ESATED: Leveraging Extra-weak Supervision with Auxiliary Task for Enhanced Non-intrusiveness in Energy Disaggregation

Non-intrusive load monitoring (NILM) is crucial to smart grid, which enables applications such as energy conservation and human activity recognition. As a typical task of NILM, energy disaggregation is to decompose total power consumption into appliance-level ones. Despite the remarkable achievements of deep-learning-based methods, their training phase still requires intrusively collected appliance-level power data as strong labels, which are directly used for supervising predictions. In this paper, we present ESATED, a novel energy disaggregation system which instead utilizes non-intrusively collected binary on-off states of appliances as labels, thus enhancing non-intrusiveness throughout the life cycle. However, our labels are inherently weak labels due to the weak correlation between labels (binary states) and predictions (real-valued power), thus making our model struggle in terms of feasible supervision and acceptable performance. To tackle this challenge, we first explore the feasibility of binary-state-based weak supervision, and then integrate it into an auxiliary learning system, where an auxiliary subtask (i.e., state classification) is introduced to further enhance the performance of the primary task (i.e., energy disaggregation). We conduct extensive experiments on two real-world public datasets, and also implement the prototype system in a practical scenario. Corresponding results reveal that even using weak labels, ESATED could achieve performance and transferability second only to the state-of-the-art model. This result demonstrates the effectiveness of the proposed approach to extract information and train the model from extra weak labels.