Embed Size (px)
Citation preview
1
Abstract — The data collected by smart meter contains a lot of
useful information. One potential use of the data is to track the
energy consumptions and operating status of major home
appliances. The results will enable homeowners to make sound
decisions on how to save energy cost and how to participate in
demand responses. This paper presents a new method to
breakdown the total power demand measured by the smart meter
into the individual appliances. A unique idea of the proposed
method is that it utilizes the electrical signatures associated with
the entire operating cycle of an appliance to identify and track the
appliance. As a result, appliances with multiple operating modes
can be tracked.
and system have been developed and deployed to real houses in
order to verify the proposed method.
Index Terms— Load management, Load signatures, Time-of-Use
Price, Demand response, Nonintrusive Load Monitoring.
I. INTRODUCTION
mater meter discloses the dynamic process of electricity
consumption which was not supported by an old electrical
meter. Sudden changes of overall signal can be easily observed
as rising or falling edges and those changes are usually caused
by state changes of internal appliance loads.
Time/Clock
Power/w
10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30
500
600
700
800
900
Fig. 1. Real-time data acquired via smart meter
However, this overall information cannot be directly used by
ordinary householders for energy saving purpose since the
meanings of those edges are never explained. On the other hand,
what householders can directly operate are only individual
appliance loads but not the overall signal sensed by smart
meter. To solve this contradictory problem, two independent
techniques were developed: sub-metering and Nonintrusive
Load Monitoring (NILM). Compared to the former, NILM is
much more cost-effective. It monitors variety of loads from a
single-entry point (usually is a utility revenue meter) instead of
individual sub-metering. Since all the load signals aggregate at
the entry point, NILM algorithms do the reverse---decode the
overall signal and restore the instant state Si(t)) of load i.
Combined with total power curve P(t), energy consumption of
load i over period T is solved in the form of:
(1)
Such information is essential for a household to make sound
energy saving decisions because they are now able to justify if
their appliances are less efficient or out of date and decide how
to change their electricity usage patterns according to coming
Time-of-Use rates[1-4]. Also, as for utility side, understanding
of detailed load behavior is significant for demand forecasting,
demand response programs development and price design [5].
Existing NILM algorithms [8-16] treat all the appliances as a
single-state model which has a pair of identical ON/OFF edges
and a flat middle operation process. The major drawback of
single-state model is that it cannot reflect the real operation
processes of complex loads such as continuous-varying
appliance and multi-state appliance shown in Fig.2[9].
Single-state
PowerMulti-stateContinuous
varying
Time
Fig. 2. Power curves of three types of loads
Continuous-varying appliance usually has a pair of different
ON/OFF edges and a gradual changing curve in the middle.
Multi-state is more commonly seen as heavy or complicated
loads such as furnace and washer. Furnace has more than one
working stage according to environmental temperatures and
washer has more operation steps like rinse and drainage
following an order pattern.
TABLE I
LOAD TYPE AND EXAMPLES
Process Window Based Appliance Monitoring
Technique for Smart Meters (V1.0)
M.Dong, Student Member, IEEE, W.Xu, Fellow, IEEE
S
2
Load type Examples Edge Process
Single-state Light bulb;
Toaster
ON=OFF
Flat middle
Continuous varying Fridge;
Freezer
ON≠OFF Varying
Multi-state Furnace;
Washer
Multiple edges Varying or flat
The other challenge brought in by continuous-varying and
multi-state appliances is the existence of them adds a lot more
edges (states) into system which makes original single-state
load signature indistinct and less identifiable. For example, a
100W rising edge may not necessarily be an ON of a 100W
bulb but a middle edge of a washer.
Recent researches of NILM focus on more sophisticated
algorithms such as genetic and neural-network [11-16] to
“guess” out possible combinations of a certain number of
single-state appliances from the aggregated signal. The issues
interfere with massive and time consuming training covering
all the appliances in a system and vulnerability to user
replacement or newly purchased appliance---once the inventory
is changed, training process has to be reset.
0 50 100 150 200 250 300-40
-30
-20
-10
0
10
20
30
40
50
Training
Meter side
0 50 100 150 200 250 300-15
-10
-5
0
5
10
15
0 50 100 150 200 250 300-4
-3
-2
-1
0
1
2
3
4
0 50 100 150 200 250 300-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Fridge Stove Microwave
0 50 100 150 200 250 300-25
-20
-15
-10
-5
0
5
10
15
20
0 50 100 150 200 250 300-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 50 100 150 200 250 300-3
-2.5
-2
-1.5
-1
-0.5
0
0.5
Training
0 50 100 150 200 250 300-40
-30
-20
-10
0
10
20
30
40
50
Meter sideFreezer Computer TV
……
Fig.3. Training process of recent NILM algorithms
This paper presents a novel NILM approach based on a
completely new load model---window model. This model
reflects both the terminal and middle characteristics, which can
successfully separate mixture of single-state, multi-state and
continuous-varying loads composed system with a promising
success rate. On the other hand, it avoids considerable training
process and uses an easy-accessible registration step for system
setup and signature database initiation.
In this paper, the concept of load window and window
signatures is firstly presented. Then overall procedure of
appliance event identification based on window model is
addressed in section III. Detailed signature similarity
evaluation methods are discussed in section IV. In section V, a
practical and convenient way of window signature collection is
stated. Finally, the experiment and verification results are
given.
II. LOAD WINDOW SIGNATURES
Load modeling is a process of abstracting load behavior into
quantified indices that can be utilized to describe and represent
load. In terms of NILM, loads are usually modeled via several
key signatures or signature groups used for identification
purposes. For example, using single-state model, a fridge may
be modeled as a current source injecting a current waveform of
specific magnitude, angle and shape to the entry point[6-7]. The
major drawback of single-state model is it assumes entire
process of load can be presented using a single steady
waveform, which is often not true. As discussed above, in this
paper, a more advanced modeling approach---load window is
adopted to present a load process more naturally and precisely.
A. Load window concept
Non-overlapping
window
Power
Time
Overlapping
window
Fridge
Fridge
Light
Fig.4. Non-overlapping window and overlapped window
Single-state model treats an appliance process as a pair of
identical ON/OFF edges with a flat middle. In contrast, window
model depicts the entire process of load, not only including
multiple different edges belonging to the load following a
certain sequence but also the middle curve and non-electric
characteristics inside.
As shown in the figure above, a load window is actually
defined as the aggregated signal occurring between the ON and
OFF of the load. Under different circumstances, there are
generally two types of windows: Non-overlapping window and
overlapped window. Non-overlapping window indicates a
period of aggregated signal without any changes coming from
other loads. In contrast, overlapped window not only includes
the process characteristics of the load but also distortions
caused by other loads. For example, during midnight when
most other appliances are inactivated, fridge is more likely to
occur as non-overlapping window; however, during daytime,
fridge process may often be overlapped with other appliances
such as microwave and stove, hence is often seen as overlapped
windows.
Non-overlapping window reflects the essential and unique
characteristics of a certain load itself and is especially
important. It is treated as a “standard” pattern which can be
previously stored in a database for further uses such as NILM
identification. In reality, only short duration (toaster) or
always-on appliances (fridge) have more chances to present
themselves in the form of non-overlapping windows at times.
Most of them will overlap with others. One important property
of overlapped window is it always includes the complete
characteristic pattern of a non-overlapping window. In another
word, all the signatures of a load’s non-overlapping window
(standard) are expected to be seen in its overlapped window
(real data). Non-overlapping window is a subset of overlapped
window.
3
B. Load window signatures
According to the discussion of window types, it is known
that the signatures of a non-overlapping window are unique and
representative and studies on those signatures becomes
essential.
Generally, a load window has 5 signature groups:
Edge signatures
Sequence signatures
Trend signatures
Time/Length signatures
Phase signature
Edge signatures
Normally, a load process owns at least two edges representing
its ON and OFF. From load power perspective, edges can be
either rising or falling. Each edge is related to an event: rising
edge usually means there is a certain load element (inside a
load) being turned on; in contrast, a falling edge means there is
an off of a certain load element. When there is more than one
functional element inside, the load usually presents more than
one edges and thus is categorized as multi-state appliance. In an
overlapped load window, some edges belong to other
appliances. However, a non-overlapping window only has the
edge signatures of the load itself.
Power
Time
Δ:P,Q,W
Fig.5. Edge signature subtracted from steady points
An edge can be easily labeled by the power (P), reactive power
(Q) and current waveform (W) change before and after it.
Those attributes are internally determined by the physical
characteristics of relevant load element such as its resistance
and harmonic spectrum. They almost stay the same if the
system voltage is not changing sharply (usually guaranteed by
utility companies). For example, whenever the circulating fan
in a basement furnace is switched on, a specific edge with
almost identical P-Q-W set will be seen. The signature values
of P-Q-W set can be easily obtained through subtraction of two
data points from two steady zones before and after the edge.
In single-state model, only one set of P-Q-W needs to be
considered since OFF is assumed to be identical reverse of ON.
In load window model, however, the number of P-Q-W sets
depends on how many element edges or events really happen.
Sequence signatures
Sequence signature describes the logic sequence of operations
of internal elements inside a load. In another word, it presents
the sequence of appearances of edges. In fact, edge signatures
divide a load process into a certain number of steady zones and
each signature set describes electric characteristic changes of
these steady zones. From sequence perspective, the exchanges
of those states strongly follow a certain pattern for most times.
For most simple case, a 100W bulb is expected to experience
a +100W ON and -100W OFF. Actually, all of single-state
appliances have a negative change in the end following an
opposite positive change at the beginning. For multi-state
appliances, sequence can be very important and distinctive.
Two types of basic event sequence: repetitive sequence, fixed
sequence and their mixture is discussed respectively as
following.
1200 1220 1240 1260 1280 1300 13200
500
1000
1500
Power
Data points
Fig.6. Repetitive sequence
Stoves, dryers or some coffee makers are typical multi-state
appliance with repetitive sequence. Controllers inside them
operate their heating elements with respect to their inner
temperature sensor or timing device. When the temperature is
high to some extent, heating element will be disabled for a
while and may be revoked when temperature is low again.
Their working cycle include several small but similar minor
cycles. From power curve, it is expected to own a repetitive pair
of rising and falling edges. This signature is sometimes even
more unique than edge value itself. For example, a house may
have a 500W toaster and a 500W stove. However, repetitive
edges found around evening (cooking time) have great chances
to be a stove.
Another category of multi-state loads like washer and
dishwasher usually follow fixed steps: water-fill, immerse,
rinse, drainage, spin-dry. During their working cycles, a fixed
pattern such as +50W,-50W,+100W,-80W,+480W,-500W will
be seen. This power pattern is very unique since the possibility
of random power edges showing up in such an order is almost
zero. Once this sequence pattern is found, it can almost be
guaranteed as washer.
80 100 120 140 160 180 200 2200
200
400
600
800
1000
1200
1400
1600
1800
Power
Data points Fig.7. Fixed sequence
Sometimes, a combination of repetitive and fixed sequence
occurs. The figure follows shows a furnace. It has repetitive
4
heating cycles. Each heating cycle includes a fixed sequence
pattern. According to the environment temperature, the heating
cycle may show up 2-5 times closely to each other.
100 150 200 250 300 350 400 450 500 550 6000
500
1000
1500
Iterative
FixedFixed
Power
Data points
Fig.8. Mixed sequence
Trend signatures
A trend signature refers to variation characteristic of curves
connecting edges. It also indicate behavior characteristic of
relevant load element. For example, an inductive motor often
accompanies with a rising spike at start; a speed adjustable
driver may experience a falling curve after start; a TV set may
experience a falling spike at moments of switching channels.
The table below lists up 7 types of variation which include all of
curve trends. It should be noted some appliance such as fridge
may have more than one type of trend signatures.
TABLE II (1)
TREND SIGNATURES
Type Curve example Power slope feature
Rising spike
4300 4400 4500 4600 4700 4800 4900 5000 5100 52000
100
200
300
400
500
600
700
800
900
A large negative
slope following a
large but smaller
positive slope
Falling spike
0 20 40 60 80 100 1200
100
200
300
400
500
600
700
800
A large positive
slope following a large negative slope
Pulses
1650 1700 1750 1800 1850 1900 1950 2000800
900
1000
1100
1200
1300
1400
Continuous large
pairs of slopes
TABLE II (2) TREND SIGNATURES
Pulses
1650 1700 1750 1800 1850 1900 1950 2000800
900
1000
1100
1200
1300
1400
Continuous large
pairs of slopes
Fluctuation
0 200 400 600 800 1000 120050
100
150
200
250
300
350
400
Continuous small slopes; signs of
slopes slowly
change
Quick vibrate
0 50 100 150 200 250 300 3500
10
20
30
40
50
60
70
80
Continuous small slopes; signs of
slopes quickly
change
Gradual falling
35 40 45 50 55 60400
600
800
1000
1200
1400
1600
1800
2000
2200
Continuous small negative slopes
Flat
0 5 10 15 200
100
200
300
400
500
600
700
800
Continuous small slopes
After continuous data points are plotted in power, Trend
signatures can be represented and detected by slope variation
modes described in the table. Those slope features are also used
as a scanning method for identification purpose.
Time/Duration signatures
The time of load window appearance relates close to its
function. Statistically, microwaves are more expected to be
seen before breakfast, lunch and supper; lights are usually
turned on after dark; fridge and furnace may run throughout 24
hours.
Duration of load window is also determined by its use
characteristics. No one keeps microwave on for more than 30
mins at a time. One working cycle of fridge is barely longer
than 40 mins. As for lights, depending on its location, it might
be on from minutes to hours. Based on statistical survey, a
universal table of load window lengths is given:
TABLE III
TYPICAL LOAD WINDOW LENGTHS
Load name Min length Max length
Fridge(cycle) >10 mins <40 mins
Freezer(cycle) >10 mins <40 mins
Furnace(cycle) >5 mins <30 mins
Stove >3 mins <45 mins
Boiler >3 mins <15mins
Washer >20 mins <90 mins
Dryer > 20 mins <75 mins
Bedroom light >0 min <5 hrs
Living room light >0 min <8.5 hrs
TV >0 min <10 hrs
5
Phase signature
There are two 120V hot wires installed in a typical North
American residential house. Hereby, the two wires can be
named as A and B. Most appliances are connected between A
or B and neutral. However, some heavy appliances such as
stove and dryer are connected between A and B to gain a 240V
voltage. Inside a meter, two CTs are connected to A and B
individually. As a result, from aggregated signals of CTs, one
can tell if one appliance is phase-A, phase-B or phase A-B type.
It should be noted phase-AB appliance has symmetrical edges
detected by both CTs. For most energy consuming appliances,
once they are placed or installed in a house, they will never be
moved. Examples are stove, fridge, microwave, furnace, lights
and even large TVs. Only very a few of them has uncertain
phase signatures such as laptop.
Meter side
CT2
CT1
Kitchen
Light
Bedroom
LightStove
Phase-A
Phase-B
Neutral
Fig.9. North America residential wiring
III. LOAD IDENTIFICATION PROCEDURE
Load identification is the most essential task for NILM
algorithm. Events of interested loads are detected, identified
and even re-organized from meter side instead of direct
load-end monitoring. To achieve identification, load signatures
are usually collected ahead of time to formulate a signature
database and this is discussed in section IV. This section
discusses general load identification procedure using load
window models and signatures already discussed in section
II.As stated in section I, traditional NILM algorithms only
focus on single state/edge of appliance and thus cannot identify
appliances from entire process perspective. The proposed
NILM uses the procedure below to identify loads:
Meter Signal
Signal split by
phases
Candidate window
selection
Signature
similarity
evaluation
Candidate
appliance
Database
* Duration signature
* Edge signature * Sequence signature * Curve signature
* Time signature
Appliance
identifed Decision making
Appliance
signature
Database
* Phase signature
Fig.10. General Identification procedure
A. Signal split by phases
Two CTs inside meter naturally divide overall signal into
two phase signals: phase-A signal and phase-B signal. To
identify phase-A loads, only phase-A signal needs to be
considered. So is for phase-B signal. One phase-B connected
light bulb will never be seen from CT-A. This simple step
easily halves the number of appliance candidates so that
signature database size can greatly shrink after signal split.
Two exceptions should be addressed: for phase A-B
appliance, since any of its edges shows up simultaneously at
both phases, it will only be left as appliance candidates at
moments when the two CTs both detect two identical edges.
Since the processes at both phases are the same, any phase of
them can be chosen for identification purpose; for portable
appliance, on the other hand, since it has an uncertain phase
connection, it will be left as candidates for both phase signals.
B. Candidate window selection
After signal split, suppose a section of aggregated signal from
CT-A is given as below:
Power/w
Time/min0 20
150
270
350
380
10 12
2
4
50
1
3
Fig.11. A section of meter signal from CT-A
Firstly, applying power slope analysis to the data, 2 rising
edges and 2 falling edges can be located (two large positive
slopes and two large negative slopes). As discussed in section
6
II, the signals between ON and OFF is defined as a load
window. Since there is no way to locate ON and OFF for a
specific load before identification is completed, signals
between any pair of rising and falling edge are considered as a
candidate window. In the figure shown above, there are in total
4 candidate windows 1-3, 2-4, 1-4, 2-3. Those candidate
windows are waited to be compared throughout candidate
appliances one by one.
Typically, a residential house may have more than 500 rising
and falling edges per day. It indicates the potential number of
candidate windows per day can be 250,000 in maximum. This
will bring too much computing burden. One way to reduce
window number is to refine these candidate windows through
appliances’ possible window lengths.
TABLE IV
CANDIDATE WINDOWS VS. CANDIDATE APPLIANCES(1)
Candidate
Appliance
Candidate
window 1-3
Candidate
window 2-4
Candidate
window 1-4
Candidate
window 2-3
Boiler
Fridge
Light
Furnace
According to table III, candidate window 2-3 is too short to be
possible for boiler, fridge and furnace. Candidate 1-4 is also too
long for boiler. Those windows are firstly ruled out before they
even enter into the next step. In fact, this window length limit
has much greater effect on refining longer period data. For a
day period, only 120-200 candidate windows will be left based
on multi-case studies.
C .Candidate window evaluation
This is the core step of identification. In this step, each of the
rest candidate windows will be compared through each of the
rest appliance candidates after step A and B according to their
rest signatures on edge, sequence, trend and time. At first, all of
those 4 signature groups will be compared respectively to get
similarity index on each one ( . Then
those 4 groups will be synthetically considered to judge if this
candidate window is matching a candidate appliance. The
mathematical calculation is completed through a linear
discriminate classifier.
, (2)
x includes similarity indices on each signature. Their
determination will be elaborated in section IV. is the weight
vector since for different types of appliances, importance of
different signatures is different. is a strictness indicator. It can
be used as a threshold. This classifier deals with two classes:
when g(x) 0, this window is determined as this appliance;
otherwise not. can be adjusted to achieve a balance between
identification rate and accuracy. The table below lists up typical
and values for some appliances.
TABLE V
EXAMPLES OF LOAD AND
Load name Distinctive signatures
Fridge Edge, trend [0.5 0.2 0.20.1] 0.85
Microwave Edge, time [0.60.20 0.2] 0.85
Furnace Edge, sequence [0.5 0.5 0 0.1] 0.85
Stove Edge, sequence, time [0.5 0.3 0 0.2] 0.85
Washer Edge, sequence [0.5 0.5 0 0] 0.8
Boiler Edge [0.8 0.2 0 0] 0.85
Laptop Edge, trend [0.5 0.2 0.3 0] 0.8
Average --- [0.56 0.29 0.07 0.08] 0.85
Generally, edge signature is always important since it
determines the electric characteristics of a window. Sequence
signature is important too, especially for multi-state appliances.
Trend signature is important for motor related and some
electronic appliances. Time signature functions accessorily and
is more effective for time-oriented loads such as kitchen
appliances. Usually weight vector stays the same for the same
type of load even when moving from one house to another. It
can be easily set based on common experience and knowledge.
The row “Average” in Table IV gives a rough setting without
load type known, which can be used to cope with unfamiliar
loads. The advantage of the weight vector is that when
comparing, there is no absolutely strong signature---various
signatures are bonded together to ensure fairness and accuracy
of system.
Identification threshold is normally set as 0.85 for most
cases. It can be lowered if imposed data noises are significant.
D. Decision making
In the end, table IV is calculated according to (1) and g(x)
values are filled in as below.
TABLE VI
CANDIDATE WINDOWS VS. CANDIDATE APPLIANCES (2)
Candidate Appliance
Candidate window 1-3
Candidate window 2-4
Candidate window 1-4
Candidate window 2-3
Boiler 0.15 -0.45
Fridge -0.3 0.15
Light -0.85 -0.85 -0.85
Furnace -0.75 -0.75 -0.75
From the signs of classifier values, window 1-3 is determined
as a boiler while window 2-4 is a fridge since their values are
greater than 0. If no positive value is found, then it means the
edges are caused by an unknown appliance probably not
registered in database (maybe not interested by users either.)
This linear classifier can also be substituted by more
advanced classifiers such as neural networks or decision tree.
Those variations are not discussed here.
IV. SIGNATURE COMPARISON
This section addresses on how to compare signatures to
obtain similarity indices .Those
7
quantities should be able to effectively reflect how similar a
window signature is with respect to a database signature of a
certain appliance.
A. Edge similarity
From the signature database, a candidate appliance only
includes its own P-Q-W edge sets. In contrast, a candidate
window may include other edges caused by overlapped
appliances. The comparison is trying to answer if this
candidate window includes all of the candidate appliance’s
edges. Thus the process is like this: each of registered edges
will be compared throughout all the edges in candidate
window one by one. Then:
(3)
: Number of edge types defined in candidate appliance.
: Recognized number of appliance edge types in candidate
window.
Power/w
Time/min
B D
0
110
10
80
Power/w
Time/min0 20
150
270
350
380
10 12
B
D
50
A
C
Candidate Window
Candidate
Appliance
Ne=2Ne
’=2
Fig.12. Edge similarity comparison
As shown in Fig.12, both of the two registered edges B-D are
found in the window ( However, if only B
exists, it is very likely the candidate window is only one part
of the appliance process and its
( .
P, Q can be easily compared since they are quantitative
values. As for current waveform W, one way is to directly
make comparison based on point-point differential of
waveforms. The other way is to quantify waveform into
magnitude and angle spectrums of key harmonic orders [6].
Selecting proper harmonic orders can also eliminate the
impact from noises and dc offset. The figure below provides
an example of harmonic spectrum obtained from an appliance
waveform:
Harmonic order
Angle/degree
1 3 5 7 9 11 13-200
-150
-100
-50
0
50
100
150
200
1 3 5 7 9 11 130
20%
40%
60%
80%
100%
Harmonic order
Magnitude/%
Fig.13. Magnitude and angle spectrum of a laptop
Since the value of is determined by three sub-attributes, as
discussed in section III, different weights can be set to those
attributes: for linear and active load such as stove, P should be
emphasized; for non-linear load such as microwave, W
should be emphasized; for reactive load such as fridge, Q
should be emphasized. Those weights can be pre-defined for
candidate appliances. Synthesizing them together, two edges
can be determined as identical or non-identical.
Overall, indicates the existence of edges of candidate
load in candidate window.
B. Sequence similarity
For ON/OFF type appliance, it has a fixed sequence of
edges; for multi-state appliance, as discussed in section II,
fixed sequence and repetitive sequence may either be found.
For fixed sequence edges, they always follow a certain
position pattern. For a candidate window, its edge positions
should comply with the position pattern defined in candidate
appliance. For example, a space heater has 5 edges in the
sequence of A-B-C-D-E. It is expected to find A-B-C-D-E in
the window. On the other hand, an A-B-D-C-E sequence may
imply a different appliance process and B-C-A-D-E is even
more different. The similarity on fixed sequence can be
quantified through a simple correlation method shown as
below:
Power/w
Time
100Candidate
Appliance
A:+100
B:+200
C: -50
D:-150
E:-100
300
B
C
D
E
Candidate
Window 1
D
A
250
A
BC
E
DC A
Candidate
Window 2E
Power/w
Time
B
Edge index
Power/w
1 2 3 4 5-150
-100
-50
0
50
100
150
200
A-B-C-D-E
B-C-A-D-E
A-B-D-C-E
Correlation factor=0.88
Correlation factor=0.44
Fig.14 Correlation factor of two power tracks
The power tracks of candidate appliance sequence and
candidate window sequence are respectively plotted in the
same graph. Correlation coefficient of two tracks can
be directly used as indicator.
(4)
X: the value vector of power track 1.
Y: the value vector of power track 2
8
As can be seen, A-B-D-C-E (window 1) is more close to
pre-defined appliance pattern A-B-C-D-E with a larger
correlation factor 0.88 than 0.44 for B-C-A-D-E (window 2).
One exception is if is found smaller than 1, then
is zero due to mismatch in the number of edge indices.
The appearances of repetitive edges are also counted in the
step of determining and only if its number is more than
one, it is recognized as an repetitive edge.
(4)
:Number of repetitive edge types defined in candidate
appliance.
: Recognized number of repetitive edge types in candidate
window.
As for a mixed sequence load such as furnace, can be
calculated based on two sub indices and which can
respectively evaluate similarities of repetitive and fixed
sequence characteristics.
(5)
C. Trend similarity
As discussed in Table II, using slope based methods can
effectively scan the candidate window and further determine
the existences of trend signatures with respect to candidate
appliance.
(6)
:Number of trend signature types defined in candidate
appliance.
: Recognized number of trend signature types in candidate
window.
D. Time similarity
In the end, the moment of appearance of candidate window t
is also compared with time signature defined in candidate
appliance. Usually, the time signature of candidate appliance
is defined as one or several hour ranges T such as
{17-23},{6-8,11-13,16-18}
(7)
V. SIGNATURE DATABASE BUILDING
As discussed above, NILM needs a customized signature
database to realize identification since appliance characteristics
are usually different from house to house. Traditionally, it may
involves a complicated and very time-consuming training
process [11-16], which may prevent uses by ordinary
householders. This paper presents an easy-accessible method to
minimize the time of building database down to half to one hour
for all major energy-consuming appliances.
It uses a unique device named register. Register is a
double-plug based device. It is firstly plugged into a regular
outlet and then interested appliance can be plugged into its
other side. The figure below shows an example. Inside register,
it has a wireless transmitter and current sensor.
Fig.15. Connection of register and system
The working principle is very easy: once it detects a current
change (means an edge) of the appliance, it sends a signal to
NILM integrated meter side. In the experiment stage, this meter
is replaced by a laptop equipped meter-data acquisition system.
Once the meter side receives a signal, it checks its CT to
determine to which phase this appliance is connected. In
another word, phase signature of this appliance is determined.
At the same time, it captures the P-Q-W signatures of appliance
edge. As time progresses, edge signatures of interested
appliance will be collected one by one.
For time signature, it can be simply completed when
householder names this appliance through computer since
appliance name usually implies time information. However, the
system also enables users’ manual input if it is unusual.
Similarly, appliance window length can also be determined
by the length between the first and the last received signals with
a reference to appliance type.
As for sequence signature, the program will automatically
analyze on number of identical edges to figure out repetitive
sequence. The rest edges are considered as fixed sequence
edges.
Trend signature can be easily extracted by slope scans
between detected edges.
So far, the five groups of signatures are all collected. The
meter-register-user input structure is unique to fast determine
the whole signature envelop. And since the electric signatures
are directly captured from meter-side, its harmonic and
waveform signatures are close to real operation under which
cancellation and attenuation effect may be concerned [17-18].
The time consumed is roughly equal to its typical working
cycle. After one appliance is registered, the register can be
moved to the next targeted appliance right away.
VI. APPLICATION AND VERIFICATIONS
The above algorithms were tested in two real residential
9
houses for several weeks with no special intention from the
landlords. The laptop based data acquisition system was
hooked to the electricity panel. A portable Zig-bee transmitter
was connected to its USB port to bridge the communication
with the appliance register. After registration was finished, a
metering signal decomposing program based on proposed
algorithm was launched and kept running. In real time, the
overall power was decomposed into appliance level. Interested
appliance events are firstly identified. Then their energy
consumption is calculated through area integration. This
method is much more accurate than traditional single-state
based energy calculation since the detailed power variation of
process can now be taken into account [15]. This is especially
significant for continuous-varying and multi-state appliance
energy.
Power/w
Time/min0 20
150
270
350380
10 12
4
50
2 A1
A2
1
3
KWhFridge=A1+A2
Fig.16. Segment area based energy calculation
The results are updated every half an hour and displayed in a
interface designed as below. As can be seen, appliance
electricity consumption information is formatted into the table
and charts. The table summarizes the total energy counted from
a certain date and converted expenses with respect to local
electricity rates. The pie chart presents the percentage
composition of individual appliance so users can be aware of
the significance of reducing a certain appliance’s consumption.
Finally, from the time distribution of energy chart, users can
understand his energy usage pattern statistically with respect to
hours. This information is quite essential for residential house
owners to adopt proper demand response strategies such as load
shifting to utility TOU rates.
Fig.17. Appliance energy decomposer software(This figure will be further
improved)
To verify the identification rates, controlled trials were
conducted continuously for days. Interested appliance events
were manually checked and recorded as comparison. The
results listed in the table below are very satisfactory for most
appliances including continuous-varying and multi-state
appliances.
TABLE VII. IDENTIFICATION RATE VERIFICATION
(THIS TABLE WILL BE FURTHER IMPROVED)
Appliance
Name
Actual
operation times
Identified
operation times
Identification
accuracy(%)
Freezer 937 874 93.3
Fridge 683 654 95.8
Furnace 82 82 100
Stove/Oven 45 45 100
Microwave 84 80 95.7
Washer 3 3 100
Dryer 4 4 100
Water boiler 47 44 93.7
VII. CONCLUSIONS
This paper systematically illustrates the process
characteristics of loads, the concept of load process window
and how to utilize process characteristics and divide windows
to detect and identify appliance events. Based on this
procedure, interested appliance information is finally extracted
from the disaggregated signal. The proposed approach paves
the way of monitoring complicated appliances and makes
NILM algorithm more universally applicable. Also this paper
proposes a convenient method and device to facilitate
user-tailored load signature database establishment. All of
those efforts are aiming to bring a really feasible NILM
solution into residential house and utility. Realistic tests and
data were used to verify the proposed method and developed
system.
REFERENCES
[1] S. Massoud Amin and Bruce F. Wollenberg, “Toward a Smart Grid”, IEEE Power & Energy Magazine, Sept/Oct 2005, pp. 34-41.
[2] H. Farhangi, “The Path of the Smart Grid”, IEEE Power & Energy
Magazine, Jan/Feb 2010, pp. 18-28. [3] Litos Strategic Communication, “The Smart Grid: An Introduction”,
United States Department of Energy, 2004.
[4] S. Rahman, “Smart Grid Expectations”, IEEE Power & Energy Magazine, Sept/Oct 2009, pp. 83-85.
[5] A. Mahmood, M. Aamir and M.I. Anis, “Design and Implementation of AMR Smart Grid System”, IEEE Electrical Power and Energy
Conference, 2008, pp. 1-6.
[6] Wilsun Xu, Alex Nassif, and Jing Yong,“Harmonic Current Characteristics of Home Appliances”, Report to CEATI International
Inc., 60 pages, March 2009.
[7] A.B. Nassif, J. Yong, Wilsun Xu and C.Y. Chung, “Comparative Harmonic Characteristics of Home Appliances”, submitted to IEEE
Transactions on Power Delivery.
[8] Sultanem, F.; , "Using appliance signatures for monitoring residential loads at meter panel level," Power Delivery, IEEE Transactions on ,
vol.6, no.4, pp.1380-1385, Oct 1991
[9] Hart, G.W. , “Non-intrusive Appliance Load Monitoring”,Proceedingsof the IEEE, vol. 80, No 12, December, pp. 1870 - 1891,1992
[10] Norford L.K., Leeb S.B., “Non-intrusive Electrical Load Monitoring in
Commercial Buildings based on Steady-state and Transient Load-detection Algorithms”.Energy and Buildings 24, pp. 51 – 64,1996
[11] Baranski, M.; Voss, J.,“ Genetic algorithm for pattern detection in
NIALM systems”, Systems, Man and Cybernetics, 2004 IEEE
10
International Conference on, Volume 4, 10-13 Oct. 2004, Page(s):3462 -
3468 vol.4 [12] Duan, J.; Czarkowski, D.; Zabar, Z., “Neural network approach for
estimation of load composition”, Circuits and Systems, 2004. ISCAS '04.
Proceedings of the 2004 International Symposium on, Volume 5, 23-26 May 2004 ,Page(s):V-988 - V-991 Vol.5
[13] Srinivasan, D.; Ng, W.S.; Liew, A.C., “Neural-network-based signature
recognition for harmonic source identification”, Power Delivery, IEEE Transactions on, Volume 21, Issue 1, Jan. 2006 ,Page(s):398 – 405
[14] Jian Liang; Ng, S.; Kendall, G.; Cheng, J.; , "Load Signature Study—Part
I: Basic Concept, Structure, and Methodology," Power Delivery, IEEE Transactions on , vol.25, no.2, pp.551-560, April 2010
[15] Jian Liang; Ng, S.K.K.; Kendall, G.; Cheng, J.W.M.; , "Load Signature
Study—Part II: Disaggregation Framework, Simulation, and Applications," Power Delivery, IEEE Transactions on , vol.25, no.2,
pp.561-569, April 2010
[16] Ruzzelli, A.G.; Nicolas, C.; Schoofs, A.; O'Hare, G.M.P.; , "Real-Time Recognition and Profiling of Appliances through a Single Electricity
Sensor," Sensor Mesh and Ad Hoc Communications and Networks
(SECON), 2010 7th Annual IEEE Communications Society Conference on , vol., no., pp.1-9, 21-25 June 2010
[17] Nassif, A.B.; Acharya, J.; , "An investigation on the harmonic attenuation
effect of modern compact fluorescent lamps," Harmonics and Quality of Power, 2008. ICHQP 2008. 13th International Conference on , vol., no.,
pp.1-6, Sept. 28 2008-Oct. 1 2008
[18] Nassif, A.B.; Wilsun Xu; , "Characterizing the Harmonic Attenuation Effect of Compact Fluorescent Lamps," Power Delivery, IEEE
Transactions on , vol.24, no.3, pp.1748-1749, July 2009
1
Abstract—The data collected by smart meters contain a lot of
useful information. One potential use of the data is to track the
energy consumptions and operating statuses of major home ap-
pliances. The results will enable homeowners to make sound de-
cisions on how to save energy and how to participate in demand
response programs. This paper presents a new method to break-
down the total power demand measured by a smart meter to those
used by individual appliances. A unique feature of the proposed
method is that it utilizes diverse signatures associated with the
entire operating window of an appliance for identification. As a
result, appliances with complicated middle process can be
tracked. A novel appliance registration device and scheme is also
proposed to automate the creation of appliance signature data-
base and to eliminate the need of massive training before identi-
fication. The software and system have been developed and de-
ployed to real houses in order to verify the proposed method.
Index Terms— Load management, Load signatures,
Time-of-Use Price, Demand response, Nonintrusive Load Moni-
toring.
I. INTRODUCTION
he increased public awareness of energy conservation in
recent years has created a huge interest in home energy
consumption monitoring. According to a recent market re-
search report [1], consumers show substantial interest in tools
that can help them manage their household energy use and
expenses. A critical link to address this need is the smart me-
ters. However, the smart meters currently available in the
market can only provide the energy consumption data of a
whole house. They cannot tell which appliances in the house-
hold consume the most energy or are least efficient. Also, to
take full advantage of Time-of-Use rates, householders need to
be informed of their usage patterns. Such information is essen-
tial for a household to make sound energy saving decisions and
participate in utility demand response programs [2-3].
In response to this need, two research directions have
emerged. One is to connect energy monitors to individual ap-
pliance of interest and to communicate the recorded data to a
data concentrator [4]. While such a sensor network based sys-
tem can provide accurate measurement of appliance energy
The authors gratefully acknowledge the support provided by Natural Sci-
ences and Engineering Research Council of Canada (NSERC) on this work.
M.Dong and W.Xu are with the Department of Electrical and Computer
Engineering, University of Alberta, Edmonton, AB T6G 2V4,Canada (email: [email protected])
consumption, it can be costly and complex to implement. The
second direction is to identify and track major home appliances
based on the total signal collected by utility meters, which is
called Nonintrusive Load Monitoring (NILM) method [5].
Compared to the former, the NILM direction is more attractive
to customers and utilities due to its high cost efficiency and less
effort on installation.
The problem to be solved by NILM approach can be stated as
follows: All the load signals aggregate at the entry point of a
house as ( )P t and NILM algorithms do the reverse---decode
the overall signal into various components ( )iP t that are at-
tributed to specific loads (appliances) i.:
1 2( ) ( ) ( ) . ( )nP t P t P t P t (1)
It must be noted that the goal of the above approach is to
extract the ( )iP t trends of large appliances in a home. It is not
intended to and there is no need to track small devices such as
phone chargers. Such devices don’t consume significant
amount of the energy and signatures of them can be “sub-
merged” in aggregated signal; however, signatures of major
appliances are better kept in aggregated signal. In a typically
home, there are about 10 to 20 large power consuming appli-
ances. The decoding process makes use of the unique signa-
tures of such appliances observable at the smart meter location
of a house to extract the ( )iP t trends.
Time/Clock
Power/w
10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30
500
600
700
800
900
Fig. 1. Real-time data acquired via smart meter
Example of smart metering data is shown in Fig.1. Appli-
ance events are observed as ON/OFF edges (arrows). Existing
NILM algorithms [5-11] treat all the appliances as a single-state
model which has a pair of identical ON/OFF edges and a con-
stant power demand between them. This is because original
studies mainly aim to help utilities conduct load studies without
An Event Window Based Load Monitoring
Technique for Smart Meters (Final Version)
Ming Dong, Student Member, IEEE, Paulo C. M. Meira, student Member, IEEE, Wilsun Xu, Fellow,
IEEE, Walmir Freitas, Member, IEEE
T
2
intrusion [5-6] and thus adopt simplified models. However, for
accurate energy monitoring purpose, real operation processes
of complex loads such as continuous-varying appliances and
multi-state appliances shown in Fig.2 need to be captured and
treated.
Single-state
PowerMulti-stateContinuous
varying
TimeFig
. 2. Power curves of three types of loads
Continuous-varying appliance usually has a pair of different
ON/OFF edges and a gradual varying power demand in the
middle. Multi-state is more commonly seen as heavy or com-
plicated loads such as furnace and washer. Furnace has more
than one working stage according to environmental tempera-
tures and washer has more steps like rinse and drainage fol-
lowing a certain operation pattern.
TABLE I
LOAD TYPE AND EXAMPLES
Load type Examples Event Power demand
Single-state Light bulb;
Toaster
ON=OFF
Flat
Continuous varying Fridge; Freezer
ON≠OFF Varying
Multi-state Furnace;
Washer
Multiple events Varying or flat
The other challenge faced by some of the published works is
that they need a time-consuming training/learning process to
support their algorithms such as genetic and neural-network
before they can work [7-9]. Such combination based ap-
proaches are vulnerable to changes in the appliance inventory.
Once a major appliance is replaced, re-training has to be con-
ducted.
To address these issues, this paper presents a novel NILM
technique to identify all three types of loads. The key idea is to
use the various signatures of the entire operating window of an
appliance for identification. Since an event window contains
various operating states and other information of the appliance,
the proposed technique is also much more reliable. Further-
more, a convenient appliance registration method is proposed
to automate the creation of appliance signature database, which
reduces users/algorithms’ efforts from training.
II. THE CONCEPT OF EVENT WINDOWS
An event window is defined as the collection of all signatures
between any pair of rising/falling step-changes (events) of the
power demand as measured by the smart meter. Sample load
windows are shown in Fig.3. Window 1 contains one ON and
one OFF event associated with one appliance. There is no ac-
tivation of other appliances in between. This is called the
non-overlapping window. Non-overlapping window contains
complete signature information about an appliance. Window 2
is called overlapping window as it contains an ON event asso-
ciated with another appliance. In reality, only short duration
(toaster) or always-on appliances (fridge) have more chances to
present themselves in the form of non-overlapping windows.
Most of them will overlap with others. The main idea of the
proposed technique is to identify and pick out the right win-
dows that are represented by interested appliances. This is
accomplished with the assistance of window signatures or
characteristics. Each window contains five types of signatures
listed as below.
Non-overlapping
window 1
Power
Time
Overlapping
window 2
Fridge
Fridge
Light
Fig.3. Non-overlapping window and overlapping window
Edge signatures
Sequence signatures
Trend signatures
Time/Duration signatures
Phase signature
A. Edge signatures
An edge refers to the event of the operating state of an ap-
pliance, which can be seen as a step change in its power de-
mand. The edge can be either rising or falling. Each edge can
be characterized by the changes in power (P), reactive power
(Q) and current waveform (W) as shown in Fig.4 [10,12-14].
Those attributes are generally fixed for each appliance if the
system voltage does not change sharply. In single-state model,
only one set of P-Q-W needs to be considered since OFF is
assumed to be identical reverse of ON [5]. In event window
model, however, the number of P-Q-W sets depends on how
many events really happen.
Power
Time
Δ:P,Q,W
Falling edgeRising edge
Fig.4. Edge signature subtracted from steady points
B. Sequence signatures
Sequence signature describes the logical sequence of opera-
tion events of a load. In another word, it represents the sequence
of appearances of edges. For example, a washer usually follows
the following operating modes: water-fill, immerse, rinse,
drainage and spin-dry. In a cycle, a fixed pattern such as
+50W,-50W,+100W,-80W,+480W,-500W will be seen. This
power pattern, the sequence signature, is very unique and is
essential for identifying multi-state appliances. There are three
3
types of basic event sequences: repetitive sequence, fixed se-
quence and the combination of the two.
1200 1220 1240 1260 1280 1300 13200
500
1000
1500
Power
Data points
(a) Repetitive sequence
80 100 120 140 160 180 200 2200
200
400
600
800
1000
1200
1400
1600
1800
Power
Data points (b) Fixed sequence
100 150 200 250 300 350 400 450 500 550 6000
500
1000
1500
Repetitive
FixedFixed
Power
Data points
(c) Combination
Fig.5. Different sequence patterns
Stoves, dryers or some coffee makers are typical multi-state
appliance with repetitive sequence due to their integer-cycle
controllers [15]. An example of fixed sequence is washer.
Sometimes, a combination of repetitive and fixed sequence
occurs. Fig.5 (c) shows a furnace. It has repetitive heating cy-
cles. Besides, each heating cycle includes a fixed sequence
pattern. According to the environment temperature, the heating
cycle may show up 2-5 times closely to each other. Table II
shows some examples of appliances with the sequence patterns
as discussed above measured through experiment.
TABLE II
SEQUENCE PATTERN AND EXAMPLES
Load type Examples
Repetitive sequence Dryer; Stove; Some coffee makers
Fixed sequence Incandescent light bulb; Fluorescent light bulb; Kettle; Microwave;
Toaster; Oven; Fridge; Freezer; Computer
Combination Furnace, Some dishwashers
C. Trend signatures
A trend signature refers to variation of power demand be-
tween two edges. For example, an inductive motor often ac-
companies with a rising spike at start due to its large inrush
current; after start, as the motor speed increases, the current
drawn may decrease and form a gradual falling curve; some
electronic devices may experience an instant interruption. A
TV set may experience a falling spike at moments of switching
channels; pulses are usually caused by electronic switches. A
lot of stoves have pulses because they have an integer-cycle
controller in it. It prevents itself from overheating. Another
example is an inverter based motor device that adjusts its fre-
quency all the time; a lot of appliances have a negligible
transient term and present as almost flat curve; in contrast,
some appliances may have continuous fluctuations all the time
instead of a steady state. Table III lists up 7 types of variation
which include all types of curve trends seen in appliances. It
should be noted some appliance such as fridge may have more
than one type of trend signatures. Those trends are not only
found in one type of appliance but usually several types of
appliances due to their common electrical characteristics. TABLE III
TREND SIGNATURES
Type Curve example Power slope feature
Rising spike
4300 4400 4500 4600 4700 4800 4900 5000 5100 52000
100
200
300
400
500
600
700
800
900
A large negative slope following a
larger positive slope
Falling spike
0 20 40 60 80 100 1200
100
200
300
400
500
600
700
800
A large positive
slope following a
large negative slope
Pulses
1650 1700 1750 1800 1850 1900 1950 2000800
900
1000
1100
1200
1300
1400
Continuous pairs of
large slopes
Fluctuation
0 200 400 600 800 1000 120050
100
150
200
250
300
350
400
Continuous small
slopes; signs of
slopes slowly change
Quick vibrate
0 50 100 150 200 250 300 3500
10
20
30
40
50
60
70
80
Continuous small
slopes; signs of
slopes quickly change
Gradual falling
35 40 45 50 55 60400
600
800
1000
1200
1400
1600
1800
2000
2200
Continuous small
negative slopes
Flat
0 5 10 15 200
100
200
300
400
500
600
700
800
Continuous small
slopes
4
From continuous power points measured from smart meter,
trend signatures can be represented and detected by slope
( / )P t variation modes described in Table III. Those slope
features are also used as a scanning method for identification
purpose.
D. Time/Duration signatures
The time of load window appearance relates close to its
function. There are some statistical studies on residential load
modeling which present typical load on-hours shown in Fig.6
[19-20]:
240 2 4 6 8 10 12 14 16 18 20 22 hr
Microwave
Kettle
Toaster
Fridge
Lighting
PC
Fig.6. Typical appliances on-hours for weekends
As can be seen, microwaves are more expected to be seen
before breakfast, lunch and supper; lights are usually turned on
in the early morning or after dark; fridge and furnace are likely
to run throughout 24 hours.
Duration of load window is also determined by its function
characteristics. No one keeps microwave on for more than 30
mins at a time. One working cycle of fridge is barely longer
than 40 mins. As for lights, depending on its location, it might
be on from minutes to hours. Based on statistical survey, some
universal load window lengths are given in Table IV:
TABLE IV
TYPICAL LOAD WINDOW LENGTHS
Load name Min length Max length
Fridge(cycle) >10 mins <40 mins
Freezer(cycle) >10 mins <40 mins
Furnace(cycle) >5 mins <30 mins
Stove >3 mins <45 mins
Kettle >3 mins <15mins
Washer >20 mins <90 mins
Dryer > 20 mins <75 mins
Bedroom light >0 min <5 hrs
Living room light >0 min <8.5 hrs
TV >0 min <10 hrs
E. Phase signature
There are two 120V hot wires installed in a typical North
American residential house as shown in Fig.7. Hereby, the two
wires can be named as A and B. Most appliances are connected
between A or B and neutral. However, some heavy appliances
such as stove and dryer are connected between A and B to gain
a 240V voltage. Inside a meter, two CTs are connected to A and
B individually. As a result, from aggregated signals of CTs, one
can tell if one appliance is phase-A, phase-B or phase A-B type.
It should be noted phase-AB appliance has symmetrical edges
detected by both CTs. For most energy consuming appliances,
once they are placed or installed in a house, they will never be
moved. Examples are stove, fridge, microwave, furnace, lights
and even large TVs. Only very a few of them has uncertain
phase signatures such as laptop.
Meter side
CT-B
CT-A
Kitchen
Light
Bedroom
LightStove
Phase-A
Phase-B
Neutral
Fig.7. North America residential wiring
III. LOAD IDENTIFICATION PROCEDURE
Load identification is the most essential task for NILM al-
gorithm. Events of interested loads are detected, identified and
even re-organized from meter side instead of direct load-end
monitoring. To achieve identification, load signatures are usu-
ally collected ahead of time to formulate a signature database
and this is discussed in section V. This section discusses gen-
eral load identification procedure using event window models
and signatures already discussed in section II.As stated in sec-
tion I, traditional NILM algorithms only focus on single
state/edge of appliance and thus cannot identify appliances
from entire process perspective. The proposed NILM uses the
procedure from Fig.8 to identify loads:
Meter Signal
Split signal by
phases
Select
window candidates
Evaluate similarity
between window and
appliance candidates
Appliance
candidates * Duration signature
* Edge signature * Sequence signature * Trend signature
* Time signature
Make decision
Appliance
signature
database
* Phase signature
Fig.8. General Identification procedure
A. Split signal by phases
Two CTs inside meter naturally divide overall signal ac-
quired by smart meter into signals of two phases: phase-A
signal and phase-B signal. Accordingly, to deal with phase-A
5
signal, only phase-A loads will remain as candidates. So is for
phase-B signal. Normally, a phase-B connected light bulb will
never be seen from CT-A.
Two exceptions should be addressed: for phase A-B appli-
ance, since any of its edges shows up simultaneously at both
phases, it will be left as appliance candidates only if two CTs
can detect two identical edges at the same time. Since the
processes at both phases are the same, any phase signal can be
chosen for identification purpose; for portable appliance, on the
other hand, since it has an uncertain phase signature, it will be
left as candidates for both phase signals.
B. Select window candidates
After signal split, suppose a section of aggregated signal
from CT-A is measured as shown in Fig.9:
Power/w
Time/min0 20
150
270
350
380
10 12
2
4
50
1
3
Fig.9. A section of meter signal collected from CT-A
Firstly, applying power slope analysis to the data, 2 rising
edges and 2 falling edges can be located (two large positive
slopes and two large negative slopes) and labeled. As discussed
in section II, signal collection between any pair of rising and
falling edge is considered as a window candidate. In Fig.10,
there are in total 4 window candidates: 1-3, 2-4, 1-4, 2-3. Those
window candidates are waited to be compared throughout ap-
pliance candidates one by one.
Typically, a residential house may have more than 500 rising
and falling edges per day. It indicates the potential number of
window candidates per day can be 250,000 in maximum. This
will bring too much computing burden. One way to reduce
window number is to trim these window candidates through
appliances’ possible window lengths.
TABLE V
WINDOW CANDIDATES VS. APPLIANCE CANDIDATES (1)
Appliance
candidate
Window
candidate
1-3
Window
candidate
2-4
Window
candidate
1-4
Window
candidate
2-3
Kettle
Fridge
Light
Furnace
According to Table V, window candidate 2-3 is too short to
be possible for kettle, fridge and furnace. Candidate 1-4 is also
too long for kettle. Those windows are firstly ruled out even
before they enter into next evaluation step. In fact, this window
length limit has much greater on refining longer period data.
For a day period, only 120-200 window candidates will be left
based on multi-case studies.
C .Evaluate similarity between candidates
This is the core step of identification. In this step, each of the
rest window candidates will be compared through each of the
rest appliance candidates after step A and B according to their
rest signatures on edge, sequence, trend and time. At first, all of
those 4 signature will be compared respectively to get similarity
indices on each one ( , , ,edge seq trd timeS S S S ). Then those 4 indices
will be synthetically considered to judge if this window can-
didate is matching an appliance candidate. The mathematical
calculation is completed through a linear discriminate classifi-
er.
( ) Tg x x (2)
with
,
edge edge
seq seq
trd trd
time time
S
Sx
S
S
(3)
x includes similarity indices of each signature. Their de-
termination will be elaborated in section IV. is the weight
vector since for different types of appliances, importance of
different signatures is different. is a qualification threshold.
This classifier deals with two classes: when ( ) 0g x , this
window is determined as this appliance; otherwise not. can
be adjusted to achieve a balance between identification rate and
accuracy. Table VI lists up typical and values for some
appliances.
TABLE VI
EXAMPLES OF LOAD AND
Load name Distinctive signatures T
Fridge Edge, trend [0.5 0.2 0.2 0.1] 0.85
Microwave Edge, time [0.6 0.2 0 0.2] 0.85
Furnace Edge, sequence [0.5 0.5 0 0] 0.85
Stove Edge, sequence, time [0.5 0.3 0 0.2] 0.85
Washer Edge, sequence [0.5 0.5 0 0] 0.8
Kettle Edge [0.8 0.2 0 0] 0.85
Laptop Edge, trend [0.5 0.2 0.3 0] 0.8
Average --- [0.56 0.29 0.07 0.08] 0.85
Those weights are firstly estimated based on observation
and analysis of appliances. For example, knowing furnace and
washer have unique sequence signatures, will be empha-
sized; knowing microwave is often used before meals, is
emphasized. Generally, edge signature is always important
since it determines the electric characteristics of a window.
Sequence signature is important too, especially for multi-state
appliances. Trend signature is important for motor related and
some electronic appliances. Time signature functions accesso-
rily and is more effective for time-oriented loads such as
kitchen appliances. After weights are pre-defined, their values
will be optimized and verified through a simulation program.
This program generates numerous testing windows based on
existing load signatures and then it adjust values of and to
ensure that maximum number of correct identification can be
made for each type of load.
6
Usually weight vector stays the same for the same type of
load even when moving from one house to another. The row
“Average” in Table VI gives a rough setting without load type
known, which can be used to cope with unfamiliar loads. The
advantage of the weight vector is that when comparing, there is
no absolutely strong signature---various signatures are bonded
together to ensure fairness and accuracy of system.
Identification threshold is normally set as 0.85 for most
cases. It can be lowered if imposed signal noises are significant.
D. Make decision
In the end, Table V is calculated according to equations (2-3)
and the ( )g x values are filled in as below.
TABLE VII
CANDIDATE WINDOWS VS. CANDIDATE APPLIANCES (2)
Appliance
candidate
Window
candidate
1-3
Window
candidate
2-4
Window
candidate
1-4
Window
candidate
2-3
Kettle 0.15 -0.45
Fridge -0.3 0.15
Light -0.85 -0.85 -0.85
Furnace -0.75 -0.75 -0.75
From the signs of classifier values, window 1-3 is determined
as a kettle while window 2-4 is a fridge since their values are
greater than 0. If no positive value is found, it means the edges
are caused by an unknown appliance not registered in database
yet (maybe not interested by users either.)
This linear classifier can also be substituted by more ad-
vanced classifiers such as neural networks or decision tree.
Those variations are not discussed here.
IV. SIGNATURE COMPARISON
This section addresses on how to compare signatures to ob-
tain similarity indices , , ,edge seq trd timeS S S S .Those quantities
should be able to effectively reflect how similar a window
signature is with respect to a database signature of a certain
appliance.
A. Edge similarity edgeS
From the signature database, an appliance candidate only
includes its own P-Q-W edge sets. In contrast, a window can-
didate may include other edges caused by overlapped appli-
ances. The comparison is trying to answer if this window can-
didate includes all of the appliance candidate’s edges. Thus the
process is like this: each of registered edges will be compared
throughout all the edges in window candidate one by one. Then:
'
eedge
e
NS
N (4)
where eN is number of edge types defined in appliance candi-
date and '
eN is recognized number of appliance edge types in
window candidate.
Power/w
Time/min
B D
0
110
10
80
Power/w
Time/min0 20
150
270
350
380
10 12
B
D
50
A
C
Candidate Window
Candidate
Appliance
Ne=2Ne
’=2
Fig.10. Edge similarity comparison
As shown in Fig.10, both of the two registered edges B-D are
found in the window ( ' 2e eN N However, if only B exists,
it is very likely the window candidate is only one part of the
appliance process and its 0.5edgeS ( '2, 1e eN N ).
P, Q can be easily compared since they are quantitative
values. As for current waveform W, one can conduct compar-
ison in either time-domain or frequency domain [12]. Selecting
proper harmonic orders can also eliminate the impact from
noises and dc offset.
Since an edge is determined by three sub-attributes, again,
different weights can be set to those attributes: for linear and
active load such as stove, P should be emphasized; for
non-linear load such as microwave, W should be emphasized;
for reactive load such as fridge, Q should be emphasized. Those
weights can be pre-defined for appliance candidates. Synthe-
sizing them together, two edges can be determined as identical
or non-identical.
Overall, edgeS indicates the existence of edges of appliance
candidate in window candidate.
B. Sequence similarity seqS
For ON/OFF type appliance, it has a fixed sequence of edges;
for multi-state appliance, as discussed in section II, fixed se-
quence and repetitive sequence may either be found.
For fixed sequence edges, they always follow a certain order
pattern. For a window candidate, its edge order should comply
with the order pattern defined in appliance candidate. For ex-
ample, a space heater has 5 edges in the order of A-B-C-D-E. It
is expected to find A-B-C-D-E in the window. On the other
hand, an A-B-D-C-E sequence may imply a different appliance
process and B-C-A-D-E is even more different.
Power/w
Time
100Appliance
candidate
A:+100
B:+200
C: -50
D:-150
E:-100
300
B
C
D
E
Window
candidate 1
D
A
250
A
BC
E
DC A
Window
candidate 2E
Power/w
Time
B
Fig.11. Sequences of two candidate windows compared to the appliance
candidate
To quantify the difference of two sequences, a simple
method based on calculating the position changes of letters is
proposed. Suppose the appliance candidate above has a se-
7
quence labeled using letters A-B-C-D-E. Window candidate 1
has A-B-D-C-E; window candidate 2 has B-C-A-D-E; window
candidate 3 has C-B-A-D-E. Then we have the table below:
TABLE VIII
EXAMPLE OF POSITION CHANGE
Window candidate Position change of letters
Length of changed position
A-B-D-C-E C:34
D:43
|4-3|+|3-4|=2
B-C-A-D-E A: 13 B: 21
C: 32
|3-1|+|1-2|+|3-2|=4
C-B-A-D-E A: 13 C: 31
|3-1|+|1-3|=4
It is easily known that B-C-A-D-E and C-B-A-D-E are more
disordered than A-B-D-C-E compared to the original sequence
A-B-C-D-E based on their lengths of changed positions. For a
given sequence composed of n letters/edges, the maximum
possible length of changed position is:
0
[ (2 1)]L
k
M n k
, 1
2
nL
(5)
From (5), it can be calculated that:
for ON/OFF appliance, n =2,M=2 (ABBA);
for three-edge appliance, n =3, M=4 (ABCCBA);
for four-edge appliance, n =4,M=8 (ABCDDCBA) ;
for five-edge appliance, n =5,M=12(ABCDEEDCBA).
Based on the discussion above, edgeS for appliance with fixed
sequence can be quantified as
1f
trd
NS
M (6)
where f
N is the length of changed position of a window can-
didate as calculated in Table VIII.
For example, sequence C-B-A-D-E’s 0.67trdS ( 4fN )
while sequence E-D-C-B-A’s 0trdS since it is completely
opposite to the original sequence A-B-C-D-E ( 12fN ).
One exception is if edgeS is already found smaller than 1,
seqfS will be automatically set to zero due to mismatch in the
number of relevant edges.
The appearances of repetitive edges are also counted in the
step of determining edgeS and only if its number is more than
one, it is recognized as an repetitive edge. '
rseqr
r
NS
N (7)
where rN is number of repetitive edge types defined in ap-
pliance candidate and '
rN is recognized number of repetitive
edge types in window candidate.
As for a combination sequence load such as furnace, seqS can
be decided based on its two sub-indices seqfS and seqrS which
can respectively evaluate similarities of repetitive and fixed
sequence characteristics.
C. Trend similarity trdS
As discussed in Table III, power slope based scanning can
effectively scan the window candidate and further determine
the existences of trend signatures with respect to appliance
candidate. '
ttrd
t
NS
N (8)
where tN is the number of trend signature types defined in
appliance candidate and '
tN is the recognized number of trend
signature types in window candidate.
D. Time similarity timeS
In the end, the moment of appearance of window candidate t
is also compared with time signature defined in appliance
candidate. As shown in Fig.6, the time signature of appliance
candidate is defined as one or several hour ranges T such as
{17-23},{6-8,11-13,16-18}.
1,
0,time
t TS
t T
(9)
V. CREATION OF SIGNATURE DATABASE
NILM needs a customized signature database to realize
identification since appliance characteristics are usually dif-
ferent from house to house. In this paper, since more process
signatures are involved, convenient creation of signature da-
tabase for ordinary householders is really important. In this
work, we propose to create a small signature database tailed for
each home utilizing a device called appliance register.
An appliance register is a device inserted between the ap-
pliance to be registered and the electric outlet the appliance is
plugged in originally (Fig.12). The device contains a current
sensor and a wireless transmitter. Once a current change is
detected (an event), the device will send a signal to the smart
meter (or the device which does appliance identification).Smart
meter does two things: capture the event window of this ap-
pliance and determine the signatures of event window.
Firstly, phase signature can be determined by the smart meter.
Then captured event window will be scanned through and all
events associated with the appliance (labeled by the register
device) are picked out. Edge signatures can be directly ex-
tracted. Sequence signature can be determined based on ap-
pearance number of edge types. Trend signature can be de-
tected based on slope modes explained in Table III. In the end,
time/duration signatures are automatically learned when the
appliance is named by users.
8
Appliance Register
Meter sideWireless
Fig.12. Connection of register and system
After waiting for one or two operating cycles of the appli-
ance, all signatures of the appliance will have been collected.
The register is then removed. This approach has another ad-
vantage in term of privacy: the customer can control which
appliances are to be registered for identification.
VI. APPLICATION AND VERIFICATIONS
A. Application and verification based on real life data
The above algorithms and devices were tested in two real
residential houses for several weeks with no special intention
from the owners. A laptop based data acquisition system was
hooked to the electricity panel and behave like a smart meter. A
Zig-bee transceiver was connected to its USB port to bridge the
communication with the appliance register. After registration
was finished, a computer program based on proposed algo-
rithms was launched and kept running. In real time, the overall
power was decomposed into appliance level. Interested appli-
ance events are firstly identified. Then their energy consump-
tion is calculated through area integration illustrated in Fig.13.
This method is much more accurate than traditional single-state
based energy calculation since the detailed power variation of
process can now be tracked and counted. This is especially
significant for continuous-varying and multi-state appliance
energy.
Power/w
Time/min0 20
150
270
350380
10 12
4
50
2 A1
A2
1
3
KWhFridge=A1+A2
Fig.13. Area integration based energy calculation
The results are updated every half an hour and displayed in
the interface shown in Fig.14. As can be seen, appliance elec-
tricity consumption information is formatted into the table and
charts. The table summarizes the total energy counted from a
certain date and converted expenses with respect to local elec-
tricity rates (say,7.4¢/KWH). The pie chart presents the per-
centage composition of individual appliance so users can be
aware of the significance of reducing a certain appliance’s
consumption. Finally, from the time distribution of energy use
chart, user can understand his energy usage pattern statistically
with respect to hours. This information is quite essential for
residential house owners to adopt proper demand response
strategies such as load shifting according to utility’s TOU rates.
Fig.14. Appliance energy decomposer software
To verify the identification rates, controlled trials were con-
ducted continuously for 5 days. Interested appliance events
were manually checked and recorded as comparison. The re-
sults listed in Table VII are very satisfactory for all types of
appliances including continuous-varying and multi-state ap-
pliances. The column “False identified operation times” indi-
cate the mistakenly identified times caused by events of other
loads.
TABLE IX.
IDENTIFICATION RATE VERIFICATION
Appliance
Name
Actual
operation
times
Correctly
identified
operation times
False
identified
operation times
Identification
accuracy(%)
Chest Freezer 178 163 2 90.5
Fridge 213 203 4 93.4
Furnace 185 172 1 92.4
Stove/Oven 16 16 0 100
Microwave 23 22 0 95.7
Kettle 12 11 0 91.6
Toaster 6 6 0 100
Washer 3 3 0 100
Dryer 4 4 0 100
B. Comparison with other solutions of NILM
This paper also presents a detailed comparison between
proposed solution and traditional combination based solution
such as discussed in [8-9]. According to [9], a two-layer
feed-forward network is adopted here for comparison. Other
similar solutions are not discussed here.
Firstly, specific appliances were measured in the lab and
their harmonic signatures were collected. Not like in the pro-
posed approach, no process related signatures is considered by
neural networks. Harmonic contents of aggregated signal are
used as input layer while appliance composition list as output
layer. Since both magnitude and phase of a certain harmonic
order are considered, the input layer has 16 nodes of up to 15th
harmonic content (odd ones). Hidden nodes are set to be 20.
According to [9], numerous training sets are generated math-
9
ematically by adding up harmonic contents (waveforms) of
designated individual appliances. Also, to make training more
reliable, a less than 10% deviation is added to original magni-
tude as noises.
0 50 100 150 200 250 300-40
-30
-20
-10
0
10
20
30
40
50
Generate
Training set
Meter side
0 50 100 150 200 250 300-15
-10
-5
0
5
10
15
0 50 100 150 200 250 300-4
-3
-2
-1
0
1
2
3
4
0 50 100 150 200 250 300-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Fridge Stove Microwave
0 50 100 150 200 250 300-25
-20
-15
-10
-5
0
5
10
15
20
0 50 100 150 200 250 300-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 50 100 150 200 250 300-3
-2.5
-2
-1.5
-1
-0.5
0
0.5
0 50 100 150 200 250 300-40
-30
-20
-10
0
10
20
30
40
50
Meter sideFreezer Computer TV
……
Generate
Training set
Fig.15. Examples of generated training sets based on harmonic (waveform)
signatures of appliances
For test stage, a bottom-up based aggregating program turns
on/off each load according to their regular behavior. The ag-
gregated meter side signal is formed this way. Then both of the
two approaches were tested to decode the overall meter signal
and their performances are discussed as below.
Comparison is firstly conducted when there are only sin-
gle-state type loads. This is because single-state type loads only
have a steady-state harmonic content. The results are shown in
Table X. TABLE X
COMPARISON FOR ONLY ON/OFF TYPE LOADS
Loads Identification accuracy(%)
NN based approach Proposed approach
Microwave 97.9 99.9
Monitor 98.3 98.3
TV 99.2 98.1
Vacuum 97.6 98.6
Monitor 99.9 98.5
Incandescent light bulb 98.9 98.5
Fluorescent light bulb 99.0 99.2
As can be seen, for a system composed of only single-state
type loads, proposed approach has a performance similar to NN
based approach. This is because there is no change in each
appliance’s operation process. However, results are heavily
changed when complicated loads are brought in. TABLE XI
COMPARISON WITH COMPLICATED LOADS
Loads Identification accuracy(%)
NN based approach Proposed approach
Microwave 99.3 97.1
Monitor 97.6 98.9
TV 84.4 98.2
Vacuum 85.0 99.3
Monitor 79.8 97.7
Incandescent light bulb 97.5 99.2
Fluorescent light bulb 95.6 98.5
Fridge 63.7 97.9
Freezer 68.5 95.3
Washer 73.4 97.1
Furnace 57.2 98.4
As can be seen, NN based approach is significantly affected
by the introduction of multi-stage loads (furnace and washer)
and continuous varying loads (fridge and freezer). This draw-
back is actually discussed in [9] due to the lack of a steady-state
harmonic content in those appliances. Their harmonic contents
can vary tremendously with time. Sometimes, harmonic con-
tents of different operational stages of the same load cannot
even be comparable such as in furnace. To cope with this
problem, NN based approach has to average the harmonic
contents and use the average value for training. This will in-
troduce not only large error to those complicated loads them-
selves but also to those single-state loads if they are turned on at
the same time. For example, for a given point, if the aggregated
waveform is composed of fridge and microwave, identification
of microwave may fail due to error from fridge. In contrast,
proposed approach captures event window and utilizes process
signatures to identify. In theory, the more complex the process
is, the more unique its window can be and the easier it can be
identified. This is the reason proposed approach has a much
better performance. In the meanwhile single-state appliances
will not be affected by complicated appliances since they have
different edges.
Another obvious advantage of proposed approach is it only
identifies loads users are interested in and willing to register. In
contrast, NN based approach’s training process has to cover all
major appliances. Also, once user purchases another heavy load
such as a stove, the accuracy of identification will become not
reliable at all. This is because the trained network itself has
been changed due to the newly added element. Any aggregated
signal that has stove on at the same time will become uniden-
tifiable (this is especially severe for other kitchen appliances).
However, stove hardly has any impact to identification of reg-
istered loads using proposed approach since proposed approach
is based on searching relevant edges of registered appliances.
Those non-relevant edges from stove will be ruled out of the
window candidates of registered loads. TABLE XII
COMPARISON WHEN STOVE IS NOT TRAINED OR REGISTERED.
Loads Identification accuracy(%)
NN based approach Proposed approach
Microwave 78.0 94.5
Monitor 77.8 94.8
TV 76.6 94.2
Vacuum 65.2 96.1
Monitor 95.8 95.8
Incandescent light bulb 64.1 95.1
Fluorescent light bulb 38.6 95.8
Fridge 51.3 96.1
Freezer 56.3 90.6
Washer 45.4 92.7
Furnace 44.3 95.6
Stove --- ---
To conclude, the proposed approach has the following ob-
vious advantages:
Process based signatures and event window make identifi-
cation of complicated loads possible. However, combina-
tion based approaches such as NN cannot effectively
identify those loads.
Composition of appliances is not only judged by an inde-
pendent point of meter side signal but also events before
and after this point. Association of load states is much
more strengthened. A load’s OFF event can only be con-
firmed if its ON event is found within a time window.
Training process does not need to cover all major appliances
10
any more. Users only need to register their interested
loads they want to track down.
Nearly no additional effort if load inventory is partially
changed.
VII. CONCLUSIONS
This paper has presented a new method to identify and track
home appliance loads using the smart meter data. The main idea
of the proposed method is to model the entire operating cycle of
a load and make identification based on event windows. A set
of algorithms has been developed for this purpose. Another
contribution of this paper is the proposition of a novel method
for creating signature databases tailored for individual houses.
Tests conducted in two houses have shown that the rate of
successful identification is above 90% for all types of appli-
ances. Although through detailed comparison with traditional
neural network based solution, several advantages of proposed
approach are revealed. It is believed that the proposed tech-
nique makes non-intrusive monitoring more applicable for
complicated loads; it can also reduce ordinary house owner’s
efforts to apply NILM.
REFERENCES
[1] David Schatsky and Clint Wheelock, Home Energy Management. [On-line]. Available: www.pikeresearch.com/research/ [Jan 22,2011].
[2] S. Massoud Amin and Bruce F. Wollenberg, “Toward a Smart Grid”,
IEEE Power & Energy Magazine, pp. 34-41, Sept/Oct. 2005.
[3] United States Department of Energy. The Smart Grid: An Introduction. [On-line].Available:www.oe.energy.gov/SmartGridIntroduction.htm
[Apr 16,2011].
[4] John N. Sharood, Graham Bailey, D. Mitchell Carr, James Turner and David Peachey. “Appliance retrofit monitoring device with a memory
storing an electronic signature”, U.S. Patent 6 934 862, Aug 23, 2005.
[5] G.W.Hart, “Non-intrusive Appliance Load Monitoring”, Proceedings of the IEEE, vol. 80, No 12, December, pp. 1870 - 1891,1992.
[6] F. Sultanem, "Using appliance signatures for monitoring residential loads
at meter panel level," Power Delivery, IEEE Transactions on , vol.6, no.4, pp.1380-1385, Oct 1991
[7] M. Baranski and J.Voss, “ Genetic algorithm for pattern detection in
NIALM systems”, Systems, Man and Cybernetics, 2004 IEEE Interna-tional Conference on, Volume 4, 10-13 Oct. 2004, Page(s):3462 - 3468
vol.4
[8] J. Duan, D. Czarkowski and Z.Zabar, “Neural network approach for estimation of load composition”, Circuits and Systems, 2004. ISCAS '04.
Proceedings of the 2004 International Symposium on, Volume 5, 23-26
May 2004 ,Page(s):V-988 - V-991 Vol.5 [9] D.Srinivasan, W.S.Ng and A.C.Liew, “Neural-network-based signature
recognition for harmonic source identification”, Power Delivery, IEEE
Transactions on, Volume 21, Issue 1, Jan. 2006 ,Page(s):398 – 405 [10] Jian Liang, S.Ng , G.Kendall and J.Cheng , "Load Signature Study—Part
I: Basic Concept, Structure, and Methodology," Power Delivery, IEEE
Transactions on , vol.25, no.2, pp.551-560, April 2010 [11] Jian Liang; S.K.K.Ng, G.Kendall and J.W.M.Cheng , "Load Signature
Study—Part II: Disaggregation Framework, Simulation, and Applica-
tions," Power Delivery, IEEE Transactions on , vol.25, no.2, pp.561-569, April 2010
[12] Wilsun Xu, Alex Nassif, and Jing Yong, “Harmonic Current Character-
istics of Home Appliances”, Report to CEATI International Inc., 60 pages, March 2009.
[13] A.B.Nassif and J.Acharya "An investigation on the harmonic attenuation
effect of modern compact fluorescent lamps," Harmonics and Quality of
Power, 2008. ICHQP 2008. 13th International Conference on , vol., no.,
pp.1-6, Sept. 28 2008-Oct. 1 2008 [14] A.B.Nassif and Wilsun Xu; ,"Characterizing the Harmonic Attenuation
Effect of Compact Fluorescent Lamps," Power Delivery, IEEE Transac-
tions on , vol.24, no.3, pp.1748-1749, July 2009 [15] N. Barry, and E.McQuade, "Temperature control using integer-cycle
binary rate modulation of the AC mains," Industry Applications, IEEE
Transactions on , vol.31, no.5, pp.965-969, Sep/Oct 1995 [16] A. Capasso, W. Grattieri, R. Lamedica and A. Prudenzi, “A bottom-up
approach to residential load modeling,” Power Systems, IEEE
Transactions on, vol. 9, no. 2, 1994, pp. 957-964. [17] S.W. Heunis, and R. Herman, “A probabilistic model for residential
consumer loads,” Power Systems, IEEE Transactions on, vol. 17, no. 3,
2002, pp. 621-625.
Ming Dong (S’08) received his B.Eng. degree in Electrical Engineering from
Xian Jiaotong University, China in 2008. He is currently pursuing his Ph.D. degree with Electrical and Computer Engineering, University of Alberta,
Canada. His research covers smart grid and power quality.
Paulo C. M. Meira (S’09) graduated in electrical engineering from University
of Campinas (UNICAMP), Brazil, in 2007. He received the M.Sc. degree in
electrical engineering from UNICAMP in 2010. His research interests include power systems reliability, protection, distributed generation and visualization.
Wilsun Xu (F’05) received his Ph.D. degree from the University of British
Columbia, Vancouver, BC, Canada, in 1989. From 1989 to 1996, he was an Electrical Engineer with BC Hydro, Vancouver and Surrey, respectively.
Currently, he is with the Department of Electrical and Computer Engineering,
University of Alberta, where he has been since 1996. His research interests are power quality, distributed generation and smart grid.
Walmir Freitas (M’02) obtained the Ph.D. degree in Electrical Engineering
from the University of Campinas, Campinas, Brazil in 2001, where currently
he is an Associate Professor. His areas of research interest are analysis of distribution systems and distributed generation.
Response to Editor and Reviewers (Paper: TSG-00189-2011: An Event Window
Based Load Monitoring Technique for Smart Meters) :
authors : Ming Dong, Paulo C. M. Meira, Wilsun Xu, Walmir Freitas
We would like to thank the reviewers for their constructive comments and suggestions which have
helped us improve the quality of the manuscript.
Response to Editor:
Comments:
The reviewers have a series of specific and valid observations and comments that the authors must
address. Furthermore, I would like the authors to compare (in quantitative and qualitative terms)
the proposed methodology and results with other approaches proposed in the literature including a
discussion of their advantages/disadvantages and a summary table that makes such comparison
evident.
The followings are the summarization of the major changes made in the revised edition.
1. A new paragraph is added into Section I to explain the different effects NILM have on small
appliances and heavy appliances.
2. Table II is added into Section II to provide examples of sequence patterns.
3. For trend signatures (Part C) in Section II, more explanations are given to address why there
are seven basic trend signatures.
4. More explanations are given to address how to determine the weights and thresholds and
5. Further research was conducted for calculation. A new method is explained to replace the
old method based on power tracks (Part B, Section IV.)
6. Page 8, Column 2, Line 3: “users can understand his energy usage pattern” is replaced with
“user can understand his energy usage pattern”.
7.Changes are made in Table IX to include false identification operation times.
8. A new part (Part B, Section VI) is written to address comparison of proposed solution and other
solutions of NILM.
Response to reviewer 1:
Comment 1: The core of the algorithm is the accurate determination of the linear discriminate
classifier in Equation (2). Please explain:
the method used to evaluate omega and gamma;
the trade-off between accuracy, identification accuracy, threshold, and computing time;
the value of the threshold used and how it affects the results.
At the beginning, the values of omega and gamma are roughly estimated by their electrical and
non-electrical characteristics of appliances. We did a lot of measurements on different appliances
in different houses and the following features are discovered. In a house, appliance can be roughly
divided into several types based on their power factors or harmonic contents:
Resistive appliances such as stoves, incandescent light bulbs and kettles. Those appliances
roughly have a power factor of 1.0;
Reactive appliances, mostly inductive motor based appliances such as fridges, freezers.
Those appliances usually have a considerable reactive power Q compared to their active
power P. Those appliance’s power factors are smaller than 1.0;
Non-linear appliances such as microwave, TV and PC. Those power electronics based
appliances usually have a heavily distorted waveforms and significant harmonic contents.
This is caused due to their internal rectifier/inverter circuits.
Linear appliances such as resistive appliances or motors. Those appliances only have linear
components such as resistors or inductors and thus their waveforms are nearly perfect
sinusoidal.
There are a few exceptions which may be both non-linear and reactive, it is not commonly
seen.
Those differences are discussed in Section IV of paper that for different appliance, will be
determined by one or more of edge signatures P/Q/W. For resistive appliance, P is used; for
reactive appliance, P and Q are both used; for non-linear appliances, unique harmonic contents
should be emphasized. This is how we determine . This is mainly determined by appliance’s
electrical characteristics.
Similarly, rough estimation on weights can be completed if non-electric characteristics are also
considered. For example, all fridges have a very unique “going-down” trend like shown in the
following figure. This is because when the motor is firstly started, a big in-rush current will occur,
however, after the speed of fridge gradually increases, the current will gradually drop. In another
word, we are expecting to see a “rising spike” and “gradual falling” trends as discussed in Table
II. Also another interesting characteristic is that during mid-night, fridge is almost the only
appliance still running. This is why some credit should also be given to .
Fig 1: Four fridge operations captured from meter side
The analysis above only gives us a rough guide of how to set the values. It seems is very
important since fridge has its unique P and Q; is necessary since we are expecting two edges
of different magnitudes to show up in a fixed order; is also important but definitely not as
important as We can’t say if there is a spike and a falling curve, it must be fridge since it
could be caused by other motor devices too. Reversely, once its P and Q meet, it is already very
likely to be a fridge even if its spike can be occasionally “erased” due to other activities of
appliances or noises. Then based on P and Q, with the other information to assist, final conclusion
of identification can be drawn.
Above discussion defines the weights analytically but it should also be optimized mathematically
through simulation. After acquiring enough data, by adding noises to the samples of appliance
windows, we can generate numerous testing windows shown like below. Identification based on
particular values of and is tested. Their values are further adjusted through iteration
method or manually to the point that fridge can be identified correctly for the maximum number of
cases. Each of the appliance’s and are determined through this way. This is how we get Table
VI.
When applied to a real house, the weights of appliances are still the same with the pre-defined
values in the lab. This is based on the assumption that same type of appliance in different houses
can use the same set of weights due to their similar working principles and characteristics. For
example, almost all fridges consume reactive power, have the similar trend and are running
throughout 24hrs. The , however, is properly lowered compared to the value of simulation
because of noises and variants in real houses. The aggregated signal is usually worse than
generated testing window used in simulation.
Fig 2: Flowchart of weight and threshold determination
The threshold is a key factor to determine the identification accuracy. It cannot be set too high or
too low. If it is too high, the classifier becomes too strict to accommodate samples that have errors;
however, setting it too low might misidentify non-relevant appliances. Its value is also calculated
through above simulation program. Once weights are determined, threshold is adjusted to
ensure the maximum identification rate. In a real house, a discount of 5%-10% is put on top of the
original value based on simulation. This is due to the error between a simulation window and a
real possible window. Based on a lot of tests, the threshold within 80%-85% is used for most of
appliances to achieve a balance between identification rate and accuracy.
To address this concern, changes are implemented in the revised version to make the discussions
of more clear. However, it is a little difficult to provide more details to readers since it
might distract readers from explanations of overall procedure and other important issues covered
in this paper.
Comment 2: As mentioned, the use of a linear classifier is considered as a simple choice for the
appropriate criterion. A closed loop control system seems more appropriate. This may affect the
identification accuracy.
It is true that linear discriminate classifier is a simple choice. And it is also very easy to switch to
other advanced classifiers such as neural networks, decision tree or even SVM. If using neural
networks, instead of setting weights and thresholds for microwave, we can use generated testing
windows mentioned above to train the neural networks. Thus any appliance such as microwave
could have its own neural network or decision tree. When applied to a house, a real window still
represented by its can be fed into the trained NN. And then each
appliance’s NN can tell us if it is truly a microwave or not.
However, the reasons to address the linear discriminate classifier in this paper are: a). It is very
straightforward and easy to explain the concept of decision making by multiple signatures; b) In
terms of implementation, it is also very convenient and practical; c) In the introduction section,
this paper discussed the previous work using NN or other classifiers. And if NN is also addressed
here, readers might get confused (actually, they are totally different concepts. The NN approach
mentioned in the beginning is based on waveform/harmonic combination. Please see newly added
section Part B: Comparison with other solutions of NILM.) This is why in the end of Section III,
there is a paragraph saying: “This linear classifier can also be substituted by more advanced
classifiers such as neural networks or decision tree. Those variations are not discussed here.”
As for a closed loop control system, we think reviewer 1 means the values of and can be
modified or updated according to user’s feedback automatically. This is a creative and interesting
idea. But it may be difficult to obtain feedback in practice. It is often troublesome to let ordinary
house owners know/remember their appliance operations and provide the info back for system
improvement. The other choice might be installing extra logging devices, however, they add to the
cost.
Comment 3: Section IVB: Although the section explains the sequence similarity Sseq, this
subsection needs further research. The metric of the correlation gives a representative image rather
than a secure solution.
To quantify the difference between two sequences is difficult. The original approach discussed in
paper is based on an assumption that two same appliance windows should have similar power
change processes (tracks) and those changes should correlate well with each other. Please be
noted that the original Fig.11 is just an illustration of three power tracks. When calculating
correlation factor, actually, only 5 values (A,B,C,D,E) are put into equation (5).
Another simple approach is discussed in this revised paper. It is based on calculating the position
changes of letters. Suppose the appliance candidate has a sequence of A-B-C-D-E. Window
candidate 1 has A-B-D-C-E; window candidate 2 has B-C-A-D-E; window candidate 3 has
C-B-A-D-E. Then we have the table below:
Table I: Example of calculating length of changed positions
Window candidate Position change of letters Length of changed
positions
A-B-D-C-E C:34;
D:43
|4-3|+|3-4|=2
B-C-A-D-E A: 13
B: 21
C: 32
|3-1|+|1-2|+|3-2|=4
C-B-A-D-E A: 13
C: 31
|3-1|+|1-3|=2
The lengths of changed positions actually reflect how big the change is between window
candidate’s sequence and the original one. Compared to the original power track based method,
this method better focuses on pure sequence and more convincing.
This change is implemented in the Part B, Section IV of the revised edition.
Comment 4: The creation of signature database using wireless devices and smart meters is
characterized by a complicacy for ordinary householders. A flexible statistical signature database
with regional attributes may be an easier solution.
The idea of using flexible statistical database with regional attributes is very interesting. Actually,
if users’ choices of appliances are closely related to region factors, it is much more convenient.
Unfortunately, in Edmonton (Canada) area, the consistence of appliance signatures is not so well.
For example, in our investigated houses, the microwaves being used have a power range from
800-1200W (one made in Japan and the others made in US). Selecting the average 1000W as a
standard signature may cause failures of recognition. However, as discussed above, those
microwaves do have things in common such as a heavy 3rd harmonic content and similar usage
pattern. This is why they can share the same weights.
Comment 5: Page 8, Column 1, Line 10: Please replace “users can understand his energy usage
pattern” with “user can understand his energy usage pattern”.
This change is implemented in the revised edition.
Comment 6: In my opinion, the appliance energy decomposer software is useful for industrial
energy analysis where the energy consumption is higher and the load types of signatures may be
more deterministic.
We fully agree that appliance energy decomposer is probably more useful for industrial energy
analysis. More user-friendly interface such as a web-portal based platform may be better. More
study should be done on this in the future. However, currently it is convenient to use the software
to get quick results and verify algorithm, this is why we used software to support its application in
this paper.
Response to reviewer 2:
Comment 1:
I do not see a clear advantage of your system compared to other solutions based on genetic and
neural network, the introduction says: "The other challenge faced by some of the published works
is that they need a time-consuming training/learning process to support their algorithms such as
genetic and neural-network before they can work [7-9]. Such combination based approaches are
vulnerable to changes in the appliance inventory. Once a major appliance is replaced, re-training
has to be conducted."
My question is : Does not your proposal have the same problems? If your proposal does not have
these problems you should explain how you solves these problems.
The main advantages of proposed solution are:
It makes identification of multi-stage and continuous varying appliances possible. This is due
to the exploit of process based signatures and window based identification procedure. In
other solutions, this information is not utilized.
It reduces the training process: Users only need to register their interested loads to track
down.
There is no additional effort if load inventory is changed.
To address this concern, in this revised paper, a new part “Comparison with other solutions of
NILM” in Section VI is written.
Comment 2: I think the results are quite good, but the evaluation might be a bit more concise.
In Table VII the authors indicate the times that an appliance is identified, but not if this
identification is correct or not. It is possible to mistakenly identify an appliance and this could
significantly worse the results. Which is the possibility of a false identification?
Firstly, the “identified operation times” has been changed to “correctly identified operation
times”. Actually the numbers listed in this column only indicate the activities correctly identified
with respect to “actual operation times.”
Secondly, to address this concern another column “false identified operation times” is added. The
numbers in this column indicate the false ones caused by other appliances..
It is found the false identification rate is very low. I think it is because the listed appliances are not
very similar to each other. The only similar appliances are chest freezer and fridge, however they
are connected to different phases (one in kitchen and the one in basement) which make them
naturally separated.
Response to reviewer 3:
Comment 1: if each individual home appliance consumes powers according to Gaussian
distribution, then the summation (total power consumption) of many individual Gaussian
distributions follows a Gaussian distribution, and so, we cannot identify each individual
power-consumption from total power consumption. Furthermore, when (1) we have many
identical home appliances, such as iPhones or TVs, or (2) many independent home appliances and
each home appliance has an arbitrary distribution of power consumption, the resulting distribution
of total power consumption follows a normal distribution according to the well-known central
limit theorem. In that case, we wonder how we can identify each home appliance’s power
consumption. We guess, in order to address this issue, the authors should prove that the power
consumption distribution of each home appliance follows non-Gaussian distribution.
First of all, NILM aims to identify appliance activities within in a single house. The house owner
needs to provide tailored signatures such as power of his own appliances through training process
or proposed registration method mentioned in Section V.
Secondly, in a single house, the number of appliances and especially major appliances in terms of
energy consumption is not large. The power and other signatures of a certain appliance are almost
fixed and do not follow Gaussian distribution. The figure below shows the fridge’s activities
during several hours. It can be told that those fridges’ activities are almost identical and present
the same power jumps and other signatures such as the spike.
Fig 3: Example of power curves captured from a real meter.
Thirdly, those switch-on and off events can be observed in a house since the number of appliances
is limited. If there are too many appliances, say 1000 houses, the edges may not be obvious and
they may be submerged by background. However, in a single house, those edges can be very
clearly seen and captured. The only exception is small appliances that only consume very tiny
amount of power such as stand-by power. Those appliances may fail to be identified, however, in
terms of energy consumption, they are not usually a big concern by house owners.
The entire process of NILM is similar to the process of speech recognition. Each sentence has a
start and an end. Different people have different volume, tones and frequencies.
To address reviewer’s concern, a new paragraph is added into section I to better define the
purpose of NILM.
Comment 2: Furthermore, when (1) we have many identical home appliances, such as iPhones or
TVs…we wonder how we can identify each home appliance’s power consumption.
In one house, if you have 2 identical TVs which have exactly the same power, reactive power,
harmonic contents and are even connected to the same electricity phase, theoretically any NILM
approach is not able to differentiate them. However, there are not many houses having two
identical major appliances like this.
As for some appliances such as iPhone chargers, this is possible. However, since they consume
fairly small power, usually less than 20 watts, they are not that important. Besides, even if we
identify two iPhone chargers as one iPhone charger and tell the house owner how much his
iPhone charger together consumes, the information is still quite enough for him.
Comment 3: Basically, the proposed window-based method needs to establish a database which
describes the power consumption behavior of each home appliance, then it extracts a sequence of
possible power consumption windows from the power consumption trace, and after that, it
identifies what home appliance matches with each identified window. So, we would like the
authors to compute the overall complexity of this procedure as the number of home appliances
increases. Considering the authors’ initial motivation that the recognized instantaneous
Power
Time
power-usage pattern can exploit the Time-of-Use rates, the complexity is important in verifying if
the proposed method can be used to determine the current usage in real-time.
Once again, a single NILM program is not aiming to identify appliances or loads in a large
number of houses. It usually aims to one single house. In each house, NILM can be integrated into
its electricity meter [1] and realize identification locally with respect to the meter data.
In one house, the house owner might have 10-20 frequently used appliances. He could be
interested to register some of them. Using the register device, it will not take more than 2 hours to
complete the database.
Then based on an Intel Dual Core CPU, 3GB DDR3 memory PC, it may take the appliance
decomposer software 5-10 minutes to process one day’s data and identify his interested
appliances.
Based on the feedback from the houses we investigated before, the complexity of this approach is
acceptable. In the market, as discussed in introduction section of this paper, some houses even
connect energy monitors (sub-meters) to individual appliances and compose a local sensor
network [2]. Compared to this method, we believe NILM is much more convenient and less
complex.
Comment 4: Additionally, we can hardly accept that there are only three types of power
consumption curves and there are only seven types of power trends for all possible home
appliances. In order to convince the audience to accept this claim, the authors should present more
concrete evidences, including enumerated literature and empirical analysis.
Please note that in the manuscript, it is not saying there are only 3 types of power consumption
curves. It is saying there are 3 types of edge sequences: Repetitive sequence, fixed sequence and
their combination.
This is based on a lot of measurements and analysis of existing appliances. First of all, more than
60% of the major appliances are ON/OFF type appliances. For example, an incandescent light
bulb is just a resistor. There are always two edges. Those belong to the fixed sequence.
For a multi-stage appliance such as a washer, it usually has a fixed sequence for one of its specific
working modes. For example, for “quick wash” mode, your washer will always experience a fixed
power pattern such as +50W,-50W,+ 100W,-80W,+480W,-500W. This is because the micro
controller in the washer automatically launches different components one by one.
The second type of appliance has repetitive sequence such as stove. This is because it has an
integer-cycle controller, it will interrupt heating element once in a while to prevent overheating
and keep desired temperature.
The last type is the combination of fixed sequence and repetitive. For example, a furnace has a
fixed sequence within each of its repetitive cycle. Based on measurements, most of appliances
belong to one of the three types.
As for the trend signatures, the typical ones are listed up in Table III. Explanations are given as
below:
Rising spike--- In a house, there are a lot of inductive motor devices such as a fan and a
fridge. The reason we will see a rising spike is when an inductive motor starts, its initial
speed is zero and it will generate a huge current called inrush current. And this is the rising
spike.
Gradual falling--- After the motor speeds up, the current will gradually fall.
Falling spike---Falling spike is not commonly seen. It happens when a component is
immediately shut down and then started again. TV is an example: when you switch among
different channels, the screen will be black for a sec.
Pulses---Pulses are usually caused by power electronic switches. For example, some stoves
have pulses because they have an integer-cycle controller in it. It prevents itself from
overheating. Another example is an inverter based motor device.
Fluctuation/Quick vibrate--- Those are due to vibrations and distortions of motor device and
electronic devices. For example, when a PC is running, high frequency noises can often be
seen.
Flat---A lot of appliances will have a flat power curve once they enter into steady mode after
a very short starting transient process (negligible).
Combination--- Also in Part C, Section II, I mentioned “it should be noted some appliances
such as fridge may have more than one type of trend signatures.”
The above are all the major trends we found through measurement in houses we investigated.
Those trends are not only found in one type of appliance but usually several types of appliances
due to their common electrical characteristics.
To address reviewer’s concern, Table II and more explanations are provided in Part B and Part C
of Section II.
Comment 5: Moreover, according to the empirical study (Section VI), we can observe that the
accuracy of identification rate verification is between 90 – 100%. So, we would like to know
how and to what extent the false identification rate (which is 0 – 10%) impacts on the
power-consumption estimation and subsequent demand responses in smart grids. Also we want to
observe a rich set of application and verification results to evaluate the proposed window based
methods in various ways.
The key of demand response is to manage customers’ consumption of electricity in response to
supply conditions, such as having electricity customers reduce their consumption at critical times
or in response to market prices. Today, this is completed through an incentive approach: smart
meters combined with “Time-of-Use” prices bill customers based on 3 different prices in a day---
On-peak, Mid-peak and Off-peak. Those prices are designed based on utilities’ supply conditions
[3].
Now based on TOU price, if users want to save money on electricity bill, they will naturally shift
their loads or replace their loads at critical times (On-peak). In another word, demand response
can be automatically/interactively realized by users if they are aware of their approximate energy
uses. [4] is a demo provided by Canada’s largest utility (one of the biggest in North America)
Hydro one. It illustrates how an ordinary user can respond to electricity demand side in smart
grid by shifting his appliances.
In terms of load shifting, the identification rate 90-100% is very enough to make users aware of
his usage pattern approximately. Suppose a customer knows that he uses his stove 18 KWH at
On-peak times this month through proposed method/software, even if he actually uses 20 KWH at
On-peak times, he would still obtain a great sense on if he should cook his meals 1 hour earlier (at
Mid-peak times) or change his stove to a greener one to save money. In contrast, without NILM,
ordinary house holders have barely any knowledge about their appliances energy uses.
As for other applications of NILM, the application to load study is mentioned in [1]; the
application as a surveillance tool is mentioned in [5]; the application for abnormality alarm is
mentioned in [6]. Also, NILM used for demand response and other applications is already
commercialized as products in [7] and [8] by a US company.
Please be noted this paper cannot cover all of those applications above since the main concern of
this paper is to address how it can improve identification rate on multi-stage and
continuous-varying appliances compared to other NILM solutions and how it can reduce the
training process for ordinary users. Besides, a lot of those applications above are already
discussed in previous NILM publications [1,5-8]. To make a more clear comparison with other
NILM solutions, a new part (Part B, Section VI) is written to address comparison of proposed
solution and other solutions of NILM.
Reference:
[1] G.W.Hart, “Non-intrusive Appliance Load Monitoring”, Proceedings of the IEEE, vol.80,No
12, December, pp, 1870-1891, 1992.
[2] http://www.theenergydetective.com/
[3] http://www.hydroone.com/TOU/Pages/Default.aspx
[4] http://www.ieso.ca/house/hydroone/
[5] Hart, G.W., "Residential energy monitoring and computerized surveillance via utility power
flows," Technology and Society Magazine, IEEE , vol.8, no.2, pp.12-16, Jun 1989
[6] US. Patent: 6,816,078
[7] http://www.enetics.com/prodmainSPEED.html
[8] http://www.enetics.com/downloads/SPEED%20Brochure.pdf
