Leveraging Data From Environmental Sensors to Enhance Electrical Load Disaggregation Algorithms

Abstract

The idea of sustainable or green buildings generally stops after the design and construction phases.
Little effort is made to continuously monitor and control the energy profile throughout the life-cycle
of these facilities. To effectively identify opportunities for consumption reduction, measurement and
feedback of current energy use is necessary. Monthly utility bills are inadequate for planning
conservation programs, or even for assessing their effectiveness once implemented. Extensive
hardware sub-metering, although very expensive, is sometimes used to obtain more granular
feedback. Non-Intrusive Load Monitoring (NILM), another method that has been studied for the past
two decades, follows an inexpensive approach for obtaining appliance-specific consumption
information. The idea behind this technique is that operation of individual appliances generates a
distinct signature in the power distribution system of the building, which can be detected by carefully
analyzing the overall voltage and current of the building. However, two of the main challenges
keeping the technology from reaching wide adoption are: (a) finding simple ways to train the
algorithms; and (b) obtaining robust appliance signatures that form spread-out clusters in the feature
space, especially for small loads.

In this paper we explore the feasibility of utilizing data from separate environmental sensors (e.g.,
light intensity, sound level, etc.) present in the building, for improving the training process by
enhancing the appliance signatures and providing an independent and trusted source of information
about the operation of appliances. We exploit the fact that the operation of appliances will likely be
reflected in both the power and environmental data streams. We present initial results from a case
study where a prototype NILM system was deployed in an occupied apartment building, along with a
number of environmental sensors. We also suggest two approaches for leveraging the environmental
data and provide descriptions for possible future research in the area.