Optimization of In-House Energy Demand - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 269

be classified into two primary techniques [ 4 , 21 ]: distributed direct sensing and
single-point sensing.
Distributed direct sensing typically requires a current sensor to be installed in-line
with every device and is therefore often referred to as intrusive load monitor-
ing. Although intrusive load monitoring easily achieves a consumption breakdown,
deploying a large number of sensors in the residential environment quickly leads to
high cost and discouraging high usage barrier [ 21 ].
Single-point sensor systems are easier to deploy and are typically subsumed under
the concept of nonintrusive load monitoring (NILM) [ 21 ]. Energy disaggregation is
the task of using an aggregated energy signal, such as that coming from a single-point
sensor or rather whole-home power monitor, to make inferences about the different
loads of individual appliances [ 10 ]. However, single-point sensor systems require
knowledge about the household devices and their electrical characteristics [ 21 ]. The
challenges in energy disaggregation are mainly due to appliances with similar energy
consumption, appliances with multiple settings, parallel appliance activity, and envi-
ronmental noise [ 19 ]. Recent studies [ 8 , 10 - 13 , 20 ] have shown that machine learning
techniques represent a suitable solution to recognize appliances in such dynamic and
unpredictable environments.
In this work, we consider energy disaggregation techniques to derive occupancy
states from appliance usage data in order to use this information in smart heating con-
trol strategies [ 9 ]. Heating control is of particular importance, since heating accounts
for the biggest amount of total residential energy consumption and recent studies
have shown that up to 30% of the total energy can be saved by turning the heat-
ing off when the occupants are asleep or away [ 14 ]. Existing work on the inference
of occupancy states in residential environments includes statistical classification of
aggregated energy data [ 9 ], hot water usage [ 3 ] as well as human motion and activity
[ 17 ]. Our own approach to infer occupancy states differs in that we consider appli-
ance usage, which gives more detailed information about the present context in a
household and the devices which suggest user activity. Furthermore, our proposed
framework does not require any additional infrastructure, and, therefore, is more
likely to be accepted by residents.
For the evaluation of our approach, we consider the REDD dataset [ 10 ], which
consists of whole-home and device-specific electricity consumption for a number
of real houses over the period of several month. In our experiments, we compare
the performance of different models for the appliance/occupancy state classification
task. We use cross-validation (training on all houses and leave-one-out for testing) to
evaluate how well the different models generalize. Our results suggest that the Naive
Bayes classifier is suitable for the prediction of occupancy/appliance states and fits
the problem of real-time heating control.
The rest of the chapter is structured as follows. In Sect. 10.2 , we give some back-
ground on recent advances in energy disaggregation. Section 10.3 introduces the
formal notation of our appliance state classification task. Our proposed framework
for heating control and scheduling by means of energy disaggregation techniques
is described in Sect. 10.4 . The experimental design and results on our approach are
presented in Sect. 10.5 . A practical application for our approach, named SOE, is