The Neural Energy Decoder: Energy Disaggregation by Combining Binary Subcomponents

In this paper a novel approach for energy disaggregation is introduced that identifies additive sub-components of the power signal in an unsupervised way from high-frequency measurements of current. In a subsequent step, these sub-components are combined to create appliance power traces. Once the subcomponents that constitute an appliance are identified, energy disaggregation can be viewed as non-linear filtering of the current signal. The approach introduced here tries to avoid numerous pitfalls of existing energy disaggregation techniques such as computational complexity issues, data transmission limitations and prior knowledge of appliances. We test the approach on a publicly available dataset and report an overall disaggregation error of 0.07.