*1.2. Scope*

The scope of this review are NILM approaches based on DNNs using low frequency data. In the remainder of this text, we use the term DNN-NILM to designate the corresponding approaches. The choice to focus on low frequency data in our review is motivated by our strong belief that many applications could benefit from NILM, coupled with our observation that low frequency data will most likely be the only one available at scale in the near future. In our vision, all households equipped with smart meters will soon be able to become fully energy aware, informing their inhabitants of which appliances are being used, how they are being used, and even whether they are behaving abnormally or about to fail. This latter point is known as predictive maintenance and is currently applied in industrial settings, but being able to detect billions of appliances which consume an abnormal amount of power would have a beneficial impact of our society and its carbon footprint. With our review, we therefore try to make a contribution to push forward the development and understanding of low frequency NILM.

The focus on DNNs is motivated firstly by their proven success in other domains, and secondly by their good performance in the NILM domain: Recently, traditional and DNN-NILM approaches have been compared under identical conditions in two works [23,24]. The authors found that each of the compared DNN approaches—with few exceptions— performed better than each of the classical approaches. In particular for multi-state appliances, the performance gap was found to be "rather discernible" [24]. Publications that use shallow neural networks with only a single hidden layer such as, e.g., [27–29], are not included in our review. We restricted ourselves to approaches that train neural networks with back-propagation, excluding alternative approaches such as, e.g., [30,31]. Since the scope involves DNNs and NILM, we assume that the reader is familiar with the general concepts of the two fields, and we will merely introduce the basic NILM problem formulation in Section 2.1. With respect to DNNs and deep learning, we will refer the reader to comprehensive books on the topic in Section 2.2.

*Energies* **2021**, *14*, 2390

**Table 2.** Reviewed references. Publications are sorted by year. Except for the starred publications, the sorting within a year is arbitrary. The table is available in Excel format on our GitHub account, see 'Supplementary Materials' for the link. Explanations with respect to specific columns follow. **Best**: Best performing, according to Section 4.1. **Datasets**: See Table 3 for details. **Setting**: *R* → residential, *I* → industrial, *C* → commercial. The columns FGE to WDR indicate if the specific appliance has been disaggregated in the reference. **FGE**: fridge, **DWE**: dishwasher, **MWV**: microwave, **WME**: washing machine, **KET**: kettle, **SOC**: stove/oven/cooker, **TDR**: tumble dryer, **HPE**: heat pump, **WDR**: washer-dryer. **Further Appliances**: Additional appliances not listed in the previous columns. **E.Sce.**: Evaluation Scenarios; *sn* → only seen scenario evaluated, *usn* → additionally unseen scenario evaluated, *ctl* → also cross-domain transfer learning evaluated. **Aug.**: Data Augmentation; *dn* → use synthetic training data, *yes* → data augmentation employed. **Input**; *P* → active power, *Q* → reactive power, *I* → current, *S* → apparent power, *P*2*D* → active power window transformed into 2D representation, *P-S* → difference between active and apparent power, Δ*P* → first-order difference of the active power signal, *PF* → power factor, *TofD* → time of day, *WE* → week or weekend day, *DofW* → day of week, *MofY* → month of year, *Tout* → outdoor temperature, *Pvar* → variant power signature [32,33], *na* → see Section 3.3.1. **DNN Elements**: See Section 3.4.1 for the meaning of the various employed abbreviations. Comma separated descriptions refer to different trained models. **Output**: Comma separated descriptions refer to different trained models. Elements connected with an & indicate that a DNN has several outputs of a different type. Similarly, the subscript *m* means that the DNN provides the identical output for *multiple* different appliances. *on/off* → on/off status of appliance, *Pclass* → class of active power. Please refer to Section 3.5 for details concerning *Papp*, *Ptotal*, *Prest*, *location*, and *stateChange*. **Code**: (electronic version) link to code repository as indicated in the reference or found through a very shallow google search.


8


*Energies* **2021**, *14*, 2390

*Energies* **2021**, *14*, 2390







1 GitHub page of first author; 2 Experimental framework available upon request.

**Table 2.** *Cont.*

*Energies* **2021**, *14*, 2390

As the DNN-NILM literature reviewed contains only three publications using data from non-domestic settings (two commercial, one industrial), this distribution means that our review concentrates mainly on domestic NILM.
