Neural NILM Learning Paradigms: From Centralised to Decentralised Learning

Centralised vs collaborative learning

Non-intrusive Load Monitoring (NILM) has become a paramount in both industrial and residential sectors to achieve efficient energy consumption. Deep neural networks have been gaining the highest interest from the research community, commonly referred to as neural NILM. In most cases, neural NILM models follow a centralised based learning scheme, where the energy data is assumed to be available in a central node for training. This practice can, however, raise privacy and security concerns from the consumer’s side since energy data can reveal in-home activities and occupancy records if intercepted. In response, Federated Learning (FL) has been suggested as a viable solution to address these issues. In the paper "Neural NILM Learning Paradigms: From Centralised to Decentralised Learning", an overview of neural NILM models following both a centralised and a federated learning paradigm was presented while also identifying the main challenges with regard to both learning paradigms and potential future research directions for more robust, secure and privacy-preserving models in the neural NILM industry. Overall, as any other new technology, FL has its merits and limitations. Typically, FL provides promising perspectives to solve the privacy issues of energy disaggregation. However, it also opens doors for new challenges, especially those related to the (i) low disaggregation performance of FL-based NILM algorithms, (ii) susceptibility to noise, (iii) lack of labeled sub-metered data at the customer’s level, and (iv) need to adopt robust security mechanisms.

Further information can be found in the paper:

Hafsa Bousbiat, Christoph Klemenjak, Yassine Himeur, Wilfried Elmenreich, Abbes Amira, Wathiq Mansoor, and Shadi Atalla. Neural NILM learning paradigms: From centralised to decentralised learning. In Proceedings of the 2022 5th International Conference on Signal Processing and Information Security (ICSPIS), pages 138–142, December 2022. (doi:10.1109/icspis57063.2022.10002485)

The paper also won the best paper award at the 5th International Conference on Signal Processing and Information Security (ICSPIS) in December 2012.