Enhanced Expandable Representation Learning for Non-Intrusive Appliance Incremental Identification

The monitoring and analysis of appliance-level usage status and power consumption behaviors of end users can provide data support for user-side demand response, energy efficiency management, and price policy, and it is crucial to realizing flexible, interactive, intelligent power consumption. Existing non-intrusive load monitoring (NILM) methods only identify appliances in fixed scenarios. When new appliances are added to the user, traditional NILM methods may need to retrain the model, which leads to monitoring interruptions and requires cumulative storage of old appliance data samples. In this paper, we propose a new expandable and incremental representation learning approach based on the V-I trajectory feature that allows the NILM model to be improved incrementally to adapt to new appliance data without interrupting monitoring. Specifically, at each incremental step for new appliances, we freeze the previously learned representation and enhance it with additional feature dimensions of the new learnable feature extractor. That allows us to combine the added new appliance knowledge with retaining the old appliance knowledge. In addition, we introduce auxiliary losses to encourage the model to learn the diversification and differentiation characteristics of new appliances. We have conducted comparison experiments on publicly available data sets to verify the effectiveness and advantages of the model.