Disaggregating Convolutional Analysis Sparse Coding

Abstract

This work proposes a convolutional analysis sparse coding based formulation for energy disaggregation. The resulting technique is shift-invariant and hence can learn to represent different devices through very few filters (compared to sparse coding based disaggregation). Consequently, this is less prone to over-fitting and hence improves disaggregation results. The technique is very fast, owing to closed form updates in the operational stage. Comparison has been carried out with some well known benchmarks on the REDD dataset. Results show that our method yields the most accurate results and is faster than most benchmarks.