MANTRA: a multi-appliance transformer for Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM) enables the estimation of individual appliances' energy footprints from the total household consumption without the cost and complexity of installing a specific smart meter per device. By providing users with real-time feedback on their energy consumption, NILM is known to foster conservation habits. The widespread diffusion of smart home meters has enabled the unprecedented availability of a vast amount of consumption data observed at short intervals, posing an urgent need for designing deep learning models that can capture insightful long-range dependencies and overcome the computational cost of training and fine-tuning the same neural network for each device one at a time, as currently performed by single-appliance solutions for NILM. To cope with these challenges, we propose a sequence-to-sequence Multi-AppliaNce TRAnsformer (MANTRA) model that extracts long-term information and simultaneously estimates individual appliance powers, providing a scalable solution for NILM in real-world settings. Experiments on several time resolutions prove the robustness of MANTRA for multi-target energy disaggregation over other deep learning methods.