Metadata-enhanced NILM for appliance-level disaggregation across diverse residential buildings

Abstract Non-Intrusive Load Monitoring (NILM) offers a cost-effective way to enhance smart meter data by providing behind-the-meter insights into appliance-level energy use, without invasive monitoring. However, NILM models face challenges in scalability and generalization due to diverse household and appliance characteristics. This study develops a structured framework to evaluate the generalizability of NILM algorithms across varied residential contexts, using controlled experiments with the REFIT dataset. The framework integrates household metadata to assess whether contextual features improve model transferability across homes. Training configurations varied by: (i) training set size (number of houses), (ii) metadata-based grouping (low/high occupancy, small/large house size, low/high appliance ownership vs. baseline (random)), and (iii) seen vs unseen house configurations. Disaggregation performance was assessed for two contrasting appliances: fridge (consistent usage) and kettle (occupant-driven). Results show that increasing training set size improved or stabilized performance initially, but beyond a certain point, additional data provided no further gains, highlighting the limitations of dataset size alone for model generalization. Metadata-based grouping improved accuracy for kettles, particularly in occupancy-based groups, but had limited effect for fridges. Models consistently performed better on seen houses, emphasizing ongoing challenges in generalizing to unseen households, even with metadata.