Disentangling Time Series Representations via Contrastive Independence-of-Support on l-Variational Inference

We introduce a dataset comprised of energy consumption data from smart meters in French households, capturing detailed, disaggregated time series for various home appliances. This dataset covers a six-month period with a one-minute sampling rate across five different households. The objective of this dataset is to support the development of models that learn disentangled representations of time series energy data, which can significantly enhance model generalization across both in-distribution and out-of-distribution scenarios.

Our approach to disentangled representation learning does not assume independence among attributes, a common assumption that can lead to non-identifiability in real-world applications. Instead, we address the inherent correlations among time series attributes by employing contrastive learning and a singular autoencoder to establish a smooth bijection. This methodology, which we have named Disentanglement under Independence-Of-Support via Contrastive Learning (DIOSC), leverages l-variational inference layers with self-attention mechanisms. These layers are specifically designed to handle the complex temporal dependencies found in bottom-up and top-down network structures, enhancing the interpretability and fairness of the resulting models. Our findings indicate that DIOSC effectively improves the representation of time series electricity consumption, offering promising avenues for further research and practical applications in energy management.