LOW COST FRAMEWORK FOR NON-INTRUSIVE HOME
ENERGY MONITORING AND RESEARCH
Lucas Pereira and Nuno Jardim Nunes
Madeira Interactive Technologies Institute, University of Madeira, Funchal, Portugal
Keywords: Sustainability, Home Energy Monitoring, Non-intrusive Load Monitoring, Eco-feedback, Low Cost.
Abstract: This paper presents a low-cost framework for non-intrusive home energy monitoring and research built on
top of Non-Intrusive Load Monitoring (NILM) concepts and techniques. NILM solutions are already
considered low-cost alternatives to the big majority of existing commercial energy monitors but the goal of
this work is to make its cost even lower by using a mini netbook as a packaged solution. The mini netbook
is installed in the home’s main circuit breaker panel and computes power consumption by reading current
and voltage through the built-in sound card. At the same time, feedback to the users is provided using the
11’’ LCD screen as well as other built-in I/O modules. The meter is also capable of detecting changes in
power consumption and tries to find out which appliance lead to that change. It is believed that such a
system will not only be important as a tool for energy monitoring and feedback, but also serve as an open
system that can be easily changed to accommodate and test new or existing non-intrusive load monitoring
techniques.
1 INTRODUCTION
Back in 1992 world leaders got together in Rio de
Janeiro for the United Nations Conference on
Environment and Development (UNCED). Two of
the issues addressed were, the use of alternative
sources of energy to replace fossil fuels and the
growing scarcity of water.
Twenty years later many actions have been taken
to face those issues, with a big focus on improving
and creating alternative sources of energy.
The building industry has also been shifting
gears towards more environmentally friendly
practices. Energy and water efficiency are two key
points of the so-called green buildings and many
technological solutions have been implemented to
improve these. Yet, although well intentioned, green
buildings are still expensive to the average
homeowner, and it did not take a lot of time to
realize that representative savings come from a more
efficient use of the building’s utilities and not from
the building itself. But are humans ready to assume
this major role in contributing to a more sustainable
use of natural resources? The short answer is NO.
And even if human beings are at the center of
concerns for sustainable development, they are not
really aware of how their actions and behaviors can
affect sustainability.
Electricity is a paradigmatic example of this lack
of awareness and this is shown in a series of studies
that present significant contradictions between
consumer perceptions and their knowledge of energy
efficiency. For example in (Attari, 2010) authors
show that most humans have a wrong perception of
the most effective thing to do when trying to be
energy efficient. While there is strong evidence that
generally efficiency-improving actions save more
than reducing the usage of inefficient equipment,
only 11.7% of participants refer to the former while
55.2% pointed out the later.
These wrong perceptions were also the subject of
Chisiks’ work (Chisik, 2011), which focus on
understanding how people perceive electricity. The
findings are quite informative about the lack of
perception regarding how much electricity is
consumed by a particular device, which users tend to
associate with the frequency and duration as well as
with the size of the device.
This working hypothesis, that most people lack
awareness and understanding of how their everyday
behaviors affect the environment, is the base for eco-
feedback technology, which is defined as technology
that provides feedback on individual or group
behaviors with a goal of reducing environmental
191
Pereira L. and Jardim Nunes N..
LOW COST FRAMEWORK FOR NON-INTRUSIVE HOME ENERGY MONITORING AND RESEARCH.
DOI: 10.5220/0003950701910196
In Proceedings of the 1st International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2012), pages 191-196
ISBN: 978-989-8565-09-9
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
impact (Froehlich et al., 2010). Eco-feedback
technology has been around for more than 40 years,
and literature shows that providing feedback to the
consumers, even at a low level of disaggregation,
may result in savings between 10% and 15% (Parker
et al., 2006). However there are other studies that
show that this effect is not long lasting, and that
consumers tend to return to previous consumptions
values is a few weeks (Peschiera et al., 2010).
Today the advances in sensing technology that
promote the ability to disaggregate power
consumption at a low-cost, combined with the
widespread use of internet based social networks and
the dissemination of handheld devices, open the
potential of eco-feedback to millions of households.
It is therefore important to understand how people
will react to the feedback, and to what extent they
are willing to change their behaviors in favor of a
more sustainable lifestyle.
This paper presents a low cost framework for
non-intrusive home energy monitoring and research,
which is capable of monitoring and disaggregating
the electricity energy consumption from a single
sensing point and at the same time provide eco-
feedback to the consumers using different
communication channels.
2 HOME ENERGY MONITORING
As seen above, humans deeply misunderstand
energy consumption, and perhaps its invisible nature
is one of the main reasons for this. After all the task
of quantifying something that hides from the human
senses is merely impossible. Lets face it, everybody
knows electricity, but nobody has actually been in
direct contact with it.
The role of quantifying electric energy
consumption is delegated to smart meters, which are
electric devices that record the electric energy
consumption in pre-defined intervals and
communicate the measured results back to the
utility. It is possible to find all kinds of smart meters.
Single point (plug-level) meters are probably the
easiest to find and their mode of operation is very
simple. Basically the appliance is connected to the
meter that in turn is connected to the outlet. Multi
point (whole house) meters provide measurements
at the service entrance and have extra channels to
track sub-panels or larger electrical loads. These are
installed in the main entry feed, and the feedback to
the user can be provided in several ways, e.g.
portable displays and http via built-in webservers or
online services. Finally, the Circuit panels
(Circuit-level) offer the possibility of measuring
each individual circuit in the house, with up to 12 or
16 circuits in each meter. These are considered, by
far, the most expensive and difficult to install
requiring the presence of a professional.
Despite the fact that owning a smart meter will
not necessarily decrease the energy consumption it is
strongly believed that the ability to reason on top of
power consumption data would be of great interest
for consumers and would be a huge help in the
process of engaging the consumers into having a
more energy efficient behavior. “Does my new
microwave spend more than my previous one?”,
“Why do I spend so much electric energy at night if I
am sleeping?”, “How much do I spend cooking
dinner?” are just a few examples of possible
questions that consumers would like to see answered
by their smart meters.
However this is not what smart-meters do
because, although they can provide several different
power metrics, their level of information
disaggregation is not enough to answer such
questions.
It is therefore safe to say that future power
meters must provide their information with very
high levels of disaggregation, that go beyond the
overall consumption and the time of the day.
3 NON-INTRUSIVE LOAD
MONITORING
The process of measuring and disaggregating,
electricity consumption from a single sensing point,
is called Non-Intrusive Load Monitoring (NILM).
NILM is not a new subject, its origins go back to
the late 1980s, early 1990s (Hart, 1992), and it is
built on top of the premise that every change in the
power consumption is due to some appliance
changing its state (either turning on, off or going to a
different working mode), and that by analyzing these
changes it is possible to determine the appliance that
was responsible for them.
The NILM process can be explained as the
combination of six consecutive steps. First sensors
measure the current and voltage signals at the main
circuit breaker. Second, the acquired current and
voltage signals are converted into traditional power
metrics like real and reactive powers. Next an event
detection algorithm is applied to the computed
metrics and load changes are flagged as power
events for further processing by the feature extractor
that will extract a set of generalized features that can
mathematically characterize the event. The set of
SMARTGREENS2012-1stInternationalConferenceonSmartGridsandGreenITSystems
192
features that describe an event is called the event
signature. In the next step previously trained
machine-learning algorithms are applied to the
unclassified signatures to obtain a classification.
Finally in the last step it is possible to estimate how
much energy each appliance is using by keeping
track of all its events and associated power levels.
Much research has undergone in this field
throughout the years after Harts first approach of
analysing real and reactive power steady state
changes at the fundamental frequency and the
advances in sensing technology allowed researchers
to greatly improve the classification accuracies by
using “microscopic” features such as current
harmonics. For example, in (Laughman et al., 2003)
authors have used harmonics as complementary
features, in addition to changes in real and reactive
power, and found that this would help to distinguish
loads otherwise indistinguishable. In (Berges et al.,
2009) the authors have applied supervised machine
learning algorithms, e.g. k-NN and decision trees, to
classify the loads under different feature sets, which
included harmonics powers. They have reported
classification accuracies between 67% and 100% for
different sets of appliances.
In a very different approach (Patel et al. 2007)
proposed that by monitoring electric noise in a
socket for transient signals they could detect most
appliances that were connected to other sockets in
the house. The same authors also presented
ElectriSense (Gupta, Reynolds and Patel, 2010), a
system that focuses on sensing very high frequency
(36-500 kHz) electromagnetic interference (EMI),
which is constantly generated by switch mode power
supplies (SMPS) which are present in most modern
consumer electronics, as well as fluorescent
lightning.
4 NON-INTRUSIVE HOME
ENERGY MONITORING
From a technical standpoint a non-intrusive energy
monitor needs to commit to a set of requirements: 1)
it has to sample both current and voltage from a
single sensing location, 2) the data needs to be
available for both offline and online analysis, and 3)
it has to allow different representations of the
measured energy trough different kinds of feedback.
Additionally, and for research purposes, it also
needs to be able to sense when humans are exposed
to the feedback, and possibly their interactions with
the feedback interfaces. A final requirement, which
is also due to the research purposes of this monitor,
is that the final solution must be very cost effective.
Otherwise it will become too expensive to conduct
research with a fair amount of simultaneous
installations.
To cope with these requirements, one opted to
use a netbook as a whole-in-one solution. The laptop
audio input Analog-to-Digital Converter (ADC) is
used to sample current and voltage, the display and
the speakers are used to provide the interactivity,
while the Wi-Fi card enables communication over
the internet and the built-in camera and microphone
act as low-cost sensors for human activity.
4.1 Eco-feedback User Interfaces
The eco-feedback interfaces of this system were
built on top of those studies, presenting consumption
(in kW/h, € and CO
2
emissions) over hour / day /
week / month / year, total consumption of the day /
week / month / year and also showing comparisons
between different months / weeks and days.
The interface also presents real time data to the
user, namely power consumption in watts and power
events. Figure 1 shows a snapshot of the eco-
feedback user interface.
Figure 1: Eco-feedback user interface snapshots, from left
to right: month view, year view, real time view
(CO2/Month, kW/h and Euros/Month).
Another very important feature of the user
interface is the fact that it stores the user navigation
history (mouse clicks) as well as the instant when
motion is detected (using the webcam as a motion
detector).
4.2 System Architecture
The system architecture is based on the “pipe-and-
filter” software architecture. Figure 2 shows the
current system architecture.
Current and voltage are continuously sensed and
sent to the data acquisition filter to be sampled. As
these are sampled they are sent to the power
calculations filter. This filter is responsible for doing
the power calculations and driving the resulting data
to the splitter, which is an active filter that is
responsible for sending the power samples to the
LOWCOSTFRAMEWORKFORNON-INTRUSIVEHOMEENERGYMONITORINGANDRESEARCH
193
Figure 2: Framework architecture.
filters that are connected to it. The GUI is
responsible for plotting the power as it is being
calculated and the Request / reply socket server
provides real time information about the system
measurements to external applications. The power
storage filter role is to average the power samples,
based on a predefined number of samples value, and
drive the resulting sample to the database. The
median filter is used to apply a median filter to the
power samples, also based on a predefined window
size, and send the filtered samples to the power
event detector filter that will apply a detection
algorithm to the filtered power samples, and trigger
a programmable event when a power event is
detected. The disaggregation filter is a composite
filter that captures the events triggered by the power
event detector and is composed by two filters that
work together to disaggregate the load. The feature
extractor is used to extract the features that will be
used by the power event classifier to classify the
power event. The event is then sent to the database
and streamed to the Internet using the streaming
socket filter.
4.3 Data Acquisition
In order to measure the power being consumed by
the house two sensors are installed at the main
breaker circuit: one split-core current transformer to
be placed around the cable that carries the current
and one voltage transformer to be connected to one
of the existing voltage sources.
These two sensors
are then connected to the netbook built-in sound
card using an audio splitter jack. Custom made
software is used to sample the acquired signal using
the sound cards’ ADC.
4.4 Power Calculations
The power metrics, real power, reactive power,
voltage, amperage and power factor are computed
by applying a Fast Fourier Transform (FFT) to each
period of the current and voltage waveforms, which
are represented by 160 samples each (considering a
sampling frequency of 8000 Hz and a 50 Hz mains
frequency).
4.5 Event Detection
In the current system, the event detector is a
modified change of mean detector that uses a log
likelihood ratio test (Luo, Norford and Shaw, 2002).
In its essence the change of mean detector works
with one sliding window, referred to as detection
window that is used to calculate the likelihood of a
change of mean in each sample, and a second sliding
window, called voting window, that is used to select
the edges with the highest likelihood.
The detection window
l, k
can been seen as
having two windows,
l, j
and
j, k
, pre-event and
post-event respectively. The former is used to
achieve a stable mean as the reference for coming
events, while the later is intended to be very
sensitive to events but yet robust to disturbances.
For each sample in the power signal the
likelihood is calculated according to equation 1:
=
×(
−
)
−
2×
(1)
Where
is the value of the mean change (
−
) at which
reaches its maximum. Additionally
a minimum change of interest
can be set, hence
discarding changes in mean that are below this
value.
is given by equation 2:
=
<
1
−+1
×
∑|
−
|
≥
(2)
The magnitude of l
will increase with the change in
power and abruptness of the change, hence
indicating the presence of a potential event of
interest.
There are five tunable parameters: Minimum step
change
; pre and post event windows lengths,
and
respectively; voting window length
; and the minimum votes an edge needs to be
considered an event of interest
.
4.6 Feature Extraction
The set of extracted characteristics is known as a
power event signature and at the time of writing a
very straightforward signature is being used,
Data
Acquisition
Power
Calculations
Splitter
GUI
Median
Filter
Power
Storage
Power
Event
Detector
Feature
Extractor
Power Event
Classfier
Listen for event
DB
Power
Power Event
Request/
Reply
Socket
Stream
Socket
Power Event
Features
Sampled
Current
&
Vo lt a g e P ow e r
Power
Power
Power
Power
Filtered
Power
Main
electric
feed
Sensed
Current
&
Vo lt a ge
Disaggregation
SMARTGREENS2012-1stInternationalConferenceonSmartGridsandGreenITSystems
194
consisting of four features extracted from both real
and reactive power: real and reactive power mean
change; and 4) real and reactive power polynomial
coefficients (3
rd
degree polynomial).
The mean power change refers to the amount of
change that just happened, and it is calculated by
simply subtracting the average power before from
the average power after the change. This value will
be either positive or negative, hence indicating if
there was an increase or decrease in consumption.
The polynomial coefficients are obtained by
finding the best-fitting curve to the power samples
using a least squares fitting procedure. It is believed
that similar appliances will generate similar
coefficients and that these will become very good
features to add to the event signature.
4.7 Event Classification
Once the power event signature is extracted it is time
to learn what appliance lead to such event. In this
work a supervised learning method is being used. In
the case, and due to good results reported in (Berges
et al., 2009) the k-NN was chosen.
The supervised learning algorithm analyzes the
training data and produces a classifier that will then
be used to assign class labels to future instances
where the values of the predictor features are known
but not the value of the class label.
5 METER VALIDATION
In order to test the event detection and load
disaggregation algorithms an experimental setup
consisting of 8 appliances was put together. The
used appliances where: 1) Compact Fluorescent
Lamp (CFL) - 10W; 2) Fan - 50W; 3) Hand blender
- 250W; 4) Hand mixer - 250W; 5) Kettle 2kW; 6)
LCD Monitor - 130W; 7) Microwave oven - 1,2kW;
8) Toaster - 900W. For these appliances only two
state transitions were considered, OFF to ON and
ON to OFF.
5.1 Event Detector
The event detector was applied, to two consumption
scenarios. In the first scenario three appliances
where used: CFL, LCD monitor and FAN. In the
second scenario five were used: CFL, Fan, Kettle,
LCD Monitor and Microwave oven. Because very
low consumption appliances are present, the
minimum power change of interest was set to 15
watts. The windows sizes are integer values
referring to the amount of power samples. For
example, at 50 Hz 150 samples represents 3 seconds.
From these 150, 100 are used in the pre-event
window and 50 in the after event.
In the first simulation the algorithm was able to
detect 5 of the 6 transitions, only the CFL being
turned off was not detected (1 false positive)
because the step is of about 10 Watts. Still, it is
interesting to notice that the CFL turning ON was
detected even though the minimum step change was
of 15W, and this is because although the average
consumption of the CFL is 10W, when turning ON
the step change reaches more than 15 Watts that then
go back to 10W. As for the second simulation data
the results were as expected. The algorithm is still
able to detect all the appliances (except the CFL
being turned OFF), however, the number of false
positives greatly increases. 22 false positives were
found, from which 13 happened when the
microwave oven was working.
5.2 Event Classifier
To test the classification algorithm the first step was
to collect and classify 10 ON and 10 OFF power
events from the appliances under test, except for the
CFL that was excluded due the difficulties in
detecting its ON and OFF events. Apart from these
20 signatures, 4 others where extracted from them:
1) averaging all the points, 2) selection the median
among all the points, which are referred to as
“Jokers”. In total there are 168 classified signatures,
24 for every appliance (12 ON and 12 OFF), and all
these are used as learners to k-NN classifiers. In total
6 classifiers were created, 3 for each set of features,
for 1, 5 and 9 nearest neighbors.
The learners were tested using the leave-one-out
cross validation. The results of the classification
process are shown in table 1:
Table 1: Results from classification using leave-one-out
cross validation.
Features
Accurac
y
(%)
1-NN 5-NN 9-NN
Step chan
g
e P 90.48 91.67 92.26
Step change Q 64.29 67.26 70.24
P&Q 100 100 100
Pol
y
P 98.02 98.40 99.78
Poly Q 98.94 99.40 99.40
Pol
y
P&Q 99.51 99.51 99.51
Results show, in the first place, that there is not
much variation when changing the number of
neighbors. The second thing to notice is the low
LOWCOSTFRAMEWORKFORNON-INTRUSIVEHOMEENERGYMONITORINGANDRESEARCH
195
classification accuracy that is obtained when using
only reactive power as a feature. The results for just
the real power step change are better, with an overall
classification rate that is above 90%. And the
explanation for these results lie in the fact that using
just one metric does not work properly for
appliances with similar power consumption.
Nevertheless, the combination of real and reactive
power yields very good results, 100% for this set of
appliances, which is in accordance to results in
previous research.
As for the polynomial coefficients the
classification accuracy is very high using any of the
features, and the first thing that this shows is that
this solution can overcome the difficulty of
separating appliances with close consumption levels.
6 CONCLUSIONS
In this paper a low cost framework for non-intrusive
home energy monitoring and research was
presented. The final system is a very cost-effective,
(less than 300 Euros), energy monitor with some
load disaggregating capabilities and at the same
time, provides a very flexible research platform for
non-intrusive load monitoring.
Despite the promising results of the implemented
algorithms there is still a lot of room for
improvement and the flexibility of this framework
will allow the testing of different algorithms with
bigger sets of appliances, and, if possible, in
different houses.
NILM offers a big field of research, for example,
its concepts and techniques can be used in a lower
scale to create a smart power strip that would be able
to detect and turn-off appliances that are found to be
in stand-by mode. Also, it cannot be forgotten that
NILM can be easily exported to other domains,
opening the possibility of creating lower-cost sensor
networks. The ability to sense a whole house
together with the possibility of inferring human
activity will open various windows of research
opportunities. For example home automation
ambient intelligence and smart-grids are just three
fields that can greatly benefit from NILM.
Finally, it is also believed that there is still a lack
of services that use this technology making it
appealing not only for the consumers but also for the
electric companies and appliance manufacturers.
Which also opens a window of opportunity in the
area of service design, where researchers can aim at
creating innovative services on top of low-cost
technologies.
REFERENCES
Attari, S. Z. et al., 2010. Public perceptions of energy
consumption and savings. Proceedings of the National
Academy of Sciences of the United States of America,
107(37), p.16054-16059.
Bergés, M. et al., 2009. Learning Systems for Electric
Consumption of Buildings. Computing in Civil
Engineering, p.1-1.
Bergés, M., Soibelman, L. & Matthews, H. S., 2010.
Leveraging data from environmental sensors to
enhance electrical load disaggregation algorithms. In
proceedings of the International Conference
Computing Civil and Building Engineering.
Chisik, Y., 2011. An Image of Electricity : Towards an
Understanding of How People Perceive
Electricity. Ifip International Federation For
Information Processing, p.100-117.
Froehlich, J., Findlater, L. & Landay, J., 2010. The design
of eco-feedback technology. Proceedings of the 28th
international conference on Human factors in
computing systems CHI 10, 3, p.1999.
Gupta, S., Reynolds, M. S. & Patel, S. N., 2010.
ElectriSense: single-point sensing using EMI for
electrical event detection and classification in the
home. Computer Engineering, p.139-148.
Hart, G. W.,1992. Nonintrusive appliance load
monitoring. Proceedings of the IEEE, 80(12), 1870-
1891
Hohpe, G. and Woolf, B., 2003. Enterprise Integration
Patterns: Designing, Building and Deploying
Messaging Solutions. Addisson-Wesley Professional
Laughman, C. et al., 2003. Power signature analysis.
Computer, 1(2), p.56-63.
Luo, D., Norford, L. K. & Shaw, S. R., 2002. High
Performance Commercial Building Systems
Monitoring HVAC Equipment Electrical Loads from a
Centralized United Technologies Corporation.
ASHRAE Transactions, 108(1), p.841-857.
Parker, D. et al., 2006. How Much Energy Are We Using?
Potential of Residential Energy Demand Feedback
Devices. Solar Energy, p.1665-06.
Patel, S. N. et al., 2007. At the Flick of a Switch:
Detecting and Classifying Unique Electrical Events on
the Residential Power Line. In J. Krumm et al.,
eds.UbiComp 07 Proceedings of the 9th international
conference on Ubiquitous computing. Springer-
Verlag, pp. 271-288.
Peschiera, G., Taylor, J. E. & Siegel, J. A., 2010.
Response–relapse patterns of building occupant
electricity consumption following exposure to
personal, contextualized and occupant peer network
utilization data. Energy and Buildings, 42(8), p.1329-
1336.
SMARTGREENS2012-1stInternationalConferenceonSmartGridsandGreenITSystems
196
