BOLT: Energy Disaggregation by Online Binary Matrix Factorization of Current Waveforms

In this paper we introduce BOLT, a novel approach for the problem of energy disaggregation that performs online binary matrix factorization on a sequence of high frequency current cycles collected in a building to infer additive subcomponents of the current signal. The system learns these constituent current waveforms in an unsupervised fashion and, in a subsequent step, seeks to find combinations of these subcomponents that constitute appliances. By doing so, points in time when appliances are active and, to some degree, their power consumption can be estimated by BOLT. Our system treats energy disaggregation as a binary matrix factorization problem and uses a neural network, with binary activations in the one but last layer and a linear output layer, to solve it. The algorithmic performance of the proposed method is evaluated on a publicly available dataset. Furthermore, we show that, once the model is trained, the algorithm can perform inference in real-time on inexpensive off-the-shelf and general purpose hardware which allows leveraging high-frequency information without having to explicitly transmit and store large amounts of data to a centralized repository.