A First Approach using Graph Neural Networks on Non-Intrusive-Load-Monitoring

Abstract

Non-Intrusive Load Monitoring (NILM), equally referred as energy disaggregation, aims to identify the individual power of each appliance by relying exclusively on the aggregated household signal. Within this paper we propose a new seq2seq approach which uses Graph Neural Networks (GNN) as encoder and a Transformer based decoder. To the best of our knowledge this is the first attempt which models the problem of NILM using graphs. Specifically, our approach uses Graph Convolutional Networks (GCN) to encode the aggregated signal into distinct graph nodes based on the entropy metric of the corresponding appliance's signal representing its different operational states. Using this approach we create encodings that incorporate time-invariant dependencies within the aggregated signal forming a more holistic and meaningful representation of our data. Experimental results indicate that our approach produce satisfying or even better results on multi-state appliances like washing machine compared to state of the art methods. Although our approach performs well only on multi-operational devices, it gives an interesting dimension to the problem and proposes the first graph based approach towards NILM leading the way for applying more GNN approaches on this domain.