Energy Disaggregation via Deep Convolutional Dictionary Learning

In nonintrusive load monitoring, the objective is to estimate the power consumption of individual appliances given the total power reading from the smart-meter. Mathematically, it is a highly underdetermined problem with infinitely many solutions. Furthermore, practical constraints like low sampling frequency and continuously varying power loads make the problem even more difficult. This work introduces a practical disaggregation approach based on deep convolutional dictionary learning (DL). It uses multiple layers of convolutional filters as the basis for modeling appliances. The ensuing formulation is solved using the alternating direction method of multipliers. Comparison with the benchmarks (published last year) on the Reference Energy Disaggregation Dataset (REDD) dataset shows that our method improves over the state-of-the-art.