Review on Deep Neural Networks applied to Low-Frequency NILM

: This paper reviews non-intrusive load monitoring (NILM) approaches that employ deep 1 neural networks to disaggregate appliances from low frequency data, i.e. data with sampling rates 2 lower than the AC base frequency. The overall purpose of this review is ﬁrstly, to gain an overview 3 on the state of the research up to November 2020 and secondly, to identify worthwhile open research 4 topics. Accordingly, we ﬁrst review the many degrees of freedom of these approaches, what has 5 already been done in literature, and compile the main characteristics of the reviewed publications in 6 an extensive overview table. The second part of the paper discusses selected aspects of the literature 7 and corresponding research gaps. In particular, we do a performance comparison with respect to 8 reported mean absolute error (MAE) and F 1 -scores and observe different recurring elements in the 9 best performing approaches, namely data sampling intervals below 10 s, a large ﬁeld of view, the 10 usage of generative adversarial network (GAN) losses, multi-task learning, and post-processing. 11 Subsequently, multiple input features, multi-task learning and related research gaps are discussed, the 12 need for comparative studies is highlighted, and ﬁnally, missing elements for a successful deployment 13 of NILM approaches based on deep neural networks are pointed out. We conclude the review with 14 an outlook on possible future scenarios. 15

and all conclusions are deduced solely from these publications. The aggregate active power x a t of a set of appliances measured at time t can be formally defined as: where y m t are the contributions of individual appliances m that have been metered at the time of data 146 acquisition and M is their total number. The sum over k corresponds to the contribution of K further 147 appliances w k t not sub-metered during the measurement campaign. t is a noise term originating 148 from the measurement equipment. In the literature, the NILM problem is typically stated such that 149 the noise term e t includes the sum over non measured equipment. We explicitly separate the two 150 contributions as their nature is quite different. We can assume that the measurement noise t is well 151 behaved, i.e., it follows approximately a standard distribution and is small compared to the actual 152 signal. On the contrary, no such assumption can be made about the term ∑ w k t . The contribution from 153 non sub-metered appliances w k t typically amounts to a major part of x a t and the power distribution is   used in the reviewed DNN-NILM literature, see  table 3. Datasets closer to the top have been employed in more studies. Type indicates the type of the dataset: R → residential, R s → 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 [40]. The authors plan to release the dataset. dataset. This scenario tests the transferability of the tested approach to an even more diverse setting 235 as in the unseen case: Data could have been metered by different electrical meters or could originate 236 from a different country. To our best knowledge, this scenario has only been investigated in [55][56][57][58][59]. 237 The different scenarios are illustrated in Fig. 2. The column 'Evaluation Scenario' in table 3 lists the 238 scenarios employed for the reviewed references. Figure 2. Different NILM evaluation scenarios: seen: the algorithm is evaluated on new data from a house that was already available during training. unseen: the algorithm is evaluated on data from a house not seen during training. cross-domain transfer learning: the algorithm is evaluated on data from a different dataset.   samples hinder training and the author used synthetic training data composed from sets of more than 406 7 appliance sub-meters.

407
A key element of DNN optimization is the loss function that guides the optimization process.

408
The vast majority of works employ either the mean absolute error (MAE) or the mean squared error The performance of NILM algorithm is assessed in various ways. The interested reader is referred to [18,54]: [18] provides a comprehensive review and discussion of employed metrics and [54] proposes a set of metrics to assess the generalization ability of NILM aglorithms. In the following, we only repeat the definition of the mean absolute error (MAE) and the F 1 -score. In the reviewed literature, these were the most encountered metrics to assess the estimated energy consumption and on/off status of an appliance and we use them for our comparison in section 4.1.
where the sum goes over T time steps and y t ,ŷ t correspond to the measured and estimated power 461 consumption, respectively. In this publication, we use Watts as the unit for the MAE.

462
F 1 = 2 · P · R P + R      One of the basic questions that accompanied us throughout this literature review was: "What 469 is the most promising approach or classes of approaches?" As the last section hints, there is no 470 straightforward answer to that. Too many degrees of freedom (see Fig. 1) make the approaches differ      In some European Countries, such as e.g. Switzerland, residential power supply arrives in 731 three phases at the master distribution board (breaker panel) and is then split into single phases.

732
As a consequence, measurements from the electrical metering infrastructure are in principle also 733 available on three phases. With respect to a practical NILM application, this additional information 734 makes the problem at first glance easier to solve, as there are on average one third as many devices 735 connected to each phase compared to households attached to a single phase. However, the challenge 736 comes in the form of multi-phase appliances such as heat pumps, pool pumps, electrical heat storage 737 radiators, or charging station for electrical vehicles. These appliances require NILM algorithm to 738 combine information from all three phases. When considering an approach that should perform on any