Transfer learning for multi-objective non-intrusive load monitoring in smart building

Buildings represent 39% of global greenhouse gas emissions, thus reducing carbon emissions in buildings is of importance to greenhouse gas emissions reductions. This requires understanding how electricity is utilized in the buildings, then optimizing electricity management to seek conservation of energy. Non-intrusive load monitoring (NILM) is a technique that disaggregates a house’s total load to estimate each appliance’s electric power usage. Several strategies for estimating one appliance at a time (one-to-one model) have been presented and experimentally proven to be effective, with two mainstream trends: appliance transfer learning and cross-domain transfer learning. The former refers to the transfer between different types of appliances in the same data domain, while the latter refers to the transfer between different data domains for the same type of appliance. Different from the previous work, this paper explores the approach of adopting one model for all appliances (one-to-many model) and proposes a novel transfer learning scheme, that incorporates appliance transfer learning and cross-domain transfer learning. Thus, a well-trained model can be transferred and utilized to effectively estimate the power consumption in another data set for all appliances, which demands fewer measurements and only one model. Three public data sets, REFIT, REDD, and UK-DALE, are used in our experiments. Further, a set of smart electricity meters was deployed in a practical non-residential building to validate the proposed method. The results demonstrate the accuracy and practicality compared to start-of-the-art one-to-one NILM transferred models. • An adaptive one-to-many NILM model is designed (one model for all appliances). • A novel transfer one-to-many NILM strategy is proposed for the first time. • The number of parameters for NILM model is significantly reduced. • A novel metric, Overall Disaggregation Proportion Error (ODPE), is used. • The proposed model is validated on both public and self-collected datasets.