**5. Outlook**

Looking into the future, we can imagine different scenarios and directions for the (DNN-)NILM field. With the rapid development of the Internet of Things, we can well think of future appliances which are aware of their own current (and possibly future) energy consumption and feature a communication interface to relay this information to the outside world. In this scenario of energy-aware appliances, NILM would become obsolete. As this scenario would require a business case for appliance manufacturers and standards for interfaces and protocols, chances are good that this state will not be reached in the near future. We believe that the rapid increase of computing power in edge devices will have a much more immediate impact. Edge nodes will soon be able to perform DNN-NILM close to the meter without the need to transfer data to a cloud computing service. The culmination of this trend would be complex NILM algorithms that run directly on meter hardware, maybe even on the raw high frequency measurement data. Developing this scenario even further, one could imagine that NILM algorithms learn and improve on local data. For this to work, the learning problem will first have to be formulated in a way that the data available on the meter can be used for further improvements. A standard supervised training approach does not seem to be feasible. Furthermore, local improvements of the model will ideally also be made available to other smart meters. This concept of local learning with global exchange of improvements is a nascent research field called Federated Learning, see [189,190].
