A comparison of non-intrusive load monitoring methods for commercial and residential buildings

TL;DR

This paper evaluates the applicability of existing residential NILM methods to commercial buildings, highlighting differences in load characteristics and providing benchmark results using a new dataset from an educational campus in Delhi.

Contribution

It provides the first empirical analysis of NILM in commercial buildings, introduces the COMBED dataset, and assesses the performance of residential NILM algorithms on commercial data.

Findings

Commercial loads differ significantly from residential loads.

Residential NILM algorithms perform poorly on commercial data.

The COMBED dataset enables future research in commercial NILM.

Abstract

Non intrusive load monitoring (NILM), or energy disaggregation, is the process of separating the total electricity consumption of a building as measured at single point into the building's constituent loads. Previous research in the field has mostly focused on residential buildings, and although the potential benefits of applying this technology to commercial buildings have been recognised since the field's conception, NILM in the commercial domain has been largely unexplored by the academic community. As a result of the heterogeneity of this section of the building stock (i.e., encompassing buildings as diverse as airports, malls and coffee shops), and hence the loads within them, many of the solutions developed for residential energy disaggregation do not apply directly. In this paper we highlight some insights for NILM in the commercial domain using data collected from a large smart meter deployment within an educational campus in Delhi, India, of which a subset of the data has been released for public use. We present an empirical characterisation of loads in commercial buildings, highlighting the differences in energy consumption and load characteristics between residential and commercial buildings. We assess the validity of the assumptions generally made by NILM solutions for residential buildings when applied to measurements from commercial facilities. Based on our observations, we discuss the required traits for a NILM system for commercial buildings, and run benchmark residential NILM algorithms on our data set to confirm our observations. To advance the research in commercial buildings energy disaggregation, we release a subset of our data set, called COMBED (commercial building energy data set).