**1. Introduction**

Non-Intrusive Load Monitoring (NILM)—equally referred to as load disaggregation— aims to identify the individual power consumption or on/off state of electrical loads by relying exclusively on the measurement of the aggregated consumption of these loads. The term was coined by Hart in their seminal works [1,2], that initiated the NILM research field. As the term non-intrusive suggests, NILM is motivated by applications where metering of single appliances is necessary, but not feasible with conventional measurement devices, e.g., because of cost considerations or obtrusiveness. Potential NILM applications include, for example, user feedback for energy saving purposes, overseeing activities of daily living for the health assessment of elderly people, or demand managemen<sup>t</sup> (see, e.g., [3–6]).

Before proceeding, it is essential to clarify some nomenclature typical of the NILM jargon and how this relates to the data used. Electric meters internally sample voltage and current signals at frequencies significantly greater than the base frequency of alternate current (AC). Meters can either output these raw data directly or averaged values such as, e.g., root mean square (RMS) voltage, current, power, or total harmonic distortion (THD) are calculated and output at lower frequencies. In the NILM literature, the following terms are used (We are aware that the given definition is slightly diverging from the conventionally used threshold of 1 Hz (e.g., [3]). However, we feel 1 Hz is somewhat arbitrary, whereas the provided definition seems to be a natural splitting point.):

**Citation:** Huber, P.; Calatroni, A.; Rumsch, A.; Paice, A. Review on Deep Neural Networks Applied to Low-Frequency NILM. *Energies* **2021**, *14*, 2390. https://doi.org/10.3390/ en14092390

Received: 9 February 2021 Accepted: 16 April 2021 Published: 23 April 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).


The advantage of using high frequency data should be quite obvious, since these preserve the entire signals and therefore allow us to extract the maximum information content. This is confirmed by various works, e.g., ref. [7] showed that, using raw data sampled at 1 MHz, it is even possible to distinguish between two appliances of the same type. Nevertheless, the cost of gathering high frequency data constitutes a major obstacle: Dedicated infrastructure is needed, and this is expensive both in terms of the hardware itself and the extra installation effort. On the other hand, the loss in information intrinsic in low frequency features is offset by the tremendous ease with which those data can be collected. Indeed, around the year 2010, the European Union and the US started to mandate and actually roll-out smart meters [8]. The advanced versions of these meters can export low frequency data to the outside (e.g., the meter E450 from Landis+Gyr has this capability). The pervasive roll-out of smart meters will therefore unlock all those applications which can benefit from NILM and which can be tackled using low frequency approaches.

Since its inception, the NILM research field has become quite diverse. The classic approach to low frequency NILM—as proposed in [1,2]—is event based. Simply stated, this means that the aggregate power series is first analyzed to find raises or drops that indicate device switching events, and these events are subsequently assigned to certain appliances. Later, one major avenue of NILM research used different variants of Hidden Markov Models, e.g., [9–12]. Over the years, the set of methods that have been applied to NILM has become extremely rich. A comprehensive overview is given in [13]. With the recent enormous success that deep learning has found in the vision and natural language processing domains, it was only a matter of time until deep neural networks (DNN) were also applied for the first time to NILM; this started in 2015 [14,15]. Since then, the number of DNN approaches to solve the NILM problem increased rapidly, as can be seen in Table 1.

**Table 1.** Number of DNN-NILM publications based on low frequency data per year. Numbers are compiled based on Table 2. That means the number for the year 2020 corresponds to the publications until end of November, see Section 1.3.


Over the years, different publications have surveyed the NILM literature from various angles: In [3], the authors focused on NILM feature types for low and high frequency data and touched algorithms and evaluation metrics. The authors of [16] proposed a taxonomy of appliance features and compared the reported performance of six classes of supervised and unsupervised NILM approaches. The authors of [17] surveyed unsupervised NILM algorithms and discussed the reported performance of eleven approaches. A very comprehensive review on inductive algorithms, employed feature sets, and the state of multi-label NILM classification approaches has been compiled by [13]. More recently, the authors of [18] discussed available public NILM datasets, and employed NILM performance metrics, tools, and frameworks, and corresponding limitations and challenges. Ref. [19] touched on dataset complexity and compared the reported performance of eight NILM approaches under this viewpoint. Three of these approaches employed DNN approaches on low-frequency data. Finally, the authors of [20] summarized advancements on HMM and DNN approaches. With respect to DNN approaches, they summarized the work performed in [21,22]. While the previously mentioned survey papers compared performance as it was reported by the original authors, very recently, two works compared classical and DNN approaches under identical conditions [23,24]. The authors of [23] presented an extension to the NILM tool kit (NILMTK)

library [25,26], with the exact purpose of simplifying direct comparisons, and [24] used this new functionality for an extensive comparison of eight available algorithms.
