Scalable Energy Disaggregation Via Successive Submodular Approximation

PHASED: Phase-Aware Submodularity-Based Energy Disaggregation

“…where the first term represents the least-squares data fit over all phases (lines); and the second term is a smoothness-inducing regularizer that seeks to maximize the similarity between the states of an appliance over consecutive time instants as in [6], [13]; and λ i ∈ R + is a regularization parameter (we set it to 1 in the experiments). The constraints in (3) guarantee the selection of only one state for each appliance at a time.…”

“…We compare PHASED to four quite different baselines to ensure broad evaluation. The baseline methods (explained in Section I) are: (i) DSC (discriminative sparse coding) [5], (ii) NMF [8], (iii) seq2p [12], and (iv) BSMA (block successive modular approximation) [13]. We measure the percentage of energy deviated (P ED) from the true consumption of appliance i in a house h at a time t using: where x i (t, h) and xi (t, h) are the true and inferred power consumption for appliance i at time t in house h, and y(t, h) is the aggregated power at t in h. Then, we present the average of P ED (AP ED) over the total time ticks in all the houses:…”

“…Recently, the work in [13] demonstrated that the energy disaggregation can be posed as a constrained set-function maximization problem, which is NP-hard in its general form. The authors proposed a discrete block successive approximation algorithm that exploits the fact that the cost function is block-submodular [13].…”

“…Recently, the work in [13] demonstrated that the energy disaggregation can be posed as a constrained set-function maximization problem, which is NP-hard in its general form. The authors proposed a discrete block successive approximation algorithm that exploits the fact that the cost function is block-submodular [13]. Building on this line of work, we propose PHASED 1 , a supervised framework for energy disaggregation that leverages the connectivity structure of the power distribution networks.…”

“…Leveraging the special properties of submodular functions [14], [15], we devise an efficient successive approximation algorithm for computing high-quality, albeit suboptimal solutions for the problem. In contrast to [13], we establish that the cost function can be expressed in DSF form over the entire time horizon, which results in a discrete approximation algorithm that features more attractive "all-at-once" updates. PHASED improves the error of four distinct classes of state-of-the-art approaches by up to 61% when averaged over appliances.…”