University of Alberta

Decomposition Techniques for Power System Load Analysis

by

Ming Dong

A thesis submitted to the Faculty of Graduate Studies and Research

in partial fulfillment of the requirements for the degree of

Doctor of Philosophy

in

Energy Systems

Department of Electrical & Computer Engineering

© Ming Dong
Fall 2013
Edmonton, Alberta

Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis
and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is
converted to, or otherwise made available in digital form, the University of Alberta will advise potential users
of the thesis of these terms.

The author reserves all other publication and other rights in association with the copyright in the thesis and,
except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or
otherwise reproduced in any material form whatsoever without the author's prior written permission.
 Abstract

In recent years, the increased public awareness of energy conservation has

attracted serious attention to detailed energy consumption monitoring and

management for the end users of power system. Load decomposition is a

technique that can extract detailed sub-load information from compound load

information. This technique decomposes a compound load such as an entire

residential house into specific sub-load levels such as different home appliances

by using only the aggregated metering data of the compound load. Through load

decomposition, users can better understand the usage patterns of individual loads

or load groups and therefore decide on how to save energy. On the utility side,

load decomposition can be very helpful for load forecast, demand response

program development, and Time-Of-Use price design.

In the past, traditional methods are either too costly or inaccurate. Therefore,

some researchers proposed a non-intrusive load monitoring (NILM) approach that

can identify and track major sub-loads based on only the total signal collected

from the meter-side with acceptable error. Recently, the vast deployment of smart

meters has raised considerable interests in this approach. However, many critical

problems still need to be solved before it truly becomes technically available for

ordinary end-users.

To solve the above problems, at the beginning, this thesis presents a novel

NILM method based on event detection and load signature studies. The key idea

is to model the entire operating cycle of a load and make identification based on
event-window candidates. The proposed technique makes NILM more applicable

for complex loads, more robust for load inventory change and can also simplify

the training process; on the other hand, the thesis addresses a new and critical

problem that previous researchers ignored---the non-intrusive extraction of load

signatures. The proposed approach is an unsupervised non-intrusive approach

which can automatically extract load signatures by using the meter-side data and

requires almost zero effort from users. This thesis also discusses how to estimate

the energy for several key components in a residential house such as the NILM

identified appliances, load groups and background power. Based on estimation,

residential energy characteristics are discussed with respect to the Time-of-use

price.
 Acknowledgement

I express my sincere appreciation to Dr. Wilsun Xu. This research and

dissertation would not have been possible without his patient guidance and great
supervision. I will always be proud to say I had the privilege of being his student.
I also thank my parents, Liping Guo and Ziqing Dong, for supporting me in so
many different ways during my past 27 years. My parents turned my desire to do
my graduate studies at a great Canadian university into a reality.
Finally, I also thank my workmates Charles Shi, Pengfei Gao, Shane Long,
Kirash Shaloudegi and Mostafa Tabatabaei, who helped me in both working and
living.
Contents

Chapter 1 Introduction ..................................................................................... 1

1.1 Background .............................................................................................. 1

1.2 Problem definition .................................................................................... 3
1.3 Review of statistic data based methods .................................................... 4
1.4 Review of measurement based methods .................................................. 8
1.5 Review of non-intrusive load monitoring based methods ...................... 10
1.6 Thesis scope and outline ........................................................................ 13

Chapter 2 Event detection .............................................................................. 16

2.1 Overview ................................................................................................ 16

2.2 Data segmentation based methods ......................................................... 20
2.2.1 Segment-conjunction method .......................................................... 20
2.2.2 Slope method ................................................................................... 22
2.2.3 Comparison of the two proposed methods ...................................... 25
2.3 Special issues with event detection ........................................................ 27
2.3.1 Detect double-phase events............................................................. 27
2.3.2 Adjacent events ............................................................................... 28
2.3.3 Event overlap .................................................................................. 30
2.4 Summary ................................................................................................ 32

Chapter 3 Event-Window Load Model and Load Signatures .................... 33

3.1 Overview ................................................................................................ 33

3.1.1 Review of single-state load model and its signatures ..................... 33
3.1.2 Review of transient load model and its signatures ......................... 35
3.1.3 Proposed event-window model and its signatures .......................... 36
3.2 Event Signatures ..................................................................................... 38
3.2.1 Real Power signatures .................................................................... 39
3.2.2 Reactive Power signatures .............................................................. 41
3.2.3 Harmonic signatures ....................................................................... 42
 3.3 Event Pattern Signatures ........................................................................ 45
3.4 Power Trend Signatures ......................................................................... 47
3.5 Time/Duration Signatures ...................................................................... 51
3.6 Phase connection Signatures .................................................................. 52
3.7 Summary ................................................................................................ 53

Chapter 4 Event-window based Load Identification ................................... 55

4.1 Overview ................................................................................................ 55

4.1.1 Review of signal-combination based algorithms ............................ 56
4.1.2 Review of event based algorithm .................................................... 59
4.1.3 Proposed event-window based algorithm ....................................... 61
4.2 Event-window based algorithm.............................................................. 62
4.2.1 Load identification procedure......................................................... 62
4.2.2 Individual signature scoring ........................................................... 67
4.2.3 Optimization of weights .................................................................. 72
4.3 System implementation .......................................................................... 74
4.3.1 Data acquisition .............................................................................. 74
4.3.2 Data preprocessing ......................................................................... 76
4.3.3 User interface.................................................................................. 79
4.4 Verification using real house data .......................................................... 80
4.4.1 Verification based on House #1 ...................................................... 81
4.4.2 Verification based on House #2 ...................................................... 87
4.4.3 Verification based on House #3 ...................................................... 90
4.4.4 Verification based on public dataset............................................... 91
4.4.5 Observations and findings .............................................................. 92
4.5 Comparative studies with neural networks based method ..................... 93
4.5.1 Implementation of neural networks based method ......................... 93
4.5.2 Simulation based verification ......................................................... 94
4.5.3 Observations and findings .............................................................. 96
4.6 Summary ................................................................................................ 97
Chapter 5 Non-intrusive Signature Extraction for Major Residential Loads
......................................................................................................... 99

5.1 Overview ................................................................................................ 99

5.1.1 Review of existing intrusive signature extraction methods ............. 99
5.1.2 Proposed intrusive event-window signature extraction system .... 101
5.1.3 Proposed non-intrusive signature extraction method ................... 102
5.2 Event Filtration..................................................................................... 104
5.3 Event Clustering ................................................................................... 110
5.3.1 Definition of event clustering ........................................................ 110
5.3.2 Selection of clustering method ...................................................... 112
5.3.3 Feature selection for mean-shift clustering .................................. 116
5.4 Event Association ................................................................................. 118
5.5 Verifications and Discussions .............................................................. 123
5.5.1 Verification and discussions based on real house #1’s data ........ 123
5.5.2 Verification and discussions based on real house #2’s data ........ 131
5.5.3 Verification and discussions based using MIT public dataset ...... 133
5.5.4 Verification of event association based on laboratory data ......... 135
5.6 Summary .............................................................................................. 139

Chapter 6 Energy Estimation of Residential House .................................. 141

6.1 Overview .............................................................................................. 141

6.2 Energy estimation methods for ordinary appliances ............................ 142
6.3 Energy estimation method for incandescent lights .............................. 146
6.3.1 Event filtration .............................................................................. 147
6.3.2 Event clustering ............................................................................ 148
6.3.3 ON-OFF match ............................................................................. 149
6.3.4 Energy calculation and the distribution plot ................................ 151
6.3.5 Results ........................................................................................... 152
6.4 Energy estimation method for background energy .............................. 154
6.4.1 Minimal power based method ....................................................... 154
6.4.2 Results ........................................................................................... 154
 6.5 Energy estimation of residential houses ............................................... 156
6.5.1 House #1 ....................................................................................... 156
6.5.2 House #2 ....................................................................................... 158
6.5.3 Seasonal changes of house #2....................................................... 159
6.6 Residential energy characteristics and its implications to TOU price . 161
6.7 Summary .............................................................................................. 164

Chapter 7 Conclusions and Future work .................................................... 165

7.1 Thesis Conclusions and Contributions ................................................. 165

7.2 Suggestions for future work ................................................................. 167

Chapter 8 References .................................................................................... 169

Appendix ............................................................................................................ 183

List of Tables

Table 3-1 Load type and examples ................................................................................... 37

Table 3-2 Typical values of real power of residential loads ............................................. 40
Table 3-3 Sequence pattern and examples ........................................................................ 47
Table 3-4 Trend signatures and slope characteristics ....................................................... 51
Table 3-5 Typical load window Lengths .......................................................................... 52
Table 4-1Window candidates vs. appliance candidates (1) .............................................. 63
Table 4-2 Examples of load  and  ............................................................................... 65
Table 4-3 Windows candidate vs. Appliances candidate (2) ............................................ 67
Table 4-4 Example of position change ............................................................................. 69
Table 4-5 Example of event matrix................................................................................... 76
Table 4-6 Example of trend matrix ................................................................................... 79
Table 4-7 Identification rate accuracy for house #1(7 days)............................................. 81
Table 4-8 Identification rate accuracy for house #2 (8 days)............................................ 87
Table 4-9 Identification rate accuracy for house #3 (7 days)............................................ 90
Table 4-10 Identification rate accuracy for house #4 (7 days).......................................... 92
Table 4-11 Comparison for only ON/OFF type loads ...................................................... 94
Table 4-12 Comparison with complex loads .................................................................... 94
Table 4-13 Comparison when stove is not trained or registered. ...................................... 95
Table 5-1 Appliance categories and examples ................................................................ 106
Table 5-2 Example OF ON-Event Filtration Condition Table........................................ 107
Table 5-3 Example of composition of suspect events..................................................... 111
Table 5-4 Average duration and data segment length for typical appliances ................. 120
Table 5-5 Theoretical criteria for association determination .......................................... 121
Table 5-6 Criteria for association determination ............................................................ 122
Table 5-7 Example of event association judgment ......................................................... 123
Table 5-8 Search window and results of event clustering............................................... 124
Table 5-9 Electric signature error between reconstructed cycles and reference cycles for
house #1 .......................................................................................................................... 130
Table 5-10: 3-stage time required for the most time-consuming appliances .................. 131
Table 5-11 Electric signature error between reconstructed cycles and reference cycles for
house #2 .......................................................................................................................... 132
Table 5-12 Electric signature error between reconstructed cycles and reference cycles for
house #3 .......................................................................................................................... 134
Table 5-13 Event association judgment For heater(b=0.3,c=0.8) ................................... 139
Table 6-1 3 possible ON-OFF matches of light A .......................................................... 150
Table 6-2 Energy consumption for house #1 .................................................................. 157
Table 6-3 Energy consumption for house #2 in spring ................................................... 158
Table 6-4 Energy consumption for house #2 in fall........................................................ 160
List of Figures

Figure 1.1: Example of compound load power ................................................................... 3

Figure 1.2: Time use probability profile for stove .............................................................. 5
Figure 1.3: Flowchart of DOE-2 ......................................................................................... 7
Figure 1.4: Comparison of direct and indirect sensing from [49] ....................................... 9
Figure 1.5: Typical NILM procedure ................................................................................ 11
Figure 1.6: Chart of current NILM researches .................................................................. 12
Figure 2.1: Real-time power data acquired from meter-side ............................................ 16
Figure 2.2: Detecting an event in sample data from [58] .................................................. 17
Figure 2.3: Examples of remaining challenges for event detection .................................. 18
Figure 2.4: Example of data segmentation based event detection .................................... 19
Figure 2.5: Flowchart of segment-conjunction method .................................................... 20
Figure 2.6: Example of good and bad segment ................................................................. 21
Figure 2.7: Merge good segments together based on SD and range calculation .............. 22
Figure 2.8: Slopes of a data period ................................................................................... 23
Figure 2.9: Flowchart of slope algorithm.......................................................................... 23
Figure 2.10: Noisy points in a data segment. .................................................................... 24
Figure 2.11: Comparison of the two methods on dealing with noises .............................. 25
Figure 2.12: The two methods on dealing with spike-type event and slow event ............ 26
Figure 2.13: Histograms of captured data segments using the two proposed methods .... 27
Figure 2.14: Examples of double-phase events ................................................................ 28
Figure 2.15: Example of adjacent events .......................................................................... 28
Figure 2.16: Result when the sampling rate is doubled .................................................... 29
Figure 2.17: Solving adjacent event problems by separating a slow event....................... 30
Figure 2.18: Example of event overlap ............................................................................. 30
Figure 3.1: Example of single-state signatures from [59] ................................................. 34
Figure 3.2: Examples of real power turn-on transients from [62]..................................... 35
Figure 3.3:Power curves of three types of loads ............................................................... 36
Figure 3.4: Non-overlapping window and overlapping window .................................... 38
Figure 3.5: An illustration of event signatures .................................................................. 39
Figure 3.6: Two-port network representation of load and its voltage, current and power 39
Figure 3.7: Relations of P,Q,S and  .............................................................................. 42
Figure 3.8: Example of a power electronic circuit--- Three phase SCR rectifier ............. 43
Figure 3.9: Distorted harmonic current waveform of a typical rectifier ........................... 44
Figure 3.10: Harmonic spectrum of a typical rectifier ...................................................... 44
Figure 3.11: Examples of V vs. I plots of linear and non-linear loads ............................. 45
Figure 3.12: Repetitive sequence ..................................................................................... 46
Figure 3.13: Fixed sequence ............................................................................................. 46
Figure 3.14.Combination sequence................................................................................... 47
Figure 3.15: Trend signature 1---rising spike (Fridge) ..................................................... 48
Figure 3.16: Trend signature 2---gradual falling (Dryer).................................................. 48
Figure 3.17: Trend signature 3---Falling spike (TV) ........................................................ 48
Figure 3.18: Trend signature 4---Pulses (Washer) ............................................................ 49
Figure 3.19: Trend signature 5---Flat (Kettle) .................................................................. 49
Figure 3.20: Trend signature 6---Fluctuation (Freezer) .................................................... 50
Figure 3.21: Trend signature 7---High frequency noise (Laptop)..................................... 50
Figure 3.22: Typical appliances on-hours for weekends .................................................. 51
Figure 3.23: North America residential wiring ................................................................. 53
Figure 4.1: The structure of NILM algorithms based on [94]........................................... 56
Figure 4.2: Example of the training process for signal-combination based algorithm ..... 57
Figure 4.3: Training inputs from [69] ............................................................................... 57
Figure 4.4: Training inputs from [78] ............................................................................... 58
Figure 4.5: Training inputs from [59] ............................................................................... 58
Figure 4.6: Event based algorithm from [55] .................................................................... 60
Figure 4.7: Event based algorithm from [87] .................................................................... 60
Figure 4.8: General Identification procedure .................................................................... 62
Figure 4.9: A section of meter signal collected from CT-A ............................................. 63
Figure 4.10: Event signature scoring ............................................................................... 68
Figure 4.11: Sequences of two window candidates compared to the appliance candidate69
Figure 4.12: Data flow chart of the NILM system ............................................................ 74
Figure 4.13: Data acquisition system at the meter-side .................................................... 74
Figure 4.14: Example of a data snapshot .......................................................................... 75
Figure 4.15: Data acquisition of load signatures............................................................... 76
Figure 4.16: Approach 2 for P,Q and Ih calculation ......................................................... 78
Figure 4.17: Appliance energy decomposer software ...................................................... 80
Figure 4.18: Examples of identification for fridge ........................................................... 83
Figure 4.19:Examples of identification for microwave .................................................... 84
Figure 4.20: Examples of identification for washer .......................................................... 84
Figure 4.21: Examples of identification for dryer............................................................. 85
Figure 4.22: Examples of identification for stove elements using low power .................. 85
Figure 4.23: Examples of identification for stove elements using high power ................. 85
Figure 4.24: Examples of identification for coffee maker ................................................ 86
Figure 4.25: Examples of identification for kettle ............................................................ 86
Figure 4.26: Examples of identification for heater ........................................................... 87
Figure 4.27: Examples of identification for waffle iron ................................................... 87
Figure 4.28: Examples of identification for furnace ......................................................... 88
Figure 4.29: Examples of fridge identification under noisy condition ............................. 89
Figure 4.30: Examples of freezer identification when operations overlap with other
appliances.......................................................................................................................... 89
Figure 4.31: Example of microwave identification when it overlaps with fridge............. 90
Figure 4.32: Examples of identification for TV ................................................................ 91
Figure 5.1: Three power meters based signature extraction ........................................... 100
Figure 5.2: Relations of magnetic field, electric field and the EMF event detector ....... 101
Figure 5.3: Smart phone and human confirmation based signature extraction system ... 101
Figure 5.4: Intrusive event-window signature extraction system ................................... 102
Figure 5.5: Flowchart of proposed approach versus corresponding data flow ............... 104
Figure 5.6: Example of data piece connection for kettle ................................................ 107
Figure 5.7: Example of K-means algorithm.................................................................... 112
Figure 5.8: Example of Mean-shift clustering applied to image segmentation .............. 114
Figure 5.9: Effect of feature selection ............................................................................. 117
Figure 5.10: Example of single events. ........................................................................... 118
Figure 5.11: Example of repetitive events ...................................................................... 119
Figure 5.12: Example of occasional events .................................................................... 119
Figure 5.13: Example of unrelated events ...................................................................... 120
Figure 5.14: An example of event association from 4 data segments ............................. 122
Figure 5.15.The total power on a typical day from house #1 ......................................... 124
Figure 5.16.Reconstructed cycles (red) vs. Labeled real cycles (blue) in house #1 ....... 128
Figure 5.17: The total power data on a typical day from house #2 ................................. 131
Figure 5.18: Reconstructed cycles (red) vs. l with labeled real cycles (blue dash) for the
top -load washer in house #2 .......................................................................................... 132
Figure 5.19. The total power data of the first 86400 points from house #3 .................... 133
Figure 5.20: Reconstructed cycles (red) vs. real cycles (blue dash) for washer in house #3
........................................................................................................................................ 134
Figure 5.21: Laboratory switching experiment---Scenario 1 .......................................... 135
Figure 5.22: Laboratory switching experiment---Scenario 2 .......................................... 136
Figure 5.23: Laboratory switching experiment---Scenario 3 .......................................... 136
Figure 5.24: Laboratory switching experiment---Scenario 4 .......................................... 137
Figure 5.25: Clustering results of all evens in space heater’s 12 segments .................... 138
Figure 6.1: Example of energy estimation using all window events .............................. 145
Figure 6.2: Flowchart of Energy estimation methods for incandescent lights ................ 146
Figure 6.3: Example of clustering results of IL events. .................................................. 148
Figure 6.4: ON/OFF pattern of light A ........................................................................... 150
Figure 6.5: Energy blocks of light A .............................................................................. 151
Figure 6.6: Energy distribution of IL in house #1........................................................... 152
Figure 6.7: Energy distribution of IL in house #2........................................................... 153
Figure 6.8: Example of background power ..................................................................... 154
Figure 6.9: Background power extracted from 2 houses ................................................ 155
Figure 6.10: Energy consumption pie-chart for house #1 ............................................... 156
Figure 6.11: Energy consumption pie-chart for house #2 in spring ................................ 158
Figure 6.12: Energy consumption pie-chart for house #2 in fall .................................... 160
Figure 6.13: A summary of TOU prices in spring, 2013 by Hydro One ........................ 162
Figure 6.14: Electricity billing example by Hydro One.................................................. 164
Chapter 1

Chapter 1

Introduction

This chapter clarifies several basic questions regarding the thesis’s research
subject: decomposition techniques for power system load analysis. Firstly, the
background and importance of this research are introduced. Secondly, the
research problem is properly defined. Thirdly, three existing research directions
that may solve the defined problem are individually reviewed by investigating the
literature details. Finally, this chapter discusses the scope of the thesis and
presents the thesis outline.

1.1 Background

The increased public awareness of energy conservation in recent years has

created a huge interest in energy consumption monitoring at the end-user side of
power system. Conventionally, meters installed in the downstream of power
system can obtain only the aggregated compound load information. However,
according to a recent market research report [1], end-users have become interested
in tools that can help them understand and manage the details of energy use and
its expense. This trend is especially obvious for the residential end-users [1]-[9]
since commercial and industry end-users may already have advanced energy
auditing tools and protocols.

For residential customers, a critical link to address the above need is smart
meter. According to [2], smart meter is defined as an electric meter that records
the consumption of electric energy in very short intervals such as an hour or less
and communicates that information at least daily back to the utility for monitoring
and billing purposes. The above two main features of smart meter enable a deeper
and clearer view of home electricity usage by the end-users. However, the smart
meters currently available in the market can provide only the compound load

1
Chapter 1

information or the information about an entire residential house. They cannot tell
which appliance loads in the household consume the most energy or are least
efficient. If households can also understand their usage patterns of concrete
appliances, the following benefits can be achieved:

Recently, instead of using constant retail price for electricity, substantial

variable electricity rates have been proposed and started to be used in some areas
of North America. The reason to adopt such pricing structures is that they can
more closely reflect the actual cost of electricity at a given time or period, which
may potentially lead to user’s adjustment of electricity usage accordingly. One
example is the hourly Real-Time Pricing (RTP) used in Illinois [12]-[13]. In
Canada, a lower-resolution variable price rate---Time of Use (TOU) price has
been widely used in the province of Ontario [14]-[19]. TOU is the electricity price
that is pre-set for a specific time period on an advance or forward basis [15]. In
Ontario, instead of varying the price hourly, only three TOU prices are applied to
three corresponding periods in a day. Overall speaking, to take full advantage of
the above variable rates, householders need to be informed of the schedule
patterns of individual appliance activities. Therefore, the compound load
information needs to be decomposed into the sub-load level.

Besides, even with flat electric pricing, breaking compound load down to
individual component level can still help customers control and save their energy
usage better. For example, customers can compare the efficiency of a certain load
with his neighbors in the same geographic area. Also, they will naturally pay
more attention to heavy power consumers after identifying them. A preliminary
study from a pilot program between IBM and the City of Dubuque, Iowa has
indicated strong engagement by residents and energy savings of up to 11% by
making comparison among residential profiles [20].

For the utility side, understanding the decomposed and detailed usage patterns
can be very helpful for load forecast, load settlement and other load studies [10].
Also, it can be used for electricity rate design and residential demand response

2
Chapter 1

management planning [10],[91]. In addition, potentially reducing power demand

in peak hours can lead to a cost reduction for utility companies [11] .

1.2 Problem definition

Active Power

10000

8000
Power(W)

6000

4000

2000

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Hours(h)

Reactive Power

Figure 1.1: Example of compound load power

2000

In Figure 1.1, the compound load power of a residential house throughout a

Power(Var)

1500

day1000
is shown. The compound power is acquired from the meter-side, and as can
be seen,
500
all different appliances’ powers aggregate together. This is because for
any circuit,
0
power is physically additive regardless of the individual component
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
being resistive, inductive or capacitive Hours(h)
[98].The problem to be solved can be
stated as follows: individual loads’ signals aggregate at the entry point of a house
as P (t ) and the intent of this research is to do the reverse or to decode the overall
signal into various components Pi (t ) that are attributed to specific loads
(appliances) i.

P(t )  P1 (t )  P2 (t ) .  Pn (t ) (1.1)

Assuming each load has its specific operation state, the above problem is
equivalent to the following equiation:

P(t )  P1 (t )  P2 (t ) .  Pn (t )
(1.2)
 f1 ( S1 (t ), p1 )  f 2 ( S2 (t ), p2 )  ...  f n ( S n (t ), pn )

3
Chapter 1

where Sn (t ) is the operation state of individual load n and pn is the power drawn
from a certain state of load n. Equation (1.2) actually implies the problem of load
decomposition can be converted to the identification of activities or states of
individual loads based on the aggregated signal. For example, an incandescent
light bulb has only two states: ON and OFF. For a given instant, once its state is
known, the instant power the bulb is drawing can also be determined. This is a
very important conclusion because many previous methods that will be reviewed
in this chapter are actually based on equation (1.2).

In addition, when a time period is interested instead of a particular time point,

the problem can be defined by using the following formula:

E (t0 , t1 )  E1 (t0 , t1 )  E2 (t0 , t1 ) .  En (t0 , t1 ) (1.3)

where t0 and t1 are the beginning and ending time of a specific time period and

En is the energy consumed by load or load group n during this period.

Equation (1.3) is especially suitable when dealing with certain load groups or
long-term energy components such as the energy consumed by all the in-home
lights in a day and the stand-by energy. In these cases, the instant power of the
sub-load is not of interest whereas the energy consumed during this period as a
whole is of more concern.

1.3 Review of statistic data based methods

This research direction aims to provide the estimate of sub-level loads based
on statistical samples and certain given parameters of compound loads. The
common procedure is as follows: As prior knowledge, statistical data such as
house occupant behavior, demographic characteristic, proclivity characteristic and
appliance power characteristics have already been collected from extensive
surveys and measurements done by previous organizations and researchers [21]-
[30]. Secondly, for a new compound load of interest, specific parameters such as
the number and type of occupants and building structure information are also

4
Chapter 1

given as calibration inputs. Then, the above two parts of information are both fed
into a simulation module. Finally, its decomposed energy information can be
estimated [31]-[35].

For example, in [21], a nationwide time use survey was conducted by

ISTAT(Italian Central Institute of Statistics) covering a sample of 40000 persons
between June 1988 and May 1989. Participants were asked to keep a diary listing
all their activities for a specific day. The survey’s questions concern about the
starting and finishing times, brief description and place etc. for daily activities of
a household. Based on the survey results and appliance penetration analysis, [33]
generated different home load activity probability profiles such as housework
loads (clothes-washer, dish-washer), cooking loads, leisure time loads (TV).

[23] is the UK 2000 Time Use Survey which collected 20,981 1-day diaries
recorded at a 10-min resolution. 19,898 of them were considered qualified for
statistical analysis. This survey includes detailed questions regarding how people
spend their time at home. For example, for home cooking category, the related
questions involve “unspecified food management”, “food preparation”, “baking”,
“dishwashing”, “preserving” and “other unspecified food management”. Based on
this survey, researchers constructed high-resolution daily activity profiles [31] and
these profiles for different appliances can be downloaded online [24]. One
example is shown in Figure 1.2.

Figure 1.2: Time use probability profile for stove

5
Chapter 1

[26] discusses the derived profiles of electric loads in Canadian houses. It

adopts the statistical study results obtained by Pacific Northwest National
Laboratory in 1989 [27]. Energy use profiles of air infiltration, ventilation,
lighting equipment, appliances and miscellaneous electric loads were obtained
from this study in which measurements in some benchmark houses were
conducted.

Besides statistical time use profiles of loads, other statistical data such as
appliance duration characteristics can be found or derived from Canadian Center
for Housing Technology [25] and standard appliance test methods of the
Canadian Standards Association [28]-[29] .Appliance power characteristics can
be obtained from available measurement data [30],[102].

After statistical data are obtained as inputs, concrete compound load inputs are
also provided to “calibrate” the statistical models. For example, in [31]-[33],
inputs such as the number and type of household members, availability of
household member during a day, appliance ownership are taken into consideration.

There are already a few commercial programs based on the above procedure.
For example, DOE-2 is a widely used building energy simulation analysis
program [34]. Its flowchart is illustrated as Figure 1.3. Its library is a huge
database which stores information of specific sub-loads, compound loads and
relevant statistical parameters such as regions, weathers, building structure,
occupant behavior and even materials of enclosures. With user input, specific
building descriptions of compound load can be generated. Along with statistical
weather data, complicated and dynamic simulations can be performed. This
program is especially useful for energy estimation of heating, ventilating, and air-
conditioning (HVAC) equipment. Finally, the output of different sub-loads and
other detailed energy predictions can be obtained.

6
Chapter 1

Figure 1.3: Flowchart of DOE-2

Similarly, [35] is a recently developed web-based service aiming to provide
detailed consumption information for ordinary residential houses in the US. To
accurately profile the house, users are inquired with numerous questions related to
their house structures, insulations, family compositions and major appliances.
Examples are questions about the roof, ceiling, siding, floor materials,
foundations of houses, types of heating and cooling system, typical hours of use
of stove, oven and dishwasher and so on. In total, more than 100 questions have to
be answered by an ordinary householder. In the end, energy consumption
estimates on different sectors such as water heating, major appliances, other
appliances and lighting are generated.

From the above examples, the advantages and disadvantages of the method
based on survey data can be seen:

 The estimate is based only on statistical data and simulation. No real

consumption information about the compound load is measured and
taken into account. Conceptually, the error between simulation result
and the real data is likely to be fairly large and is difficult to be
interpreted and reduced.

7
Chapter 1

 Values collected from survey data cannot represent individual samples

accurately. Variations of individual samples such as uncommon types
of appliances, uncommon materials used for doors and windows, and
different indoor preferences such as temperatures cannot be avoided.
Sometimes, the variations can be huge.

 The prior parameters of compound load are required and have to be

inputted to the system. However, it is very difficult for ordinary users
to collect such parameters. Sometimes, the collecting of some
parameters even requires special expertise.

 The advantage is that this solution is cheap: no additional hardware

installation is needed for this method.

1.4 Review of measurement based methods

The measurement based methods require additional sensing devices connected

to individual sub loads of interest. Generally, the two different streams are direct
sensing, which senses the current drawn by the load and indirect sensing, which
senses the other physical quantities related to the acoustic, light or magnetic field.

In residential houses, three methods are usually used to implement direct

measurement: "Smart plugs" [36]-[37] are devices connected between appliances
and electricity outlets. They have measurement circuits inside which can acquire
appliance's currents flowing via and the voltage from the outlet. Thus, smart plugs
can measure the power of connected appliances in real-time. The data from
different appliances can be collected either through wireless or power-line based
communication; some devices [39]-[40] are installed inside house electricity
panels. Pocket-size current sensors are connected to individual circuit branches in
panels. These branches are either wired to specific appliances such as a dryer or to
specific rooms such as a kitchen; "smart appliances"[41]-[43] are appliances that
have their own measurement circuits inside and are also able to communicate with
a smart meter or other devices via a Home Automation Network [44]-[48].

8
Chapter 1

However, such appliances are not commonly seen in the market yet at present. In
addition, the direct measurement based method can also be applied to specific
loads such as HVAC and VFD in commercial and industry buildings.

Figure 1.4: Comparison of direct and indirect sensing from [49]

Sometimes direct current sensing is not available. For example, the wiring of
residential lights is often through the wall and current meter cannot be connected
without cutting out the wire. Thus, [49] proposed and developed the “indirect”
sensors (acoustic, light and magnetic sensors) and put them near the appliances to
estimate their power consumptions. No in-line sensor such as current-plug is
needed. Also, each sensor has a calibration process so that the sensors can
establish a specific relationship between the measured physical quantities and the
power. Direct and indirect sensing are compared shown in Figure 1.4.

[50]-[51] are also “indirect” sensors but they measure only the state transitions
or ON/OFF changes of appliances. To estimate the energy, the state recordings
need to be associated with the meter-side power data. The working principle can
be explained by using formula (1.2). [3] uses a clamp based magnetic
sensor ,which can be clamped around the supply cable of an appliance. Although
the summation of the cable’s flow-in and flow-out currents is equal to zero, after
signal amplification, a minor magnetic difference can still be observed. In [51],

9
Chapter 1

the concept “binary sensor” is proposed and it can be used to detect ON/OFF
states of appliances.

Overall, the characteristics of measurement based methods can be summarized

as below:

 They require additional hardware measurement devices.

 The cost of a system can be very high in terms of installation and

maintenance.

 In some cases, a communication system is also necessary. Moreover,

the customer acceptance of Home Automation Networks is still low, in
spite of the government and media efforts.

 The accuracy of the consumption data collected through measurement

is much higher than that of the statistic data based method, especially
when using direct sensing devices. This is because the consumption of
individual load is directly measured.

1.5 Review of non-intrusive load monitoring based methods

The research on non-intrusive load monitoring (NILM) originates from MIT

[53]: The goal of the work was to develop a method for power companies to study
the residential load characteristics without having to enter the residences.

Unlike the measurement based method, a NILM system uses only the
information acquired from the main breaker level or meter-side. This system is a
viable alternative to the Home Automation Network.

Recently, with the fast development and vast deployment of smart meters,
more research attention has been drawn into this area. Smart meter’s high-
resolution data acquisition capability, communication capability along with its
computation capability [52]-[54] can provide sufficient support for the
implementation of NILM in residential houses. On the other hand, NILM can add

10
Chapter 1

great intelligence and value to the smart meters and make smart meters a truly
“smart” solution for residential energy management.

Aggregated
meter-side signal

NILM classification/ Prior signatures of

identification individual loads
module

Decomposed signal of
individual loads or load groups

Output

Applications

Figure 1.5: Typical NILM procedure

A typical NILM method and system comprise the steps as shown in Figure 1.5
[55]-[79]. The aggregated meter-side signal is acquired from meter-side through
either smart meters or additional data acquisition devices. Also, specific appliance
features or signatures are collected and mathematically characterized [94].
Afterwards, both the aggregated signal and signatures are fed into the core step:
NILM classification/identification module. In this step, the aggregated signal are
classified or identified, and the signal is decomposed into the individual load or
load group level. Finally, the results are formatted as output, and different
applications such as energy estimation, demand response, and condition
monitoring can be implemented.

Generally, all the NILM studies can be divided into the categories shown in
Figure 1.6. Detailed literature reviews will be presented in the overview sections
of relevant chapters. Here, a rough division is given:

11
Chapter 1

NILM
Researches

Signatures Algorithm Application

studies studies studies

Figure 1.6: Chart of current NILM researches

 Signatures studies. In the time domain, the studies include steady-state

signatures [55]-[60] and transient signatures [61]-[66]. The other
studies also involve signatures after particular transforms such as
Fourier transform [59]-[60],[69] and wavelet transform [68]. These
studies will be reviewed further in Chapter 3.

 Algorithm studies. According to [94], two main algorithm approaches

are studied: the signal-combination based method [59]-[60],[69]-[79]
and the event based method [55]-[58],[61]-[66]. The algorithm studies
along with the signatures studies have been the main focuses in NILM
research area. It should be noted that the two are often tied together
since sometimes different signatures will result in different algorithms.
These studies will be reviewed further in Chapter 4.

 Application studies. [84]-[86] discusses how to use NILM to estimate

the power consumption of variable-speed drives. [87]-[90] presents on
how to use NILM for condition monitoring and fault diagnostics. In
[91]-[92], a load-shedding strategy is proposed based on NILM.
However, traditional application such as energy estimation for common
residential energy components has not been fully explained before.
This issue will be discussed further in Chapter 6.

Compared with the survey based method and the measurement based method,
the characteristics of NILM based method are as follows:

12
Chapter 1

 Low cost: no additional hardware needed except for the aggregated

meter-side signal acquisition device. In many cases, the meter-side
acquisition can be achieved by using an existing smart meter.

 High accuracy: the accuracy may be lower than that from the direct
measurement based methods but is much higher than that from the
survey based method and indirect measurement based methods.

Overall, the NILM based method provides good balance between cost and
accuracy and is therefore the most promising load decomposition technique so far.

1.6 Thesis scope and outline

The purpose of this thesis research is to solve the remaining but critical
challenges and limitations related to the existing NILM based methods:

1. Unable to deal with complex loads effectively. Complex loads such as

continuous-varying loads and multi-state loads have not been sufficiently
researched. However, in fact, they are an important portion of residential loads
and need to be addressed.

2. Time-consuming training process. Many NILM methods require a time-

consuming training/learning process to establish the map between specific
activated appliances and their aggregated signal. Moreover, after the load
inventory is changed, training process has to be redone. This requirement is a
critical obstacle that prevents NILM from being applied to the ordinary
households.

3. Lack of research on signature extraction. This is an important research area

that has been neglected by previous NILM researchers. The existing measurement
based signature extraction methods are actually intrusive. This is another critical
obstacle that prevents NILM from being practically applied.

13
Chapter 1

4. Insufficient research on event detection. The previous event detection

methods are very simple and can lead to detection error in some cases.

5. Insufficient research on energy estimation methods. How to estimate energy

based on NILM results and how to estimate energy for specific energy
components such as load groups and background power have not been addressed
before.

The thesis is organized to present different studies to tackle the above listed
problems. The outline is below:

 Chapter 1: Clarifies the basic questions of this work.

 Chapter 2: Presents a study of event detection and tackles problem 4.

The study consists of a review of the existing methods, explanations of
two new sophisticated methods, comparative studies based on real field
data and discussions on the special issues of event detection.

 Chapter 3: Presents a study of event-window based load signatures and

prepares for tackling problems 1&2. The study consists of a review of
the existing load signatures, explanations of the proposed load model
and discussions on different event-window signatures.

 Chapter 4: Presents a study of a proposed NILM identification

algorithm and tackles problems 1&2. The study consists of a review of
the existing algorithms, explanations of the proposed algorithm, a
discussion on the system implementation and thorough verification and
comparative studies.

 Chapter 5: Presents a study of non-intrusive signature extraction and

tackles problem 3. The study consists of a review of intrusive signature
extraction methods, explanations of the proposed algorithm and
thorough verification studies.

14
Chapter 1

 Chapter 6: Presents a study of energy estimation for residential house

and tackles problem 5. The study consists of explanations of estimate
methods for different energy components, an interpretation of the
residential energy characteristics and their implications to the Time-of-
Use price.

 Chapter 7: Presents the main conclusions from this work, and

suggestions for future studies and improvements.

 Appendix: Presents a preliminary study of a load decomposition

technique specified for a commercial compound load. The appendix
also addresses problem 5.

15
Chapter 2

Chapter 2

Event detection

A load event is defined as the transition of a load’s state. Event detection is an

essential pre-processing step for most of the methods and algorithms proposed in
this thesis. The quality of event detection has a direct impact on the final results of
event-window based load decomposition and signature extraction.

This chapter presents elaborate discussions on event detection issues. It first

reviews the existing event detection method and identifies the potential challenges.
To deal with these challenges, this chapter proposes two data-segmentation based
methods---segment-conjunction method and slope method. The basic idea is that
instead of directly seeking for status-transitions, one can find out all the
continuous data segments and then the portions between two neighboring
segments can be considered as the events. The two proposed methods are
compared according to the tests on actual field data.

In addition, this chapter also addresses the special problems of event detection-
--double-phase event detection, adjacent event handling and event overlap. Some
contents in this chapter have been submitted as publication [103].

2.1 Overview

Power/w
900

800

700

600

500

10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30

Time/Clock
Figure 2.1: Real-time power data acquired from meter-side

16
Chapter 2

According to [55], a load event is defined as the transition of a load’s operation

state. As can be seen from Figure 2.1, although aggregated signal acquired from
meter-side contains the summation of individual signals of all currently activated
appliances, an event which is relevant to a certain appliance’ operation state
change such as its ON/OFF usually associates with a power jump and thus can be
observed as an independent “edge” from meter-side signal. To capture such edges,
event detection methods are needed.

Figure 2.2: Detecting an event in sample data from [58]

The previous researches on event dection are limited and most of them rely on
a certain threshold based step-change detection [55]-[59]. In [57], a revoving
memory keeps the last measured successive real power values P1 and P2.  is the
detection threshold and whether P2  P1   is continuously tested. Once
P 2  P1   is found, the corresponding step change is captured as an event.
Based on interested load power level, different thresholds can be adopted. For
example, in [59], 100W is used as the threshold. In [58], a similar method is
proposed with extra points considered to judge if a steady period starts right after
a potential event. Also, both real power and reactive power are considered
together for determination. A period of change is detected if the site-specific
thresholds dPtres and dQ are exceeded while the tolerances dPtol and dQtol are not
tres

17
Chapter 2

exceeded according to the following equation (Pi , Pi+1 and Qi , Qi+1 are successive
samples). The process is shown in Figure 2.2.

| ( Pi  Pi 1 ) / 2  Pi 3 | dPtres

| (Qi  Qi 1 ) / 2  Qi 3 | dQtres
 (2.1)
| Pi  4  Pi 3 | dPtol
| Qi  4  Qi 3 | dQtol

The above step-change methods can deal with simple events. However, there
are some remaining challenges that need to be coped with.

1000

800
Power(W)

600

400

200

0
2.295 2.3 2.305 2.31 2.315
Time(sec) x 10
4

(a) Spike-type event

2000

1500
Power(W)

1000

500

0
3.143 3.144 3.145 3.146 3.147 3.148
Time(sec) x 10
4

(b) Slow event

450

400

350
Power(W)

300

250

200

150
5.08 5.09 5.1 5.11 5.12 5.13 5.14 5.15 5.16 5.17 5.18
Time(sec) 4
x 10

(c) False event (signal noise)

Figure 2.3: Examples of remaining challenges for event detection

18
Chapter 2

As can be seen from Figure 2.3, (a) is a spike-type event that can be often seen
when a motor device is started. This event is composed of a sharp positive step
change and a smaller negative step change that follows right after the positive
change. The traditional step-change based methods may consider this event as two
independent ON/OFF events; (b) shows a slow ON event from a microwave.
However, it actually has two small step changes followed by a big step change.
The traditional method such as [58] will discard this event
because | Pi  4  Pi 3 | dPtol ; (c) shows the mis-detection case---these quick step
changes are actually caused by signal noises and should not be considered as
events.

To deal with the above challenges, this chapter proposes two novel data-
segmentation based methods. As shown in Figure 2.4, the new philosophy behind
is that instead of looking for step change like the previous methods do, the
proposed methods look for all the continuous data segments and then the data
portions between two neighboring segments are considered as the events no
matter how complicated they are. Besides, the continuous data segment should be
capable to include limited abnormal points as long as the continuity of the data
segment is not significantly broken by them.

Figure 2.4: Example of data segmentation based event detection

19
Chapter 2

In addition, other special issues about event detection such as double-phase

event detection, adjacent event handling and event overlap will also be discussed
in the end of this chapter.

2.2 Data segmentation based methods

As shown in Figure 2.4, data segment is defined as a segment of which the

continuity is good. It should include three scenarios: 1) a steady segment that has
a constant power level and does not change over time. It can be seen as a straight
line; 2) a continuous varying segment that has inconstant but slowly varying
power level over time. It can be seen as a curve; 3) a segment with acceptable
noises like the one shown in Figure 2.3 (c).

Two methods that can detect the above defined data segments are proposed.
The data portions between two neighboring segments are considered as events.
They are explained as follows.

2.2.1 Segment-conjunction method

As shown in the flowchart below, the segment-conjunction method has 4 steps.

Divide entire dataset

X into small data
segments
X={X1.X2.X3…}

Segment continuity judgment:

“Good” segment à Calculate its average

Merge neighboring “good” segments to

longer segment through judging SD of
average values

Output

Figure 2.5: Flowchart of segment-conjunction method

20
Chapter 2

Step 1: Divide the entire dataset X into small continuous data segments {X1,
X2, X3…} with a preset minimal data length such as 5 points per segment. Since
the data is acquired every 1 second, 5 points represents 5 seconds.

Thus, we have X= {X1, X2, X3…..} X: aggregate of magnitudes. X1, X2,

X3… follows the original sequence in the dataset.

Step 2: Judge each segment to see whether it is a “good” segment with

acceptable variation.

If the variation among each segment is beyond a certain threshold, it is marked

as a “bad” segment; if the variation is below the threshold, it is marked as “good”
and its average value is calculated. Standard deviation (SD) is used as an index
for variation evaluation.

Examples of good and bad segments are shown in Figure 2.6. Each circle
represents a data point. Compared with (a), in (b) the variation of five points is
too big or in other words, its SD exceeds the predefined threshold, hence this
segment is not considered acceptable.

Figure 2.6: Example of good and bad segment

After marking all these segments, calculate the average value of each “good”
segment. E.g. If X3, X4, are good segments, Ave3=Average(X3),
Ave4=Average(X3)…

Step 3: Try to merge neighboring “good” segments: Calculate the standard

deviation (SD) of the average values of neighboring segments. If it is less than a
threshold we set, like 2.0, join them together as longer segment.

21
Chapter 2

Figure 2.7: Merge good segments together based on SD and range calculation
For example, assume X3,X4,X5,X6, X7 segments are “good” segments and
they are continuous. If SD (Ave3, Ave4, Ave5) <2.0 then {X3,X4,X5}=LX1.
LX1 is the merged longer segment. Similarly, if SD(Ave3, Ave4, Ave5,
Ave6)>2.0, X6 is excluded from LX1 and LX1 is finalized as a whole segment
that includes X3,X4 and X5.

Step4: Output. Label data segments and events in different colors

2.2.2 Slope method

Slope algorithm focuses on a group of data points’ slopes, which is considered

as an effective index to judge the continuity of data segment.

As is known to us, slope dx / dt can be used to evaluate the velocity of

variation. Points in the segments that have good continuity should have small
slopes. In reality, the approximation below is used:

dx x
 (2.2)
dt t

Here, x is the difference between two successive acquired data points. t is

the time difference between two sampling points, which is actually the acquisition
interval. Since t is fixed, only x needs to be taken into account. In other words,
x of each point can represent the point’s slope or velocity of variation.

Figure 2.8 below shows the slope xn for a specific data period. Here
xn  xn  1  xn .Power curve is shown in dash line and the slope line below it
shows the corresponding the slope values of each point.

22
Chapter 2

2000
Slope
Power
1500

1000
Powe(W)

500
Data segment Data segment

-500

Event Event
2.6 2.65 2.7 2.75 2.8 2.85
Time(sec) x 10
4

Figure 2.8: Slopes of a data period

From this figure, it can be seen that many slope values are close to zero, which
indicate small variations. Some points, however, are very large, which can reach
up to 50 or even 1000. Those points express large variations, which could be an
event or part of an event. The basic idea is to join continuous points that all have
slope values close to zero as a data segment.

The flow chart of the proposed slope algorithm is shown in Figure 2.9.

Calculate slopes:
xn  xn  1  xn

Set the next data point as

the starting point

Join the points with close-to-

zero slopes as a segment

N
Reach the end?

Output

Figure 2.9: Flowchart of slope algorithm

23
Chapter 2

Step 1: Subtract each data point from the next data point following it.
xn  xn  1  xn . In other words, the slope values of all data points are calculated.

Step 2: From a starting location, try to join the following points with small
slopes. A special concern here is the occasional noisy points shown in Figure 2.10.

Figure 2.10: Noisy points in a data segment.

For this situation, since slopes of noisy points are big, they will not be
automatically connected with the points in front. However, in reality, one long
segment with few noisy points included is still acceptable. In order to prevent
cutting off a segment too early due to the noisy points, the method will make an
additional check on the steady points behind the noisy peak at the same time, to
see if they are close to the values in front of the short peak. If they are close, these
noisy points are “bypassed” and the connection with the following points will
continue.

The connection stops when both of the following two conditions are met:

 The slope value of the point is beyond the threshold;

 The difference of the average of the steady points after this point and
before the point is also beyond the threshold

Step 3: After the interruption point, start the processing of a new segment from
the first point that has a close-to-zero slope value after the interruption point.
Redo step 2. The whole iteration ends until all points of the dataset have been
processed.

Step 4: Output.

24
Chapter 2

2.2.3 Comparison of the two proposed methods

700

600

500
Power(W)

400

300

200

100

0
4.9 5 5.1 5.2 5.3 5.4 5.5 4
Time(sec) x 10

(a) Segment-conjunction method

700

600

500
Power(W)

400

300

200

100

0 4.9 5 5.1 5.2 5.3 5.4 5.5 4

x 10
Time(sec)

(b) Slope method

Figure 2.11: Comparison of the two methods on dealing with noises

First of all, the performance of the two proposed methods on noisy point
handling is compared. Figure 2.12 shows four consecutive fridge’s operations.
They are all polluted by signal noises from the meter-side. The two proposed
methods are applied to the four cycles respectively. As can be seen from (a), the
data segments (in red color) captured by using the segment-conjunction method
are interrupted many times by noises; while in (b), there are only 8 big data
segments that represent the ON and OFF states of fridge. Accordingly, 8 events

25
Chapter 2

are captured (in blue color) between neighboring data segments. Hence, the slope
method seems much more powerful when dealing with noisy segments.
Referring to the other challenges mentioned in 2.1, both of the two methods
can effectively deal with the spike-type event and slow event that contains more
data snapshots. Examples are shown in Figure 2.12.

1400

1200

1000
Powe(W)

800

600

400

200

4.93 4.935 4.94 4.945 4.95

Time(sec) x 10
4

(a) Spike-type event

2000

1500
Power(W)

1000

500

3.143 3.144 3.145 3.146 3.147 3.148 3.149

Time(sec) x 10
4

(b) Slow event

Figure 2.12: The two methods on dealing with spike-type event and slow event

The two methods have also been applied to the data acquired from a single hot
phase of a local residential house on a typical day. The histograms of captured
data segments using the two proposed methods are shown in Figure 2.13. As can
be seen, the number of segments captured from the slope method is smaller but
the average length of data segment is longer. Again, it proves that the slope
method can catch longer segments containing noises. Accordingly, the number of

26
Chapter 2

events captured by the slope method will also be smaller but much more
reasonable.

2000

1500
Number of segments

1000

500

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Length of segments

(a) Segment-conjunction method

600

500
Number of segments

400

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Length of segments

(b) Slope method

Figure 2.13: Histograms of captured data segments using the two proposed
methods

2.3 Special issues with event detection

2.3.1 Detect double-phase events

As shown in Figure 2.14, some events happen simultaneously on both of the

two hot phases in a residential house. It is because some appliances are connected
between the two opposite hot phases to gain a higher voltage. Therefore, there is a
need to recognize these double-phase events and label them. The following
formula can be used to effectively make judgment on double-phase events. At a
given instant when simultaneous events are found at Phase A and B, if

dPA / dPB  Thresmin and dPA / dPB  Thresmax (2.3)

27
Chapter 2

then the event is labeled as a double-phase events. The power differentials

between dPA and dPB should be very close since they have almost identical

currents. Typical values of Thresmin and Thresmax are 0.9 and 1.1.

Phase A
5000

4000
Power(W)

3000

2000

1000

0
4.4 4.45 4.5 4.55 4.6 4.65 4.7 4.75 4.8 4.85 4.9
Time(sec) x 10
4

Phase B
3500

3000

2500
Power(W)

2000

1500

1000

500

0
4.4 4.45 4.5 4.55 4.6 4.65 4.7 4.75 4.8 4.85 4.9
Time(sec) x 10
4

Figure 2.14: Examples of double-phase events

2.3.2 Adjacent events

Actual
power curve

Acquired
data points

dP

Figure 2.15: Example of adjacent events

28
Chapter 2

Sometimes adjacent events may lead to event misdetection due to low

sampling rate. As can be seen from Figure 2.15, the actual power curve has two
adjacent events with a short interval in between. The acquired data points (circles)
are shown under the actual curve as comparison. For each interval, one data point
is acquired. Based on the proposed event detection methods, during this period,
the first 3 points and the last 4 points will be considered as two segments and the
3 points in between will be detected as a slow event. This obviously violates the
truth. However, when the sampling rate is low this problem is evitable.

One solution is to increase the sampling rate. For example, doubling the
sampling the rate will result in different results as shown in Figure 2.16. In this
case, three segments can be captured and thus two independent events dP1 and
dP2 can be found.

Actual
power curve

Acquired
data points
dP2

dP1

Figure 2.16: Result when the sampling rate is doubled

However, when the sampling rate cannot be increased due to hardware
limitations, the suspected slow event that could comprise multiple events can be
broken into pieces. This can often help solve the misdetection of events without
adding cost on higher sampling rate. Still taking the period in Figure 2.15 as
example, if an event is found too long (4 intervals in this example), a method is
that one manually divide it in half such as the steps shown in Figure 2.17. With
respect with Figure 2.15 and Figure 2.15, the two events dP1 and dP2 can still be
obtained even with original data acquisition capability.

29
Chapter 2

Actual
power curve

Acquired
data points
dP2

dP1

Figure 2.17: Solving adjacent event problems by separating a slow event

2.3.3 Event overlap

Also, in some cases, simultaneous occurrence of more than one appliance

events can be encountered. This does not happen frequently but does exist.
Theoretically, if two events fall in the same time interval, any event detection
method will result a misdetection---only a single event can be captured. One
example is shown in the Figure 2.18. The ON event in overlapping is composed
of two single events belonging to the fridge and furnace. Its power magnitude is
roughly the summation of the two.
1200

1000
Power(W)

800
Fridge
600

400 2000

Overlapping
Power(W)

200
0 50 100 150 200 250 300 350
1500
Time (Second)
1200 1000

1000
Furnace 500
Power(W)

800

0 50 100 150 200 250 300 350 400

600
Time (Second)
400

200

0 50 100 150 200 250 300 350 400

Time (Second)

Figure 2.18: Example of event overlap

30
Chapter 2

This problem is mathematically analyzed as a probability problem. Now

assume a period of M seconds (intervals) is studied and N appliances that each
show up K times on average are included in this period. Therefore, for each time,
N different appliance events can overlap with each other. The total number of
possible combinations for events occurring at different seconds is M N and the
number of non-overlapping combinations is AMN . AMN can be calculated based on
factorial. Assuming all the events are randomly distributed, the probability of
overlap occurrence is:

 AMN K
 P  1  ( )
MN
 (2.4)
 AN  M!
 M
( M  N )!

In practice, for a specific hour, assume 3600 points are captured, 5 appliances
are involved and each appliance shows up 3 times on average. According to (2.4),
the probability of event-overlap occurrence in this hour is

5
A3600
Ph  1  ( 5
)3  0.84% (2.5)
3600

Based on Ph , for an entire day, the probability of event-overlap occurrence is

Pday  1  (1  Ph ) 24  18.3% (2.6)

The above results may be higher or lower than the ones in reality because
appliances events are not uniformly distributed along the time line. In some hours,
different appliance operations are more expected to be seen while in some hours
are not. Also the event frequency of some particular appliances such as stove and
washer within certain hours could be extremely high.

It is also found that increasing sampling rate can also help solve the problem.
For example, now assuming the sampling rate is increased to 0.5sec/snapshot,
there will be 7200 snapshots in an hour. Therefore, we have:

31
Chapter 2

5
A7200
Ph  1  ( )3  0.42%
72005 (2.7)
Pday  1  (1  Ph ) 24  9.5%

The event overlap probability is decreased to approximately half of the original

value.

2.4 Summary

To summarize, this chapter firstly reviewed the existing event-detection

methods and identified several challenges that need to be solved. They are spike-
type event, slow event and signal noises. To cope with these problems, two data
segmentation based methods were proposed. Instead of looking for state
transitions directly, the proposed methods look for stable data segments in which
a certain level of noise is also acceptable. The first method bases on the
conjunction of small segments while the second method studies the slope pattern
of data points. Detailed comparison of the two methods was also presented and
overall speaking, the slope method has a better performance, especially when
dealing with noisy signals. In other words, it is more suitable for practical
applications. It was also found both methods can effectively detect spike-type
event and slow event.

In addition, specific common issues with event detection were also discussed:
a simple method to detect double-phase event was proposed; misidentification due
to adjacent events was also discussed. Both hardware and software solutions were
proposed and explained; in the end, the problem of event overlap was also
introduced and the method to calculate its probability was addressed. Based on a
hypothetical calculation, it was also found higher sampling rate can effectively
reduce the probability of overlap.

32
Chapter 3

Chapter 3

Event-Window Load Model and Load Signatures

This chapter presents detailed discussions on a specific load modeling

approach---event-window modeling and different types of load signatures
associated with this model. With this model, loads are treated as a time window
that consists of all the events such as turn ON/OFF events and other middle
changes during its operation process. The signatures associated with the event-
window are studied. These signatures include both electrical and non-electrical
ones that can accurately and completely describe the operation process of any
particular type of load.

The aim of this chapter is to establish a technical foundation for the proposed
event-window based load identification method which will be further discussed
afterwards. Some contents in this chapter have been submitted as publication
[104].

3.1 Overview

3.1.1 Review of single-state load model and its signatures

Most of the existing non-intrusive load monitoring methods treat all the
appliances as a single-state model which has a pair of identical ON/OFF edges
and a constant power demand between them [55]-[60]. In other words, it assumes
the operation of appliance has a constant steady point which does not change its
electric status such as real power, reactive power and harmonic content
dramatically. This assumption simplifies the load models but unfortunately, it
cannot accurately reflex the reality--- many complex loads cannot be represented
by this model and thus large errors could occur for the future identification steps.

33
Chapter 3

Representative signature studies of single-state model can been seen in [59]

and [60]. New single-state signatures such as “instantaneous power waveform”
and “instantaneous admittance waveform” were discussed. Its assumption is still
that each load only has a constant set of signatures that is not changing with time.
Examples are shown in Figure 3.1.

Figure 3.1: Example of single-state signatures from [59]

As can be seen, the waveforms of all loads look almost the same for the period
0-10 ms and 10-20ms. However, this is only true for the following scenarios:

1. When signatures are only compared in a short time scale such as within a
second. In this case, 10 ms is only half cycle (50Hz system).

2. When the studied loads have only single operation state. In Figure 2-1, only
water boiler can fall in this category. Air conditioner, TV and induction cooker
are all complex loads which have multiple or continuous varying states. They
should not to be treated as single-state appliances.

34
Chapter 3

3.1.2 Review of transient load model and its signatures

Some other researches [61]-[66] focus on the transient signatures of loads at

the moments when the studied load is turned ON. It is believed that different
loads create consistent observable turn-on transient profiles suitable for
identifying specific load classes. The turn-on transients associated with a
fluorescent lamp and an induction motor, for example, are distinct because the
physical tasks of igniting an illuminating arc and accelerating a motor are
fundamentally different. Examples are shown as below:

(a) Florescent Lamp (b) Measured induction motor

Figure 3.2: Examples of real power turn-on transients from [62]

The drawbacks of modeling a load using only its turn-on transients are listed as
below:

1. Some loads do not have significant turn-on transients such as resistive loads.
The loads with observable turn-on transients are limited to florescent lights,
motors and some electronic appliances. Some major loads such as stove and
incandescent lights may not contain observable transients.

2. The NILM based on turn-on transient models can only determine the
operating schedule of loads but cannot track the whole process of loads. It is
because a large portion of load information is not considered except for the turn-
on transient. For example, turn-off events usually do not contain sufficient
transient information as turn-on events and cannot be easily identified in this way.

35
Chapter 3

Also, similar to single-state load model, complex operation process of loads

cannot be identified. NILM application such as load energy cannot be accurately
calculated.

3. In order to acquire and process transient signatures, both high-resolution

data-acquisition system and high-speed processing units are needed. Existing
smart meters may have to be significantly modified to support this kind of
application. To meet this requirement, particular hardware solution such as using
multi-processing unit is proposed in [63]. However, it also increase the total cost
significantly.

3.1.3 Proposed event-window model and its signatures

To solve the drawbacks of steady-state load model and transient load model,
event-window load model is proposed and its signatures are studied.

Power
Continuous Multi-state
Single-state varying

Time

Figure 3.3:Power curves of three types of loads

Generally, the loads in the residential houses can be divided into three types as
shown in Figure 3.3: single state appliance has an identical pair of ON/OFF
events. During its whole operation process, the electric signatures of this load stay
almost constant; 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 complex loads such as furnace and washer.
Furnace has more than one working stage according to current environmental
temperatures and washer has many steps like rinse and drainage that follows a
certain operation pattern. The examples and characteristics of house loads are
listed in Table 3-1.

36
Chapter 3

Table 3-1 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
Hence, the approximation of assuming all appliances have single states can
lead to failure or significant error of identification for continuous-varying
appliances and multi-state appliances. Besides, for combination based approach,
the error on complex loads can further affect the identification of single-state
appliances heavily. For accurate energy monitoring purpose, real operation
processes of such complex loads need to be captured and treated specifically.

Therefore, a more generic model---event-window is proposed. 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 Figure 3.4. Window 1 contains
one ON and one OFF event associated with one appliance. There is no activation
of other appliances in between. This is called a non-overlapping window. Non-
overlapping window contains complete signature information about an appliance.
Window 2 is called an overlapping window as it contains an ON event belonging
to 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. Many of them will overlap with others. The main idea of
the proposed load identification method in Chapter 4 is to identify and pick out
the right windows that represent the interested appliances. This is accomplished
with the help from window signatures or characteristics that will be discussed
later in this chapter.

37
Chapter 3

Power

Fridge

Fridge
Light
Non-overlapping Overlapping
window 1 window 2

Time

Figure 3.4: Non-overlapping window and overlapping window

Each window contains five types of signatures that are listed as below:

 Event signatures
 Event pattern signatures
 Power trend signatures
 Time/Duration signatures
 Phase signature

Comprehensively, these signatures are able to describe the load operation from
a more complete and accurate perspective. They allow the load model to contain
multiple steady states because different events’ signatures and event pattern
signatures are now included. In the meanwhile, power variation which can be
often seen in continuous-varying loads is also described. Besides, time/duration
signatures are proposed because they can reflect the usage behaviors of appliances.
Finally, phase signature is also explained to describe the electric “location” of the
load in a residential house. These signatures are respectively discussed as below.

3.2 Event Signatures

An event refers to the change of the operating state of an appliance, which can
be often seen as a step change or edge in its power, reactive power or harmonic
content versus time plot. 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 Figure 3.5. For a certain appliance, its attributes stay
generally fixed for each time operation if the system voltage does not change
sharply. In single-state model, only one set of P-Q-W is considered since the OFF

38
Chapter 3

event is assumed to be the identical reverse of the ON event [55]. In event-

window model, however, the number of P-Q-W is equal to the events that really
happen inside this event-window.
Power

Δ:P,Q,W

Rising edge Falling edge

Time

Figure 3.5: An illustration of event signatures

3.2.1 Real Power signatures

As shown in Figure 3.6, a load is connected to the power system and it can be
represented by using a two-port network [98].

Figure 3.6: Two-port network representation of load and its voltage, current and
power

For residential loads, the port voltage u  2U sin(t  u ) .U is the magnitude

of residential system voltage and is usually around 120V in North America.  is

the fundamental frequency of power system and is usually around 60Hz in North
America. u is the phase angle of voltage and its value is dependent on the
reference such as the swing bus voltage in system.

39
Chapter 3

The current flowing into the load is i  2I sin(t  i ) . I is the magnitude of

load current and is dependent on u and the equivalent impedance of load. u is

the phase angle of current and its value is also dependent on the a certain
reference in the system.

Real power is defined as the average power absorbed by the above network
within a cycle. Its value is:

1 T
P  pdt  UI cos(u  i )  UI cos 
T 0
(3.1)

Heavily affected by the designed functions, different loads can have very
distinct real powers. For example, a stove usually has a large power since it is a
cooking device and a lot of heat is converted from electricity; a LED light on the
other hand normally consumes a little power since it only generates a little light.
Typical values of real powers of different residential load types are shown in
Table 3-2.

Table 3-2 Typical values of real power of residential loads

Appliance Type Power (W)

Incandescent light 120
Fluorescent light 40
TV 500
Stove 2000
Oven 3000
Kettle 1500
Washer 500
Microwave 1500
Fridge 200
Freezer 100

40
Chapter 3

3.2.2 Reactive Power signatures

In order to introduce the concept of reactive power of loads, the equation (3.1)
can be rewritten as [98]:

P  UI cos(u  i )  UI cos(2t  u  i )
 UI cos(u  i )  UI [cos(u  i ) cos(2t  2i )
 sin(u  i ) sin(2t  2i ) (3.2)
 UI cos [1  cos((2t  2i )]  UI sin  sin(2t  2i )
 p1  p2

In the above equation, p1 does not change its direction of power transferring
and thus represent the actual power consumed by the two-port network. On the
other hand, p2 ’s frequency is 2 and in a cycle under the actual frequency , its
average value is zero. In reality, this part of power is transferring back and forth
between the network source and network load. Its magnitude UI sin  is defined
as the reactive power of load in power system and its unit is “var”:

Q  UI sin  (3.3)

As can be seen from (3.3), when  >0, Q  UI sin  >0 and it means the load is
inductive and can absorb reactive power; when  <0, Q  UI sin  <0 and it
means the load is capacitive and can generate reactive power. cos  is defined as
the power factor of load.

Compared (3.3) with (3.2), UI can be viewed independently and it is defined as

the apparent power of load. Thus we have:

 S  UI  P 2  Q 2

 P  S cos  (3.4)
 Q  S sin 


Their relations can be graphically shown as the triangle below:

41
Chapter 3

Figure 3.7: Relations of P,Q,S and 

In power system, many resistive loads such as incandescent light, stove, oven
and kettle has a power factor which is close to 1. It also means they have
negligible Q or p2 . For motor based loads such as fridge, freezer, washer,
dishwasher and dryer, their equivalent circuits are similar to inductors. Thus they
have a lower power factor and have a notable Q or p2 . Hence, reactive power
levels can be very useful signatures when distinguishing between active loads and
reactive loads.

3.2.3 Harmonic signatures

Harmonic contents are sinusoidal components of a periodic waveform with a

frequency that is an integral multiple of the fundamental power frequency [99].
When harmonics are combined with the fundamental frequency component,
sinusoidal waveform distortions are caused. In power system, current waveform
distortion often accompanies with the use of power electronic loads, which
includes high frequency switching circuits such as AC-DC, DC-DC and DC-AC
conversion circuits. An example of a power electronic circuit is shown in Figure
3.8.

42
Chapter 3

Figure 3.8: Example of a power electronic circuit--- Three phase SCR rectifier
Circuits like the above can generate distorted current waveforms which can be
described by a function of f(t) with period T and fundamental frequency of f0=1/T.
f(t) can be expressed by a Fourier series:


f (t )  c0   ck cos(2 hf 0t  h ) (3.5)
h 1

T
2
1
where ck h 
T  T
f (t )e  j 2 hf0t dt .

2

The Fourier series decomposes the original waveforms into a series of

sinusoidal components with different frequencies. The component of f0 is called
the fundamental component, and the hf0 component is called the h-th harmonic of
the periodic function. As an example, Figure 3.9 shows the distorted harmonic
current of a typical DC power supply load, and Figure 3.10 presents the harmonic
spectrum of this waveform [101].

43
Chapter 3

Figure 3.9: Distorted harmonic current waveform of a typical rectifier

Figure 3.10: Harmonic spectrum of a typical rectifier

Residential loads can be divided into two types --- linear loads and non-linear
loads. Linear load barely causes any distortion on its current waveform and thus
has no harmonic contents on its spectrum after applying Fourier transform to
examine their current waveforms. Examples of linear loads are incandescent light,
fans, stove, oven and dryer; non-linear load can cause significant distortion on its
current waveforms and large harmonic contents can be seen after applying Fourier
transform to its current waveforms.

Another equivalent way to analyze harmonics is using V vs. I plot. Examples

can be seen in Figure 3.11. As can be seen, linear load has a “linear” relation
between its voltage and current (a) while non-linear load has “non-linear” relation
between its voltage and current (b).

44
Chapter 3

(a) Linear-load

(b) Non-linear load

Figure 3.11: Examples of V vs. I plots of linear and non-linear loads

3.3 Event Pattern Signatures

Event pattern signature describes the logical sequence of operation events of a

load. In other words, it represents the sequence of appearances of edges. For
example, a washer usually follows the below 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 or the event
pattern signature is very unique and is very important for identification of multi-
state appliances. There are three types of basic event sequences: repetitive
sequence, fixed sequence and the combination of the two.

45
Chapter 3

Stoves, dryers or some coffee makers are typical multi-state appliance with
repetitive sequence due to their internal integer-cycle controllers [96]. Figure 3.12
shows an operation process of stove.

Figure 3.12: Repetitive sequence

An example of fixed sequence is washer. It is shown in Figure 3.13.

Figure 3.13: Fixed sequence

Sometimes, a combination of repetitive and fixed sequence occurs. Figure
3.14 shows a furnace. It has repetitive heating cycles. 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.

46
Chapter 3

Figure 3.14.Combination sequence

Table 3-3 shows some examples of appliances with the sequence patterns as
discussed above measured through experiment.

Table 3-3 Sequence pattern and examples

Load type Examples

Repetitive Dryer; Stove; Some coffee makers
sequence
Fixed Incandescent light bulb;
sequence Fluorescent light bulb; Kettle; Microwave;
Toaster; Oven; Fridge; Freezer; Computer
Combination Furnace, Some dishwashers

3.4 Power Trend Signatures

A trend signature refers to the variation of power demand between two

neighboring events. For example, an inductive motor often accompanies with a
rising spike right after being turned on due to its large inrush current. This kind of
transient can be seen from Figure 3.15.

47
Chapter 3

Power 900

(W) 800

700

600

500

400

300

200

100

0
4300 4400 4500 4600 4700 4800 4900 5000 5100 5200

Time(sec)
Figure 3.15: Trend signature 1---rising spike (Fridge)
After start, as shown in in Figure 3.16, with the motor speed gradually
increases, its current drawn may sometimes decrease and form a gradual falling
curve.

Power 2200

(W) 2000

1800

1600

1400

1200

1000

800

600

400
35 40 45 50 55 60

Time(sec)

Figure 3.16: Trend signature 2---gradual falling (Dryer)

Some electronic devices may experience an instant interruption. A TV set may
experience a falling spike at moments of switching channels. An example is
shown in Figure 3.17.
Power 800

(W) 700

600

500

400

300

200

100

0
0 20 40 60 80 100 120

Time(sec)
Figure 3.17: Trend signature 3---Falling spike (TV)

48
Chapter 3

Pulses are usually caused by electronic switches. A lot of stoves have pulses
because they have integer-cycle controllers inside [96]. It prevents itself from
overheating. Another example is an inverter based motor device that adjusts its
frequency/speed all the time. An example can be seen from Figure 3.18.

Power 1400

(W)
1300

1200

1100

1000

900

800
1650 1700 1750 1800 1850 1900 1950 2000

Time(sec)
Figure 3.18: Trend signature 4---Pulses (Washer)
A lot of appliances have negligible transient characteristics and present almost
flat power curves during operation. Actually, this is the only type of appliance
which can be ideally represented by the old single-state load model used in
previous researches. An example can be seen as below:

Power 800

(W) 700

600

500

400

300

200

100

0
0 5 10 15 20

Time(sec)
Figure 3.19: Trend signature 5---Flat (Kettle)

In contrast, as can be seen from below, some appliances may have continuous
low-frequency fluctuations all the time instead of a steady state. An example can
be seen from Figure 3.20.

49
Chapter 3

Power 400

(W)
350

300

250

200

150

100

50
0 200 400 600 800 1000 1200

Time(sec)
Figure 3.20: Trend signature 6---Fluctuation (Freezer)
In addition, some audio and video electronic devices such as audio box, laptop
and desktop PC may have high-frequency noises which are shown in Figure 3.21.
It is because their power consumptions constantly vary with the sound, visual
display or other tasks their internal analogue or digital circuits are processing.

Power 80

(W)
70

60

50

40

30

20

10

0
0 50 100 150 200 250 300 350

Time(sec)
Figure 3.21: Trend signature 7---High frequency noise (Laptop)

It should be noted some appliances such as fridge may have more than one
type of trend signatures. Also, some trends are not only found in one type of
appliance but usually in several types of appliances due to their common electrical
characteristics.

50
Chapter 3

Since continuous power points are being measured from smart meter, trend
signatures can be further represented and defined by several slope
(P / t ) variation modes described in Table 3-4. Also, these slope modes can be

used as a scanning method for identification purpose.

Table 3-4 Trend signatures and slope characteristics

Trend signatures Slope characteristics

Rising spike A large negative slope
following a larger positive slope
Gradual falling Continuous small negative
slopes
Falling spike A large positive slope following
a large negative slope
Pulses Continuous pairs of large slopes
Flat Continuous small slopes
Fluctuation Continuous small slopes; signs
of slopes slowly change
High frequency noise Continuous small slopes; signs
of slopes quickly change

3.5 Time/Duration Signatures

The time of load window appearance relates closely to its function. There are
some statistical studies on residential load modeling which present typical load
on-hours such as shown in Figure 3.22 [31],[97].

Microwave

Kettle

Toaster

Fridge

Lighting

PC

0 2 4 6 8 10 12 14 16 18 20 22 24 hr

Figure 3.22: Typical appliances on-hours for weekends

51
Chapter 3

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 each use, it
might be on from minutes to hours. Based on statistical survey, some typical load
window lengths are given in Table 3-5.

Table 3-5 Typical load window Lengths

Load name Min Max length

length

Fridge(cycle) >10 mins <40 mins

Freezer(cycle) >10 mins <40 mins

Furnace(cycle) >5 mins <30 mins

Stove >5 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

3.6 Phase connection Signatures

There are two 120V hot wires installed in a typical North American residential
house as shown in Figure 3.23. 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

52
Chapter 3

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 that can be detected by both CTs. For most 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 have uncertain phase signatures such as laptop.

Meter side
Phase-A

CT-A
Phase-B

CT-B
Neutral

Kitchen Bedroom Stove

Light Light

Figure 3.23: North America residential wiring

3.7 Summary

This chapter proposed a novel load model--- event-window model. Compared

with the existing single-state and transient load models, the proposed event-
window model can accurately depict the process of complex loads such as
continuous varying loads and multi-stage loads which in fact take up a large
portion of end-user loads in power system.

Based on the proposed load models, five categories of signatures were

discussed in detail. Different events’ signatures and event pattern signatures were
included so that multi-stage appliance operation can be accurately described. In
the meanwhile, power variation which can be often seen in continuous-varying

53
Chapter 3

loads was explained. Besides, based on the behavior statistics of different types of
appliances, time/duration signatures were presented. Finally, phase signature was
introduced to label the electric “location” of the load in a North American
residential house.

The proposed event-window model along with the above load signature
studies established a solid foundation for proposed event-window based load
identification method that will be discussed in Chapter 4.

54
Chapter 4

Chapter 4

Event-window based Load Identification

This chapter presents the details of event-window based load identification

method. Chapter 3 has discussed all types of event signatures. How to achieve
load identification based on these signatures is therefore the main problem this
chapter attempts to solve.

Firstly, an overview of the existing NILM algorithms is given. Secondly,

theoretical details such as the proposed event-window load identification
procedure and calculation of individual signature scores are elaborately explained.
Thirdly, the implementation of a practical NILM system in a real residential
house is presented. The details related to the practical side are such as data
acquisition and preprocessing also addressed. Fourthly, the proposed algorithm is
verified extensively based on the real data from several local houses and a public
dataset. Finally, to further clarify the advantages of the proposed algorithm, a
previous signal-combination based method is also implemented using neural
network technique. With a bottom-up based simulation, comparative studies of
the proposed method against the signal-combination based method are conducted.
The findings are discussed and concluded in the end. Some contents in this
chapter have been submitted as publication in [104].

4.1 Overview

Load identification is the most essential task for NILM algorithm. Researchers
have attempted to solve the problem using different algorithms. In [55], the
authors suggest dividing these algorithms into two categories---signal-
combination based algorithm and change of appliance state or event based
algorithm. This chapter intends to introduce a new category---event-window
based algorithm. The structure of NILM algorithms is shown in Figure 4.1. The

55
Chapter 4

existing signal-combination based and event based algorithms are respectively

reviewed. Based on the reviews and identified problems, event-window based
algorithm is proposed eventually.

Figure 4.1: The structure of NILM algorithms based on [94]

4.1.1 Review of signal-combination based algorithms

In [94], the first major research stream is signal-combination based algorithms

[59]-[60],[67]-[79]. This stream looks for a combination of appliances that the
resultant aggregate signal is as close to the observed signal as possible. In other
words, different appliance combinations are matched simultaneously to the
disaggregated signal and the best match is selected as the desired combination.
Typically, it is achieved through the prior training process on a limited number of
combinations first. Afterwards, it is believed the trained model can also support
future classification on untrained combinations.

One example is shown in Figure 4.2, in an enclosed system with limited and
fixed number of appliances, specific appliances are turned on and their aggregated
signals at the meter-side are respectively recorded and labeled. In other words,
training sets can be generated this way and each training set is relevant to a
specific combination. Since each appliance has an ON and OFF state (if not
considering the middle states), for a system that consists of N appliances, there
are in total 2 N combinations. It is therefore impossible to try all the possible of
combinations when N is big. However, limited number of tries may be enough

56
Chapter 4

already to set up the classification model. Then the classification model should be
able to predict future combinations.

Fridge Stove Microwave Meter side

4 15 2.5 50

2 40
3
10

5
1.5

1
Generate 30

20
1 0.5
10
0 0 0
0
-0.5
-1
-5 -10
-1
-2
-1.5 -20

-3
-10
-2

-2.5
Training set -30

-4 -15 0 50 100 150 200 250 300 -40

0 50 100 150 200 250 300 0 50 100 150 200 250 300 0 50 100 150 200 250 300

20
Freezer 1
Computer 0.5
TV 50
Meter side
15 0.8 40

10
0.6

0.4
0

-0.5
Generate 30

5 20
0.2
-1
0 10
0
-5 -1.5 0
-0.2
-10
-0.4 -2
-10

-15

-20
-0.6

-0.8
-2.5 Training set -20

-30

-25 -1 -3
0 50 100 150 200 250 300 0 50 100 150 200 250 300 0 50 100 150 200 250 300 -40
0 50 100 150 200 250 300

……

Figure 4.2: Example of the training process for signal-combination based

algorithm
Typical pattern recognition approach such as neural network and support
vector machine can be used. [69] uses the real and imaginary parts of harmonic
contents shown as below as training inputs and a single-hidden-layer MLP model
for training.

Figure 4.3: Training inputs from [69]

To enhance the accuracy, [78] further combines other features such as
surrounding conditions (temperature, humidity and so on) with various electric
features together into training process. The entire training process becomes very
complicated. The features are shown in Figure 4.4 below:

57
Chapter 4

Figure 4.4: Training inputs from [78]

Some optimization based algorithms such as least residue, integer
programming [72], and genetic algorithms [73] can also be used together with
classifiers such as neural networks. An example is shown in [59]. As can be seen,
different features such as current waveform (CW), Eigenvalues (EIG) are fed in
using either least residue or neural networks into the “Committee Decision
Mechanism” (CDM).

Figure 4.5: Training inputs from [59]

In practice, there are several fatal drawbacks for this signal-combination based
stream:

58
Chapter 4

1. Since this stream directly deals with aggregated or combined signal, all
loads in the system have to be covered. For example, in a 20-appliance house,
even if only 3 appliances are interested to be identified, the other 17 appliances
still need to participate in the training process because they can significantly
affect the aggregated signal when they are turned on together with the interested 3.

2. Due to the number of appliances and the number of possible combinations,

to obtain a reliable classification model, very extensive training process is needed.
For example in [70], the basic training set includes 300 combinations for only 9
loads. This is not very practical for ordinary household owners to perform.

3. The obtained model is very vulnerable to inventory change. If the user

changes the load inventory, for example after he purchases a new appliance or
replaces his old appliance, the previous trained model becomes not reliable and
the training process has to be redone. It further makes the algorithm less practical.

4. The algorithm is not likely to be able to deal with multi-state or continuous

varying loads. If different states of appliances are all taken into consideration, the
training process will be over-complicated due to the large number of
combinations. For most of the existing researches, only a single state is
considered for training. This will lead to large error when dealing with a system
that owns complex loads.

Generally speaking, the signal-combination based methods only work well for
system consisting of ON/OFF appliances. In the meanwhile, the studied system
should have a fixed load inventory. Besides, the users have to be knowledgeable
and patient enough to go through such a long initial process. In order to more
clearly reveal the above drawbacks, a signal-combination based algorithm based
on [69] is implemented and compared with the proposed method in Section 4.5.

4.1.2 Review of event based algorithm

According to [94], the other algorithm stream is based on the change of

appliance operation state or event. The first NILM publication done by MIT [55]

59
Chapter 4

is based on this approach. As shown in Figure 4.6, only real power and reactive
power are considered as the event signatures due to the hardware limit and
processing speed at that time. When an event is detected, it is compared to the P-
Q map and identified as the name of the closest “circle” in the map.

Figure 4.6: Event based algorithm from [55]

As can be seen from Figure 4.6, some appliances have very close signatures
and overlap “circles” on the P-Q map. [61]-[64],[87] have attempted to enhance
the separation by including transient characteristics into consideration. As shown
in Figure 4.7, different sections of turn-on transient are enveloped and used as
signatures for event match (identification).

Figure 4.7: Event based algorithm from [87]

60
Chapter 4

Agreed with the comments given in [94], this research stream is more robust
and practical. The event based algorithms only focus on interested target loads
and do not need to cover all the loads in the system. It can also resist the change
of load inventory. However, it has the following drawbacks that need to be
improved:

1. The existing event-based algorithm looks into “isolated” events at different

time points but not all the events of a load as a whole. When the number of events
increases, with only isolated pieces of information, errors or mis-identification
can easily occur.

2. Since events are separately identified, the whole operation process and
energy of load cannot be re-organized and tracked. In [62], it is claimed that only
operating schedule of loads can be obtained. Also, it cannot effectively deal with
multi-state and continuous varying loads.

4.1.3 Proposed event-window based algorithm

To solve the problems of the above algorithms, the event-window based

algorithm is proposed as a new attempt. Instead of looking into the aggregated
signal or separated events, it looks into the related events of a certain load as a
whole. In other words, it takes the entire operation process of a load into
consideration.

Its procedure is shown in Figure 4.8. The events of interested loads are
detected, identified and finally re-organized as output. Firstly, meter-side current
and voltage are acquired and transferred to event detector. Related data
acquisition will be explained in Section 4.3.1. Secondly, as explained in Chapter 2,
event detector detects all the events of all the loads. Thirdly, through specific data
preprocessing, real power, active power, harmonic of all events are retrieved and
organized. Also trend signatures are scanned along the power curves. This will be
addressed in Section 4.3.2. The next steps are the core of load identification
procedure--- through signal split, window candidate selection and evaluation,

61
Chapter 4

decision of identification can finally be made with respect to the prior collected
appliance signatures. These signatures are used as benchmarks and they are
named as appliance candidates. The details of this core step will be discussed in
Section 4.2. Finally, output results are organized and displayed on user interface.
This will also be presented in Section 4.3.3.

Meter-side signal

Event
detector

Data
preprocessing

Appliance Signal split

by phases * Phase signature
candidate
* Interested appliance
database
name

Window candidate * Length signature

* Phase signature Appliance selection
* Interested appliance candidates
name
* Edge signature
Signature * Sequence signature
similarity * Curve signature
scoring * Time signature

Appliance
identifed Decision making

Output

Figure 4.8: General Identification procedure

4.2 Event-window based algorithm

4.2.1 Load identification procedure

A. Split signal by phases

62
Chapter 4

Two CTs inside meter naturally separate the aggregated signal acquired by
smart meter into signals of two phases: phase-A signal and phase-B signal.
Accordingly, to deal with phase-A signal, only phase-A loads will remain as
candidates. So is for phase-B signal. Normally, a light bulb connected to Phase B
will never be seen from CT-A. This is a very important step because many
candidates can be ruled out very easily.

Two exceptions should be addressed: for phase A-B appliance, since any of its
edges shows up simultaneously at both phases, it will be treated 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 Figure 4.9.

Power/w
380
350
2
270
3

150
1
4
50

0 10 12 20
Time/min

Figure 4.9: A section of meter signal collected from CT-A

Table 4-1Window candidates vs. appliance candidates (1)

Appliance Window Window Window Window

candidate candidate candidate candidate candidate
1-3 2-4 1-4
2-3
Kettle
Fridge

63
Chapter 4

Light
Furnace

Firstly, applying power slope based event detection 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 edges is considered as a window candidate. In Fig.9,
there are in total 4 window candidates: 1-3, 2-4, 1-4, 2-3. Those window
candidates are waited to be compared with appliance 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 cause too much computing burden. One way to
reduce the window number is to discard some window candidates by using
appliances’ possible window lengths.

According to Table 4-1, 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 thrown away even before they enter into next evaluation step.
In fact, this window length limit has stronger effect on refining longer period data.
For a day period, only 120-200 window candidates will be left based on multi-
case studies.

C. Evaluate window candidate with respect to appliance candidate

This is the core step of identification. In this step, the collected signatures of
appliance candidates in prior are set as initial benchmarks. Then each of the left
window candidates will be compared to these benchmarks and obtain their
individual scores on event, event pattern, trend and time signatures
( Sevt , S ptn , Strd , Stime ). Then those 4 scores will be synthetically considered to get an
overall score. Afterwards, the overall score is used to judge if this window
candidate is matching the benchmark---the appliance candidate. The mathematical
judgment is completed by using the following linear equation.

64
Chapter 4

g ( x)   T x   (4.1)

with

 Sevt   evt 
S   
x   ptn  ,    ptn  (4.2)
 Strd   trd 
   
 Stime  time 

x includes the scores of each signature with respect to the prior benchmarks.
The determination on individual scores will be elaborated in Section 4.2.2.  is
the weight vector since for different types of appliances, different signature sub-
scores are not equally important. T x is therefore the overall score.  is the
qualification threshold. The result g(x) has two scenarios: when g ( x)  0 , this
window candidate is determined as this appliance; otherwise not.

Overall, the above process is equivalent to a linear discrimination classifier.

This classifier is a two-class classifier. Each specific load has its independent
weights and classifier. Its classifier judges if a certain event window matches
itself or not. Table 4-2 lists up some typical  and  values of several appliances.

Table 4-2 Examples of load  and 

Load name Distinctive T 

signatures
Fridge event, trend [0.53 0.17 0.22 0.08] 0.85
Microwave event, time [0.65 0.19 0 0.16 ] 0.85
Furnace event, pattern [0.53 0.47 0 0] 0.85
Stove event, [0.51 0.3 0 0.19] 0.85
pattern, time
Washer event, pattern [0.55 0.45 0 0] 0.8
Kettle event [0.86 0.14 0 0] 0.85
Laptop event, trend [0.5 0.25 0.25 0] 0.8
Average --- [0.59 0.28 0.07 0.06] 0.85

65
Chapter 4

The weights  are firstly estimated based on the observation and analysis for
different appliance types. For example, knowing furnace and washer have unique
event pattern signatures, S ptn will be emphasized; knowing microwave is often

used before meals, Stime is emphasized. Generally, event signature is always

important since it determines the electric characteristics of a window. Event
pattern signature is important too, especially for multi-state appliances. Trend
signature is important for motor related and some electronic appliances which
contain transient characteristics. Time signature functions accessorily 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. This is
elaborated in Section 4.2.3.

Usually weight vector  does not change much for the same type of loads from
a house to another. This is because the same type of loads normally share very
similar working principles and follow a certain standard in a particular region.
The row “Average” in Table 4-2 gives a rough setting when the load type is not
known. This can be used to cope with unfamiliar loads whose weights have not be
studied and optimized. The advantage of the weight vector is that when
comparing, there is no absolutely dominant signature---various signatures are
bonded together to ensure fairness and accuracy.

Identification threshold  is normally set as 0.8-0.85 for most cases. It should

be noted all the weights and thresholds can still be adjusted manually and locally
for a specific house in a small scale such as ±0.15 to achieve the optimum
accuracy. For example,  can be lowered if imposed signal noise is significant.
Weights can also be tested and adjusted accordingly. This is also explained in the
end of Section 4.2.3.

D. Make decision

66
Chapter 4

In the end, Table 4-3 is calculated according to equation (4.1) and the g ( x)
values are filled in Table 4-3.

Table 4-3 Windows candidate vs. Appliances candidate (2)

Appliance Window Window Window Window

candidate candidate candidate candidate candidate
1-3 2-4 1-4
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 events could be due to mis-match of events or an unknown
appliance not registered in the database yet (maybe not interested by users either).

This linear classifier can also be substituted by more sophisticated classifiers

such as neural networks or decision tree. Those variations are not discussed here.

4.2.2 Individual signature scoring

This section addresses on how to calculate individual signature scores

Sevt , S ptn , Strd , Stime according to the signature benchmarks from the appliance
candidates.

A. Event signature score Sevt

From the signature database, an appliance candidate only includes its own P-
Q-W event sets. In contrast, a window candidate may include other events caused
by overlapped appliances. The comparison is trying to answer if this window
candidate includes all of the appliance candidate’s events. Thus the process is like

67
Chapter 4

this: each of registered events will be compared throughout all the events in
window candidate one by one. Then:

Ne'
Sevt  (4.3)
Ne

where Ne is number of event types defined in appliance candidate and N e' is

recognized number of appliance event types in window candidate.

As shown in Figure 4.10, both of the two registered events B-D are found in
the window ( Ne  Ne'  2 However, if only B exists, it is very likely the window

candidate is only one part of the appliance process and its Sevt  0.5 ( Ne  2, Ne'  1 ).

Power/w Power/w

380
350
B Ne’=2
Ne=2
270
C

A Candidate
150 110 Appliance
D
Candidate Window 80
50 B D
0 10 12 20 0 10
Time/min Time/min

Figure 4.10: Event signature scoring

P, Q can be easily compared since they are quantitative values. As for current
waveform W, one can conduct comparison in either time-domain or frequency
domain. Selecting proper harmonic orders can also eliminate the impact from
noises and dc offset.

Since an event 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 different appliance types. Synthesizing them together, two events can
be determined as identical or non-identical.

68
Chapter 4

Overall, Sevt indicates the existence of events of appliance candidate in window

candidate.

B. Sequence signature score S seq

For ON/OFF type appliance, it has a fixed sequence of events; for multi-state
appliance, as discussed in section II, fixed sequence and repetitive sequence may
either be found.

For fixed sequence events, they always follow a certain order pattern. For a
window candidate, its event order should comply with the order pattern defined in
appliance candidate. For example, as shown in Figure 4.11, a space heater has 5
events 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 Power/w

300
C A:+100
D 250
B:+200
B C A
D B C: -50
B D D:-150
C E:-100
Window Window
A candidate 1
E candidate 2 E 100
Appliance
A candidate E

Time Time

Figure 4.11: Sequences of two window candidates 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 sequence 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 4-4.

Table 4-4 Example of position change

Window Position Length of

69
Chapter 4

candidate change of changed

letters position
A-B-D-C-E C:3à4 |4-3|+|3-4|=2
D:4à3
B-C-A-D-E A: 1à3 |3-1|+|1-
B: 2à1 2|+|3-2|=4
C: 3à2
C-B-A-D-E A: 1à3 |3-1|+|1-3|=4
C: 3à1

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/events, the
maximum possible length of changed position is:

L n 1
M   [n  (2k  1)] , L  (4.4)
k 0 2

From (5), it can be calculated that:

for ON/OFF appliance, n =2,M=2 (ABàBA);

for three-event appliance, n =3, M=4 (ABCàCBA);

for four-event appliance, n =4,M=8 (ABCDàDCBA) ;

for five-event appliance, n =5,M=12(ABCDEàEDCBA).

Based on the discussion above, Strd for appliance with fixed sequence can be

quantified as

Nf
Strd  1  (4.5)
M

where N f is the length of changed position of a window candidate as shown in

Table 4-4.

70
Chapter 4

For example, sequence C-B-A-D-E’s Strd  0.67 ( N f  4 ) while sequence E-

D-C-B-A’s Strd  0 since it is completely opposite to the original sequence A-B-C-

D-E ( N f  12 ).

One exception is if Sevt is already found smaller than 1, Sseqf will be

automatically set to zero due to mismatch in the number of relevant events.

The appearances of repetitive events are also counted in the step of

determining Sevt and only if its number is more than one, it is recognized as an

repetitive event.

N r'
Strdr  (4.6)
Nr

where Nr is number of repetitive event types defined in appliance candidate and

N r' is recognized number of repetitive event types in window candidate.

As for a combination sequence load such as furnace, Strd can be decided based
on its two sub-indices fixed pattern score Strdf and repetitive pattern score Strdr .

C. Trend signature score Strd

As discussed in Section 3.4, power slope based scanning can effectively scan
the window candidate and further determine the existences of trend signatures
with respect to appliance candidate.

Nt'
Strd  (4.7)
Nt

where Nt is the number of trend signature types defined in appliance candidate

and N t' is the recognized number of trend signature types in window candidate.

D. Time signature score Stime

71
Chapter 4

In the end, the moment of appearance of window candidate t is also

compared with time signature defined for 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}. The judgment is made by equation (4.8).

1, t  T
Stime   (4.8)
0, t  T

4.2.3 Optimization of weights

As explained in Section 4.2.1, for different loads, their weights on event, event
pattern, trend and time are different. In addition, their events also have different
weights on real power, reactive power and harmonic contents. Rough ranges of
weights are pre-defined based on the experience and understanding of the
electrical/physical attributes of different loads, but their values should be further
fine-tuned and verified through a specific program. Once the weights are
determined, they do not need to be changed from one house to another. This
advantage also implies no local training process is necessary when the proposed
NILM is applied to a new system.

To better explain, the above process is compared to the “face recognition”

feature supported by many digital cameras. The “face recognition” feature can
automatically recognize human faces from different photo backgrounds. Similar
to the identification of a certain type of load, face recognition also have internal
weights on different facial attributes such as color/shape of eyes, nose, mouth and
so on. However, when taking a picture, no matter the photographed person is
standing with whatever objects such as animals, cars and other people, the
weights of facial attributes do not need to be re-adjusted. The reason behind this is
that the “face recognition” weights have been already optimized by the
manufacturer with different photo backgrounds and thus the optimized weights
are hardcoded into the cameras before they are sold to customers.

72
Chapter 4

Similar to the “face recognition” process, for a specific load, this optimization
program generates numerous pre-labeled event-windows based on the signatures
of this load and also other types of loads that can mimic different “backgrounds”
(other appliances). Besides, certain amount of signal noises (<10%) are added
onto these windows. By doing this, the large number of pre-labeled testing event-
windows includes not only the actual objective loads but also the other loads.

According to equation (4.9), different set of  and  values will result

different judgments ---correction identification and false identification. If the most
appropriate values of  and  for this specific load can be found, maximum
correct identification rate can be achieved. The optimization goal can be described
as below:

max f ( ,  )   correct match ( T x  )   mis match ( T x  )


 s.t.0    1 (4.9)
 0   1


The process of the above optimization is based on empirical initial ranges and
method of exhaustive search. Since the accuracy needed for  and  is not
extremely high, for each try, the step change is set as 0.02. By using this simple
optimization method, optimal  and  values can be found for each load type.

On top of the above automatic process, another kind of optimization that is

based on the local data collected from the target house can be further applied to
fine-tune the values. This can be understood as a local calibration process. For a
specific appliance, if its activities have been labeled or manually identified for a
short period such as a single day, its weights can be further increased or decreased
by values within a range of ±0.15 according to the labels, until a maximum
identification rate is achieved. This testing procedure is convenient because for
most of the cases, weights will not leave far apart from the results obtained from
(4.9).

73
Chapter 4

4.3 System implementation

To validate the proposed event-window based algorithm, practical system has

been implemented and taken into field test. This section will discuss the key
components of the system and the feasible solutions. The data flowchart of this
NILM system is shown as below. Current and voltage signals at different phases
are acquired; after data preprocessing, event matrix and trend scan table will be
generated and fed into the identification module; finally, reports are generated and
pushed up to the specific user interface for display.

Data acquisition Data preprocessing

Event-based User Interface
3 event matrices; identification
Current and
Voltage 3 trend scan Event windows Report
table

Figure 4.12: Data flow chart of the NILM system

4.3.1 Data acquisition

Figure 4.13: Data acquisition system at the meter-side

74
Chapter 4

The setup of the data acquisition system is shown in the above picture:
Electricity panel of a residential house is first opened; two current transducers are
respectively connected to two phases. For each phase, the current transducer is
clamped around its live wire; also a voltage transducer is connected between one
of the live wires and the neutral wire; currents and voltages are being transferred
to the National Instrument (NI)-DAQ system. The NI-DAQ system supports
simultaneous inputs from multiple channels and has a high-resolution A/D
converter in it; finally, digital signals of voltage and current are sent to the
connected laptop via a USB port. Generally speaking, this data-acquisition system
behaves like a smart meter.

The data is acquired at per second basis. For each second, a snapshot
composed of 6 cycles of synchronous currents and voltage are recorded. For each
cycle, 256 points are acquired. Example of a snapshot is shown in Figure 4.14.
3
Phase A
Current(A) 2
1

-1

-2

-3
0 200 400 600 800 1000 1200 1400 1600

Phase B
6

Current(A)4
2

-2

-4

-6
0 200 400 600 800 1000 1200 1400 1600

Voltage(V) 200
150

100

50

-50

-100

-150

-200
0 200 400 600 800 1000 1200 1400 1600

Sampling point(256 pts/cycle)

Figure 4.14: Example of a data snapshot

Besides, similar NI-DAQ based acquisition system is applied to acquire load
signatures. The setup is shown as below: only one current transducer is used and

75
Chapter 4

connected to the live or neutral wire of the load supply cable. The voltage
transducer is connected to a vacant electric outlet adjacent to the outlet the load is
plugged into. The data format and sampling rate stay the same as the meter-side.

Figure 4.15: Data acquisition of load signatures

4.3.2 Data preprocessing

The amount of data acquired from the previous step is usually very large and
not convenient for processing by the identification module. Hence, a data
preprocessing step that can compress the data but keep the important information
becomes necessary. The purpose of this step is to generate event matrix and trend
scan table.

A. Generation of event matrix

Table 4-5 Example of event matrix

Event Start End Rising/Falling Real Reactive Fundamental
No. sec sec Power Power (Var) Current
(W)
1 6 7 2 28.7 0.9 0.3
2 38 40 … … … …
… 3rd 5th 7th
harmonic Harmonic Harmonic current
current Current
… 0.1 0.0 0.1
… … … …

76
Chapter 4

An example of event matrix is shown in Table 4-5: event No., time points,
rising or falling type and electric signatures of all the events are calculated and
recorded for the data of a given period such as a day. As discussed in Chapter 2,
events can be divided into 3 groups: phase-A events, phase-B events and phase A-
B events or double phase events. Accordingly, 3 event matrices are generated
respectively for the above three groups. Later on, in the identification step,
different event matrices can be selected according to the phase signature of the
objective load.

To improve the signature’s accuracy, electric signatures are calculated based

on the subtraction of the average of three consecutive points after a certain event
and the average of three consecutive points before this event. For example, a data
series have t1-t10 ten data points and a load state change or an event happens at t4-
t5. In this case, the electric signatures of this event are calculated using the
following equations:

P  ave( Pt 5 , Pt 6 , Pt 7 )  ave( Pt 2 , Pt 3 , Pt 4 )


Q  ave(Qt 5 , Qt 6 , Qt 7 )  ave(Qt 2 , Qt 3 , Qt 4 ) (4.10)
 h
I  ave( I t 5 , I t 6 , I t 7 )  ave( I t 2 , I t 3 , I t 4 )
h h h h h h

It should also be especially noted that since only 6-cycles snapshots are
acquired within each second, P,Q and Ih from different snapshots have to be
aligned up before subtraction. For alignment, voltage can be used as reference
since the phase angle of voltage stays constant before and after a load state
transition. Two approaches are proposed for the calculations of P,Q and Ih. The
snapshot shown in Figure 4.14 is used as an example.

 I 1  I 1 ( I1  V1 )

 I h  I h ( h  h   1 )
 I V
 1
V  V V
1 1
(4.11)

 P  I V cos(V   I )
1 1 1 1


Q  I 1 V 1 sin(V1   I1 )

77
Chapter 4

Approach 1: Apply Fourier transform to the 6 cycles of two-channel current

and one channel-voltage. Then use equation (4.11) to calculate P,Q and Ih.

Figure 4.16: Approach 2 for P,Q and Ih calculation

Approach 2: Voltage is assumed to be completely sinusoidal. The first step is
to locate the first maximum point of voltage from its 6 cycles. Then only the five
left cycles after this point are considered for calculation. This is shown in Figure
4.16. By doing this, currents from different snapshots are aligned with respect to
the voltage. In other words, V1 in equation (4.11) is set to 0 degree. Therefore,
the above equation can be simplified as (4.12) :

Compared with Approach 1, Approach 2 is much faster since there is only one
time Fourier transform on 5-cycle currents. However, it is less accurate since not

78
Chapter 4

all cycles are fully utilized and the assumption that voltage has no harmonic
content is not always true.

 I 1  I 1  I1

 I h  I h  h
 I
 (4.12)
 P  I V cos( I )
1 1


Q  I 1 V sin( I1 )

B. Generation of trend scan table

The above 3-phase event matrices have included event, event pattern,
time/duration and phase connection signatures. However, as discussed in Chapter
3, trend signatures should also be fed into load identification module for judgment.
The other task in the data preprocessing stage is to generate a trend scan table
shown as below:

Table 4-6 Example of trend matrix

Window Rising Falling Pulses Fluctuation Quick Gradual Flat
No. spike spike vibrate Falling
3-4 T F F T F T F
4-5 F F F F F F T
…-… … … … … … … …

As can be seen from Table 4-6, for a given window, different trend slope
signatures are recognized according to the slope characteristics listed in Table 3-4.
The above process is conducted throughout all types of phase connections for the
studied period. Finally, 3 trend scan tables that each represents one type of phase
connection can be generated.

4.3.3 User interface

After identification decisions are making from the event-based identification

module, activity report along with energy estimation results are generated and
displayed on a graphic interface shown as below.

79
Chapter 4

The results are updated every half an hour and displayed in the interface shown
in Figure 4.17. 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
(say,7.4¢/KWH). 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 use chart,
user can understand his identified appliance activities 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. More discussion can be found in Chapter 5.

Figure 4.17: Appliance energy decomposer software

4.4 Verification using real house data

The above algorithms and devices were tested in several real residential houses
in Edmonton, Canada for several weeks. A laptop based data acquisition system
was hooked to the electricity panel and behaved 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
the proposed algorithms was launched and started to process. Interested appliance

80
Chapter 4

activities were identified. To verify the identification rates, appliance activities

under track were either recorded manually or labeled through human experts’
inference for comparison.

In addition, verification was conducted based on a public dataset provided by

MIT [95]. In this dataset, several appliances were labeled using extra data-logging
devices that were directly connected with objective appliances. The results and
discussions are respectively shown in the following sections.

4.4.1 Verification based on House #1

Table 4-7 Identification rate accuracy for house #1(7 days)

Appliance Actual Correctly False Identification

Name operation identification identification accuracy(%)
times times times
Fridge 312 253 0 81
Microwave 54 50 0 93
Washer 5 5 0 100
Cooktop 1 1 0 100
Stove elements 16 16 0 100
(low power)
Stove elements 18 18 0 100
(high power)
Kettle 20 17 0 85
Dryer 5 5 0 100
Heater 248 223 0 90
Waffle iron 13 13 1 92
Average --- --- --- 94

The results for house #1 are listed in Table 4-7 (this house has no furnace). The
definitions of all the table columns are explained as below:

 Actual operation times (AOT): the actual observed and inferred operation
times by manual identification.

 Correctly identification times (CIT): the correctly detected times by the

algorithm that matches the actual operation times. It should be noted, for

81
Chapter 4

some action-intensive appliances such as stove, washer and dryer, if their

detected events are close to each other (say shorter than 30 mins), there
should be combined together as one event. It is because those appliances
usually have repetitive events that actually come from single time
operation.

 False identification times (FIT): Times misidentified by the algorithm

which cannot match the actual window observed or inferred manually.

 Identification accuracy: Defined as

CIT  FIT
Accuracy  100% (4.13)
AOT

Therefore, using equation (4.13), false identified time has a negative impact on
identification accuracy and thus the index can be more objective. In order to more
clearly present the results and discussions, examples of identified event-windows
are automatically marked in red by the algorithm and some failures are circled in
green. These examples are shown as below:

1. Fridge (Identification accuracy: 81%)

(a) Success identification

82
Chapter 4

(b) Failure (green circle)

Figure 4.18: Examples of identification for fridge

Most events of fridge can be easily identified. Some failures are due to the
overlapping with pulse-based appliance such as a stove. One example is shown as
the green circle part in Figure 4.18 (b). Its ON event is ruined by the “needles” of
stove which cannot be detected by the event detector any more. However based
on human inference, this could highly possible be a real fridge event. As shown in
Figure 4.18 (b), under this circumstance, many fridges can still be correctly
identified, which shows the robustness of the proposed algorithm.

2. Microwave (Identification accuracy: 93 %)

(a) Success identification

83
Chapter 4

(b) Failure

Figure 4.19:Examples of identification for microwave

As can be seen from Figure 4.19, (a) shows the example of successful
identifications; (b) shows a failed case. The reason for failure is that the OFF
event of this microwave operation overlaps with another appliance.

3. Washer (Identification accuracy: 100%)

Figure 4.20: Examples of identification for washer

4. Dryer (Identification accuracy: 100%)

84
Chapter 4

Figure 4.21: Examples of identification for dryer

5. Stove elements using low power (Identification accuracy: 100%)

Figure 4.22: Examples of identification for stove elements using low power

6. Stove elements using high power (Identification accuracy: 100%)

Figure 4.23: Examples of identification for stove elements using high power

85
Chapter 4

7. Cooktop (Identification accuracy: 100%)

Figure 4.24: Examples of identification for coffee maker

8. Kettle (Identification accuracy: 85%)

(a) Success identification

(b) Failure
Figure 4.25: Examples of identification for kettle

86
Chapter 4

Similar to the fridge, sometimes when the events of kettle overlap with other
frequent appliances, identification failure may occur due to event detection failure.

9. Heater (Identification accuracy: 90%)

Figure 4.26: Examples of identification for heater

10. Waffle iron (Identification accuracy: 92%)

Figure 4.27: Examples of identification for waffle iron

4.4.2 Verification based on House #2

Table 4-8 Identification rate accuracy for house #2 (8 days)

Actual Correctly False identified Identification

Appliance Name operation identified operation times accuracy(%)
times operation times
Freezer 576 523 0 91
Fridge 417 371 0 89
Furnace 28 27 0 96

87
Chapter 4

Microwave 72 71 0 99
Kettle 14 14 1 93
Washer 4 4 0 100
Dryer 5 5 0 100
Stove elements 8 8 0 100
(low power)
Stove elements 5 5 0 100
(high power)
Average --- --- --- 96

The results are shown in Table 4-8. As a multi-state appliance, the

identification example of furnace is shown as below:

(a) Zommed-out view of furnace identfication

(b) Zommed-in view of furnace identification

Figure 4.28: Examples of identification for furnace

As can be seen, even if the furnace’s operations sometimes overlap with the
fridge’s operations, identification is almost not affected. This is because fridge is

88
Chapter 4

different from stove---although it has frequent operations throughout a day, it

does not generate high-frequency pulses and it has a much lower possibility to
cause an event overlap problem.

Also, it is found the proposed algorithm can withstand certain level of signal
noises such as the figure shown as below. As can be seen, even if the steady state
of two later fridge operations have significant noises, they can still be correctly
identified as long as their events ON/OFF are not heavily polluted.

Figure 4.29: Examples of fridge identification under noisy condition

Figure 4.30: Examples of freezer identification when operations overlap with

other appliances

89
Chapter 4

Figure 4.31: Example of microwave identification when it overlaps with fridge

In addition, it is also found that the proposed algorithm is resistant to the
overlap of different types of appliances. In Figure 4.30, the first freezer’s
operation does not overlap with other appliances. However, the three freezer’s
operations behind it all overlap with some other appliances. But they are still
correctly identified just like the first one. Another example is shown in Figure
4.31. As can be seen, the microwave’s operation can still be identified even if it
overlaps with the fridge’s operation.

4.4.3 Verification based on House #3

Table 4-9 Identification rate accuracy for house #3 (7 days)

Appliance Actual Correctly False Identification

Name operation identification identification accuracy(%)
times times times
Fridge 25 22 0 88
Furnace 60 55 0 92
Microwave 40 36 1 90
Washer 15 15 0 100
Dryer 8 8 0 100
Stove 14 13 0 93
TV 8 7 1 75
Kettle 23 23 1 96
Average --- --- --- 92

The results are shown in Table 4-9. It is found TV has a lower identification
rate. This is because TV has a flexible duration. The duration can be very long

90
Chapter 4

such as hours but can also be shorter than half an hour. The identification rate is
therefore reduced since many more window candidates will be included which
lead to potential interference.

Figure 4.32: Examples of identification for TV

4.4.4 Verification based on public dataset

REDD, a public dataset available for NILM research was released by MIT in
2011 [95]. The data was acquired from the greater Boston area in US. The dataset
can be used to validate the NILM algorithm since several appliances have been
labeled using extra data-logging devices that are directly connected to the target
appliances. The No.3 house from the dataset is selected for the validation purpose.
In general, the dataset contains 1427284 seconds or roughly 16 days and the
dataset is not continuous. Many appliances are pre-labeled. For some unknown
reason, the washer is not pre-labeled by MIT, however, it is labeled by us through
realistic inference with respect to dryer activities. Several major appliances have
been identified and the results have been listed in Table 4-10.

91
Chapter 4

Table 4-10 Identification rate accuracy for house #4 (7 days)

Appliance Actual Correctly False Identification

Name operation identification identification accuracy(%)
times times times
Fridge 661 628 5 94
Furnace 34 32 1 91
Microwave 95 95 4 96
Washer 7 6 0 86
Dryer 13 13 0 100
Dishwasher 5 5 0 100
Electronics 10 10 0 100
Average --- --- --- 95

4.4.5 Observations and findings

Overall, the algorithm has excellent performance in terms of identification rate.

It is capable for the following cases:

 Dealing with all types of appliances including single-state appliances

(microwave, kettle and waffle iron etc.), continuous-varying appliances
(fridge and freezer) and multi-state appliances (furnace, washer and
dryer). The average identification rates in the four houses are all above
90%;

 When the load window duration is not changing significantly or not too
long;

 Even when the operations of different appliances overlap;

 Even when there is a certain level of signal noise.

However, the performance can be affected on the following cases:

 If the appliance has a very long and flexible duration, it can be

troublesome such as the TV in house #3;

92
Chapter 4

 If a certain appliance can generate high-frequency pulses, it might

affect the identification of other appliances because the pulses may
cause the event overlap problems.

4.5 Comparative studies with neural networks based method

4.5.1 Implementation of neural networks based method

This chapter also presents a detailed comparison between the proposed

solution and previous signal combination based solutions such as the one
discussed in [69]-[70].According to [69], a two-layer feed-forward network is
adopted here for comparison.

Firstly, specific appliances were measured in the lab and their harmonic
signatures were collected. Not like the proposed 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 (only odd ones). Hidden
nodes are set to be 20. As shown in Figure 4.2, according to [69], numerous
training sets are generated mathematically 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 magnitude as noises.

For test stage, a bottom-up based aggregating program turns on/off each load
according to a certain operation probability. The aggregated 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.

93
Chapter 4

4.5.2 Simulation based verification

Table 4-11 Comparison for only ON/OFF type loads

Loads Identification accuracy(%)

NN based Proposed
approach 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

Table 4-12 Comparison with complex loads

Loads Identification accuracy(%)

NN based Proposed
approach approach
Microwave 92.3 99.9
Monitor 93.6 94.9
TV 84.4 94.2
Vacuum 85.0 96.3
Monitor 79.8 97.7
Incandescent light bulb 97.5 98.5
Fluorescent light bulb 95.6 99.2
Fridge 63.7 97.9
Freezer 68.5 95.3
Washer 73.4 97.1
Furnace 57.2 98.4

Comparison is firstly conducted when there are only single-state type loads.
This is because single-state type loads only have a steady-state harmonic content.
The results are shown in Table 4-11.

94
Chapter 4

As can be seen, for a system composed of only single-state type loads, the
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 complex loads are brought in.

As can be seen from Table 4-12, NN based approach is significantly affected

by the introduction of multi-stage loads (furnace and washer) and continuous
varying loads (fridge and freezer). This drawback is actually discussed in [69] due
to the lack of a steady-state harmonic content in those appliances. Their harmonic
contents can vary tremendously with time. Sometimes, harmonic contents of
different operational stages of the same load are not even 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 introduce not
only large error to those complex loads themselves 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 the error from fridge. In contrast, the 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 events.

Table 4-13 Comparison when stove is not trained or registered.

Loads Identification accuracy (%)

NN based Proposed
approach approach
Microwave 78.0 94.5
Monitor 77.8 94.3
TV 76.6 94.2
Vacuum 65.2 96.1
Monitor 95.8 95.8
Incandescent light 64.1 95.1

95
Chapter 4

bulb
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 --- ---
Another obvious advantage of proposed approach is it only identifies loads that
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 neural network model needs to be changed due
to the newly added element. As shown in Table 4-13, any aggregated signal that
has stove on at the same time will become unidentifiable (this is especially severe
for other kitchen appliances). However, stove hardly has any impact to the
identification of registered loads using proposed approach because the proposed
approach is based on searching relevant events of registered appliances only.
Those non-relevant events from stove will be excluded from the window
candidates of the registered loads.

4.5.3 Observations and findings

To summarize, compared to the signal-combination based approach, the

proposed approach has the following obvious advantages:

 Window based signatures make identification of complex loads

possible. In contrast, signal-combination based approaches such as NN
cannot efficiently identify those loads;

 Composition of appliances is not only judged by an independent point

of meter-side signal but also events before and after this point. Hence,
the association of load states is much more strengthened. A load’s OFF

96
Chapter 4

event can only be confirmed if its ON event is found within a time

window;

 Training process does not need to cover all major appliances 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.

4.6 Summary

Both the theoretical and practical sides of event-window based load

identification method have been discussed in this chapter. The overview and
classification of existing NILM algorithms, load identification procedure,
signature score calculation, weight optimization, system implementation,
extensive real house based validations and comparative studies with previous
signal-combination based approach were all addressed in details.

Generally, the proposed event-window based algorithm has the following

advantages:

 It can effectively deal with complex loads including continuous-

varying and multi-state loads;

 It does not require a local training process since the weights of different
signatures for different loads can be optimized in the lab before taking
to the field. Only simple optimization process is needed;

 It can resist the change of load inventory;

 The average identification rates in the four tested houses are all above
90%. The overall accuracy is very satisfactory;

 It can resist a certain level of signal noises;

 It can deal with the operation overlaps of different loads.

97
Chapter 4

However, the proposed algorithm does have some disadvantages for the
following scenarios:

 If the appliance has a very long and flexible duration, the number of
window candidates will increase dramatically which could lead to
larger identification error.

 Sometimes load events can be ruined by the events of other loads,

especially when the other loads generate continuous high-frequency
pulses such as stove and washer. If the events are ruined, they cannot
be correctly detected by the event detector and this will further also
lead to identification error.

Overall speaking, the proposed algorithm makes a good balance between being
effective and being practical.

98
Chapter 5

Chapter 5

Non-intrusive Signature Extraction for Major Residential

Loads

This chapter presents a technique to extract load signatures non-intrusively by

using the smart meter data. Load signature extraction is different from load
activity identification. It is a new and important problem to solve for non-
intrusive load monitoring (NILM) applications.

For a target appliance whose signatures are to be extracted, the proposed

technique first selects the candidate events that are likely to be associated with the
appliance by using generic signatures and an event filtration step. It then applies a
clustering algorithm to identify the authentic events of this appliance. In the third
step, the operation cycles of appliances are estimated using association algorithms.
Finally, the electric signatures are extracted from these operation cycles.

The results can have various applications. One is to create signature databases
for the NILM applications. Another is for load condition monitoring. Validation
results conducted on the data collected from three actual houses and a laboratory
experiment have shown that the proposed method is a promising solution to the
problem of non-intrusive load signature collection. Some contents in this chapter
have been submitted as publication in [105].

5.1 Overview

5.1.1 Review of existing intrusive signature extraction methods

To perform load decomposition, all NILM techniques must rely on the unique
signatures of individual loads. In order to obtain these signatures, the existing

99
Chapter 5

NILM researches generally require measurement steps [55]-[83], which are

supervised and intrusive.

For example, in [69], signatures such as harmonic contents of all appliances in

the house have to be collected in advance. Only after this, training sets can be
generated based on the combinations of above signature collections. In [81], it is
proposed that three power meters are used to acquire the signatures of a single
load at a time. The hardware devices are shown as below:

Figure 5.1: Three power meters based signature extraction

In [82] and [50], electric/magnetic field sensor based event detectors are
developed to label the events of a single appliance. Then electric signatures of the
appliance can be extracted by comparing its labeled events with the metering side
aggregated data. One example is shown in Figure 5.2 [50] to explain its working
principle: the current of an overhead fan changes when its operation state changes.
And the current will cause the changes of both electric field and further magnetic
field. The EMF detector is a piece of PCB board composed of signal amplifier
circuits. It can process both the electric/magnetic field signals in real-time. Based
on the distinction between the electric and magnetic fields, events of the overhead
fan can be labeled such as b and c. When the time points of b and c are located on
the aggregated meter-side signal, electric signatures of the overhead fan can be
extracted for NILM applications.

100
Chapter 5

Figure 5.2: Relations of magnetic field, electric field and the EMF event detector
In [83], a smart phone based signature extraction system is developed. It
requires manual confirmations from human experts or house residents to label the
events of appliances. The process is shown in Figure 5.3.

Figure 5.3: Smart phone and human confirmation based signature extraction
system

5.1.2 Proposed intrusive event-window signature extraction system

As explained in Chapter 3 and Chapter 4, an event-window consists of various

signatures. In order to acquire these signatures, we firstly propose to create a
small signature database tailed for each home utilizing an intrusive piece of
hardware called appliance register.

As shown in Figure 5.4, an appliance register is a device installed between the

appliance to be registered and the electric outlet the appliance is supposed to be
plugged in. 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

101
Chapter 5

meter (or the device which does appliance identification).Smart meter does two
things: capture the event window of this appliance and determine the signatures
of event window.

Figure 5.4: Intrusive event-window signature extraction system

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. Event signatures can be
directly extracted. Event pattern signature can be determined based on the
appearance number of event types. Trend signature can be detected based on
slope modes explained in Section 3.4. In the end, time/duration signatures are
automatically selected after the appliance is named by users. After waiting for one
or two operating cycles of the appliance, all signatures of the appliance will have
been collected. The register is then moved to another appliance interested for
registration. This approach has another advantage in term of privacy: the customer
can control which appliances are to be registered for identification.

5.1.3 Proposed non-intrusive signature extraction method

In fact, the above signature extraction processes discussed in 5.3.1 and 5.1.2
have the following common disadvantages:

1. They are intrusive methods. They all require additional hardware devices to
either directly measure load signatures or label load events. Fundamentally, this

102
Chapter 5

intrusive step is opposite to the basic principle and intention of non-intrusive load
monitoring research. In other words, it makes NILM not purely non-intrusive.

2. They are supervisory methods. No matter it is based on measurement or

labeling, human efforts are required to be involved. This is fairly inconvenient
and less practical for ordinary household residents to perform. This disadvantage
can significantly affect the wide application of NILM.

3. They are hardware based methods. Extra cost will be added into the
solutions. Unfortunately, so far, there is no other pure-software based approach on
signature extraction for NILM.

It is clear that the above intrusive ways of signature collection are not desirable
by ordinary house owners. Therefore, there is a need for methods that can collect
the appliance signatures non-intrusively. If successful, unique signatures that are
specific to an appliance in a particular home can be extracted and archived. To
solve the above drawbacks, this chapter proposes a novel unsupervised non-
intrusive signatures extraction (NISE) approach which does not require any
measurement or direct input from users and is purely software based. Given the
meter-side data of a certain days, specific major appliance events can be located,
associated and then their complete operation cycles can be reconstructed
automatically. For the proposed NISE method in this chapter, the key is to
intelligently study these events and establish knowledge of appliances
automatically.

As shown in Figure 5.5, firstly, all appliance events are captured through the
event detection method discussed in Chapter 2. Secondly, event filtration is
adopted to identify the suspect events of interested appliances. Thirdly, event
clustering is utilized as a tool to pick out the appliance’s authentic events from its
suspect events. Fourthly, based on the authentic events, event association
algorithm is further applied to reconstruct complete operation cycles of appliances.
Finally, both electric and event pattern signatures of appliances can be extracted
from the reconstructed cycles to fit any existing signal combination based NILM

103
Chapter 5

approach [69]-[78] and event-window based NILM approach [104]. In the

following sections of chapter, event filtration, clustering and association are
discussed in detail respectively. Also, the proposed approach is validated using 3
real houses and laboratory data. The accuracy of signatures extracted is compared
and analyzed. Conclusions are presented in the end.

Proposed
Data flow
flowchart
Start Data

Edge detection Events

Event filtration Suspect events

Event clustering Authentic events

Representative
Event Association
operation cycle

Knowledge
End
(signatures)

NILM

Figure 5.5: Flowchart of proposed approach versus corresponding data flow

5.2 Event Filtration

One operation cycle of each appliance may have two or more events depending
on if it is an ON/OFF type appliance or a multi-state appliance. An ON/OFF type
appliance such as a light bulb has only a pair of ON and OFF events while multi-
state appliance such as furnace may have a series of operation state changes in the
middle. An appliance’s events are heavily characterized by its function and
physical electric attributes and can be roughly located. For example, a fridge
usually has an ON event with real power of 70-300 W and a reactive power of 30
to 200 Var. Due to the cooling function of a fridge, the ON events can be

104
Chapter 5

observed during 24 hours even when users are sleeping; A microwave usually has
an ON event that has an real power of 800 to 2500 W and a heavy third harmonic
content. As a cooking device, a microwave is likely to be observed before
mealtime.

In this chapter, the events that match the specific conditions of a certain
appliance are defined as the “suspect events” of this appliance. The aim of event
filtration is to locate the suspect events of a given appliance that may lead to the
reconstruction of its entire operation cycle by carrying out future steps. To
implement filtration, the conditions which can restrict the captured events are the:
real power range, reactive power range harmonic content range, with or without
spike, single phase or double-phase and searching time.

The real power, reactive power and harmonic content ranges are closely
determined by the electric attributes of specific appliances. Residential loads can
be roughly divided into four categories based on their linearity and reactivity:
linear/active appliance, linear/reactive appliance, non-linear/active appliance and
non-linear/reactive appliance. Depending on the type of category a certain
appliance belongs to and its designed function, the numeric ranges of the above
conditions can be quantified. For example, as a linear/active appliance, a resistive
kettle has a very low reactive power and almost zero harmonic contents. Also, due
to its water-heating purpose, its real power may statistically range from 1300 to
3000 W. Motor based appliances such as fridges can be viewed as inductors
which lead to large reactive power. Another category is the switch-mode power
supply based electronic appliances such as TVs and computers. They are neither
inductive nor capacitive, but produce a large amount of harmonic contents. In
addition, some appliances are both non-linear and reactive. Table 5-1 lists some
examples of the above four appliance categories.

105
Chapter 5

Table 5-1 Appliance categories and examples

Appliance category Examples

linear/active Stove, Kettle, Toaster
linear/reactive Fridge, Freezer, Furnace, Dryer
Non-linear/active TV, Desktop PC, Laptop
Non-linear/reactive Washer, Microwave

A Spike can be another ancillary condition that helps locate some induction-
motor based appliance events. A large inrush current occurs at the first moment of
operation when the rotor is triggered from the station into movement. This unique
feature is accompanied by an ON event, which can be reflected as a sharp edge on
the appliance’s real power curve.

The phase condition can easily separate some events from other appliances’
events. In North America, some appliances are connected between two hot phases
to gain a 240V voltage while others are connected between a single hot phase and
a neutral to gain a 120V voltage. From identification perspective, the events of
double-phase appliances can be observed simultaneously at both hot phases,
whereas the events of single-phase appliances occur at only one of them. In a
residential house, only a few heavy-consumption appliances are connected to
double-phase electric outlets. These appliances are the stoves, ovens and clothes
dryers.

The search window is another very important restriction that greatly reduces
the searching space of suspect events. Some statistical studies are available on
residential load behaviors which present typical appliance runtimes [97].For
example, microwave’s operations are more expected to be seen before breakfast,
lunch and supper; lights are usually turned on in the early morning or after dark;
fridges and furnaces are likely to run day and night. These occurrences show that,
to locate the suspect events of specific appliances easily, it is best to search their
less overlapped time ranges. For example, for a fridge, it is best to search from
2:00 AM-5:00 AM while many other appliances are generally inactive. For

106
Chapter 5

microwave, it’s best to search the periods before meals. By doing so, the
interference from other appliances can also be minimized.

In addition, to achieve the knowledge discovery purpose through massive

amount of data, sufficient event samples must be obtained. Thus the data for
multiple days need to be provided. According to the search window conditions,
for different appliances, data pieces are cut from multiple days and re-joined
together as a large data piece as shown in Figure 5.6. It is expected that the
objective appliances’ events will have a much higher density in the joint data
pieces.

Hours

Mon. 14-20 20-24

0-5 5-9 9-14

Tues. 14-20 20-24

0-5 5-9 9-14

Weds. 14-20 20-24

0-5 5-9 9-14
.
.
.
.
.
.

Day
Kettle’s joint data piece

Figure 5.6: Example of data piece connection for kettle

Table 5-2 Example OF ON-Event Filtration Condition Table

With Phase
Appliance P(W) Q(Var) THD (%)
spike? Condition
Fridge 70-300 30-200 0-20 Yes Single
Furnace 120-800 200-800 0-20 Yes Single
Microwave 800-2500 80-500 20-50 No Single
Stove (big
1800-3000 0-30 0-5 No Double
element)

Stove (small 1000-2000 0-30 0-5 No Double

107
Chapter 5

element)
Oven 2200-3600 0-30 0-5 No Double
Kettle 1300-3000 0-30 0-5 No Single
Clothe dryer 3000-6000 60-250 0-5 Yes Double
Washer
80-300 <100 65-95 Yes Single
(Front-load)
Washer
300-1000 300-1200 0-20 Yes Single
(Top-load)

Based on the above discussion, Table 5-2 presents an ON-Event filtration

condition table for 10 major appliances. The listed P/Q/THD values can be
understood as the generic ranges of these appliances in the geographic area of our
research location (Edmonton, Canada). The signatures are based on the
measurements of different brand/models of appliances from several residential
houses. At least 4 appliances of the same kind were measured. It should be noted
that measuring all models/brands is impossible due to our limited resources.
However, to compensate, the signature ranges are all expanded by at least 20%
from our measurement values. For example, during the measurement, the power
range of microwave was 1200-2000W, and in Table 5-2, it is modified to 800-
2500W in order to be more inclusive. Since the electric attributes of appliances
are essentially determined by their functions, they will generally fall into the
above ranges. However, in different countries or regions, different electricity
voltage levels, climates and even cultures may affect the above signature ranges.
Thus, a more local ON-Event filtration table can be updated according to the local
measurements.

Appliances may have different working modes. For example, a stove usually
has 4 heating elements (two big ones and two small ones) on its panel. The small
elements of a stove consume only approximately half the power of the big
elements. Since their signatures are quite different, they should be treated as two
different types of appliances.

108
Chapter 5

It should also be noted that some appliances may have different working
principles which result very different signatures. For example, most modern front-
load washers are controlled by a variable-speed drive which outputs significant
harmonic contents. In contrast, an old top-load washer behaves more like a
regular big motor which outputs higher P/ Q and little THD. These two washers
should also be treated as two different types of appliances.

In Table 5-2, P and Q are calculated based on the fundamental components of

the current and voltage so that heavy harmonic contents will not affect their
values [99]. For harmonic contents, the total harmonic distortion (THD) of current
is used. Only the odd orders of the current harmonic contents(ik) are considered
due to their significance and usually a larger order than 9 does not need to be
considered [99]-[100].

THD 
 ik2
, k  3,5, 7...
i1 (5.1）

According to the conditions listed in Table I, the suspect ON events of an

appliance can then be identified by inspecting the events one by one in its joint
data piece. To improve processing efficiency, only the ON events are verified and
then considered as potential suspect events after passing inspection, because many
transient features such as spikes accompany only ON events.

In order to guarantee that the events of an appliance can be included, the

filtration conditions should not be set too strictly. In other words, the suspect
events located by using the conditions in Table I may not be able to exclude the
non-relevant events caused by other appliances; however, the conditions must
allow enough of the events caused by the appliance to be included, because for a
specific appliance, without enough event samples, the following procedures will
fail or lead to inaccurate results.

109
Chapter 5

5.3 Event Clustering

5.3.1 Definition of event clustering

As discussed in Section II, the suspect events of appliance X are a group of

ON-events that are possibly but not necessarily caused by X. Since the conditions
are defined as loose ranges and the data for multiple days are used, other events
which do not belong to X can be easily included in the suspect events
occasionally. The aim of event clustering is to determine which suspect events are
the real ones belonging to X. In this chapter, the real events are named as the
“authentic events” of X.

One basic assumption is that inside a suspect event group of appliance X, the
number of events belonging to appliance X should be much larger than the
number of events belonging to other appliances. The following examples below
explain this assumption: suppose 50 suspect events of a fridge are located from a
data piece for 2:00-5:00 AM of a whole week. It is possible that the user wakes up
between 2:00-5:00 AM on Tuesday and for an unknown reason, uses an appliance
such as a ceiling fan that accidentally meets the same filtration conditions as those
of the fridge. Thus this fan’s event will be mis-recognized as the fridge’s suspect
event and included in the fridge’s 50-suspect-event group. This scenario might
occur occasionally on certain days; however, this scenario is highly unlikely to
occur as frequently as the fridge since such operations would be abnormal---he
would have to turn on/off the fan not only every day but also as frequently as the
fridge’s kickins during sleeping time. Another example is a microwave. The user
might turn on/off an unknown appliance that draws similar P/Q/Harmonic signals
as a microwave does before mealtime, but it is unlikely to occur as often as a
microwave. For any of these abnormal scenarios, the proposed unsupervised
approach is not intended for and not able to deal with.

Furthermore, in some cases, corrupted events may occur that are usually due to
simultaneous occurrences of more than one appliance event. This rarely happens

110
Chapter 5

but is possible. For example, the accidental overlapping of two single events
belonging to a fridge and furnace will result in a power jump that is roughly the
summation of the two events. Normally, the frequency of these corrupted events
is much smaller than the authentic events.

Based on the above assumption, if the number of groups of events and the
number of events each group possesses are known, the authentic events can be
determined as those in the event group with the maximum number of members.
Clustering is an effective tool to obtain the information. A clustering algorithm
determines which events are roughly the same and can be grouped together as one
cluster. After event filtration, since the clustering space is greatly reduced, many
uncertainties and noises can be ruled out and clustering can be done much faster
and more accurately compared to applying clustering to all the events of a day in
the beginning. For instance, as Table 5-3 shows, suppose 88 suspect events of a
fridge are located and after applying clustering algorithm, only four clusters are
found. Then from the number of events each cluster owns, the dominant cluster
which has 75 event members can be identified as having been caused by the
fridge and these 75 event members can be labeled as the fridge’s authentic events.
On the other hand, smaller clusters are discarded because they are likely to be
from other appliances or noises and cannot represent the objective appliance. The
clusters and their mean values are listed in Table 5-3. Two types of clustering
methods are applicable to our study.

Table 5-3 Example of composition of suspect events

Mean Mean Mean Number Physical

Cluster
Q of
Index P (W) THD(%) cause
(Var) Events
1 100.3 76.2 10.6% 75 Fridge
2 87.7 67.9 10.3% 10 Fan
3 73.6 58.8 2.2% 2 Motor X
4 189.6 138.5 9.9% 1 Corrupted

111
Chapter 5

5.3.2 Selection of clustering method

Our proposed approach selects the mean-shift algorithm for clustering. Before
introducing mean-shift clustering, it is necessary to explain another related
clustering method named K-means clustering. Comparison of the two will also be
given afterwards.

(a) (b)

(c) (d)

Figure 5.7: Example of K-means algorithm

K-means clustering is a classic method of clustering analysis that can partition

n samples into k clusters in which each sample belongs to the cluster with the
nearest mean. It is firstly proposed in [106]. An illustration of its standard
algorithm is shown as follows:

112
Chapter 5

As shown in Figure 5.7, before the procedure begins, the number of clusters K
must be initialized. In this example, K is set as three. Accordingly, in step (a),
three “means” are randomly chosen from the given dataset. In (b), three clusters
are created by grouping every sample with its nearest mean. In (c), based on the
three clusters that are determined from (b), centroids of the three clusters are
calculated and the three old means are now “shifted” towards the three new
centroids. In the end, three new clusters are formed in (d). After this, (b) and (c)
are repeated until means are not shifted (convergence has been reached).

The above K-means procedure is straightforward. However, it is difficult to be

applied to our application---load event clustering. The reason is K-means
algorithm normally requires the number of clusters K as a priori that is actually
the number of load event types in the objective house. However, as a non-
intrusive approach, the number of loads in the house is unknown. To cope with
this challenge, another clustering approach---mean-shift clustering is adopted.
Mean-shift clustering is a density based clustering algorithm. It has been widely
applied to the areas of computer vision and pattern recognition, such as
intelligence fusion, target following, and image segmentation (Figure 5.8) [107]-
[108]. In [109], the most widely used mean-shift clustering was proposed.

The comparison of K-means and mean-shift clustering in general can be found

in [110]. It has been mentioned that one advantage of mean-shift over K-means is
that there is no need to know and choose the number of clusters before clustering
starts. Instead, mean-shift requires the bandwidth parameter that restricts the
variance of samples belonging to a certain cluster with respect to the cluster mean.
For residential appliance loads, the variation of electric signatures is determined
by the system voltage fluctuation level and the variation range can be pre-
estimated. In the earliest publication of NILM [55], this phenomenon has already
been explained in detail. In other words, the bandwidth required by mean-shift
can be easily obtained while the number of clusters required by k-means is
difficult. [110] also mentioned mean-shift might be slower than K-means but this

113
Chapter 5

is not a strong concern because non-intrusive signature extraction is an off-line

process that does not require fast computation.

Figure 5.8: Example of Mean-shift clustering applied to image segmentation

Furthermore, [111] explained the above difference and specifically discussed

the feasibility of applying mean-shift to residential appliance identification. It
demonstrated a simple attempt of mean-shift algorithm on clustering a few
appliance operations (the "triangle"-shaped operation and "rectangle"-shaped
operations) in a residential house. The trail in this paper is not a complete
approach but has proven the feasibility of mean-shift clustering applying to NILM
related applications.

For mean-shift clustering, assuming the sample data is a finite set S in the n-
dimensional Euclidean space, X, its kernel function is shown as below:

114
Chapter 5

1, x  
K ( x)   (5.2)
0, x  

The mean of the cluster is

m( x ) 
 sS K ( s  x)s (5.3)
 sS K (s  x)

The difference m( x)  x is defined as the mean shift. The steps of mean-shift

clustering algorithm are listed as below:

Step 1: Initialize random samples as cluster centers;

Step 2: Within the ranges defined by a specific radius or bandwidth, calculate

the means of each cluster ;

Step 3: For all the clusters, shift the cluster centers towards the new means
calculated in step 2;

Step 4: Repeat step 2 and 3 until the centers of all the clusters are not moving
anymore. In other words, the means of all the clusters are not updated anymore;

Step 5: Merge similar clusters that have close means together.

In our study, P/Q/THD are selected as three attributes used for clustering
purpose. The “mean” linkage type is selected to determine the distance between
two clusters. All features are firstly normalized to the range of 1 to 100 through
min-max scaling [113]. For mean-shift clustering, its bandwidth parameter is set
from 5-20 due to the estimated variation of the electric signatures caused by the
local system voltage fluctuations. Generally, for the above two clustering methods,
when the power variance inside the suspect event group is large, the bandwidth of
mean-shift clustering can be relatively lowered in order to strengthen the cluster
differentiation; when the power variance inside the suspect event group is small,
the bandwidth of mean-shift clustering can be relatively increased to strengthen
the cluster fusion.

115
Chapter 5

5.3.3 Feature selection for mean-shift clustering

The above standard mean-shift clustering has one drawback due to the
Euclidean space it uses where different attributes will be treated as equally
important. However, as explained in Table 5-1, depending on which category the
appliance type belongs to, different electric features should not be treated equally
because some features may be almost noise features. For example, for a fridge, as
a linear/reactive appliance, it produces considerable active/reactive powers. But it
barely produces any harmonic content and thus the THD feature is actually a
noise feature that might be caused by misreading from the meter-side; for a stove,
both of its reactive power and harmonic contents should be disabled because it is
basically a resistor.

To cope with the above noise feature problem, P/Q/THD features can be easily
selected according to Table 5-1. This is because for the proposed non-intrusive
signature extraction algorithm, the type of appliance is already known in advance
and the selection of features can be therefore determined based on the category
the appliance type falls into. The effect of feature selection is similar to using a
Minkowski metric weighted space explained in [112] where the “relative
distance” on low-weight features contributes little to the total distance.

To make a comparison, both of the mean-shift clustering methods without

feature selection and with feature selection were applied to the 87 suspects
discussed in Table 5-3. They consist of 75 events from a fridge (1), 10 from a
ceiling fan (2), 2 from a motor X (3) and 1 corrupted event (4). The clustering
results are shown as P-Q 2D plots in Figure 5.9. It shows that the mean-shift
clustering method without feature selection mistakenly includes 2 fans’ events in
its authentic group (the fridge’s event group) because both the fridge and fan have
similar THD values. However, because of the fridge’s linear/reactive load type, its
harmonic content should not contribute because it is usually small and instable,
and can easily become mixed up with other linear appliances; in contrast, after
feature selection, the THD feature is ruled out and the result is able to exclude the

116
Chapter 5

2 events caused by the fan. Generally, mean-shift with feature selection has a
better performance than mean-shift clustering in the proposed application.

(a) Mean-shift clustering without feature selection

(b) Mean-shift clustering with feature selection

Figure 5.9: Effect of feature selection

117
Chapter 5

5.4 Event Association

After the event clustering, the knowledge of the appliance is still incomplete
since the other events except for the ON event in an appliance’s operation cycle
are still unknown. For accurate energy-tracking purposes, all the events and even
the event pattern signatures [104] of an appliance have to be obtained. For
example, for a fridge, its OFF event needs to be known; for multi-stage appliance
such as a furnace, it is even more important since the ON event is only a small
part of the furnace’s full operation cycle. Its middle events and OFF event also
need to be found. Knowing how these middle stage events occur in a particular
pattern is also valuable. The overall purpose of event association is to determine
the other events which also belong to the objective appliance and hence
reconstruct its full representative operation cycles. .

The proposed method counts the number of event occurrences and calculates
the frequency to judge whether an event is associated with a specific appliance
and determines the event’s association type. An operation study of actual
appliances [104] defines three association types:

1. Single event: It has a strong association with an appliance. Once the

appliance is working, this event must appear. However, it appears only once in
single operation cycle. As Figure 5.10 shows, a fridge has a fixed OFF event and
it accompanies every operation cycle of the fridge but appears only once.

Figure 5.10: Example of single events.

118
Chapter 5

2. Repetitive event: It has a strong association with an appliance. Once the

appliance is working, this event must appear. It can appear more than once in a
single operation cycle. As shows in Figure 5.11, a furnace has repetitive events
caused by its heating elements, which can be triggered multiple times according
to the environment temperature.

Figure 5.11: Example of repetitive events

3. Occasional event: It has a weak association with an appliance. Once the
appliance is working, this event might not occur. As shown in Figure 5.12, when
the door of a fridge is open, the light inside the chamber will be automatically
switched on; when the door is closed, light will be switched off. Then users can
see a small pair of power jumps between ON and OFF. An occasional event is
usually caused by an ancillary element of appliance, such as the light of a fridge
and the hood of a stove.

Figure 5.12: Example of occasional events

119
Chapter 5

4. Unrelated event: It has zero association with an appliance because it is

caused by another appliance. As shows in Figure 5.13, during a fridge’s operation,
a big incandescent lamp is also turned on. The ON event of the lamp happens
before the OFF event of the fridge and the OFF event of the lamp happens after
the OFF event of fridge.

Figure 5.13: Example of unrelated events

Table 5-4 Average duration and data segment length for typical appliances

Load name Average Data

duration segment
length
Fridge(cycle) 15 mins 22.5 mins
Furnace(cycle) 20 mins 30 mins
Microwave 4 mins 6 mins
Stove 25 mins 37.5 mins
Kettle 4 mins 6 mins
Oven 10 mins 15 mins
Washer 45 mins 67.5 mins
Clothes dryer 50 mins 75 mins

The way to locate associated events and determine their association types can
be determined by examining the data segments starting from the authentic ON
events (determined by the event clustering in the last step). Each data segment is
defined as a piece of data that has an authentic ON event in its beginning. The

120
Chapter 5

length of the piece is properly defined so it will not be too short (may fail to
include all the associated events of the appliance) while it will also not be too
long (may include too many unrelated events as interferences). The reasonable
lengths can be defined as 1.5 times the average durations of the appliance
operation. Table 5-4 shows the average durations of typical appliance operation
and their segment lengths.

Suppose there are M authentic events, there will also be M corresponding data
segments which may include not only associated events of appliance but other
irrelevant events. Applying event clustering methods discussed in Section III to
all events in all segments together, clusters of events can be determined.
According to the definitions of three association types, theoretically, the criteria in
Table 5-5 can be used to examine association of those clusters individually.

Table 5-5 Theoretical criteria for association determination

Association type No. of events the cluster include

(N)

Single event N=M

Repetitive event N>M

Occasional event bM<=N<M

Irrelevant event N<bM (0<b<1)

For single event, they should appear in each data segment and thus event
number N should be equal to data segment number M; for repetitive event, since
in some data segments, multiple events may appear, event number N should be
larger than M; for occasional event, its occurrence should be more frequent than a
threshold ruled by b. b is a factor to differentiate if an event is related to objective
appliance or not. Considering possible mistaken clustering or missing events due
to improper data segmentation, the above theoretical criteria should be corrected
using confidence c and additional conditions shown in Table 5-6. Normally,
values of b,c can be set as 0.3 and 0.8.

121
Chapter 5

Table 5-6 Criteria for association determination

Association type No. of events the cluster include (N)

Single event cM<=N<=M
Repetitive event N>=cM & n>=2 in at least one certain segment
Occasional event bM<=N<cM
Unrelated event N<bM (0<b<1)

An example of how event association works is shown in Figure 5.14 and Table
5-7. For the given appliance, ON-event “1” is the initial authentic event
determined by the event clustering step previously. Accordingly, 4 data segments
of this appliance are located and each segment has the same length determined
referring to Table 5-4. At the beginning of event association, clustering is applied
again to all the events included in these segments. As a result, each event can be
labeled by the index of the cluster it individually belongs to. After the occurrences
of each cluster are counted, the association type can be determined by using
criteria in Table 5-6. The results are presented in Table 5-7 .

Power/w Power/w

Data segment #1 Data segment #2

3 4 3 4 6 3 4

1 2 2
1

Time Time
Power/w Power/w

Data segment #3 Data segment #4

5
4
3 43 4 3
2
2
1 1

Time Time
Figure 5.14: An example of event association from 4 data segments

122
Chapter 5

In this example, the authentic operation cycle of the given appliance is

determined as 1à à2 (~~~: repetitive). An effective way to improve the

above association accuracy is to have more data segments, which may require
more data pieces from more days, say a week. The more data segments being used,
the less likely an unrelated event is to be incorrectly judged as a related one due to
the error of the confidence factors b and c; also, the impact of a possible
clustering error from the previous step or the missing events due to improper data
segmentation can be reduced to a minimum.

Table 5-7 Example of event association judgment

Item Number of Criteria used Association type

appearances (b=0.3;c=0.8)
Data segment M=4 --- ---
Event 1 N=4 cM<=N<=M Single event
Event 2 N=4 cM<=N<=M Single event
Event 3 N=6 N>=cM & n>=2 Repetitive event
Event 4 N=6 N>=cM & n>=2 Repetitive event
Event 5 N=1 N<bM Unrelated event
Event 6 N=1 N<bM Unrelated event

5.5 Verifications and Discussions

5.5.1 Verification and discussions based on real house #1’s data

The above algorithms were tested by using data acquired from a real
residential house in Edmonton, Canada for a week with no special attention from
the house owner. A laptop-based data acquisition system was hooked the house’s
electricity panel and continuously collected all the voltages and currents from the
two hot phases (A and B) inside. It behaves just like a smart meter. The data were
sampled every second. In each second, 6 consecutive cycles are acquired and each
cycle has 256 points. Data acquired was then processed using the proposed

123
Chapter 5

algorithm and the operation cycles of specific appliances were automatically

learned and reconstructed. In order to compare, operation times of appliance
activities were manually recorded. By doing this, true operation cycles of
appliances could be directly labeled as references. If the labeled events were more
than one, their average values were used for comparison. In this house, there were
3027 events (>50W) in Phase A, 2055 in Phase B and 2012 in Phase A-B. The
total power of both phases on a typical day is shown in Figure 5.15. Table I is
adopted for event filtration. In total, 10 appliances were found and two different
fridges were connected to the two different phases.

Figure 5.15.The total power on a typical day from house #1

Table 5-8 shows the search window and authentic events determined in the
step of event clustering. As can be seen, for each appliance, the number of
authentic events (the largest no. of events in a single cluster) is larger than the
average number of events per cluster. This result supports the assumption in
Section III that authentic events will dominate inside suspect events.

Table 5-8 Search window and results of event clustering

Appliance Search # of suspect # of Largest no. of

window events clusters events in single
cluster
Fridge 2:00AM- 69 10 21 (30.4%)

124
Chapter 5

(PhaseA) 5:00 AM
Fridge 2:00AM- 31 2 7(22.6%)
(Phase B) 5:00 AM
2:00AM- 4 1 4(100%)
Furnace 7:00AM
8:00PM-
0:00AM
6:00AM-
9:00AM;
Micro- 11:00AM- 54 7 27(50%)
Wave 2:00PM;
4:00PM-
8:00PM
Stove(big 4:00PM- 517 9 505(97.7%)
element) 8:00PM (Repetitive)
Stove (small 4:00PM- 77 5 60(77.9%)
element) 8:00PM (Repetitive)
Oven 8:00AM- 7 5 3(42.9%)
8:00PM
6:00AM-
9:00AM;
Kettle 11:00AM- 11 4 8(72.7%)
2:00PM;
4:00PM-
8:00PM
8:00AM- 92 5 61
Clothes dryer 11:00PM (Repetitive) (66.3%)
(Weekend)
Washer 3 hrs before 380 10 189(49.7%)
dryer (Repetitive)
As for the reconstructed operation cycles, the following two verification
methods are conducted:

125
Chapter 5

(a) Furnace

(b) Microwave

(c) Stove(small element)

126
Chapter 5

(d) Stove (big element)

(e) Oven

(g) Kettle

127
Chapter 5

(h) Dryer

(i) Washer

Figure 5.16.Reconstructed cycles (red) vs. Labeled real cycles (blue) in house
#1

1. To visualize the effectiveness of the approach, the reconstructed operation

cycles are compared with the labeled reference cycles in Figure 5.16. For the sake
of convenience, the events determined by using event clustering and association
are aligned with the original events in the labeled cycles. As Figure 5.16 shows,
they are quite similar. The only difference is that some power variations between
neighboring events are not captured, because the proposed approach focuses only
on events. However, this missing information is not important since most NILM
algorithms do not use it. Particularly, if any transient information accompanied

128
Chapter 5

with an event is needed, the corresponding event sample can be simply loaded for
study.

2. An appliance’s reconstructed operation cycle contains the electric signatures

of all its associated events such as P, Q, THD (but not limited to these, depending
on the needs of NILM).

Table 5-9 presents a detailed comparison between the signatures of individual

events in the reconstructed cycles and labeled cycles. To compare, the mean
values of event clusters are adopted. The errors are calculated with respect to the
labeled reference events as true values.

Table 5-9 shows acceptable accuracy between the reconstructed cycles and
labeled cycles. Please note the electric signatures of appliances can change within
+/- 5% due to system voltage fluctuation. As

Table 5-9 shows, most of the real power errors are lower than 5%. Some errors
of reactive power or harmonic THD are a little higher, because the corresponding
appliances produce little reactive power or harmonic contents. The true values of
these attributes are comparable to signal noises and can greatly fluctuate between
measurements, especially when they are small. For example, even if several
measurements are directly performed for the kettle’s harmonic THD, differences
at the listed level may still be observed. Another example is the washer. Since it is
controlled by a variable speed driver, which continuously generates many
repetitive event pairs with heavy noises (as shown in Fig.7), the inherent signature
consistency of these events is lower than that of other appliances. Thus, the above
errors are quite normal. Furthermore, they will not affect NILM’s identification
significantly. Very small or zero weights will be given to these unstable attributes
when making identification [104]. Similarly, classifiers such as neural networks
[69] will “ignore” these signatures automatically during their training stages
because these signatures in their training sets vary a lot too. The mean values
being used can also balance out the variations of abnormal events inside clusters.
Generally speaking, the electrical signature accuracy is very satisfactory.

129
Chapter 5

Table 5-9 Electric signature error between reconstructed cycles and reference
cycles for house #1

Appliance Power Reactive Power Harmonic THD

Event Error % Error % Error %
Fridge(PhaseA) ON 0.54 0.45 -4.41
Fridge(PhaseA) OFF 1.80 -0.39 -0.58
Fridge(PhaseB) ON 3.00 7.31 -4.75
Fridge(PhaseB) OFF -0.01 -0.04 -5.48
Furnace ON 0.73 4.11 1.94
Furnace Middle 1 1.13 -4.91 2.16
Furnace Middle 2 -1.41 1.40 -3.16
Furnace Middle 3 1.11 1.44 6.57
Furnace OFF -0.13 1.03 -2.72
Microwave ON -0.32 0.79 -0.86
Microwave OFF -0.85 -2.72 -2.68
Stove(small) ON 0.52 0.17 2.24
Stove(small) OFF -0.41 0.12 -7.46
Stove(big) ON -1.15 1.93 -0.91
Stove(big) OFF 0.14 -4.70 -0.68
Oven ON 1.21 -0.21 2.98
Oven OFF -0.72 -4.49 -3.08
Kettle ON -2.19 4.46 -5.84
Kettle OFF 0.24 -2.81 12.08
Dryer ON 0.96 0.97 -0.20
Dryer OFF 0.63 -1.41 -4.89
Washer ON -4.21 -0.54 -1.41
Washer OFF -5.31 -3.32 -7.29

Also, the speed of the proposed algorithm was found to be fast. Based on a
regular desktop PC (Intel 2-Quad CPU 2.33GHz, 4GB memory), the 3-stage time
required to process the appliances are recorded. The most time-consuming
appliances “Stove(big element)” and “Washer” in house #1 is listed in Table 5-10.

130
Chapter 5

Table 5-10: 3-stage time required for the most time-consuming appliances

Appliance Event Event Event Total

filtration(s) clustering(s) association(s) time(s)
Stove(big) 0.012243 0.441213 0.030468 0.4839
Washer 0.009009 0.359888 0.027966 0.3969

5.5.2 Verification and discussions based on real house #2’s data

Figure 5.17: The total power data on a typical day from house #2
Similarly to house 1, another real residential house with a smaller family size
in Edmonton was also measured for a week. The house contained 1451 events
(>50W) in Phase A, 1698 ones in Phase B and 564 ones in Phase A-B. The
house’s total power on a typical day is shown in Figure 5.17. It should be noted in
this house an old top-load washer is used. Its reconstructed cycle referring to a
real labeled cycle is shown in Figure 5.18, which shows that between ON and
OFF, its power gradually decreases and becomes stable after a few minutes.

By using the proposed approach, 8 major appliances were found (the house has
only one fridge). As Table 5-11 shows, the electric signature is quite accurate
with respect to the signatures contained in the labeled reference cycles.

131
Chapter 5

Figure 5.18: Reconstructed cycles (red) vs. l with labeled real cycles (blue dash)
for the top -load washer in house #2
Table 5-11 Electric signature error between reconstructed cycles and reference
cycles for house #2

Appliance Power Reactive Power Harmonic THD

Event Error % Error % Error %
Fridge ON 1.84 3.04 -5.87
Fridge ON -1.93 -1.01 3.33
Furnace ON 0.82 -2.03 -5.57
Furnace Middle 1 -1.71 -6.99 1.20
Furnace Middle 2 -4.37 -5.54 -5.39
Furnace Middle 3 1.42 -0.40 3.93
Furnace Middle 4 -0.80 -0.86 -2.55
Furnace OFF -2.73 -2.62 -1.58
Microwave ON -1.24 -8.78 -1.77
Microwave OFF 1.91 -6.46 3.25
Stove(small) ON 4.65 -5.41 2.02
Stove(small) OFF 4.14 -4.78 0.37
Oven ON -1.42 2.88 -5.58
Oven Middle -1.47 -4.60 2.85
Oven OFF 1.04 0.44 1.69
Kettle ON 1.04 0.44 1.69
Kettle OFF 0.41 2.75 -3.71
Dryer ON 1.91 -0.59 -2.99

132
Chapter 5

Dryer OFF 0.97 0.98 3.33

Washer ON 5.32 1.85 3.15
Washer OFF 0.85 0.07 3.57

5.5.3 Verification and discussions based using MIT public dataset

Figure 5.19. The total power data of the first 86400 points from house #3
To further verify the proposed approach, the data from house 3 in the REDD
dataset [95] was used. REDD, a public dataset available for NILM research was
released by MIT in 2011. The data was acquired from the greater Boston area in
US. Please note: 1. In this dataset, many middle hours in a day are missing. For
example, activities of the stove, oven and kettle were not recorded. 2. Although
REDD labels specific appliances but unfortunately, some appliances are not
clearly labeled such as the washer. In general, the dataset contains 1427284
seconds or roughly 16 days. The sampling rate was 275 points/cycle. The house
contains 1160 events (>50W) in Phase A, 2497 in Phase B and 523 in Phase A-B.
As an example, the total power of the first 86400 seconds (the length of a day) is
plotted in Figure 5.19.

In spite of the above issues in the dataset, 5 major appliances can still be
extracted using proposed approach. Surprisingly, even when the washer activities
were not labeled in the REDD dataset, they were still able to be located and
“mined” by using the proposed method. With the help of the proposed method,

133
Chapter 5

the washer’s cycles were re-labeled. Its reconstructed cycles verses real cycles are
plotted in Figure 5.20. Like the washer used in house 1, this is a front-load washer
with a variable-speed drive inside.

Figure 5.20: Reconstructed cycles (red) vs. real cycles (blue dash) for washer in
house #3

As Table 5-12 reveals, the electric accuracy of reconstructed cycles is very

good. It should be noted the filtration conditions in Table I used in this case are
based on our local measurements in Edmonton, Canada. However, it is found still
viable to deal with the house in the Boston area. This finding implies that
appliances may share common signature ranges in similar geographic regions
(such as northern part of North America).

Table 5-12 Electric signature error between reconstructed cycles and reference
cycles for house #3
Appliance Power Reactive Power Harmonic THD
Event Error % Error % Error %
Fridge ON -1.04 -1.23 -4.16
Fridge ON -0.20 -0.55 0.25
Furnace ON -2.36 -2.49 -1.92
Furnace Middle 1 -1.47 -5.46 5.39
Furnace Middle 2 7.21 -1.81 -0.96
Furnace Middle 3 2.46 3.88 -7.12
Furnace OFF 0.61 0.81 -2.68

134
Chapter 5

Microwave ON -0.83 1.95 -1.92

Microwave OFF -0.56 -3.20 -3.36
Dryer ON -1.96 -4.32 3.25
Dryer OFF -0.71 2.06 -1.17
Washer ON 1.08 4.39 1.94
Washer OFF 1.83 -1.69 -8.03

5.5.4 Verification of event association based on laboratory data

To further test the effectiveness of the proposed event association approach

specifically, a space heater was tested in the laboratory. To bring in the
interference from other appliances’ events, a lamp and a microwave were also
connected to the same power supply bar which the space heater was also
connected to. Then the aggregated signal of the power supply bar was measured
under 4 types of scenarios, each for different times:

Scenario 1: The heater ran with only its heating function on; this scenario was
conducted 5 times.

Figure 5.21: Laboratory switching experiment---Scenario 1

Scenario 2: The heater ran with both its heating function on and swaying
function on (constantly changing the wind direction); this scenario was conducted
4 times.

135
Chapter 5

Figure 5.22: Laboratory switching experiment---Scenario 2

Scenario 3: The heater ran with only its heating function on; also, the
microwave was switched on in the middle of the heater’s operation; this scenario
was conducted once.

Figure 5.23: Laboratory switching experiment---Scenario 3

Scenario 4: The heater ran with only its heating function on; also, lamp and
microwave were both switched on in the middle of the heater’s operation; this
scenario was conducted twice.

136
Chapter 5

Figure 5.24: Laboratory switching experiment---Scenario 4

The above 4 scenarios simulate not only occasional events (such as those
from the optional sway function) but also unrelated events caused by other
appliances (in this case, microwave and lamp). In addition, the heating elements
of the heater itself can be triggered repetitively. The purpose of this experiment
was to test the capability of the algorithm to distinguish all types of event
association for a given appliance. Examples of the measured power signals of the
above 4 scenarios are plotted from Figure 5.21 to Figure 5.24. All events are also
marked by different numbers according to their physical causes.

(a) Clustering results of ON events

137
Chapter 5

(b) Clustering results of OFF events

Figure 5.25: Clustering results of all evens in space heater’s 12 segments

Event clustering was applied to the above four scenarios, which included 12
data segments in total. The clustering plots for both the ON and OFF events are
shown in Figure 5.25. The numbers marked around clusters are consistent with
the numbers marked around events in Fig.12. Also, according to the number of
events each cluster owns, event association is judged as shown in Table 5-13. The
table shows that the determined event pattern 1à2à3à4à5(6à7) is
completely consistent with the physical observations of the heater shown from
Figure 5.21 to Figure 5.24 (~~~: repetitive; bracket: occasional event).

Overall, the above verification procedures and discussions show that the
proposed signature extraction approach is capable of providing accurate electric
and event pattern signatures for specific appliances automatically as long as the
amount of feeding data is sufficient(more than a week). With this knowledge,
most non-intrusive load identification algorithms can be adopted without intrusive
measurements for the training or registration stages. In other words, by combining
proposed NISE with a NILM method such as [104], a truly non-intrusive load
monitoring solution can be provided.

138
Chapter 5

Table 5-13 Event association judgment For heater(b=0.3,c=0.8)

Item Number of Criteria used from Event

apperance Table IV association type
Data segment M=12 --- ---
Event 1 N=12 cM<=N<=M Single event
Event 2 N=21 N>=cM & n>=2 Repetitive event
Event 3 N=21 N>=cM & n>=2 Repetitive event
Event 4 N=12 cM<=N<=M Single event
Event 5 N=12 cM<=N<=M Single event
Event 6 N=4 bM<=N<cM Occasional event
Event 7 N=4 bM<=N<cM Occasional event
Event 8 N=3 N<bM Unrelated event
Event 9 N=3 N<bM Unrelated event
Event 10 N=2 N<bM Unrelated event
Event 11 N=3 N<bM Unrelated event

5.6 Summary

This chapter addressed a novel problem related to the NILM research---the

non-intrusive extraction of load signatures. Although most previous NILM studies
addressed the identification or classification of load activities, the non-intrusive
signature extraction of loads haven’t obtained enough research efforts yet. In
reality, an intrusive signature extraction process of loads can significantly impact
the application of NILM to ordinary households. Thus, more research attentions
should be drawn to solve this problem. The proposed approach is an unsupervised
non-intrusive approach which can automatically extract load signatures by using
the meter-side data and requires almost zero effort from users. The intention of
this research was to eliminate or at least reduce the intrusive work load required
by most existing NILM methods. The proposed approach uses event filtration,
clustering and association to locate suspect events, determine authentic events and
associated different events together to reconstruct operation cycle for an objective

139
Chapter 5

appliance, respectively. These reconstructed cycles contain most of the electric

signatures and event pattern signatures of appliances and are enough for existing
NILM approaches for identification and energy tracking purposes. The proposed
approach was verified off-line by using the data acquired from 3 actual residential
houses and a laboratory experiment. Both electric and event pattern signatures
were tested, analyzed and discussed. Generally, the accuracy of the extracted
signatures was satisfactory.

This chapter focused on major appliances. Future research could study the
smaller or unique appliances. Meanwhile, event filtration conditions can be
updated with vast measurements in different geographic regions.

140
Chapter 7

Chapter 6

Energy Estimation of Residential House

This chapter addresses an important application of NILM--- Energy estimation,

specifically for residential loads. According to different types of energy
components in a house, energy estimation methods for ordinary appliances,
specific load group (such as incandescent lights) and background power are
respectively discussed. Based on these methods, energy characteristics of several
houses are analyzed and presented. The meaning and implications of the results
are discussed in the end.

6.1 Overview

The above chapters have systematically discussed the identification of load

activities using NILM. With the identified activities, many applications can be
achieved. Energy estimation is an important area that is worth studying. It can
help the end-users to gain an insight into the usage pattern of residential loads. A
good understanding of the usage pattern may further support the decision making
on energy management. For the utility side, with this information, the
effectiveness of demand response program and Time-Of-Use (TOU) price can
also be analyzed.

The overall residential energy consumption in a residential house can be

divided into 3 components: ordinary appliances, specific load groups and vampire
power:

 Ordinary appliance activities are usually the ones that can be tracked
based on NILM’s identification results. Thus their energy can be
estimated accordingly. This chapter firstly discusses energy estimation
methods for this type of energy component.

141
Chapter 7

 Specific load group is the group of appliances that is better to be dealt

with as a whole together. Examples are the incandescent light group
and fluorescent light group. There are so many lights in a house and it
is thus difficult to label, register and identify each individual light.
However, as a group, since they share some particular characteristics in
common, it is easier to use a specific estimation method to deal with. In
this chapter, an independent method to estimate the incandescent light
is proposed and discussed.

 Background energy refers to the electric energy consumed by

appliances while they are switched off or kept in a standby/always-on
mode. For example, some appliances have remote control and digital
clock features. Also, some small appliances such as wireless router
always stay on in a house and they should also be considered as
background energy. This chapter also proposes the estimation method
for this type of energy.

Finally, based on the above components, energy characteristics of several

residential houses are synthetically analyzed and the findings are also presented.
Their implications of the residential house energy characteristics are also
discussed with respect to the Time-of Use price and load-shift actions.

6.2 Energy estimation methods for ordinary appliances

After the activities of ordinary appliances are identified either manually or by

using NILM algorithms, their energy estimation can be achieved by using three
different ways. They will be respectively discussed.

A. Energy estimation using average energy consumption

This method takes the operations times of appliances and its average energy
consumption into consideration. The calculation is pretty simple but less accurate.
It can be presented in the following formula:

142
Chapter 7

E  nEave (6.1)

Where Eave is the average energy consumption of one time operation and n is

the operation times that are detected. Eave can be averaged based on several
operation samples.

It is a preferred method when:

 The NILM algorithm can only provide the schedule information of

appliances. For example, in [61]-[62], since the NILM identification
can only identify the ON-event transients of appliances, energy cannot
be accurately calculated without knowing the OFF event. This method,
however, can provide some estimation results instead.

 The operation process of appliance is relatively fixed or stable. For

example, some loads such as HVAC fan and refrigerator are
programmed to operate similarly each time. The actual energy
consumed within a period can be very close to the result obtained from
equation (6.1).

B. Energy estimation using average power

For a certain appliance, this method takes the ON/OFF time and the average
power of each operation into consideration. It can be mathematically presented as
below:

N
E   (tnoff  tnon ) Pave (6.2)
n 1

Where tnon and tnoff are the ON/OFF time of a certain time operation; N is the

total operation times for energy estimation; Pave is the average power and it can
be calculated based on several operation samples.

143
Chapter 7

For single-state appliance, Pave can be chosen from the real power signature
of the steady-state or the ON/OFF event; for a continuous varying ON/OFF
appliance such as a fridge, Pave can be approximated as:

( Pon  Poff )
Pave  (6.3)
2

where Pon and Poff are the real power signatures of ON and OFF events.

As for the multi-state appliances, the following universal equation can be used
no matter what type of appliance it is:

N
 En
n 1
Pave  N
(6.4)
 (t  t )
off
n
on
n
n 1

where N is the number of operation samples that are taken into consideration
in order to obtain Pave ; tnon and tnoff are the ON/OFF times of each operation
sample.

Generally, this method has a good performance for all types of appliances as
long as enough samples N are considered in (6.4). However, this method may be
less accurate than the third method as follows.

C. Energy estimation using all window events

Compared to method B, this method uses all the events belonging to a certain
appliance, which includes not only the ON/OFF events but also the middle events.
In addition, this method uses integration of power points to accurately track the
power variations between events. This method requires the time information of all
events inside an event-window and is only applicable to specific NILM methods
such as the proposed event-window based method.

As shown in Figure 6.1, the objective appliance has two related events #2 and
# 4 while #1 and #3 are events coming from other appliances. The energy of this

144
Chapter 7

appliance is therefore the area summation of A1 and A2. In this case, even if the
power curve between event #2 and event #3 is not constant, it can still be
calculated by adding all the acquired power points between event #2 and event #3
together with the 270W base power before event #2 subtracted.

Figure 6.1: Example of energy estimation using all window events

Mathematically, it can be expressed using the following equation:

m1
N M 1 tn
E    [  P(t )  Pref (tnm )(tnm1  tnm )] (6.5)
n 1 m 1 t tnm

Where:

N is the total number of operations that are taken into energy calculation;
M is the number of events that are included in an event-window. In the
example shown in Figure 6.1, it is three;
P(t) is the acquired total power at a given instant t;

t nm is the time instant where event m occurs;

tnm1 is the time instant where event m+1 occurs;

Pref(t) is the reference power level which may change with time. In this
example, at the beginning, it is the initial power before the first event (270W).
Once an unrelated event such as event #3 occurs, Pref(t) will be updated as the

145
Chapter 7

value of initial power plus this event’s power. For example, Pref(10) to Pref(12) is
270W; after 12th min, Pref(12) to Pref(20) is 270+(-220)=50W.
To summarize, method A is the simplest and fastest but also the least accurate
method. It is especially suitable when the target load has a fixed operation
process or when only its schedule information is known; method B is applicable
to all types of appliances and it makes a balance between accuracy and
simplicity; method C is the most accurate method but it requires the time instant
information of all window events. Also, it is the most complicated method.

6.3 Energy estimation method for incandescent lights

Proposed
flowchart
Start

Edge detection

Event filtration

Event clustering

ON-OFF match

Energy calculation

End

Figure 6.2: Flowchart of Energy estimation methods for incandescent lights

The intention of the proposed method is to obtain the energy consumption of

all the incandescent lights (IL) for a given period, for example, a week in a typical
residential house. It is better to treat all the IL as a special load group instead of

146
Chapter 7

treating each light individually. The main technique the method bases on is the
mean-shift clustering. The proposed method consists of 4 core steps: event
filtration, event clustering, ON-OFF match and energy calculation Some ideas
used in this method can be referred back to Chapter 4. The flow chart is shown in
Figure 6.2. Energy estimation for other load groups may also be referred to the
incandescent light that is explained as follows.

6.3.1 Event filtration

All the events in a house are firstly captured. Then according to the unique
features that most incandescent lights have in common, suspect events of all ILs
are picked out. This step is implemented through a filtering step---only the
specific events, which can satisfy a certain set of filtration conditions, are further
determined as the suspect events of IL. An example of filtration conditions used
for IL is listed as below:

 Active power: 30-300W

 Reactive power: <10 var (IL is a resistive load)

 Harmonic level: THD< 0.2 (IL is linear)

 Time of operations: 5:00PM---8:00AM of the next day

 Single-phase connection (all 120V based)

The above conditions can be adjusted according to the actual system

environment. For example, depending on different seasons and regions, the length
of daylight might be different. Daylight is closely related to the operation of
lighting devices. Also, if the power level of incandescent light bulbs in a house is
known, the active power range can be refined. For the other conditions such as
reactive power, harmonic and single-phase connection, they stay roughly the same
for all types of incandescent lights.

147
Chapter 7

6.3.2 Event clustering

Event clustering is then applied to the suspect events of IL obtained by the

previous step. The purpose of event clustering is to make clear how many
groups/types of ILs there are in this house. Since clustering algorithm is based on
common electric characteristics of lights, the events from the same type of IL can
be grouped as a “cluster”.

The model used for IL is a single-state load model which has the same absolute
electric characteristics for both its ON event and OFF event. The only difference
is that their signs are opposite to each other. It implies in an event cluster that is
caused by an IL, the number of its OFF members should be approximately equal
to the number of its ON members.

Phase-A clustering results of IL events:10 clusters

30
Event
Cluster centroid
Normalized Reactive Power(Q)

20

10

-10

-20

-30
0 10 20 30 40 50 60 70 80 90 100
Normalized Active Power(P)

Figure 6.3: Example of clustering results of IL events.

Mean-shift clustering is applied to all the ON events and OFF events together
and only their absolute event signatures are considered. In other words, both ON
and OFF events of a certain appliance should be found inside a single cluster.
During clustering, only active power feature is considered. This is because its
reactive power and harmonic contents are too small and can be noises. To rule out
the noisy impact from Q, the nominal values of Q of all events are set to zero. A
typical clustering map for one phase is shown in Figure 6.3.

148
Chapter 7

It can be seen from the clusters in different colors, there are likely 10 (kinds of)
incandescent lights at this phase. Among these clusters, only the ones which
satisfy:

 Member number is larger than 4;

 Owns both ON and OFF events.

are selected and taken into the following steps. This is because as a lighting
device, it is expected the frequency of light activities is sufficient; also, as a
single-state load, it is expected to see both ON and OFF events.

6.3.3 ON-OFF match

This step digs into a certain cluster obtained from the above step and examines
its inside ON-OFF events. The purpose is to check if they have roughly the same
number of appearances. The philosophy behind this is that theoretically, most of
incandescent lights (simply resistors) are single state appliances of which an
ON event should always match an OFF event. In reality, exceptions may occur
due to the following reasons:

1. Not all the ON and OFF events are successfully captured. Some may be lost;

2. Events coming from other appliances are mistakenly included in. For
example, for cluster A, we find 10 ON events and 12 OFF events. It can still be an
incandescent light, but two (12-10=2) OFF events could result from a third
appliance, for example the middle events of a furnace.

When one of the above is true, mismatch on numbers of ON and OFF will
happen.

Hypothetically, if all true ON and OFF events can be identified and selected,
accurate energy calculation can be done. However, it is impossible to do so.
Alternatively, as a statistic estimation method, all possible combinations of ON

149
Chapter 7

and OFF events can be calculated and an energy distribution can be obtained. An
example is shown as below:

Figure 6.4: ON/OFF pattern of light A

Light A has the above pattern in a given studied period. It has 3 possible ON
events (1,2,3) but only 2 possible OFF events (1’ and 2’). For the 3 ON events,
any two of them can be selected as a combination. Thus, there are 3 possible
combinations (1,2), (1,3) and (2,3). Each of these combinations can be matched
with its OFF event--- it only has one combination (1’,2’). The 3 ON combination
and the OFF combination can form the following 3 ON-OFF matches:

Table 6-1 3 possible ON-OFF matches of light A

Match index ON events OFF events Energy

1 (1,2) (1’,2’) 0.5 kWh
2 (1,3) (1’,2’) 0.6 kWh
3 (2,3) (1’,2’) -0.3 kWh

The above 3 matches will result different energy results. Thus, an energy
distribution can be formed.

Generally speaking, for a cluster which has m OFF events and n ON events
(m>n), the total number of possible combinations can be calculated using factorial:

m!
N cmb  (6.6)
n!(m  n)!

150
Chapter 7

It should be noted, when (m-n) is large, N cmb can be a very large number and

obtaining all the N cmb combinations can result in a heavy computation burden.
Alternatively, in actual implementation, 10000 different ON-OFF matches are
randomly generated and they can roughly represent all the possible combinations.

In addition, some ON-OFF match can result in a negative energy such as the
#3 match shown in Table 6-1. It is because the selected OFF events are before the
selected ON events. This kind of matches should be ruled out from the possible
ON/OFF matches because the energy is not allowed to be negative.

6.3.4 Energy calculation and the distribution plot

It should be noted, for a given ON-OFF match, paring of ON and OFF event
will not make a change on the energy calculation result. Still taking the above
light A as example, for match#1---(1,2)à(1’,2’), we have:

Power

1 1’ 2 2’
A1 A2 A3 3
time
Figure 6.5: Energy blocks of light A
Scenario 1: if 1 is paired with 1’ and 2 is paired with 2’, its energy is

E1  A1  A3 (6.7)

Scenario 2: if 1 is paired with 2’ and 2 is paired with 1’, its energy is

E  ( A1  A2 )  ( A3  A2 )  A1  A3  E1 (6.8)

As can be seen, it can be concluded that no matter which ON is paired with

which OFF, as long as the ON-OFF match is determined, the energy is always the

151
Chapter 7

same. In this example, match #1’s energy is always E1 no matter which ON pairs
with which OFF. But it can be different from match #2 and match #3 since events
from different matches can happen at different times.

Hence, the energy calculation is very simple:

1. For each qualified cluster, calculate the energy consumption for all possible
matches;

2. Add the energy of all the qualified clusters together to get overall IL energy
values in this house. In other words, an estimation of energy distribution can be
obtained. This is shown in the following section.

6.3.5 Results

A. House #1

In house #1, 6 days in September are processed. The estimate of energy

distribution, which is due to different possible ON-OFF matches, is shown in
Figure 6.6. In total, 10000 random generated ON-OFF matches are taken into
calculation. In this histogram, the width of per unit interval is 0.1kWh.

2.5

2.0
Probability(%)

1.5

1.0

0.5

0
6 8 10 12 14 16 18 20
Energy (kWh)

Figure 6.6: Energy distribution of IL in house #1

The minimum, mean and maximum values from the above distribution are
shown as below:

152
 Chapter 7

 Emin  Emin
phA
 Emin
phB
 1.24  3.45  4.69 kWh

 Emean  Emean  Emean  6.52  5.60  12.12kWh
phA phB
(6.9)

 Emax  Emax  Emax  11.66  10.95  22.61kWh
phA phB

The mean energy for IL is 12.12 kWh. This value is 10.54% of the total energy
of this house in the 6 days, which is about 115 kWh.

B. House #2

In house #2, 7 days in June are processed. The energy distribution that is due to
different possible ON-OFF matches is shown in Figure 6.6. In total, 10000
random generated ON-OFF matches are taken into calculation. In this histogram,
the width of per unit interval is 0.1kWh.

1.2

1.0

0.8
Probability(%)

0.6

0.4

0.2

0
10 15 20 25 30 35 40
Energy(kWh)

Figure 6.7: Energy distribution of IL in house #2

The minimum, mean and maximum values from the above distribution are
shown as below:

 Emin  Emin
phA
 Emin
phB
 5.40  5.02  10.42kWh

 Emean  Emean  Emean  10.61  12.68  23.29kWh
phA phB
(6.10)

 Emax  Emax  Emax  20.14  23.42  43.56kWh
phA phB

The mean energy for IL is 23.29 kWh. This value is 16.88% of the total energy
of this house in the 7 days, which is about 138 kWh.

153
Chapter 7

Generally, the above results comply with the statistic results obtained in
[114],[115]. Unfortunately, there is no feasible measurement method to accurately
validate the above results.

6.4 Energy estimation method for background energy

6.4.1 Minimal power based method

A simple way to estimate the background energy is proposed. Since

background power is the summation of the stand-by power and always-on load
power, the minimal power point when the other loads are all turned off is
considered as the background power.

Figure 6.8: Example of background power

As shown in Figure 6.8, during sleeping time, most of the appliances are
turned off, the minimum power found is defined as the background power.
Generally, to locate the background power, one can search the lowest power from
0:00AM to 6:00 AM. Background energy is calculated by multiplying the located
lowest power by 24 hours per day.

6.4.2 Results

The above method is applied to two houses and each for 7 days. The extracted
background power is shown as below:

154
Chapter 7

(a) House #1

(b) House #2

Figure 6.9: Background power extracted from 2 houses

As can be seen, the background power is an important energy component in a
house. Its power can be larger than 200W and can take up 20% to 40% of the
average total power.

155
Chapter 7

6.5 Energy estimation of residential houses

6.5.1 House #1

After all the energy components are estimated using the above introduced
methods, the total energy of a house can be broken down into different parts. A
pie-chart of energy consumption for a week in a residential house is plotted and
shown in Figure 6.10.This residential house is located in Edmonton, Canada and
the processed week is in June. Please note, other lights such as fluorescent lights
and compact fluorescent lights are roughly estimated as ¼ of incandescent light
according to their energy efficiency [116]-[118].

Figure 6.10: Energy consumption pie-chart for house #1

Assuming the electricity price per kWh is 9.9 Canadian cents (a typical mid-
peak price according to [16]), the electricity expenses in terms of dollars spent by
major individual appliances, appliance group and background power can also be
obtained. The results are shown in Table 6-2.

156
Chapter 7

Table 6-2 Energy consumption for house #1

Appliance Energy Consumption(kWh) Percent (%) Dollar

Fridge 11.34 8.23 1.47
Microwave 0.73 0.53 0.10
Washer 1.76 1.28 0.23
Cooktop 0.80 0.58 0.10
Stove(low power) 5.87 4.26 0.76
Stove(high power) 9.80 7.12 1.27
Kettle 3.53 2.56 0.46
Heater 1.59 1.15 0.21
Dryer 7.70 5.59 1.00
Waffle iron 1.21 0.88 0.16
Incandescent light 21.21 15.40 2.76
Other lights 5.30 3.85 0.69
Background 40.95 29.73 5.32
Others 25.95 18.84 3.37
Total 137.74 100.00 17.91

As can be seen, a single fridge takes a large portion of energy which is 8.23%.
Although the fridge’s power in this house is only about 140W, it is running
throughout 24 hours a day and the accumulated energy is significant.

Some kitchen appliances like microwave, kettle and waffle iron (in total 3.97%)
do not consume a lot although they have a large power (>1000W). This is because
for each time operation, it only lasts for minutes or sometimes even shorter.

Other kitchen appliance such as stove (15.67%) contributes a lot to the total
energy consumption because it is connected to double live phases and its cooking
time lasts much longer, for example, 30 minutes.

For laundry activities, washer does not consume a lot while dryer consumes an
important portion of energy (5.59%).

157
Chapter 7

Lighting is another significant energy component. Incandescent lights along

with other lights consume 19.25% of the total energy.

“Background energy” and “others” take almost a half of the house’s total
consumption including stand-by power, always-on appliances and all the other
appliances such as TVs, computers and dishwasher which have not been
registered or processed by NILM.

6.5.2 House #2

House #2 is another residential house located in Edmonton, Canada. A week in

May was processed. The results are shown in Figure 6.11 and Table 6-3.

Figure 6.11: Energy consumption pie-chart for house #2 in spring

Table 6-3 Energy consumption for house #2 in spring

Appliance Energy Consumption(kWh) Percent (%) Dollar

Fridge 10.98 8.81 1.09
Freezer 2.12 1.70 0.21

158
Chapter 7

Furnace 5.72 4.59 0.57

Microwave 1.01 0.81 0.10
Washer 0.32 0.26 0.03
Dryer 2.87 2.31 0.28
Coffeemaker 0.05 0.04 0.00
Oven 0.37 0.30 0.04
Stove 15.14 12.15 1.50
Waffle iron 0.06 0.05 0.01
Incandescent lights 20.18 16.19 2.00
Other lights 5.04 4.05 0.50
Background 40.95 32.85 4.05
Others 18.16 14.57 1.80
Total 124.66 100.00 12.34

As can be seen, in May, furnace is still used in this house due to the cold
temperature in Edmonton. It consumes 4.59% of the total energy. Compared to
fridge and freezer (10.51%), it consumes less because its frequency of operation is
lower than the frequency of fridge and freezer.

Similar results for kitchen appliances and lighting can be found compared to
house #1. The consumption by dryer is smaller than house #1 and it is possibly
due to smaller washing load in this house. Also, some clothes may be dried
outside in the sun.

The percentage of “background energy” and “others” (47.42%) is also similar

to house #1.

6.5.3 Seasonal changes of house #2

A week in late October from House #2 was also processed in comparison with
spring data to reveal the seasonal changes. The results are shown as below:

159
Chapter 7

Figure 6.12: Energy consumption pie-chart for house #2 in fall

Table 6-4 Energy consumption for house #2 in fall

Appliance Energy Consumption(kWh) Percent (%) Dollar

Fridge 11.50 8.65 1.14
Freezer 4.48 3.37 0.44
Furnace 14.38 10.82 1.42
Microwave 0.51 0.39 0.05
Washer 0.58 0.44 0.06
Dryer 10.74 8.08 1.06
Coffeemaker 0.01 0.01 0.00
Oven 4.66 3.51 0.46
Stove 2.30 1.73 0.23
Incandescent Light 16.75 12.60 1.66
Other lights 4.19 3.15 0.41
Background 37.54 28.24 3.72
Others 25.28 19.02 2.50
Total 132.93 100.00 13.16

160
Chapter 7

Generally, the total energy consumed in fall is higher than the energy
consumer in spring. This is partly due to the significant increase of the furnace
operation. The energy consumed by furnace increases from 5.72 kWh to 14.38
kWh. This is easy to be understood since the house temperature in October is
much lower than spring. In the meanwhile, the energy of fridge and freezer
increases from 13.1 kWh to 15.98 kWh. Also the dryer increases significantly
from 2.87 kWh to 10.74 kWh. This is probably because the washing load in fall
are generally higher than spring (more clothes). Also in late October, clothes are
not likely to be dried outside. The lighting and stove/oven energy consumed by
this particular week is smaller than spring and this is believed to be caused by
some unknown change of house owner’s behaviors such as less time in the house
and more dining in restaurants during the studied week.

6.6 Residential energy characteristics and its implications to

TOU price

Based on the above estimated results, some common characteristics of

residential electric energy consumption are summarized as below:

1. High power loads do not necessarily consume large amount of energy.

Examples are microwave and kettle. This is because their operation duration and
frequency can be very small.

2. Small power loads can still consume a lot of energy due to energy
accumulative effect. Examples are fridge and freezer.

3. “Automatic” appliances such as fridge, freezer and furnace usually consume

a lot of energy because they run throughout 24 hours per day.

4. Double-phase appliances such as stove, oven and dryer usually consume a

lot of energy. There are two reasons: their power can be extremely high, for
example, over 3000 watts; their operation durations are quite long, for example,
30 minutes to one hour.

161
Chapter 7

5. Washer does not consume a lot of energy as originally expected. This is

because modern front-load washer’s power (100-200W) is not large and it is a
VFD type device which does not generate a lot of heat like dryer.

6. The residential energy consumption changes in different seasons. One

example is the furnace. It is also expected air conditioner used in some warmer
areas has a similar change.

7. Lighting does consume a significant amount of energy. Our results are

consistent with the statistic results in[115]---lighting consumes a significant part
of total electrical energy worldwide. 20 to 50 percent consumed is due to lighting
in homes and offices.

8. Background energy also takes a large portion of energy. According to [120],

the large number of such devices and their being continuously plugged in resulted
in energy usage of 8 to 22 percent of all appliance consumption in different
countries, which is 32 to 87W. On top of the stand-by power, there are also some
always-on devices such as the wireless router and our data logging devices used at
the meter-side (about 100W)

Figure 6.13: A summary of TOU prices in spring, 2013 by Hydro One

In terms of the electricity bill cut and money saving, it seems the existing
Time-of-Use (TOU) price may not be able to create an effective and sufficient
incentive for ordinary house owners to change their electricity usage behavior.
However, before, TOU and TOU based smart meters are considered as effective

162
Chapter 7

tools to encourage load shifting and energy saving for residential customers.
According to [16], the current TOU price is shown in Figure 6.13.

Now considering house #2 as an example, in fact there are many appliances

that cannot be shifted from On/Mid-Peak hours to Off-Peak hours. For example,
fridge, freezer and furnace are automatically controlled according to the actual
environment/chamber temperature and house owners cannot easily shift them. In
addition, lighting cannot be shifted because its usage is determined by the
daylight conditions.

In Table 6-3, the only available appliances for load-shifting are microwave,
washer, dryer, coffeemaker, oven, stove and waffle iron. Hypothetically, changing
all of them from On-peak price to Off-peak price can save only
19.82  (11.8  6.3)  4( weeks)  4.36 dollars per month.

However, even for this little amount of money, the required load-shift can
greatly affect customers’ comfortableness and sometimes is even impossible. For
example, cooking is normally always before mealtime and cannot be shifted.

On the other hand, based on Table 6-3, replacing all the incandescent lights
with fluorescent/compact fluorescent lights may be able to reduce the energy by
about 12%. But this requires additional money investment since fluorescent lights
are much more expensive than incandescent lights [118].
In addition, a typical electricity bill from Hydro One [119] is shown in Figure
6.14. It can be seen that electricity charge ($81.40) is only 49% of the electricity
bill. The delivery charge, regulatory charge and retirement charge are relatively
fixed charges that do not change with the actual amount of electricity spent by the
residential customers.

Overall speaking, the effectiveness of TOU prices for residential houses in

order to encourage energy usage behavior change is suspected based on our
results.

163
Chapter 7

Figure 6.14: Electricity billing example by Hydro One

6.7 Summary

In this chapter, different energy estimation methods were proposed for three
major energy components in a residential house --- ordinary appliances, appliance
group (incandescent lights) and background energy.

Using these methods, meter-side data for several weeks which were acquired
from two local houses were processed. The energy estimation results were
presented. Also, common characteristics were revealed and summarized.

Based on the revealed characteristics, hypothetical energy savings by load shift

according to the existing TOU price was calculated and analyzed. It was found
that the existing TOU may not be able to effectively encourage residential load
shift behaviors after taking the customer’s comfortableness and the actual saving
amount into account.

164
Chapter 7

Chapter 7

Conclusions and Future work

7.1 Thesis Conclusions and Contributions

This thesis has approached topics related to event detection, load signature
studies, non-intrusive load monitoring algorithm, non-intrusive signature
extraction algorithm, energy estimate methods and the characteristics of
residential house. The thesis has presented a complete solution for power system
load decomposition, especially for residential end-users. Two major closely
related issues—non-intrusive load monitoring and non-intrusive signature
extraction were discussed in detail. But not limited to this, some other questions
related to event detection, load modeling, load group energy estimation, and
residential energy characteristics were also answered in the thesis.

The main conclusions and contributions of this thesis are summarized as

follow:

 Two data-segmentation based event detection methods were proposed and

studied. Instead of looking for state transitions directly, the proposed
methods look for stable data segments in which a certain level of signal
noise is also acceptable. The proposed slope method can effectively solve
the existing event detection challenges such as slope-type event, slow
event and signal noise caused misdetection.

 Common special issues of event detection have received attentions in this

thesis. A simple method to detect double-phase event was proposed. Also,
the causes of adjacent events and event overlap issues were studied and
possible solutions were discussed. It was found that increasing the data-

165
Chapter 7

acquisition sampling rate could effectively reduce the occurrence of the

above problems.

 Based on a review of the drawbacks of existing steady-state and transient

load models, an event-window load model was proposed. Studies for
event-window signatures were carried out. It was found that the event-
window model along with its signatures could accurately describe the
whole operation process of a complex load.

 A novel event-window based NILM algorithm was proposed and

discussed. The new algorithm:

o Can effectively identify the activities of complex loads including

continuous-varying and multi-state loads;

o And does not require a complicated local training process. In addition,

no re-training process is required after the load inventory is changed;

Extensive verifications based on three houses and a public dataset were

conducted. It was found the average identification rate is promising even
when the aggregated signal contains a certain level of noises.

 A novel NILM related problem--non-intrusive signature extraction for

major residential loads--was proposed and studied. The previous
measurement based signature extraction methods were actually intrusive.
This critical obstacle prevents NILM from being practically applied.
Unfortunately, in the past, researchers did not pay enough attentions to
this problem. In this thesis, this problem was systematically discussed, and
a clustering based solution was presented.

 Energy estimation of residential houses was studied and investigated from

the following perspectives:

o Three energy estimation methods for ordinary appliances were

proposed and compared.

166
Chapter 7

o A clustering based energy estimation method for load groups such

as incandescent lights was presented.

o An energy estimation method for background energy was

presented.

o Residential energy characteristics and their implication for the

Time-of-Use price were summarized and discussed. It was found
that the existing TOU price might not encourage residential load
shift behaviors after taking both the customer’s comfortableness
sacrifice and the actual amount saved into account.

7.2 Suggestions for future work

As with any study, something can always be done to extend and improve the
research. Several extensions and modifications of this thesis can be explored as
follows:

 Future studies can be extended into the commercial and industry

compound load area. This thesis focuses mainly on residential
compound loads since many commercial and industry loads have their
own sub-load monitoring equipment already. However, a cost-effective
non-intrusive solution may still be of interest. In the appendix, some
preliminary studies for a typical commercial building were presented.
More studies including extensive field data based verification can be
completed in the future.

 The energy estimation method for load groups can be improved:

o This thesis focused on the incandescent light load group. Other

load groups such as the florescent load group for a residential
house, the computer load group for an office building, and the
HAVC load group for an industrial building can be further
studied.

167
Chapter 7

o In this thesis, for the energy estimation method proposed for

incandescent light group, no field data based verification was
conducted because of the limited available data acquisition
devices that can be used to effectively measure all the
incandescent lights. This problem may be solved by using
advanced in-direct sensing devices, which may also be applied
to other types of load groups.

 Event-misdetection by event overlap is still a problem that could affect

the accuracy of the proposed event-window based NILM algorithm.
More sophisticated hardware and algorithm solution can be studied for
improvement.

 Comparative studies between NILM algorithm and the statistical data

based approach [34]-[35] can be conducted to clarify the advantages
and disadvantages on the accuracy and easiness aspects of each
approach.

 More research should focus on non-intrusive load signature extraction.

More types of loads, more geographic regions and different sensitivity
studies should be provided because clearly this very important area has
been neglected by the previous NILM researchers.

 The energy characteristics of commercial and industry loads can be

further studied, and the potential application of load decomposition to
demand response and load shedding can be further discussed.

168
Chapter 8

Chapter 8

References

[1] David Schatsky and Clint Wheelock, “Home Energy Management”,

available online at: http:// www.pikeresearch.com/research/

[2] David Kathan and et al. , “Assessment of Demand Response and

Advanced Metering”, Federal Energy Regulatory Commission, Dec. 2008

[3] S. Massoud Amin and Bruce F. Wollenberg, “Toward a Smart Grid”,

IEEE Power & Energy Magazine, Sept/Oct 2005, pp. 34-41.

[4] H. Farhangi, “The Path of the Smart Grid”, IEEE Power & Energy
Magazine, Jan/Feb 2010, pp. 18-28.

[5] United States Department of Energy, “The Smart Grid: An Introduction”,

available online at: http://www.oe.energy.gov/SmartGridIntroduction.htm

[6] Litos Strategic Communication, “The Smart Grid: An Introduction”,

United States Department of Energy, 2004.

[7] S. Rahman, “Smart Grid Expectations”, IEEE Power & Energy Magazine,
Sept/Oct 2009, pp. 83-85.

[8] 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.

[9] ACEEE, “Study: Smart Meters Alone Not Enough to Significantly Reduce
Power Use”, available online at:

http://tdworld.com/customer_service/aceee-smart-meter-study-0610/

169
Chapter 8

[10] Andy Berrisford, “New Technologies for End-use Load Research at BC

Hydro”, CIGRE Conference Vancouver 2010.

[11] Gridwise, “A Smart Grid:Cost Reduction” ,Gridwise Alliance, available

on line at: http://www.gridwise.org/smartgrid_costreduction.asp

[12] ComEd, “Residential Real-Time Pricing (RRTP) program overview”,

available online at: https://rrtp.comed.com/about/

[13] Ameren Illinois, “Power Smart Pricing” available online at:

http://www.powersmartpricing.org/

[14] IESO, “Time-of-Use prices”, available online at:

http://www.ieso.ca/imoweb/siteshared/tou_rates.asp

[15] Hydro One, “Introducing Time-of-Use Prices” , available online at:

http://www.hydroone.com/TOU/Documents/Introducing_TOU_Prices.pdf

[16] Hydro One , “Time-of-Use Prices” , available online at:

http://www.hydroone.com/TOU/Pages/Default.aspx

[17] Burlington Hydro, “Time of Use: A New Way to Think About

Electricity”, available online at: http://www.burlingtonhydro.com/your-
business/time-of-use.html

[18] Guelph Hydro , “Time-of-Use Rates for Electricity”, available online at:
http://www.guelphhydro.com/en/residential/timeofuserates.asp

[19] Ontario Ministry of Energy, “Smart Meters and Time-of-Use Prices”,

available online at: http://www.energy.gov.on.ca/en/smart-meters-and-tou-
prices/

[20] Milind Naphade and et al., “Smarter Electricity Portal Pilot Study Report”,
IBM Research, available online at:
http://www.cityofdubuque.org/DocumentCenter/View/15984

170
Chapter 8

[21] ISTAT, Time Use Survey, Notiziarion. 10, 1990 (in Italian).
[22] ISTAT, Household consumption: year 1987, vol. 14, 1989 (in Italian).

[23] Ipsos-RSL and Office for National Statistics, United Kingdom Time Use
Survey, 3rd ed., UK Data Archive, Colchester, Essex, September 2003

[24] Ian Richardson and Murray Thomson, “CREST Domestic Electricity

Demand Model”, available online at: https://dspace.lboro.ac.uk/dspace-
jspui/handle/2134/5786

[25] Bell, M. Swinton, M.C. Entchev, E. Gusdorf, J. Kalbfleisch, W. Marchand,

R.G. Szadkowski, F., “Development of Micro Combined Heat and Power
Technology Assessment Capability at the Canadian Centre for Housing
Technology”, pp. 48. 2003-12-08.

[26] M. Armstrong, M. Swinton, H. Ribberink, I. Beausoleil-Morrison and J.

Millette, “Synthetically derived profiles for representing occupant driven
electric loads in Canadian housing”, Building Performance Simulation,
vol.2, pp. 15–30, 2009.

[27] Pratt, R., et al. Description of Electric Energy Use in Single-Family

Residences in the Pacific Northwest – End-Use Load and Consumer
Assessment Program (ELCAP). DOE/BP-13795-21, Richland, WA:
Pacific Northwest National Laboratory, 1989.

[28] CSA, “Energy Consumption Test Methods for Household Dishwashers”,

CSA Standard CAN/CSA-C373-92, 1992.

[29] CSA, “Test Method for Measuring Energy Consumption and Drum
Volume of Electrically Heated Household Tumble-Type Clothes Dryers”,
CSA Standard CAN/CSA-C361-92, 1992.

[30] Statistics Canada, “Selected dwelling characteristics and household

equipment”, available online at:
http://www40.statcan.gc.ca/l01/cst01/famil09b-eng.htm

171
Chapter 8

[31] I. Richardson, Murray Thomson, David Infield and Conor Clifford,

"Domestic electricity use: A high-resolution energy demand model,"
Energy and Buildings, vol. 42, pp. 1878-1887, 2010.

[32] C. Walker and J. Pokoski, "Residential load shape modeling based on

customer behavior," Power Apparatus and Systems, IEEE Transactions on,
pp. 1703-1711, 1985.

[33] Capasso, A.; Grattieri, W.; Lamedica, R.; Prudenzi, A., "A bottom-up
approach to residential load modeling," Power Systems, IEEE
Transactions on , vol.9, no.2, pp.957,964, May 1994

[34] “DOE-2”, available online at: http://gundog.lbl.gov/dirsoft/d2whatis.html

[35] Microsoft Corp., “Microsoft-Hohm”, available online at:

http://www.microsoft-hohm.com/Default.aspx?wa=wsignin1.0

[36] 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.

[37] Digi Smart Plug, “XBee Smart Plug” , available online at:
http://www.digi.com/wiki/developer/index.php/XBee_Smart_Plug

[38] Blue line innovations, “BlueLine Electricity Power Cost Monitor”,

available online at:

http://www.blueline innovations.com/Products/PowerCost-Monitor/

[39] AlertMe Energy, “How it works”, available online at: http://www.alert

me.com/products/energy/how-it-works

[40] Energy Inc., “TED The Energy Detective”, available online at:
http://www.theenergydetective.com/

[41] Echelon, “Smart Metering”, available online at:

http://www.echelon.com/solutions/smartmetering/han_solutions.htm

172
Chapter 8

[42] Engadget, “LG Thinq linqs your smart appliances with WiFi and
smartphone apps“, available online at:
http://www.engadget.com/2011/01/04/lg-thinq-linqs-your-smart-
appliances-with-wifi-and-smartphone-ap/

[43] TriplePundit, "Study Shows Bright Green Future for Smart Appliances",
available online at: http://www.triplepundit.com/2010/03/smart-
appliances-could-be-the-next-big-thing/

[44] A. Zaballos, A. Vallejo, M. Majoral, and J. M. Selga,“Survey and

Performance Comparison of AMR Over PLC Standards,” IEEE
Transactions on Power Delivery, 2009, vol. 24, pp.604-613.

[45] Q. Y. Liu, B. Z. Zhao, Y. R. Wang, and J. Hu, “Experience of AMR

Systems Based on BPL in China,” IEEE International Symposium on
Power Line Communications and Its Applications, 2009, pp.280-284.

[46] IEEE Std. 802.15.4-2003, Wireless Medium Access Control(MAC) and

Physcial Layer(PHY) Specifications for Low Rate Wireless Personal Area
Networks[S], 2003.

[47] W. Qiu, E.Skafidas, and P. Hao, “Enhanced tree routing for wireless
sensor networks,” Elsevier on Ad Hoc Networks, 2009, vol. 7, pp.638-650.

[48] K. Taehong, K. Daeyoung, P. Noseong, Y. Seong-eun and T.S. Lopez,

“Shortcut Tree Routing in ZigBee Networks,” Proceedings of the Second
International Symposium on Wireless Pervasive Computing, February
2007.

[49] Y. Kim, T. Schmid, Z. M. Charbiwala, and M. B. Srivastava, "ViridiScope:

Design and implementation of a fine grained power monitoring system for
homes", in Proc. 11th Int. Conf. Ubiquitous Comp., pp. 245–254, 2009

173
Chapter 8

[50] A. Rowe, M. Berges, and R. Rajkumar, "Contactless Sensing of Appliance

State Transitions Through Variations in Electromagnetic Fields", BuildSys
2010, pp. 19-24

[51] D. Jung, and A. Savvides, "Estimating Building Consumption

Breakdowns using ON/OFF State Sensing and Incremental Sub-Meter
Deployment". Proceedings of the 8th ACM Conf. Embedded Networked
Sensor Systems, pp. 225-238, 2010

[52] Tobias Ryberg, “Smart Metering in Western Europe”, M2M Research

Series, available online at: http://www.triplepundit.com/2010/03/smart-
appliances-could-be-the-next-big-thing/ .

[53] Jeffery F. Fye and Mark A. Johnson, “Solid-state Electricity Meter”,

United States Patent, US 6,734,663 B2, May 11 2004.

[54] Aclara TWACS Technology Specification Sheets, available online at:

http://www. aclara.com/AclaraPLS/Pages/specsheets.aspx.

[55] Hart, G.W., "Non-intrusive Appliance Load Monitoring", Proceedings of

the IEEE, vol. 80, No 12, pp. 1870 - 1891, 1992

[56] C. J. Andrews, “Nonintrusive monitoring device for electric loads in

commercial buildings”, MIT Laboratory for Electrical and
Electromagnetic Systems Technical Report, 1988.

[57] F. Sultanem, "Using appliance signatures for monitoring residential loads

at meter panel level", IEEE Trans. Power Delivery, vol. 6, no. 4, pp. 1380-
1385, 1991

[58] Hannu Pihala, "Non-intrusive appliance load monitoring system based on

a modern kWh-meter", VTT Energy licentiate thesis, Technical research
center of Finland, 1998

174
Chapter 8

[59] Jian Liang, Simon K. K. Ng, Gail Kendall, and John W. M. Cheng, "Load
Signature Study—Part I: Basic Concept, Structure, and Methodology",
IEEE Trans. Power Delivery, vol. 25, no. 2, pp.551-560, 2010

[60] Jian Liang, Simon K. K. Ng, Gail Kendall, and John W. M. Cheng, "Load
Signature Study—Part I: Basic Concept, Structure, and Methodology",
IEEE Trans. Power Delivery, vol. 25, no. 2, pp.551-560, 2010

[61] R. W. Cox, S. B. Leeb, S. R. Shaw, and L. K. Norford, “Transient event

detection for nonintrusive load monitoring and demand side management
using voltage distortion,” in Proc. 21st IEEE Applied Power Electronics
Conference and Exposition (APEC), 2006

[62] S.B. Leeb, S.R. Shaw and J.L. Kirtley, "Transient event detection in
spectral envelope estimates for nonintrusive load monitoring," Power
Delivery, IEEE Transactions on , vol.10, no.3, pp.1200,1210, Jul 1995

[63] S.B. Leeb, S.R. Shaw and S.R.Shaw, “Multiprocessing Transient Event
Detector For Use In A Nonintrusive Electrical Load Monitoring System”,
United States Patent, US 5,717,325, Feb. 10,1998.

[64] S. Leeb , J. L. Kirtley Jr., M. S. LeVan and J. P. Sweeney "Development

and validation of a transient event detector", AMP Journal Technol., vol.
3, pp.69 -74 1993

[65] S.N. Patel, T. Robertson, J.A. Kientz, M.S. Reynolds, and G.D. Abowd,
"At the Flick of a Switch: Detecting and Classifying Unique Electrical
Events on the Residential Power Line", In: UbiComp 2007, pp. 271-288

[66] S. Gupta, M. S. Reynolds, S. N. Patel, “ElectriSense: Single-Point

Sensing Using EMI for Electrical Event Detection and Classification in
the Home,” Conference on Ubiquitous Computing, pp. 139-148, 2010

175
Chapter 8

[67] H.Y. Lam, G. S. K. Fung, W. K. Lee, “A Novel Method to Construct

Taxonomy of Electrical Appliances Based on Load Signatures,” IEEE
Transactions on Consumer Electron., vol. 53, no. 2, pp. 653-660, 2007

[68] Yi-Ching Su; Kuo-Lung Lian; Hsueh-Hsien Chang, "Feature Selection of

Non-intrusive Load Monitoring System Using STFT and Wavelet
Transform," IEEE 8th International Conference on e-Business
Engineering (ICEBE), pp.293-298, 2011

[69] D.Srinivasan, W.S.Ng and A.C.Liew, "Neural-network-based signature

recognition for harmonic source identification", IEEE Transactions on
Power Delivery, Volume 21, Issue 1, pp. 398 – 405, 2006

[70] 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 , vol.5, no., pp.V-
988,V-991 Vol.5, 23-26 May 2004

[71] Yu-Hsiu Lin, Men-Shen Tsai, “Application of Neuro-Fuzzy Pattern

Recognition for Non-intrusive Appliance Load Monitoring in Electricity
Energy Conservation”, WCCI 2012 IEEE World Congress on
Computational Intelligence, pp.686-692, 2012.

[72] Suzuki, K.; Inagaki, S.; Suzuki, T.; Nakamura, H.; Ito, K., "Nonintrusive
appliance load monitoring based on integer programming", SICE Annual
Conference, 2008, pp. 2742 – 2747, 2008

[73] Baranski, M.; Voss, J., "Genetic algorithm for pattern detection in NIALM
systems," Systems, Man and Cybernetics, 2004 IEEE International
Conference on , vol.4, no., pp.3462,3468 vol.4, 10-13 Oct. 2004

[74] M.L. Marceau, and R. Zmeureanu, "Nonintrusive load disaggregation

computer program to estimate the energy consumption of major end uses
in residential buildings", Energy Conversion & Management, vol. 41, pp.
1389-1403, 2000

176
Chapter 8

[75] Prudenzi, "A Neuron Nets Based Procedure for Identifying Domestic
Appliances Pattern-of-use from Energy Recordings at Meter Panel", IEEE
Power Engineering Society Winter Meeting, pp. 941-946, 2002

[76] HH Chang; Kun-Long Chen; Yuan-Pin Tsai; Wei-Jen Lee; , "A New
Measurement Method for Power Signatures of Nonintrusive Demand
Monitoring and Load Identification," Industry Applications, IEEE
Transactions on , vol.48, no.2, pp.764-771, March-April 2012

[77] Yu-Hsiu Lin, Men-Shen Tsai, “Application of Neuro-Fuzzy Pattern

Recognition for Non-intrusive Appliance Load Monitoring in Electricity
Energy Conservation”, WCCI 2012 IEEE World Congress on
Computational Intelligence, pp.686-692, 2012.
[78] A.G. Ruzzelli, G.M.P. O’Hare, A. Schoofs, C. Nicolas, "Real-Time
Recognition and Profiling of Appliances through a Single Electricity
Sensor", In 7th Annual IEEE Communications Society Conference on
Sensor, Mesh, and Ad Hoc Communications and Networks (SECON’10),
2010
[79] H. Murata , T. Onoda, "Applying Kernel Based Subspace Classification
to a Non-intrusive Monitoring for Household Electric Appliances",
Artificial Neural Networks - ICANN 2001

[80] Marchiori, D. Hakkarinen, H. Qi, and L. Earle, “Circuit level load

monitoring for household energy management,” IEEE Pervasive
Computing, vol. 10, no. 1, pp. 40-48, 2011

[81] M. Berges, E. Goldman, H. S. Matthews and L. Soibelman, “Training

Load Monitoring Algorithms on Highly Sub-Metered Home Electricity
Consumption Data,” 12th International Conference on Computing in Civil
and Building Engineering, Beijging, 2008.
[82] M. Berges, E. Goldman, H. S. Matthews, and L. Soibelman, "Learning
Systems for Electric Consumption of Buildings," in ASCE International
Workshop on Computing in Civil Engineering, Austin, TX, pp. 1-10, 2009

177
Chapter 8

[83] M. Weiss, T. Staake, F. Mattern, "Leveraging smart meter data to

recognize home appliances" IEEE Pervasive Computing and
Communication (PerCom), Switzerland, 2012

[84] K. D. Lee, S.B. Leeb, L.K. Norford, P.R. Armstrong, J. Holloway, S.R.
Shaw, "Estimation of Variable-Speed-Drive Power Consumption From
Harmonic Content," , IEEE Trans. Energy Conversion, vol.20, no.3, pp.
566- 574, 2005

[85] W. Wichakool, A.T. Avestruz, R.W. Cox, S.B. Leeb, "Resolving Power
Consumption of Variable Power Electronic Loads Using Nonintrusive
Monitoring," Power Electronics Specialists Conference, IEEE PESC 2007,
pp.2765-2771, 2007

[86] W. Wichakool, A.T. Avestruz, R.W. Cox, S.B. Leeb, "Modeling and
Estimating Current Harmonics of Variable Electronic Loads", , IEEE
Trans. Power Electronics, vol.24, no.12, pp.2803-2811, 2009

[87] S.R. Shaw, S.B. Leeb, L.K. Norford, R.W. Cox, "Nonintrusive Load
Monitoring and Diagnostics in Power Systems", IEEE Trans.
Instrumentation and Measurement, vol.57, no.7, pp.1445-1454, 2008

[88] U.A. Orji, Z. Remscrim, C. Laughman, S.B. Leeb, W. Wichakool, C.

Schantz, R. Cox, J. Paris, J.L. Kirtley, L.K. Norford, "Fault detection and
diagnostics for non-intrusive monitoring using motor harmonics", Applied
Power Electronics Conference and Exposition (APEC), pp.1547-1554,
2010

[89] J. Paris, Z. Remscrim, K. Douglas, S. B. Leeb, R. W. Cox, S. T. Gavin, S.

G. Coe, J. R. Haag, and A. Goshorn, "Scalability of non-intrusive load
monitoring for shipboard applications," in American Society of Naval
Engineers Day 2009, National Harbor, 2009

178
Chapter 8

[90] R. W. Cox, P. Bennett, D. McKay, J. Paris, and S. B. Leeb, “Using the

non-intrusive load monitor for shipboard supervisory control,” in IEEE
Electric Ship Technologies Symposium, Arlington, VA, May 2007

[91] D.C. Bergman, Jin Dong , J.P. Juen, N. Tanaka, C.A. Gunter, A.K. Wright,
"Nonintrusive Load-Shed Verification," Pervasive Computing, IEEE ,
vol.10, no.1, pp.49-57, 2011

[92] P. K. Parikh, R. W. Cox, and J. M. Anderson, "Using Voltage Distortion

Measurements for Demand-Side Management", 12th Workshop on
Control and Modeling for Power Electronics (COMPEL), pp-1-5, 2010

[93] G. Kalogridis, C. Efthymiou, S.Z. Denic, T.A. Lewis, R. Cepeda,

"Privacy for Smart Meters: Towards Undetectable Appliance Load
Signatures", First IEEE Inter©national Conf. on Smart Grid
Communications (SmartGridComm), 2010 , pp.232-237, 2010

[94] M. Zeifman, and K. Roth, “Nonintrusive Appliance Load Monitoring:

Review and Outlook,” IEEE Trans. Consumer Electronics, vol.57, no.1,
pp. 76-84, 2011

[95] J. Zico Kolter and Matthew J. Johnson. REDD: A public data set for
energy disaggregation research. In proceedings of the SustKDD workshop
on Data Mining Applications in Sustainability, 2011

[96] 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

[97] S.W. Heunis, and R. Herman, “A probabilistic model for residential

consumer loads,” Power Systems, IEEE Transactions on, vol. 17, no. 3,
pp. 621-625, 2002.

[98] Chenzhi Fan, Dun Sun and Mei Tong, “Circuit Theory” , China Machine
Press, Edition 2, 2008

179
Chapter 8

[99] "IEEE Recommended Practices and Requirements for Harmonic Control

in Electrical Power Systems," IEEE Std 519-1992
[100] IEEE Working Group on Nonsinusoidal Situations: Effects on Meter
Performance and Definitions of Power, “Practical Definitions for Powers
in Systems with Nonsinusoidal Waveforms and Unbalanced Loads: a
Discussion”, IEEE Transactions on Power Delivery, vol. 11, no. 1,
January 1996, pp. 79-101.
[101] Hooman Erfanian Mazin, Data-based Harmonic Source Identification,
PhD Thesis. University of Alberta, Fall 2011.
[102] Chen Jiang, A Probabilistic Bottom-up Technique for Modeling and
Simulation of Residential Distributed Harmonic Sources, Master of
Science Thesis. University of Alberta, Spring 2012.
[103] M.Dong, H.E.Mazin and W.Xu, "Data segmentation algorithms for a time-
domain harmonic source modeling method," Electrical Power & Energy
Conference (EPEC), 2009 IEEE , vol., no., pp.1-6, 22-23 Oct. 2009
[104] M.Dong, Paulo C.M.Meria, W.Xu and Walmir Freitas, “An Event
Window Based Load Monitoring Technique for Smart Meters”, IEEE
Transactions on Smart Grid, vol.3,No.2, pp 787-796, June 2012.
[105] M.Dong, Paulo C.M.Meria and W.Xu , “ Non-intrusive Signature
Extraction for Major Residential Loads”，IEEE Transactions on Smart
Grid, accepted and in press , Feb. 2013
[106] Hartigan, J. A.; Wong, M. A., "Algorithm AS 136: A K-Means Clustering
Algorithm". Journal of the Royal Statistical Society, Series C (Applied
Statistics) 28 (1): 100–108, 1979
[107] Comaniciu, D., Meer, P., "Mean shift: a robust approach toward feature
space analysis," Pattern Analysis and Machine Intelligence, IEEE
Transactions on , vol.24, no.5, pp.603-619, May 2002
[108] Hung-Kuo Chu, Tong-Yee Lee, "Multiresolution Mean Shift Clustering
Algorithm for Shape Interpolation," Visualization and Computer Graphics,
IEEE Transactions on, vol.15, no.5, pp.853-866, Sept.-Oct. 2009

180
Chapter 8

[109] Y. Z. Cheng, “Mean shift, mode seeking, and clustering,” IEEE

Trans.Pattern Anal. Mach. Intell., vol. 17, no. 8, pp. 790–799, 1995.
[110] M.A. Little and N.S.Jones, "Generalized Methods and Solvers for
Piecewise Constant Signals: Part I". Proceedings of the Royal Society
A.,2011
[111] Wang Zhenyu and Zheng Guilin, "The application of mean-shift cluster in
residential appliance identification," Control Conference (CCC), 2011
30th Chinese , vol., no., pp.3111,3114, 22-24 July 2011
[112] Amorim, R. C. and Mirkin, B, "Minkowski metric, feature weighting and
anomalous cluster initializing in K-Means clustering". Pattern Recognition
45 (3): 1061–1075, 2012
[113] A. P. Engelbrecht, Computational Intelligence. Chichester, UK: John
Wiley & Sons, Ltd, 2007
[114] "National Lighting Inventory and Energy Consumption Estimate",
prepared for U.S. Department of Energy, National Technical Information
Service (NTIS),U.S. Department of Commerce, Sep.2002
[115] Hawken, Paul; Lovins, Amory; Lovins, L. Hunter (2000), “Natural
Capitalism”, Back Bay Press, Time Warner Book Group. ISBN 0-316-
35300-0
[116] Design Recycle Inc., "LED Lights vs. Incandescent Light Bulbs vs. CFLs",
available online at:
http://www.designrecycleinc.com/led%20comp%20chart.html
[117] Earthfriends, “CFL Vs. Incandescent”, available online at:
http://www.energyclassroom.com/pdfs/EC_CO_AC_LightBulbCost.pdf
[118] Energyclassroom ,“Light Bulb Cost Comparison”, available online at:
www.energyclassroom.com/pdfs/EC_CO_AC_LightBulbCost.pdf
[119] “My Home Bill Explained”, available online at:
www.hydroone.com/MyHome/MyAccount/MyBill/Pages/default.aspx
[120] Energy review 2006, Department of Trade and Industry, UK

181
Chapter 8

[121] R.Fernandes, I.Silva, M.Oleskovicz , "Load Profile Identification Interface

for Consumer Online Monitoring Purposes in Smart Grids," Industrial
Informatics, IEEE Transactions on , vol.PP, no.99, pp.1,1, 0

182
Appendix

Appendix

Load Disaggregation Method for Commercial Buildings

The appendix proposes a load decomposition method specified for commercial

buildings. Limited simulations are also given for basic verification purpose. The
scope of the entire study is preliminary and the main contribution is to present a
novel thinking on non-intrusive load group monitoring. Suggestions for the future
work are also given in the end.

A.1 The proposed method

The proposed technique aims to break down the total commercial building
power usage into the level of individual load groups such as all lighting devices
and all computing devices in the building.

Two basic assumptions are:

 A commercial building has only a few appliance groups and categories

although the total number of appliances can be fairly large.

 From the statistical perspective, a certain load group which has many
individuals will follow a generally stable signature pattern and individual
appliance difference amongst it is negligible when treated collectively.

For example, an office building has 500 computers, there are two brands A and
B---300 computers with brand A and 200 with brand B. Brand A computer has a
total harmonic distortion (THD) of 30% while brand B has a THD of 40%.
However, since the mixed ratio of brand A and B generally stays the same, as a
group, all computer devices approximately always produce a stable aggregated
THD, say 37%. The prerequisite of this assumption is that brand A and brand B
computers have almost the same behavior characteristics (People do not use them

183
Appendix

differently). Statistically, for any random given time, the number of brand A
computers being used vs. the number of brand B computers being used are
generally still 3:2. Thus, the aggregated THD signature can stay approximately
the same for this load group “computer devices”. However, if for example, brand
A is only used in night-time and brand B is only used in day-time, they have to be
treated as two independent groups. In reality, this rarely happens since for most
cases the same function oriented appliances behave consistently.

Based on the above assumptions, some group signatures such as THD and
individual harmonic distortion (IHD) as well as power factor can be used to label
specific certain load groups.

The figure below shows an example of load division in an office building.

Figure A-1: Example of load division in an office building

In total, its 1500 loads can be classified into the 7 groups under 4 categories---
non-linear, reactive (linear), resistive (linear) and background. Each of these
categories has their own unique collective group signatures such as IHD ratio for
non-linear load category and power factor for reactive load category. Based on the
collective group signatures, aggregated profile acquired from the building’s

184
Appendix

meter-side can be decomposed into individual groups through mathematic

methods. Different load categories can be treated specifically as below:

A. Background load

Background load can be understood as the summation of stand-by power and

always-on loads. Its value can be approximated as the lowest point at night, say
1:00AM---5:00 AM when no people are acting in the building.

B. Non-linear load

As shown in Figure A-1, non-linear load in an office building might be

composed of the following parts:

 CFL lighting

 Computer Device (AC-DC rectifier based)

 HVAC-A (6-pulse variable frequency driver based)

 HVAC-B (12-pulse variable frequency driver based)

From harmonic perspective, the above groups have different harmonic

spectrums with each other, however, within each group, the appliances roughly
have similar harmonic patterns in terms of normalized magnitude (IHD) and
phase angles and can present a relatively stable mixed harmonic pattern. The
pattern is often determined by the power electronic circuit it uses. For example,
most modern computer devices such as desktops and laptops use single-phase
AC-DC rectifiers. HVAC equipment can have either 6-pulse or 12-pulse VFD
drives. Thus it is treated as two independent groups A and B separately.

To obtain an accurate representative harmonic pattern for each group,

measurement based on a certain ratio can be further performed for each group.
For example, suppose there are three types of computers A,B,C being used. In this
office building, they have a mixture ratio as 200:300:400 (900 computers in total).
Accordingly, 2 type-A computers,3 type-B computers and 4 type-C computers

185
 Appendix

can be randomly chosen and put together. Then their collective IHDs and phase
angles can be measured. This measured representative harmonic pattern can
approximately represent the aggregated pattern of all computer devices in this
building at any given time. The table below is an example of representative
spectrums of all four groups (CFL /Computer /HVAC-A / HVAC-B).

Table A-1 Representative Spectrums of load groups

Harmonic Representative Spectrums of CFL /Computer /HVAC-A / HVAC-B

order Normalized Magnitude (IHD) Phase Angle (degree)
1 1.00 / 1.00 / 1.00 / 1.00 21 / 0 / 0 / -28
3 0.81 / 0.79 / 0.03 / 0.02 54 / 2 / 0 / 0
5 0.57 / 0.49 / 0.61 / 0.02 106 / 4 / -175 /-128
7 0.45 / 0.20 / 0.33 / 0.01 169 / 9 / -172 /51
9 0.44 / 0.12 / 0.01 / 0.01 -135/ 0 / 0 / 0
11 … …

In the meanwhile, total harmonic signal of the commercial building after the
background load deducted can also be analyzed and contents of harmonic orders
can be extracted, which should be equal to the summation of the four non-linear
groups. The following equations can be listed:
/135.42
 I3  I CFL  I CFL 3( p.u.)  I CMP  I CMP 3( p.u.)  I HA  I HA3( p.u.)  I HB  I HB 3( p.u.)

 I5  I CFL  I CFL 5( p.u.)  I CMP  I CMP 5( p.u.)  I HA  I HA5( p.u.)  I HB  I HB5( p.u.)

I7  I CFL  I CFL 7( p.u.)  I CMP  I CMP 7( p.u.)  I HA  I HA7( p.u.)  I HB  I HB 7( p.u.)

 I9  I CFL  I CFL 9( p.u.)  I CMP  I CMP 9( p.u.)  I HA  I HA9( p.u.)  I HB  I HB 9( p.u.)
...
 (A.1)

where the left side are the harmonic contents of total aggregated signal. The
per-unit value can be read from Table A-1. For example, I CFL3( p.u.) is complex

value 0.8154 . ICFL , ICMP , I HA , I HB are unknown variables that need to be solved. If
there are only four groups, harmonic contents of four orders ( I 3 , I 5 , I 7 , I 9 ) are
sufficient to solve the variables. However, more harmonic contents can also be
used to convert the problem to an over-determined problem in which case some

186
Appendix

optimization methods such as least square method can be applied to improve the
accuracy. Once ICFL , ICMP , I HA , I HB are solved, their power and energy profiles can
be immediately obtained.

C. Reactive (linear) load

Based on the above discussions, the summation of reactive power from the
non-linear load category can also be calculated. Apart from non-linear load and
background load, the rest aggregated reactive power all comes from the reactive
(linear) loads. Similar to the measurement taken to get representative harmonic
spectrum, representative power factor can also be measured. For example, a
couple of refrigerators and fans can be measured together according to a mixing
ratio. Thus power profile of reactive load can be obtained. Also, based on the
mixing ratio, refrigerators and fans can be further separated.

D. Resistive (linear) load

In the end, the residual power apart from background load, non-linear load and
reactive load belongs to resistive load.

A.2 Simulation results

Due to limited resources, two simulations are conducted based on the bottom-
up model and program developed for multiple residential houses in [102].
However, it is still sufficient to verify the basic working principle for commercial
buildings and reveal some interesting findings. In this case, a commercial building
is assumed to include 6 groups of loads--- 800 CFL lights, 300 computers, 100
furnaces, 100 inductive motors, 2400 incandescent lights and background power.
Their powers and mixing ratios are listed in Table A-2.

The program uses Monte Carlo based simulation to generate the aggregated
signal of all the loads according to certain statistical time of use probability
profiles for different loads [102].

187
Appendix

Table A-2 Load group inventory and composition

Load group Mixture Power of Mixing ratio Total
appliances appliances Number
ID-6 29.47 W 50 % 800
CFL ID-12 14.66 W 50 %
PC ID-1 98.76 33.3%
+Monitor ID-1 W+32.43 W
Computing PC ID-2 90.85 33.3%
device + Monitor ID-1 W+25.27 W 300
PC ID-3 73.71 33.3%
+ Monitor ID-12 W+29.87 W
Furnace ID-1 519.3W 40 % 40
ID-2 648.9W 60 % 60
Inductive ID-1 5050.8W 100% 100
Motor
Incandescent ID-2 59.28 W 100% 2400
Light
Background --- 15000 W --- ---

Simulation 1: At any given moment of a day, the mixing ratios are kept fixed
and the same with the above table’s values. Also, the background power is
removed. This is a very ideal case.

Using the proposed method explained in A.1, the power curve of individual
load group can be separated from the aggregated signal. In simulation 1, it is
found the separated load group power has exactly the same power has the actual
one. An example is shown in Figure A-2.This is because the mixing ratios are
constant. When solving the problem using equations such as (A.1), the exact
results can be obtained.

188
Appendix

4
x 10 Actual power of furnace group
2.5

Power(W)
1.5

0.5

0
0 500 1000 1500
Minutes

4
x 10 Decomposed power of furnace group
2.5

2
Power(W)

1.5

0.5

0
0 500 1000 1500
Minutes

Figure A-2: Example of actual power vs. decomposed power in simulation 1

Simulation 2: At any given moment of a day, a single appliance is switched

ON/OFF randomly according to its own time of use probability profile. In other
words, the instant mixing ratio is unknown. However, statistically, instant ratio
should fluctuate around the average ratio that is the pre-defined mixing ration
listed in Table A-2. This scenario is much closer to reality, under which it is only
possible to know the static mixing ratio but impossible to know the dynamic
mixing ratio at a certain given instant.

Power(W)

Minutes

(a) Non-linear load groups

189
Appendix

Actual inductive motor group

Power(W)
Decomposed inductive motor group

Minutes

(b) Reactive (linear) load groups

Figure A-2: Examples of actual power vs. decomposed power in simulation 2

As can be seen from Figure A-2, the error on power curves becomes larger,
especially for reactive load groups. Quantitative comparison can be seen in Table
A-3.

Table A-3 Load group inventory and composition

Group type Trend Actual Decomposed Energy
correlation Energy Energy Error
coefficient (KWH) (KWH) (%)
CFL group 0.9999 70.6625 68.7657 2.7
Computing 0.9999 243.8629 245.1459 0.5
device
HVAC group 0.9981 242.8680 245.3361 1.0
Inductive 0.9853 292.0204 352.5564 20.7
Motor group
Incandescent 0.9977 571.7378 525.5483 8.1
Light

190
Appendix

As can be seen from Table A-3, the above decomposed results are close to the
actual results. The error for non-linear load category such as CFL group and
computing devices is very low. However, for reactive and resistive loads, errors
are larger. This is because they do not have as plentiful information as the
harmonic spectrum. In other words, they cannot be accurately solved or optimized
by using equations like (A.1). But the results can still be used as reasonable
references. Besides, it is found that the trend shapes of all types of load groups are
very similar to the actual ones.

A.3 Suggestions for future work

Some suggestions are given for future work on load disaggregation research
for commercial buildings:

 For different types of commercial buildings, more realistic load groups

and categories can be investigated.

 More realistic simulations that consider the actual time of use probability
profiles in commercial buildings [121] can be done.

 Field data based verification should also be performed. This is the most
challenging part since the number of loads in a building is massive and
difficult to be measured directly. Energy auditing based methods may be
considered.

 The application of the proposed method may also be extended to industry

compound loads.

191