KSCS09

KSCS09
ViridiScope: Design and Implementation of a Fine Grained
Power Monitoring System for Homes
Younghun Kim, Thomas Schmid, Zainul M. Charbiwala, and Mani B. Srivastava
Networked and Embedded Systems Lab.,
Electrical Engineering Department,
University of California, Los Angeles
{kimyh,thomas.schmid,zainul,mbs}@ucla.edu
ABSTRACT
A key prerequisite for residential energy conservation is knowing when and where energy is being spent. Unfortunately,
the current generation of energy reporting devices only provide partial and coarse grained information or require expensive professional installation. This limitation stems from the
presumption that calculating per-appliance consumption requires per-appliance current measurements. However, since
appliances typically emit measurable signals when they are
consuming energy, we can estimate their consumption using
indirect sensing. This paper presents ViridiScope, a finegrained power monitoring system that furnishes users with
an economical, self-calibrating tool that provides power consumption of virtually every appliance in the home. ViridiScope uses ambient signals from inexpensive sensors placed
near appliances to estimate power consumption, thus no inline sensor is necessary. We use a model-based machine
learning algorithm that automates the sensor calibration process. Through experiments in a real house, we show that
ViridiScope can estimate the end-point power consumption
within 10% error.
ACM Classification Keywords
H.4 Information Systems Applications: Miscellaneous
General Terms
Algorithms, Design, Experimentation, Human Factors, Measurement
Chetty et. al. [8] observed last year, researchers in ubiquitous computing have a major role to play by developing technologies that encourage people to efficiently manage energy
consumption at home. A fundamental component of such
technologies is fine grained monitoring to measure real-time
consumption of each domestic appliance [14, 18]. Many researchers and companies have begun addressing this problem. For example, Jiang et. al. [17] developed a wireless
networked sensor that measures the power consumption at
a power outlet; while other devices that sense the real-time
electricity consumption for a whole house (e.g. Cent-a-Meter), per circuit (e.g. EM-2500 [1], TED [2]) or per outlet
(e.g. Kill-A-Watt and Watts Up [11]) are now commercially
available.
Although these devices are promising, each has some drawbacks. For example, Cent-a-Meter, EM-2500 and TED devices monitor the household power consumption but do not
provide per-appliance level measurements. Kill-A-Watt and
Watts Up devices provide finer granularity but require inline installation between a standard AC plug and the outlet.
While it is possible to instrument many appliances in this
way, some of the major energy consumers cannot be easily instrumented. For example, most heating and ventilation
systems (HVAC) and electric boilers do not have standard
AC plugs, or are hard-wired to the main power lines. Ceiling lights are another example. We conclude that the current
generation of energy reporting devices can simultaneously
provide only two of the following three features essential for
true ubiquity:
Author Keywords
Adaptive Sensor Calibration, Machine Learning, Nonintrusive and Spatially Distributed Sensing
INTRODUCTION
Natural resource preservation has recently become a significant concern, and has motivated research and development
efforts to assist in both conservation and management. As
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Ubicomp 2009, Sep 30 – Oct 3, 2009, Orlando, Florida, USA.
Copyright 2009 ACM 978-1-60558-431-7/09/09...$10.00
• Comprehensive Coverage: The system should monitor all
the appliances in the home, making information complete
for users.
• Fine-grained Reporting: The system should be able to report individual consumption profiles for each appliance as
this will enable targeted conservation.
• Easy and Seamless Installation: The system should be installable by a non-professional, e. g. no modifications to
power lines or power cords should be necessary. Further,
user intervention for configuration, calibration and maintenance should be minimal.
Simultaneously satisfying the above criteria precludes the
sole use of traditional direct monitoring techniques, and a
AC Plug
AC Plug
Indirect
Sensor
Current
Meter
(Magnetic,
Light, Acoustic)
Current
Energy
Meter
Energy
Inference
Engine
Voltage
Appliance
Magnetic Field Change near a PC
Indirect Sensing
Appliance
Figure 1: Direct power monitoring devices need to be installed in
series with the appliance, whereas our indirect power monitoring
concept senses signals emitted from appliances making it less invasive
(Left adapted from [17]).
new monitoring dimension that does not require an in-line
sensor is necessary. To this end we present ViridiScope,
a spatially fine grained power monitoring system for residential spaces. The ViridiScope system provides real-time
appliance-level power estimation by extensively leveraging
ubiquitous sensing and computing devices, including magnetic, acoustic and light sensors. The modus operandi is a
network of radio-enabled distributed sensors monitoring signals that appliances emit and forwarding them to a personal
computer acting as a back-end fusion center. The fusion center collects data from the heterogeneous sensors and measurements from the main power meter, and runs a modelbased machine learning algorithm that automatically learns
and estimates power consumption of every appliance on-thefly. The contributions of this paper are three fold:
Introducing Indirect Power Monitoring Concept:
Our approach is based on the fact that an appliance emits
measurable signals when it consumes energy. By sensing
these signals we can estimate power consumption (Figure
1). The core advantage is that one does not need to install a monitoring device in-line with the electric wire. Instead an indirect sensor near a power line or the appliance
is enough for accurate power estimation. For example,
we show in the Problem Description Section that simply
placing a magnetic sensor near a power cord or a light intensity sensor near a ceiling light is sufficient to estimate
their consumption profile.
Autonomous Sensor Calibration Framework:
The application of indirect sensing introduces a crucial
technical challenge because the indirect sensors are inherently more prone to variations in ambient conditions, e.g.
the actual sensor placement, surrounding materials [4]. To
estimate the power consumption of an appliance based on
signals that an indirect sensor is reporting, we need to find
a unique mapping from the signals to power consumption. Because of uncertainties in installation, changes in
ambient conditions, and sensor variability, it is virtually
impossible to calibrate indirect sensors prior to installation. Therefore, our approach uses in-situ sensor calibration. Unfortunately, in-situ calibration often involves extensive manual interaction. To avoid this we develop an
15
Magnetic Field [mGauss]
Direct Sensing
10
5
0
−5
−10
−15
0
5
10
15
20
25
30
Time[min.]
Figure 2: The magnetic field change near a PC looks like noise,
although we can infer whether the PC is active or not. The noisy
region indicates that the PC is active.
autonomous sensor calibration framework. We formulate
a model based regression problem that combines the aggregate power consumption and signals from indirect sensors. Our framework is able to incorporate with uninstrumented appliances as well as direct monitoring devices.
Prototype Deployment in a Realistic Setting:
We conduct a small scale deployment in a 2-person apartment unit to show the feasibility of this approach. We find
that the indirect monitoring is feasible and results in less
than 10% of estimation error.
MOTIVATION
Various studies have shown the necessity of fine grained energy monitoring to encourage conservation. For example,
McMakin et. al. [21] find that to sustain changes in human behavior for energy conservation, continued awareness
is as important as incentives and disincentives. Stern [24]
suggests that information awareness has a synergistic effect
with incentives in energy conservation. He states that awareness not only encourages people to actively participate and
do the right thing, but also provides indirect monetary benefits because of reduced cost of resources. Darby [10] also
concluded that feedback about consumption is essential for
energy savings, and that immediate direct feedback can be
extremely valuable in influencing behavior with savings in
the range of 5-15%. The World Business Council for Sustainable Development in [27] noted, “Lack of awareness and
information on energy consumption and cost - people are often not aware that they are wasting energy - prevents them
from behaving efficiently.” They also note, “Technical devices to measure energy consumption and provide immediate feedback could help households cut energy consumption
by as much as 20%. Direct and immediate feedback reveals
the link between actions and their impacts. Well informed
consumers choose actions to save energy with minimal impact on their comfort.” We believe that helping individuals
know the what, when, and where of their energy usage, and
identifying wasteful patterns of usage, encourages modifying wasteful behaviors.
PROBLEM DESCRIPTION AND SYSTEM DESIGN
To show the feasibility of indirect power monitoring, we
present some experimental data that illustrates that the measured signals from magnetic, light and acoustic sensors are
Power Consumption and Magnetic Field
Power Consumption and Magnetic Field
Power Consumption and Magnetic Field
150
150
1000
0
200
400
600
800
1000
1200
1400
200
150
100
50
0
0
200
400
600
800
1000
1200
Time[s]
(a) Personal Computer Set
1400
1600
Power
600
400
200
1600
Time[s]
800
120
110
90
50
100
150
Time[s]
900
850
800
750
700
650
600
550
130
100
0
Std of Magnetic ADC Values
50
0
Std of Magnetic ADC Values
Power[W]
100
Std of Magnetic ADC Values
Power[W]
140
0
50
100
150
Time[s]
0
5
10
15
20
25
30
20
25
30
Time[s]
150
140
130
120
110
100
90
0
5
10
15
Time[s]
(b) Hot-air-gun
(c) Massage Chair
Figure 3: Correlation between power consumption and the magnetic field variation near AC power wires of a PC set, a hotairgun and a massage
chair. It is clear that they are very closely correlated, and we exploit this in our algorithm.
Refrigerator Power States
300
Power Consumption [W]
good proxies for the appliance power consumption. We then
propose a system architecture that consists of a two tiered
information and sensing hierarchy. This architecture details
the sensor configuration and information flow among ubiquitous sensing and computing devices. To make the monitoring system self-configurable, we formulate an autonomous
calibration mechanism that automatically trains the indirect
sensors so that it can compute the appliance-level power consumption in real-time.
250
Both
On
200
150
100
Door
Lamp
On
Compressor
On
50
0
10
11
12
13
14
15
16
17
18
19
Time[min.]
Theory of Operation: Indirect Sensors and Power Estimation
The prototypical ViridiScope system infers per-appliance level power consumption using three types of indirect sensors:
magnetic, acoustic, and light. A magnetic sensor near the
power cord of an appliance is a good monitor of the current
draw because the magnetic field variations near the power
wires are strongly correlated with the current flow inside the
wires [7]. In fact, magnetic field transducers are regularly
used for current sensing [16, 17].
ViridiScope exploits the noisy magnetic field changes near
an appliance (Figure 2). We use a statistical signature, the
standard deviation of the magnetic field near the power wire.
This signature has a strong correlation with the real power
consumption(W). This is intuitive because when an appliance is active, electrons are flowing and generate magnetic
field in the air [7]. Figures 3 (a), (b), and (c) demonstrate
that the standard deviation is a good proxy for the power
consumption of a device. This phenomena is universal to all
AC powered appliances. Therefore, a sensor node equipped
with a one dimensional magnetometer is adequate to capture
the variable power consumption of an appliance.
Different from a magnetic sensor, a light intensity or an acoustic sensor can detect the internal power state of an appliance. Fortunately, most appliances have a very limited number of power states. Knowing these states and their average power consumption is sufficient to describe the overall
power consumption of an appliance [15, 19, 20, 23]. For example, a simple light can only be On or Off. Detecting these
on and off states is simple, and is enough information to in-
Figure 4: A refrigerator is a good example of an appliance having
several power states. It has two main components: a door light and a
compressor making four effective power states (Both off, Light on,
Compressor on, and Both on).
fer the light’s power. A refrigerator, on the other hand, has
four power states, Compressor On, Inside Light On, Both
On, Both Off, since it has two active components: a compressor and a door light (see Figure 4). A combination of a
light and acoustic sensor can effectively detect all four states.
For example, a light intensity sensor inside the refrigerator
can easily detect the on/off status of the door light (Figure 6).
In addition, a very primitive microphone sampling at 4Hz is
adequate for detecting the power state of a compressor in a
typical refrigerator (Figure 5). All these examples support
the concept of indirect power monitoring.
Technical Challenges
The biggest technical challenge with indirect sensing is calibration. Even if the form of the calibration function is
known, the calibration parameters for each sensor must be
identified after the sensor is installed. Thus, our approach
requires in-situ sensor calibration. In-situ calibration traditionally requires a well trained technician [3, 26].
To automate sensor calibration, ViridiScope exploits an additional piece of information from the existing infrastructure: the main power meter that provides real-time aggregated power consumption for the whole household. Since
most homes have one main power-line, this information is
easily available. Using this information, we develop an au-
Sound Level on a Refrigerator
700
Compressor
On
Sound Level[ADC]
650
600
550
500
450
400
300
Magnetic Sensor Strapped
on Power Wires
Compressor
Off
350
0
20
40
60
80
100
120
140
Television Screen
Time[min.]
Figure 5: An acoustic sensor at 4Hz sampling rate can detect the
on/off status of the compressor.
Light Intensity Change inside of a Refrigerator
Refrigerator
Light Intensity[ADC]
700
Standing Light
600
500
Door
Closed
400
300
Door
Open
200
100
0
0
50
100
150
Time[s]
Figure 6: Light intensity in a refrigerator goes up when the door is
open and the inside light is on.
tonomous calibration algorithm that calibrates the indirect
sensors with minimal human intervention.
Another problem is that indirect sensors are generally more
prone to external noise, which makes reliable detection difficult. This is resolved by adding adaptive filtering mechanisms. We discuss and analyze these in the Practical Consideration Section.
System Architecture
Information Hierarchy
ViridiScope consists of a two tiered information hierarchy
that contains one utility installed meter (or a main meter
monitor), and multiple uncalibrated indirect sensors for energy consuming end-points (Figure 7). The first tier provides
the aggregated power consumption, y0 (t). The second tier
sensors monitor energy consuming related signals, si (t).
The characteristics of this hierarchy are as follows. (1) The
first tier is reliable and the utility provider maintains it. (2)
The calibration of the non-intrusive indirect sensors can be
automated using correlations between the first and second
tier sensors. This makes the monitoring system readily deployable and can be installed by a nonprofessional without any technical background. (3) The first tier is used as
“ground truth” that is always available. This allows the system to seamlessly track its performance.
Monitoring the Main Power Meter
The first tier information is the aggregated power consumption of a household. This is obtained by monitoring the
main power meter in real-time. Although the traditional
Figure 7: A prototype implementation of ViridiScope consists of
several non-intrusive sensors: a main meter monitor, light intensity
sensors, magnetometers, and microphones. This implementation uses
CrossBow MicaZ motes, HMC1002 magnetic sensors, and MTS310
sensor boards. Each battery-powered sensor node monitors energy
consumption related signals, and sends them back to the fusion center.
The fusion center is a PC that solves a numerical optimization
problem using CVX tools, combines data from the distributed sensors,
and profiles appliance-level power consumption according to unique
node IDs. It’s important to note that the system doesn’t require in-line
sensors.
main power meter may not have an interface to get real-time
power consumption, many companies sell easy-to-use devices that can monitor the main power meter in real-time.
For example, the Energy Detective (TED) and The Power
Cost monitor are readily available in the U.S. [2, 25] (See
Figure 8). Furthermore, a better and systematic solution
is expected in the very near future. The Automatic Meter Reading Association is developing a system that relays
real-time meter information wirelessly for billing purposes.
Following the trend of utility companies simplifying their
billing systems through wireless technologies, we believe
that extracting data from the main power meter of a household in real-time will soon be commonplace as the Smart
Grid technology is deployed at a large scale in the US [6]
and elsewhere.
Power Monitoring Using Indirect Sensing Subsystem
The second tier sensors monitor a signal related to energy
consumption, which is used to infer the power consumption
of an appliance. Our prototypical design uses HMC1002
magnetic sensors, CdSe photocell light intensity sensors and
simple piezoelectric microphones. Magnetic sensors monitor raw magnetic field changes near power wires and the
standard deviation of the magnetic field change is used to
estimate the power consumption. Internal power states of
appliances is inferred from acoustic and light intensity sensors that monitor noise patterns and light intensity changes
near appliances. Our implementation uses Crossbow MicaZ
wireless sensor nodes that run TinyOS, HMC1002 magnetic
sensors, and MTS310 sensor boards (Figure 7). Note that an
instantiation of ViridiScope is not limited to this particular
choice.
Standing Lamp Near Window
Real!time Power Consumption Trace of A 2!person apartment
1400
Dinner
Prep.
4500
4000
Power [W]
3500
3000
Light Intensity[ADC]
5000
Refrigerator
Activity
Evening
Activities
2500
1200
800
600
400
Misc
Lamp Off
0
5
10
15
20
1000
500
14
16
25
30
35
40
Time[min.]
1500
0
12
Lamp On
200
0
2000
Ambient
Light
Change
During
Sunset
1000
18
20
22
24
Figure 9: A simple light intensity sensor can detect the on/off status of
a light near a window even during sunset, this plot shows a sufficient
noise to signal ratio to infer on/off status very reliably.
Time[Hours]
Figure 8: Real-time power trace of a 2-person apartment: A TED
device, installed on the main power line, monitors the real-time power
consumption. Although this information gives a good sense of the
energy usage at home, it is only giving a partial view of where the
energy gets used. For example, in the evening, many appliances are
simultaneously on, and it is hard to disambiguate where the energy is
actually used.
Information Fusion via Numerical Optimization Toolbox and
Data Aggregation
The data collected by the two tiers is sent to a personal computer (fusion center) where the indirect signals and the aggregated power consumption from the main meter are “fused”. The PC fuses the data by solving a numerical optimization problem introduced in the following section. The solution calculates the calibration parameters for the second tier
sensors, which subsequently are used to compute the realtime power consumption of each appliance. Our implementation uses the CVX toolbox [9], an open source convex optimization tool, to solve the autonomous calibration problem
(Equation 5). Alternative tools, such as MOSEK [22], could
be used as well.
Problem Formulation
The ViridiScope hardware provides the total power consumption, y0 (t), and the signals, si (t), that are correlated with
the energy consumption of the individual appliances. We can
now formulate a problem statement that allows the system
to automatically learn the appliance-level power consumption. We first model the explicit calibration functions that
are mappings from the measured signals to the power consumption (fi : si (t) → pi (t)). Then we set up a numerical
optimization problem that solves for fi by exploiting the fact
that the total power consumption of a house is the sum of the
power consumption of all the appliances.
Calibration Function Modeling
Magnetic Sensors
It is very well known that magnetic fields are coupled to AC
currents (Maxwell’s Equation [7]). Many types of magnetic
sensors are used for current measurement methods [7, 16].
Power consumption can be calculated through current measurement if the consumption is purely resistive. But the
power consumption is usually both resistive and inductive,
and so both voltage and current is needed to calculate power
consumption. Also, clamp type current sensors are not adequate for our needs because they need to be clamped on
one of the two AC wires, thus requiring non-trivial power
wire modification. Clamping both wires results in the useless measurement of the net current of zero.
As shown in Figure 3, the standard deviation of the magnetic field change has a strong correlation with the power
consumption. We can setup a simple calibration curve that
is an affine mapping from the magnetic field variation to the
power consumption:
def
pi (t) = αi si (t) + βi ,
(1)
where pi (t) is the power consumption, αi , βi are calibration
parameters, and si (t) is the standard deviation of the magnetic field change. In our implementation, a node samples
the magnetic field change at 100Hz and calculates the sample standard deviation, si (t), over a 1 second sliding window.
Acoustic and Light Sensors
Because many appliances have a limited number of power
states, estimating the average power consumption of the internal states is enough. Since internal power states often
have an almost constant power consumption (Figure 4), we
can infer the power consumption if we know the states and
their respective power consumption.
Although this approach compromises time-resolution and does not capture transient power consumption, its resolution is
generally good enough for this non-critical domestic application. The main benefit comes in total system cost. For
example, a simple light sensor can detect the on/off status of
light (Figure 6) and costs much less than a magnetic sensor.
For example, a HMC1002 magnetometer is $20 whereas a
simple light sensor is less than $1. Similarly, a simple acoustic sensor is much cheaper than the magnetic sensor, and
can easily identify the internal power states of an appliances
(Figure 5).
We now define a simple calibration model. Let si,j (t) be an
indicator function indicating the j-th internal state of the i-th
appliance. This value is the output from the indirect sensors
detecting the internal power state of appliances. Then the
power consumption of an appliance can be described by the
following equation
def
pi (t) =
Ki
X
j=1
Pi,j si,j (t),
(2)
Uninstrumented Appliances
ViridiScope needs to handle the case where several appliances are not instrumented with any sensors. This is likely
in a real house, since a user might have forgotten one or more
devices, or may not want to install sensors everywhere [4].
We define an artificial appliance that is always on to account
for the uninstumented appliances, and thus takes over for all
the uninstumented appliances (also called ghost power consumer). We denote the average power consumption of this
artificial appliance as Pi . This gives:
def
pi (t) = Pi si (t),
(3)
where Pi is the power consumption for all uninstrumented
appliances and si is the indicator function that indicates if
the system has uninstrumented appliances in the configuration or not.
Direct Meters
Direct appliance power meters are commercially available
[11, 17] and it is reasonable to include these meters in the
ViridiScope system configuration. Because these meters directly monitor pi (t), we can directly use that measurement
in our formulation. For notational consistency we define:
def
pi (t) = p̃i (t).
(4)
where p̃i (t) is the data from a direct meter.
Autonomous Calibration Formulation
Equations 1, 2, and 3 state that if we know all the calibration parameters, we can calculate the individual power consumption of every appliance. Unfortunately, the calibration
parameters are not known and must be learned after the sensors are deployed. This could be done by manual intervention, measuring and calibrating each sensor. However, this
introduces a significant burden on the person deploying the
system, and thus we want to automate this step.
The total power consumption of a household is the sum of
the power consumption of all the appliances. Each of Equations 1, 2, 3, and 4 represents the power consumption of an
individual appliance. We can formulate an explicit numerical optimization problem because y0 (t) has to be equal to
PN
i=1 pi (t), where N is the total number of appliances. And
thus,
PN
min ||y0 (t) − i=1 pi (t)||
where 
αi si (t) + βi : magnetometers


 PKi
j=1 Pi,j si,j (t) : light/acoustic sensors
pi (t) =

P
s
(t) : uninstrumented

 i i
p̃i (t) : direct meter input.
(5)
Sound Level on a Refrigerator
1200
1000
Sound Level[ADC]
where Ki is the number of internal power states of the i-th
appliance, Pi,j is the average power consumption of the jth power state of the i-th appliance, and si,j (t) is a boolean
indicator function that indicates the on/off status of the j-th
power state of the i-th appliance.
800
600
400
200
0
0
20
40
60
80
100
120
140
Time[min.]
Figure 10: A single acoustic sensor is capturing ambient noise in
addition to the sound from the refrigerator compressor. This makes
detection of the compressor state difficult.
In this problem, the control variables are the calibration parameters, αi , βi , Pi and Pi,j . Note that || · || is an arbitrary
norm, making the formulation a convex program. For example, Equation 5 can be easily cast as a linear programming problem given that the norm is l1 , which can solved
in known polynomial time using very efficient numerical
solvers [5]. Or if the norm is l2 , we can solve it as a least
square problem. Equation 5 shows that a completely noncalibrated function known to have a monotonic relation with
the phenomena of interest can be calibrated with the help
of an additional sensor that provides the aggregated information, y0 . Thus, once we deploy indirect sensors near
appliances of interest the system can automatically calculate calibration parameters used to computes real-time power
consumption of the individual appliances. Although not explicit, by letting si (t) be an on/off indicator function for an
individual appliance the formulation estimates the average
power consumption of an appliance over time.
We use the l1 norm in the ViridiScope system, because it is
more robust to outliers than the l2 norm. It is well known
that the l1 regression, or Least Absolute Value Regression,
solves for the median value. The l2 regression, or Least
Mean Square Regression, solves for the average value [5].
The median operator is more robust to outliers than the mean
operator.
In addition, a real-time measure of the system performance
makes the system more autonomous and adaptive. Equation
5 implies that a performance measure can
P be the normal
def N
p̂ (t)−y (t) ized estimation error, Performance = i=1 yi0 (t) 0 ,
where p̂i (t) is the estimated power consumption of the i-th
appliance. This measure is always available because y0 (t) is
always available. By monitoring this value, the system can
adapt to performance degradation by re-solving Equation 5.
PRACTICAL CONSIDERATION
Because indirect sensors measure ambient conditions to detect power states, they are prone to external noise sources
that also influences the ambient condition. Although l1 regression is intrinsically robust to the outlier problem, it performs better when the system has a better signal-to-noise ratio. Thus, in a real system, we want to suppress such noise.
Filtered Signal
Cases
I
II
Variance of ADC
20000
15000
Compressor
On
Compressor
Off
10000
III
5000
Participating Appliances
PC, refrigerator, table lamp
PC, refrigerator, table lamp,
alarm clock, router, settop box
2 laptops, massage chair, TV,
water heater, 2 lights, 2 table lamps
0
!5000
0
200
400
600
800
1000
Table 1: Evaluation Scenarios
1200
Time[s]
Figure 11: A simple thresholding of a statistical signature with an
adaptive filter helps to detect states more reliably.
Rejecting Undesired Noise via Adaptive Filtering: Acoustic Case
Figure 6 shows that a light sensor in a dark room deems to
have a very good signal to noise ratio. Even if the light sensor is exposed to ambient light, by placing the sensor close
to the lamp we can still achieve an excellent signal to noise
ration (Figure 9). Therefore, no sophisticated filters are necessary in many cases.
A simple microphone can detect the compressor’s power state
(see Figure 5). However, it also captures other ambient noise
sources adjacent to the refrigerator, including cooking noise,
human voices, and TV sound (Figure 10). Thus, the raw
signal is not very reliable to detect the compressor’s on/off
status.
The story changes quickly when using one additional microphone. Noting a high spatial correlation among the sampled
signals, a simple adaptive filter can reject the excess noise.
For example, if one of the microphones is placed close to
the compressor and the other one on the counter top then a
simple scale estimator and subtraction of the two sampled
signals is enough to clean up the signal (Figure 11).
One possible form of an adaptive filtering for acoustic based
detection is,
s1 = s + n
s2 = κn
: on the refrigerator
: on the counter top,
(6)
where s is the sound signature from the compressor, n the
ambient noise, and κ is a scale factor. When the compressor
is off, s is 0 and simple training for κ is possible by solving
2
2 2
2
σ (s2 ) = κ σ (n) = κ (s1 )
2
2)
κ2 = σσ2 (s
(s1 ) ,
(7)
where σ 2 (·) is the variance operation. Given κ, a simple
adaptive filter effectively rejects the ambient noise and thus
allows a reliable detection of the compressor state (Figure
11).
sensor configurations: (1) A magnetometer on the power
cord to the computer, a light sensor near the table lamp, and
a light and acoustic sensor to monitor the refrigerator. (2) A
magnetometer on the power cord of each of the three appliances. For Case II, we add an alarm clock, a wireless router,
and a cable TV settop box to the Case I configuration. The
goal is to test the system with unmonitored appliances, and
thus the power states for these three additional devices was
not monitored. In Case III, we let the ViridiScope system
estimate the average power consumption of 9 different appliances by monitoring only on/off status of the appliances.
For all the experiments, the main power line was monitored
with a commercial power meter that measures the real power
consumption. We used a Crossbow MicaZ mote connected
to this power meter to send the measurements y0 (t) to the fusion center. Ground truth data was collected with a WattsUp
Pro [11] connected to every device. These measurements
were used to evaluate the accuracy of our estimation.
Case Study I: Three Appliances
Figure 12 illustrates the information processing of the ViridiScope system for the case of one magnetic, one light, and
one sound/light combination sensor. As mentioned earlier,
there is a small error introduced in the power-state estimation, because it assumes that the power consumption of a
device in one power state is constant. Figure 13 depicts this
effect. While the error for the magnetic sensor power estimation quickly varies around 0, the error for the refrigerator
and the lamp slowly varies because their power consumption
patterns are monotonic. We note that the average power consumption is accurate, and the accumulated errors are small
(see Figure 14).
When we use only magnetometers, ViridiScope performs
very well with estimation accuracy almost the same as in
Figure 12. In summary for only the magnetometers, the
mean errors(the standard deviations) for the PC, the lamp,
and the refrigerator are 1.29%(2.74%), 0.04%(0.05%), and
-1.05%(4.24%), respectively. The estimation accuracy for
the lamp is particularly good because it consumes almost
constant power.
Case Study II: With Uninstrumented Appliances
EVALUATION
Evaluation Setting
To test and validate the ViridiScope design concept, we conducted several experiments in a 2-person apartment. We
chose three cases with increasing complexity (Table 1). In
Case I, we estimate power consumption of a desktop computer, a table lamp, and a refrigerator using two different
In the case where uninstrumented appliances are present, we
have to introduce an artificial appliance that is always on.
The power consumption of this artificial appliance will contain the accumulated power consumption of all the uninstrumented devices. In our test scenario, we added an alarm
clock, a WiFi access point, and a cable TV set-top box. The
power consumption of the three appliances is about 2.7W,
Raw Sensor Reading
Total Power
100
600
PC Set [Magnetic]
50
500
y0(t) [W]
0
400
!50
300
!100
5
200
10
15
20
25
30
35
Table Lamp [On/off State]
100
0
On
5
10
15
20
25
30
35
Time[min.]
Off
5
10
15
20
25
30
35
Refrigerator [Power States]
Both On
ViridiScope
Comp. On
Light On
Both Off
5
10
15
20
25
30
35
Time[min.]
ViridiScope: Estimated Appliance Power
True Appliance Power
200
200
PC Set Power [W]
PC Set Power [W]
150
150
100
100
50
50
0
100
5
10
15
20
25
30
35
Table Lamp Power [W]
80
0
100
60
60
40
40
20
5
10
15
20
25
30
35
Table Lamp Power [W]
80
20
0
300
5
10
15
20
25
30
35
0
300
5
10
15
20
25
Refrigerator Power[W]
200
200
100
100
0
5
10
15
20
25
30
35
30
35
Refrigerator Power[W]
0
5
10
15
Time[min.]
20
25
30
35
Time[min.]
Figure 12: The ViridiScope system takes the total power consumption, magnetic field, and internal power states information from heterogeneous
sensors including magnetic, light and acoustic sensors. It then solves the l1 norm minimization problem (Equation 5) to compute calibration
parameters. The calculated calibration parameters are used to estimate appliance-level power consumption. The estimated power consumption(left
bottom) tracks the true power consumption(right bottom) very well.
Case I: Estimation Error Plot
Accumulated Energy Consumption
20
Error[%]
10
0
!10
Error[%]
!20
200
10
15
20
25
Time[min.]
30
35
40
Table Lamp Estimation Error
10
True Consumption
Estimated Consumption
100
50
0
0
!10
!20
200
Error[%]
Mean Error: 0.37%
Std: 2.53%
5
Energy Consumption[Wh]
150
PCSet Estimation Error
Mean Error: 0.14%
Std: 0.37%
5
10
15
20
25
Time[min.]
30
35
40
Refrigerator Estimation Error
10
PC
Lamp
Refrigerator
Figure 14: Case I: Comparison among true accumulated energy
consumption of the appliances and their estimation.
0
!10
!20
Mean Error: !0.36%
Std: 1.42%
0
5
10
15
20
25
30
35
40
Time[min.]
Figure 13: Case I: The estimation error of Figure 12. The estimated
system power consumption of each individual appliance has very good
accuracy, even when they are simultaneously on.
to instrument the appliances. This is a case that would work
well with the system by Patel et. al. [23]. To test the accuracy, we test the algorithm using 9 different appliances.
Table 2 summarizes the outcome. The error is consistently
less than 10%.
RELATED WORK
6.3W, and 22W respectively. This is a total of 32W. Figure
15 illustrates the estimated, and true power consumption for
this scenario. We can see that, except in some very small
transition cases, the estimated power consumption is close
to the real power consumption. As a matter of fact, ViridiScope estimates the ghost power consumption at 32.1W.
Case Study III: 9 Appliances with On/off detection
If we are only interested in the average power consumption
of each appliance, it is enough to know only the on/off status
of an appliance. Therefore, we use light or acoustic sensors
Resource Monitoring in Homes
Directly metering the consumption of every electrically operated light, device, or appliance requires current and voltage sensing at every end-point or circuit leading to that endpoint. Such sensors are costly and often require installation
by an electrician. An exception to this are plug-in power meters for appliances that are plugged into wall sockets such
as Kill-A-Watt and Watts UP that can be installed in series.
Moreover, such technology is limited in its ability to aggregate data from multiple sensors in real-time for fusion, analysis, and visualization. Products based on such technology
Estimated Power vs True Power(Stacked Plot)
Power[W]
400
Clock,AP,Settop Box
PCSet
Lamp
Refrigerator
Estimated Power
300
200
100
0
Power[W]
400
5
10
15
20
25
30
Clock,AP,Settop Box
PCSet
Lamp
Refrigerator
True Power
300
200
100
0
Appliance
Light 1
Light 2
Massage Chair
Table Lamp 1
Table Lamp 2
Water Heater
TV
Laptop 1
Laptop 2
True Power[W]
59.8
31.9
46.2
68.1
14.7
1623.6
100.1
38.9
31.4
Estimate[W](Error)
58.4(-2.32%)
34.2(7.02%)
45.8(-1%)
69.3(1.73%)
15(2.54%)
1620.6(-0.18%)
101(0.94%)
38.3(-1.45%)
29.2(-7%)
Table 2: Case III evaluation
5
10
15
20
25
30
Time[min.]
Figure 15: Case II: In the presence of uninstrumented appliances,
ViridiScope still succeeds in determining the individual appliance
power consumption. On average, the uninstrumented devices (alarm
clock, WiFi access point, and set-top box) consume 32W. ViridiScope
estimates their power consumption to be 32.1W.
are often designed as data loggers and limited to displaying and storing data on the unit itself, or sending one or two
channels of sensor data to remote display units or to a PC via
serial, USB, or Ethernet cables [1, 2]. Recently developed
wireless sensors that can measure the power consumption of
a power outlet have been presented in research setups [17].
Even though they are a promising step toward in ease of
use and management, they have the same cost and laborintensive installation problem as already existing solutions.
Another problem is that such in-line monitoring momentarily interrupts services. Even worse, appliances that do not
have standard AC connectors or that are hard-wired into the
power distribution network (e.g. ceiling lights, water heater,
etc.) can not be monitored with such direct sensors [15].
Similar to our approach is the work on non-intrusive appliance load monitoring (NALM) [12, 15]. Instead of a peroutlet sensor, a single central sensor is envisioned to monitor
a circuit with multiple electrical loads that operate independently. In the extreme case, this single sensor is situated at
the electrical meter for the entire building. Sophisticated statistical signature detection algorithms are used to analyze the
current and voltage waveforms to separate out the individual loads, their state transitions, and their energy consumptions. Models are used to describe individual appliances or
groups of appliances to assist in the disaggregation of the
total load into its constituents. Issues of cost, size due to circuit complexity, and the accuracy of appliance detection and
monitoring were identified as problems [12]. In particular,
NALM has problems in coping with multi-state appliances
that change their power profile over time, as well as certain
two-state appliances. The training phase for NALM is quite
intrusive and labor intensive in the sense that someone must
feed signature, patterns, or similar information to the system
so that the system can infer what’s going on based on the signature. A distinctive aspect of ViridiScope is that it makes
power monitoring almost autonomous, is able to monitoring
variable power consumption, and can monitor power consumption of multiple appliances that are simultaneously on.
Home Infrastructure Monitoring
Several recent studies have shown monitoring infrastructure
leads to interesting conclusions of what is going on in a
household and provides resource consumption relevant signals. Often times this information can be extracted through
simple interfaces and means. For example, Patel et. al.
[23] monitor the electrical noise within the power-lines of
a house. They exploit the fact that each appliance introduces
a unique noise signature. They can infer if an appliance is
on or off, by detecting and identifying this signature, but not
its actual power consumption. Our approach is complementary to their result, and our algorithm is able to incorporate
results from their system to infer the power consumption of
an appliance. In a different context, Fogarty et. al. [13] investigated monitoring the plumbing system by using microphones on pipes to infer water activity in a household. Both
systems [13, 23] are easy to install but employ complex calibration mechanisms in order to learn the detection patterns.
Though these systems capture user behavior, we consider
them complementary for our approach because our ViridiScope formulations can incorporate with them to estimate actual consumption numbers with a slight modification.
LIMITATIONS AND FUTURE WORK
In the current implementation, we assume that uninstrumented devices consume constant power. A better model assumes that the ghost power varies. One approach is to opportunistically calibrate the sensors so that they can track the
power consumption of the instrumented appliances. Once
they are calibrated, it is possible to infer a change in the
ghost power consumption by subtracting the sum of the power estimation from the main power consumption. The challenge is to figure out when to learn and when to estimate.
As Beckmann et. al. [4] pointed out, five design principles
for end-user sensor deployment challenges need to be considered. While the ViridiScope system design addresses the
sensor calibration issue, further researches need to be done
in regard to (1) user conceptual models for familiar technologies, (2) balancing installation usability with domestic
concerns, (3) detection mechanisms of incorrect installation
of sensors, and (4) general end-user education. A possible
first step is to exploit the performance metric. Since it is
a measure of the normalized error, this information can be
useful to users. For example, by observing this metric they
can try different types of indirect sensors on appliances, adjust sensor placement, or try to reduce the number of unin-
strumented appliances to achieve better accuracy. Moreover,
this information can be used to detect the incorrect installation of sensors and identify faults once a machine learning
mechanism is employed.
CONCLUSION
We present ViridiScope, a fine grained power monitoring
system for residential spaces using the combination of existing infrastructure and indirect sensors. By introducing indirect sensors, we extend the traditional power monitoring
dimension. Indirect sensing introduces a new sensor calibration challenge. By exploiting existing infrastructure, we
develop an autonomous sensor calibration scheme that automates the sensor calibration procedure. Experiments show
that the system tracks the per-appliance power consumption
to less than 10% error. In addition, ViridiScope can easily
monitor the power consumption of appliances with multiple
simultaneously active appliances as well as variable power
consumption.
The challenges of sustainability is an impending issue to the
global society. We believe that ViridiScope is a step towards energy efficient homes, because measurement is one
of the most critical components. ViridiScope leverages ubiquitous technologies and their networking capabilities to extract valuable information from else meaningless data. It
is just one example showing that a shrewd combination of
various information sources is the key, and ubiquitous networked sensing and computation devices make this a feasible concept.
ACKNOWLEDGEMENT
The authors like to thank to our shepherd, James Scott, for
his guidance in preparation the final version of the paper,
Roy Shea for helping revise the paper, and the anonymous
reviewers for their helpful comments. This material is based
upon work supported by the NSF under award # CNS-0627084 and CCF-0820061, and by the UCLA Center for Embedded Networked Sensing. Any opinions, findings and conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the views
of the funding agencies.
REFERENCES
1. Energy monitoring technologies inc,
http://www.energymonitor.com/.
2. TED: The Energy Detective,
http://www.theenergydetective.com.
3. L. Balzano and R. Nowak. Blind calibration of sensor
networks. In ACM IPSN 2009.
4. C. Beckmann, S. Consolve, and A. LaMarca. Some
Assembly Required: Supporting End-User Sensor
Installation in Domestic Ubiquitous Computing
Environments, In UbiComp 2004.
5. S. Boyd and L. Vandenberghe. Convex Optimization.
Cambridge University Press, 2004.
6. A. Burger. Meter, meter on the wall: Giving & taking
from a smarter grid. triplepundit, 2007.
7. D. K. Cheng. Field and Wave Electromagnetics.
Prentice Hall, 1989.
8. M. Chetty, D. Tran, and R. E. Grinter. Getting to green:
understanding resource consumption in the home. In
UbiComp 2008.
9. J. Dahl and L. Vandenberghe. CVXOPT.
http://abel.ee.ucla.edu/cvxopt.
10. S. Darby. The effectiveness of feedback on energy
consumption: A review for DEFRA of the literature on
metering, billing and direct displays. Tech. report,
Environmental Change Institute, University of Oxford,
2006.
11. Watts Up? Pro. http://www.wattsupmeters.com/.
12. EPRI. Noninstrusive appliance load monitoring system
(NIALMS): Beta-test results. Tech. report, 1997.
13. J. Fogarty, C. Au, and S. E. Hudson. Sensing from the
basement: A feasibility study of unobtrusive and
low-cost home activity recognition. In ACM UIST 2006.
14. J. Froelich, K. Everitt, J. Fogarty, S. Patel, and
J. Landay. Sensing opportunities for personalized
feedback technology to reduce consumption. In the
CHI workshop on Defining the Role of HCI in the
Challenge of Sustainability, 2009.
15. G. Hart. Nonintrusive appliance load monitoring. In
Proc. of the IEEE, 80(12):1870–1891, Dec 1992.
16. Honeywell. Hall effect sensing and application. Tech.
report.
17. X. Jiang, S. Dawson-Haggerty, P. Dutta, and D. Culler.
Design and implementation of a high-fidelity AC
metering network. In ACM IPSN/SPOTS 2009.
18. Y. Kim, T. Schmid, Z. M. Charbiwala, J. Friedman, and
M. B. Srivastava. NAWMS: Nonintrusive autonomous
water monitoring system. In ACM SenSys 2008.
19. K. D. Lee. Electric load information system based on
non-intrusive power monitoring. PhD thesis, MIT,
2003.
20. S. B. Leeb, S. R. Shaw, and J. R. Kirtley Jr. Transient
event detection in spectral envelope estimates for
nonintrusive load monitoring. IEEE Trans. on Power
Delivery, 10(3):1200–1210, Jul 1995.
21. A. H. McMakin, E. L. Malone, and R. E. Lundgren.
Motivating residents to conserve energy without
financial incentives. Environment and Behavior
Journal, 2002.
22. MOSEK Inc. http://www.mosek.com.
23. S. N. Patel, T. Robertson, J. A. Kientz, M. S. Reynolds,
and G. D. Abowd. At the flick of a switch: Detecting
and classifying unique electrical events on the
residential power line. In UbiComp 2007.
24. P. C. Stern. Information, incentives, and
proenvironmental consumer behavior. Journal of
Consumer Policy, 1999.
25. The Blue Line Company. The power cost monitor.
26. K. Whitehouse and D. Culler. Calibration as parameter
estimation in sensor networks. In WSNA 2002.
27. World Business Council for Sustainable Development.
Energy efficiency in buildings: Business realities and
opportunities. Tech. report, 2008.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising