*3.5. Output*

With respect to their output, we observe four different dimensions along which DNN-NILM approaches can be distinguished: A first dimension is the number of time steps that are disaggregated by the DNN models in a single go, be it a sequence, subsequence, or a single value. This information is available through the abbreviations *s2s*, *s2sub*, and *s2p* in the column 'DNN Elements' of Table 2, see also Section 3.4.1. The second dimension concerns the type of inferred output. With the exception of [70], where location data (*location*) are combined with the aggregated power consumption, and [95], where a DNN infers state changes in the aggregate power (*stateChange*), the goal of DNN-NILM approaches is either to infer the on/off state or the energy usage of an appliance. We mark this information in column 'Output' of Table 2 with the abbreviations *on/off* and *P*, respectively. Naturally, this dimension is mostly coupled with the third distinction, whether the learning problem is formulated as a classification or regression task. However, there are four works [33,61,68,118] where power values are clustered into groups, and the power

regression problem is recast into a classification task. These references are marked with *Pclass*. Lastly, we can distinguish between approaches that learn on a single task and those learning on multiple tasks simultaneously, i.e., perform multi-task learning. The majority of approaches train one model for each appliance to be disaggregated. A sizable number of approaches infer, however, the on/off state or power disaggregation of multiple appliances simultaneously. These cases are marked with a subscript *m* in the 'Output' column of Table 2. Where multi-task learning is done on different modes, the corresponding outputs are jointed with an '&' in Table 2: [35,63,97,106] trained networks on both on/off and active power data. Ref. [117] used both active and reactive power of an appliance as target, and [59] used three targets, i.e., the aggregate power (P*total*), the appliance power (P*app*), and the difference between the two (P*rest*). Finally, ref. [46] took multi-task learning furthest by simultaneously learning on both on/off states and active power of multiple appliances.
