*1.1. Contributions*

Some of the mentioned review papers include DNN-based NILM approaches. However, what is missing so far is a comprehensive and structured overview on the ideas and findings in the field of DNNs applied to low frequency NILM and, based on that, a discussion on the usefulness of DNN architectural elements, input features or multi-task learning as well as research gaps particularly for applied deep learning-based NILM. In this paper, we intend to fill this gap. Therefore, the main contributions of this work are:

	- **–** We compare the performance of approaches and extract common features of best performing approaches in Section 4.1;
	- **–** We discuss the possible role of multiple input features and multi-task learning on NILM performance in Sections 4.2 and 4.3, respectively;
	- **–** We illustrate the importance of parameter studies in Section 4.4, and
	- **–** We outline major research gaps concerning the application of deep learning for NILM in Section 4.5.

Thereby, we hope that the interested reader will quickly identify the relevant literature for their own research and that our contributions will inspire new research activities, and thus ultimately advance the entire research field.
