DISCOV: A Time Series Representations Disentanglement via Contrastive for Non-Intrusive Load Monitoring (NILM)

Improving the generalization capabilities of current machine learning models and improving interpretability are major goals of learning disentangled representations of time series. Notwithstanding this, methods for disentangling time series have mainly focused on identifying independent factors of variation in the data. This overlooks that the causal factors underlying real-world data are often not statistically independent. In this paper, we investigate the problem of learning disentangled representations for the electricity consumption of customers’ appliances in the context of Non-Intrusive Load Monitoring (NILM) (or energy disaggregation), which allows users to understand and optimise their consumption in order to reduce their carbon footprint. Our goal is to disentangle the role of each attribute in total aggregated consumption. In contrast to existing methods that assume attribute independence, we recognise correlations between attributes in real-world time series. To meet this challenge, we use weakly supervised contrastive disentangling, facilitating the generalisation of the representation across various correlated scenarios and new households. We show that Disentangling the latent space using Contrastive On Variational inference (DISCOV) can enhance the downstream task. Furthermore, we find that existing metrics to measure disentanglement are inadequate for the specificity of time series data. To bridge such a gap, an alignment time metric has been introduced to assess the quality of disentanglement. We argue that ongoing efforts in the domain of NILM need to rely on causal scenarios rather than solely on statistical independence. Code is available at https://oublalkhalid.github.io/DISCOV/.