**3. Literature Review**

When applying DNNs to NILM, many options are available: For example, what data to use, what DNN architecture to employ, how to evaluate the results, and so on. An illustration of these 'degrees of freedom' is given in Figure 1. The subsections below roughly follow the grouping done there. The aim of this section is to provide the reader with an overview on what has already been done in the literature in the scope of this review.

**Figure 1.** Illustration of the main degrees of freedom for DNN-NILM research. Colors indicate a loose grouping and should not be understood as a taxonomy.

#### *3.1. Datasets and Appliances*

DATASETS: The number of NILM datasets has been increasing over the last years, see [127,128] for recent overviews and [129–131] for the most recent published datasets we are aware of. In Table 3, we characterize only the publicly available datasets that have been used in the reviewed studies. The datasets at the beginning of the table are those more frequently used: Both UK-DALE and REDD were employed in approximately 40 and 30 studies, respectively, followed by AMPds, REFIT, and dataport each employed around 10 times. The ECO dataset is used three times, and other datasets are only used once or twice. The Industrial Machines Dataset (IMD) is to our knowledge the only available open industrial dataset. All remaining open datasets were measured in a residential setting. There are also a number of studies based on proprietary datasets measured in different settings: Nine residential, two commercial, one industrial, and one with dc-appliances. While not explicitly the scope of this review, the distribution of the employed datasets means that our review concerns mostly domestic NILM. Table 2 lists the datasets employed by each reviewed publication.

APPLIANCES: Appliances that have been disaggregated in the corresponding publications are listed in Table 2. The most investigated residential appliances in decreasing order are: dishwasher, fridge, microwave, washing machine, kettle, stove/oven/cooker, tumble dryer, HVAC, washer-dryer, heat pump, and light. Further electrical loads that appear fewer than ten times in the literature are given in the column 'Further Appliances' of Table 2. A few publications concentrated either on commercial or industrial applications using mostly proprietary datasets. These publications are marked in the column 'Setting' of Table 2.

**Table 3.** Main characteristics of the open datasets used in the reviewed DNN-NILM literature, see Table 2. Datasets closer to the top have been employed in more studies. *Type* indicates the type of the dataset: *R* → residential, *Rs* → synthetic residential, *I* → industrial. IMD is, to our knowledge, the only publicly available industrial dataset. '#H' and '#A' mean number of houses and appliances, respectively. '*Agg*' and '*Appl*' stand for 'aggregate' and 'appliance', respectively. For the IDEAL dataset, available information has been extracted from [98]. The authors plan to release the dataset.

