An L0-Norm Constrained Non-Negative Matrix Factorization Algorithm for the Simultaneous Disaggregation of Fixed and Shiftable Loads

Energy disaggregation refers to the decomposition of energy use time series data into its constituent loads. This paper decomposes daily use data of a household unit into fixed loads and one or more classes of shiftable loads. The latter is characterized by ON OFF duty cycles. A novel algorithm based on nonnegative matrix factorization NMF for energy disaggregation is proposed, where fixed loads are represented in terms of real-valued basis vectors, whereas shiftable loads are divided into binary signals. This binary decomposition approach directly applies L0 norm constraints on individual shiftable loads. The new approach obviates the need for more computationally intensive methods e.g. spectral decomposition or mean field annealing that have been used in earlier research for these constraints. A probabilistic framework for the proposed approach has been addressed. The proposed approach s effectiveness has been demonstrated with real consumer energy data.

💡 Research Summary

The paper addresses the problem of energy disaggregation by separating household electricity consumption into two distinct categories: fixed loads that operate continuously throughout the day and shiftable loads that exhibit distinct ON‑OFF duty cycles. Building on the non‑negative matrix factorization (NMF) framework, the authors propose a novel algorithm that treats these two categories differently. Fixed loads are modeled with conventional real‑valued basis vectors (W_f) and activation coefficients (H_f), which are learned using standard multiplicative update rules and column‑wise normalization.

Shiftable loads, on the other hand, are represented by binary basis matrices (W_js) where each column corresponds to a possible one‑minute ON interval. The activation vector h_js(n) for day n is binary, and a hard L0‑norm constraint ‖h_js(n)‖_0 ≤ L_j limits the number of ON cycles for each appliance, reflecting realistic usage limits. To enforce this constraint directly, the authors design a hill‑climbing heuristic (subroutine hillClimb). Starting from an all‑zero vector, the algorithm iteratively selects the basis column that yields the greatest reduction in the residual error ‖r_j(n) – W_js h‖_2^2, adds it to the active set, and stops when either the maximum allowed cycles L_j is reached or any further addition would increase the error. The procedure runs in O(L_j·log|D|) time per day, where D is the number of time samples, and is far more efficient than previous approaches based on mean‑field annealing or repeated SVD.

The overall objective function is the Frobenius norm of the reconstruction error, ‖X – X̃‖_F^2. The authors provide a probabilistic justification: assuming each daily sample vector follows a Gaussian distribution centered at its reconstruction, the negative log‑likelihood reduces exactly to the Frobenius norm, establishing a maximum‑likelihood interpretation of the loss.

Experiments use real 1‑minute resolution data from the Pecan Street Dataport for a single residential customer over 15 weekdays in April 2019. Four appliances (furnace, washer/dryer, oven, kitchen devices) are treated as shiftable loads, while the remaining aggregate consumption forms the fixed load. Peak powers p_j and maximum cycle counts L_j are manually estimated from the data; each shiftable load is given a binary basis of size 1440 (one column per minute). The algorithm first updates W_f and H_f via multiplicative rules, then for each day and each shiftable load updates h_js(n) using hillClimb on the residual.

Results show that the reconstructed total load matches the measured aggregate closely, reproducing morning and evening peaks. Individual appliance reconstructions accurately capture ON‑OFF intervals for the furnace, washer/dryer, and furnace, with minor discrepancies for the kitchen devices due to their lower peak power. The authors highlight that the method achieves high‑resolution disaggregation without resorting to computationally intensive techniques, and that the explicit L0 constraint yields physically interpretable cycle limits.

Contributions include: (i) a high‑resolution (1‑minute) disaggregation demonstration, (ii) a dual‑category modeling approach that leverages distinct characteristics of fixed and shiftable loads, (iii) a direct L0‑norm enforcement via a simple hill‑climbing heuristic, and (iv) a computationally efficient alternative to spectral or annealing methods. Limitations involve reliance on manually set parameters (p_j, L_j) and the heuristic’s potential to converge to local minima. Future work is suggested on automatic parameter estimation, modeling interactions among multiple shiftable loads, and extending the method to online, real‑time applications.