Unsupervised Approximate Power Trace Decomposition Algorithm

We propose the Approximate Power Trace Decomposition Algorithm for decomposing a power signal into components based on power consumption levels. Such a decomposition provides users with more detailed information about their energy consumption without attempting the difficult problem of disaggregating all of the devices in a building. Much of the NILM literature has focused on the problems of event detection and not treated the full problem of energy disaggregation; we have relaxed the problem by moving away from individual appliance disaggregation and provided a solution that makes estimating the energy used by different classes of devices feasible. We present preliminary results using phase A of the BLUED dataset.