techniques are verified across several domain specific datasets.…”

Non-Intrusive Load Monitoring Algorithm for PV Identification in the Residential Sector

“…For instance, Gillis et al presents in [8] a NILM algorithm based on discrete wavelet transformation for feature identification combined with a decision tree algorithm to identify four different loads during switching events, reporting an 96.65% accuracy. Additionally, Weisshaar et al proposed in [9] a NILM using Frequency Invariant Transformation of Periodic Signals (FIT-PS), active and reactive power as inputs of conventional classifier algorithms such as k-Nearest Neighbor, SVM and Naïve Bayes. Finally, some of the methodologies has been developed to identify photovoltaic systems in the distribution network.…”

A Nonintrusive Load Identification Model Based on Time-Frequency Features Fusion

“…In NILM, the steady-state characteristics of the active power P and the reactive power Q are commonly used in the early stage [2]. In order to offset the high-frequency information that is cancelled out when calculating active and reactive power, many features such as current waveforms [3], harmonics [4], [5], transient power waveforms [6], and switching transient waveforms [7] are combined on P and Q.…”

“…In order to offset the high-frequency information that is cancelled out when calculating active and reactive power, many features such as current waveforms [3], harmonics [4], [5], transient power waveforms [6], and switching transient waveforms [7] are combined on P and Q. In addition to combining P-Q waveform features [2], there are also other methods, such as combining VI trajectories, to convert power signals into images. Although these methods play roles in load identification, it is still necessary to improve the classification of resistive and multistate electrical appliances with similar waveforms.…”