Federated Learning Based Sparse Coding for Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM), also referred to as Energy Disaggregation, is a family of methods that predict the appliance consumption signals of a building using only its main aggregate signal. One of the methods proposed to solve the NILM problem is Sparse Coding (SC), as it is able to find hidden representations of consumption signals. Combining SC with NILM provides a model which produces feature encodings that best represent the appliance signals in their relation to the aggregate signal. However, this method requires access to both aggregate and appliance signals of the house, which can potentially risk the safety of the users' privacy since their activity can be predicted by an ill-intentioned model. This paper proposes a new framework called Federated Learning Discriminative Disaggregation Sparse Coding (FL-DDSC) based on federated learning and sparse coding. This model maintains a good level of sparse code-based disaggregation performance while maintaining user privacy. We test our model on two public datasets: REFIT and REDD. We show that our framework is able to provide similar performances to regular DDSC and only loses a small amount of information, which confirms the validity of a distributed approach to sparse coding-based disaggregation.