PONTIFICIA UNIVERSIDAD CATOLICA DE CHILE

ESCUELA DE INGENIERIA

NONINTRUSIVE LOAD IDENTIFICATION &

MONITORING: TECHNIQUES AND
APPLICATIONS FOR SMART METERS

NICOLÁS A. KIPREOS

Thesis submitted to the Office of Research and Graduate Studies in partial

fulfillment of the requirements for the Degree of Master of Science in
Engineering

Advisor:
DAVID E. WATTS

Santiago de Chile, June, 2011

! MMXI, Nicolás Andrés Kipreos de la Fuente
PONTIFICIA UNIVERSIDAD CATOLICA DE CHILE
ESCUELA DE INGENIERIA

NONINTRUSIVE LOAD IDENTIFICATION &

MONITORING: TECHNIQUES AND
APPLICATIONS FOR SMART METERS

NICOLAS A. KIPREOS

Members of the Committee:

DAVID WATTS

HUGH RUDNICK

GUILLERMO PEREZ DEL RIO

MARCELO ARENAS

Thesis submitted to the Office of Research and Graduate Studies in partial

fulfillment of the requirements for the Degree of Master of Science in
Engineering

Santiago de Chile, June, 2011

A mi papá y mi tía Tasa.
 ACKNOWLEDGEMENTS

Foremost, I would like to thank my family, for their unconditional support. I would

also like to thank my supervisor, Prof. David Watts, who shared with me a lot of his

expertise and research insight. A great mentor and professor, whose thoughtful advise

and time served to give direction during my M.Sc. studies.

Special thanks also to my uncle, José de la Fuente, whose house served as the

pilothouse for this investigation, and also to my little cousin, Ferni de la Fuente, who

was my assistant on the field tests.

I would also like to thank my friends at the electrical engineering department for

their support at the realization of my thesis, they are, Hernán, Kaipi, Fernando, Melisa,

Danilo, Lurys, Gabriel, Joaquín, Toño, Roberto, Camila, Francisco, José Luis and

Barros.

Finally I would like to thank to the staff at the electrical engineering department

they are, Carlos, Andrés, Betty, Gianina, Jessica, Vicky, Karina and the M.Sc. secretary

Danisa for her kind help.

iii
 TABLE OF CONTENTS
Pág.

Acknowledgements ..................................................................................................... iii

List of Tables.............................................................................................................. vii

List of Figures ........................................................................................................... viii

Abstract ....................................................................................................................... ix

Resumen ....................................................................................................................... x

1 Introduction .............................................................................................................. 1

1.1 Smart metering, consumption feedback and more ................................................ 1

1.2 Smart metering beyond consumption feedback .................................................... 2

2 NILM, NALM, NILI ................................................................................................ 4

2.1 Basic definitions .................................................................................................... 5

2.1.1 Load identification and monitoring .................................................................... 5

2.1.2 Load signature: Intrusive and Non-intrusive...................................................... 6

3 Families of Load Signatures..................................................................................... 8

3.1 PQ Clustering ........................................................................................................ 8

3.1.1 Hart’s algorithm ................................................................................................. 9

3.2 Load signature classification ............................................................................... 10

3.2.1 Steady state signatures ..................................................................................... 10

3.2.2 Transient state signatures ................................................................................. 15

4 Algorithm, Load Characterization and Field Tests ................................................ 18

4.1 4 State algorithm ................................................................................................. 18

iv
4.1.1 Measurement .................................................................................................... 19

4.1.2 Normalization ................................................................................................... 19

4.1.3 De-noising: Border detection ........................................................................... 19

4.1.4 Rule-based appliance identification ................................................................. 20

4.2 Dealing with residuals ......................................................................................... 21

4.3 Results: Appliances consumption disaggregation ............................................... 21

5 Conclusions and Discussions ................................................................................. 24

5.1 Self-learning limited by lack of intrusive measurements and limitations ........... 26

6 Future Work ........................................................................................................... 27

References .................................................................................................................. 30

APPENDICES............................................................................................................ 32

7 Appendix A: Algorithm Results............................................................................. 33

8 Appendix B: Types of Intrusiveness ...................................................................... 34

9 Appendix C: Large Scale Applications .................................................................. 37

10 Appendix D: References Comparisson ................................................................ 39

10.1 Signatures used.................................................................................................. 39

10.2 Sampling period and frequency ........................................................................ 40

11 Appendix E: Harmonic Distortion Index ............................................................. 42

12 Appendix F: Data Extraction and Equipment ...................................................... 45

12.1 Limited memory ................................................................................................ 45

13 Appendix G: Getting the Ground Truth ............................................................... 46

14 Appendix H: Pilot House Statistics ...................................................................... 48

15 Appendix I: Energy Share by Appliance.............................................................. 49

v
16 Appendix J: Inventory and Database ................................................................... 50

16.1 Kitchen .............................................................................................................. 50

16.2 Washing room ................................................................................................... 50

16.2.1 Others ............................................................................................................. 51

16.3 Total .................................................................................................................. 52

17 Appendix K: Reference Clasiffication ................................................................. 53

vi
 LIST OF TABLES

Pág.
TABLE 7.1. NILM Algorithm Results ...................................................................... 33

TABLE 16.1. Kitchen Inventory ................................................................................ 50

TABLE 16.2. Washing Room Inventory ................................................................... 50

TABLE 16.3. Rest of the House Inventory ................................................................ 51

TABLE 16.4. Total Inventory .................................................................................... 52

TABLE 17.1. Reference Classification ...................................................................... 53

vii
 LIST OF FIGURES

Pág.

Figure 1-1: Advance Metering Architecture ................................................................ 2

Figure 2-1: Non-Intrusive Load Monitoring Connection Architecture........................ 4

Figure 3-1: Cluster Map of Appliances Present in the Test Residence........................ 9

Figure 3 2: Classification of Energy Signatures ........................................................ 10

Figure 3-3: Difference between nonlinear and linear load......................................... 13

Figure 4-1: Algorithm's flow diagram........................................................................ 18

Figure 4-2: Signal without and with border detection ............................................... 20

Figure 4-3: Energy disaggregation and energy consumption .................................... 23

Figure 6-1: Current harmonics ................................................................................... 28

Figure 8-1: Types of Intrusiveness ............................................................................. 36

Figure 9-1: NILM algorithm with clustering integration ........................................... 38

Figure 10-1: Classification of Signatures that are actually used for LI ..................... 39

Figure 10-2: Sampling period classification .............................................................. 40

Figure 11-1: Current harmonics ................................................................................. 43

Figure 13-1: Sample worksheet for registration of events ......................................... 46

Figure 13-2: Comparison between EPP and MPP ..................................................... 47

Figure 15-1: Energy cake ........................................................................................... 49

viii
 ABSTRACT

During the last decade there has been an accelerated adoption of smart meters,

some of them with the potential to become the backbone of the Smart Grid. Several

meters have the potential to help utilities characterizing their clients, motivating the use

of Non-Intrusive Load Monitoring (NILM) and Identification (NILI). These techniques

have been the target of research for nearly 30 years, but few advances have been made

over the last decades. They have been characterized by very low accuracies, without

achieving real applications in the market. This paper offers a critical review of these

techniques and the most relevant literature, proposing a 4-stage steady state algorithm

capable of being implemented with numerous low frequency smart meters. The

algorithm shows good results that can be further improved with higher frequency

techniques. Some implementation and evaluation issues are also presented.

Keywords: Nonintrusive Load Identification, Load Signatures, Transient

Signatures, Steady State Signatures, Load Monitoring .
ix
 RESUMEN

Durante la pasada década ha habido una acelerada adopción e instalación de

diversos tipos de medidores inteligentes, algunos de ellos con el potencial de

convertirse en la piedra angular de las futuras redes inteligentes. Muchos de los nuevos

medidores tienen la capacidad de ayudar a las empresas de distribución a caracterizar a

sus clientes, motivando así el uso de tecnicas de detección de carga no intrusiva (NILI)

y monitoreo no intrusivo (NILM). Estas técnicas han sido objeto de investigación por ya

cas 30 años, pero pocos avances se han realizado al pasar los años, los cuales se han

caracterizado por muy baja presición y ninguna aplicación seria a nivel de mercado.

Esta investigación ofrece revisión crítica de las técnicas propuestas en la literatura

existente, además de proponer un algoritmo de 4 etapas facilmente implementable en

medidore inteligentes de baja frecuencia. El algoritmo muestra buenos resultados que

pueden ser mejorados con técnicas de mayor frecuencia de muestreo. Algunos

problemas de implmentación y evaluación también son presentados.

Palabras Claves: Detección de carga no intrusiva, monitoreo de carga no intrusivo,

Firmas de carga, Monitoreo de carga, Medición inteligente.

x
 1

1 INTRODUCTION
The current energy supply system is undergoing fundamental challenges. Systems

are facing strong pressure to increase energy efficiency, accommodate large amounts of

renewables, reduce local and global emissions, minimize their footprint, and at the same

time become more cost effective and reliable. Several smart grid technologies are

expected to help meet these demands in the future, and those based on smart metering

technologies are among those with the largest potential in the near term.

1.1 Smart metering, consumption feedback and more

Smart metering technologies promise to reduce residential energy consumption

between 5 and 10% and to cut down CO2 emissions by around 2.5 million tons per year

in the U.S. Advanced meters have the ability to effectively show the consumption

information to customers so they can manage their consumption patterns and save

energy (Houseman, 2005). More advanced features on meters can push these figures up

considerably. The advanced metering infrastructure (AMI) offers the possibility of

services such as demand side management (DSM), the realization of virtual power

plants VPPs (van Gerwen, Jaarsma, & Wilhite, 2006), and many others, having the

potential to boost reliability, while smoothly integrating large amounts of distributed

resources without compromising reliability, dealing economically with non-

dispatchable natural resources such as wind and solar.

Smart metering plays an important role in future power distribution networks for

the efficient and effective management of distributed power resources, and for the

control of smart houses and offices (Houseman, 2005). From telemetry to active control
 2

of resources and demand, smart metering is the door to the future of the distribution

networks. It allows the operator to manage the grid more efficiently, reduces costs and

emissions, and offers an effective way of performing demand side management by only

changing the architecture of the network to a more communicated an automated one, as

shown in Figure 1-1.

Figure 1-1: Advance Metering Architecture

1.2 Smart metering beyond consumption feedback

Smart metering technologies offer the potential ability for the utility company to

characterize and know better its clients. For example, a utility could estimate (or even

determine) which clients have a pool, and sell effective advertising on the monthly bill

for pool products. Also, they could offer or recommend distributed generation solutions

according to the consumption profile of the client to help him save money. Moreover,

the utility could measure demand price elasticity of each client type (or even individual
 3

clients) in the long and short run and then assess the sources of those changes in

consumption behavior. This could be pushed even harder and assess demand price

elasticity associated with specific residential appliances and determine how certain

policies or events could affect the energy consumption in a residential client or client

type.

While the potential for these technologies goes well beyond improving load

profiles, estimating elasticities, detecting appliance disconnection and scheduling,

pushing energy efficiency and residential energy management (Sawyer, Anderson,

Foulks, Troxler, & Cox, 2009), this article has a more limited but fundamental scope. It

focuses on the load characterization through nonintrusive load monitoring and

identification, as these techniques are needed to perform all others accurately, requiring

urgent further development.

The structure of the rest of the thesis is the following: Section 2 provides some

basic definitions; Section 3 shows the different families of load signatures and their

classification. Section 4 presents our proposed methodology and its application to our

field test, Section 5 contains the conclusion and Section 6 contains some

recommendations for future work on the subject.

 4

2 NILM, NALM, NILI

While smart meters have been widely deployed in several areas, getting good use of

their data and deriving meaningful information from it is still a pending task.

Techniques such as non-intrusive load monitoring (NILM), non-intrusive appliance load

monitoring (NALM) or non-intrusive load identification (NILI) have great potential

nowadays. They all are based on a principle called disaggregation, in which the

constituent parts of the load are extracted from an aggregated or total load signal, as

shown in Figure 2-1. This total contains the components from all the devices or

appliances within a client’s total load1. This disaggregation process is quite simple in

theory, but the diversity of appliances, their different uses and consumption patterns, the

limitations on the meters’ capabilities and on their communication and data centers

provides pending challenges such as how to perform accurate NILI and how to perform

cost-effective and accurate NILI?

Figure 2-1: Non-Intrusive Load Monitoring Connection Architecture

Non-intrusive techniques yield detailed information about the load’s energy

consumption without measuring each end device directly, so that fewer sensors are

1
While loads can be residential, commercial or industrial, residential ones are simpler to deal with, and provide a large set of
smart meters to take advantage of, but they also are the ones with the least capable smart meters.
 5

needed and monitoring, storage and transmission gets cheaper (Marceau & Zmeureanu,

2000). Another advantage of nonintrusive approaches is that consumers are not

necessarily aware that they are being monitored, so psychological effects don’t affect

their consumption profiles.

The basic principle for data collection is measuring the waveforms or RMS values

of both voltage and current, and computing other variables of interest from them

(real/reactive power, power factor, etc.). Obtaining accurate and frequent measurements

of these parameters is crucial for the identification of the load. The sampling frequency

and the level of detail of the signal (waveform or rms value) will depend on the kind of

algorithm to be implemented; the goal is often to do more with less resources.

2.1 Basic definitions

Some basic definitions are introduced next. They are necessary to allow for

understanding and classifying the methodologies presented in the literature review.

2.1.1 Load identification and monitoring

While load identification (LI) and load monitoring (LM) may seem to be

equivalent, they are not and it is worth understanding the difference:

• LI and State: The identification of a load, e.g. an electrical appliance in a

house, is the ability of a system to successfully detect the state (that could

be on/off or an intermediate state in the operation of the load) in which the

appliance is operating.
 6

• LM and Consumption: The monitoring of the load is the following of all

the loads or appliances inside an aggregated load and the estimation of the

energy consumed by each one of them, and of all the loads together.

They are closely related, as load monitoring has to use load identification to

estimate the energy consumed by each one of the appliances. Load monitoring can also

be used for characterizing load types in different profiles (Gavrilas, Sfintes, & Filimon,

2001), which could be very useful for utility companies and load forecasters.

2.1.2 Load signature: Intrusive and Non-intrusive

A load signature is a characteristic that makes a load unique (e.g. real power

consumption, in-rush current, admittance, etc.). Signatures can be classified according

to their sources (or the level of disruption required for observing the signature) into two

families, these are intrusive and nonintrusive.

Intrusive signatures require significant levels of intrusion or distortion at the

client’s facility, and they can be further classified according to the media used to assess

the signature into two families: physically intrusive and electrically intrusive.

Physically intrusive signatures can be generated by a sensor located on the electric

cord of each of the appliances in a residence or by placing presence-activated sensors in

the rooms of a house or also the generation of a unique signal that is emitted by a circuit

present in each appliance at the moment it is turned on. An electrically intrusive

signature is, for example, the injection of a signal at the energy meter to observe the

system response signal. Evaluating the distortion of the signal could give useful

information on the appliances that are operating at that time (Hart, 1992; Marceau &

Zmeureanu, 2000).
 7

Nonintrusive signatures are those signatures that don’t require intrusion or

disruption into clients’ physical and electrical facilities. There are two main families of

nonintrusive (electrical) signatures: transient and steady state ones. The former refers

to short-run transitory changes in consumption patterns associated to an electrical

appliance, while the latter refers to their steady state changes; this is when transient

changes have died out. These signature families are described in more detail next.
 8

3 FAMILIES OF LOAD SIGNATURES

Load signatures can be obtained by intrusive and nonintrusive methods. George W

Hart (Hart, 1992) was the pioneer of nonintrusive load monitoring (NILM). He

introduced the concept of load signature in the mid 80’s and developed a nonintrusive

appliance load monitor (NALM) aiming to determine the energy consumption of every

appliance inside a house. His algorithm is based on the evaluation of the steady state

change of real and reactive (admittance-corrected) power.

3.1 PQ Clustering

Every appliance has a positive and a negative real power variation, which

corresponds to it being either turned on or turned off. This is shown in Figure 3-1 for

some of the appliances we tested. Other appliances, such as a washing machine, have

multiple state changes in their cycles, and because of that do not lie fixed in the PQ

map, but cycle on it instead. This is why they are not shown in the PQ cluster map,

where they would show a denser state trajectory or cycle. This simplified cluster map

shows a group of 8 appliances, in which resistive loads, such as an iron or water heater

are near the real axis (active power axis), while devices with some reactive nature (due

to motor use for instance) move away from the real power axis, adding a significant

reactive power dimension.

 9

Figure 3-1: Cluster Map of Appliances Present in the Test Residence

3.1.1 Hart’s algorithm

Hart’s algorithm constructs on this idea. Steady state is defined (in his algorithm) as

three or more data intervals where consumption does not vary more than 15W or

15VArs. The values within the steady state period are averaged for minimizing the

effect of noise on the signal. A step change in this context is a change in the power

consumption of the appliance with respect to two consecutive data samples. Hart

proposes that step changes should be grouped into clusters, which contain the variations

of the power consumption of each individual appliance.

Since the first applications of nonintrusive load monitoring and identification,

different approaches have been proposed. The general basis for most of them follows a

simple, but economic and meaningful steady state approach as proposed by Hart. A

more formal and detailed classification is provided next.

 10

3.2 Load signature classification

Hart (Hart, 1992) made an initial classification of load signatures in the year 1992.

We extended/updated it to include two additional signatures, steady state shape and

distortion index ones. The following figure shows this classification of energy

signatures, where the main two families are steady state and transient ones, as we

explain and compare next.

Figure 3-2: Classification of Energy Signatures

3.2.1 Steady state signatures

The steady state (SS) signatures are based on changes of the power consumption

level (state) of a particular appliance in the steady state. These are mainly due to turning

on or turning off the appliance, or due to energy changes in its cycle. Steady state

changes are very informative of the nature of the load and they are easier to detect than
 11

a transient change. The sampling frequency of the SS meter needs to be far less than

the one targeted to measure transient signatures. Another advantage of steady state

signatures over transient state ones is that they are additive, that is, if two appliances of

2kW connect, then the total power consumption is 4kW. Transient state signatures are

not additive.

3.2.1.1 Fundamental frequency signatures

The measurement of the power network’s voltage and current at a fundamental

frequency of 50Hz (or 60Hz) can lead to obtaining real, reactive and apparent power, as

well as the load’s admittance.

Due to local legislation, often the variation in a low voltage network cannot exceed

±10% of the nominal voltage, which is 220V (110V in other areas). Therefore, in linear

loads a ±10% variation in the current and, consequently, a ±20% variation in the power,

is possible. Hart proposes that instead of using a signature based on the power, which

relies too much on the voltage, this signature should be corrected by the admittance,

which is a voltage invariable feature of the load. Unfortunately, the load’s admittance is

an unfamiliar characteristic if one lacks engineering intuition, so Hart proposes the

admittance-corrected power as a signature, as shown in (1):

2
! 220 $
PNorm (t) = 220 2 Y (t) = # & P(t) (3.1)
" V (t) %

This correction can be used for both active (P) and reactive power (Q). It can be

seen that (1) is only the admittance correction by a constant scale factor, where the

normalized watts and vars reflect the changes in impedance over time normalized to a

constant ideal voltage supply (Cole & Albicki, 1998b). The quadratic expression is
 12

only true when assuming that the load is linear, which is unlikely because most of the

loads in a residence are of a nonlinear nature. In the case of nonlinear loads, the formula

is corrected by a ! factor (2), which doesn’t fit nonlinear loads perfectly but it makes a

better approximation than a quadratic exponent. The problem with this is that ! factors

for P and Q must be calculated isolating every device preferably in a laboratory, which

is not possible most of the time. Due to this difficulty in determining ! factors, the

admittance corrections are usually done using a quadratic exponent or extrapolating

factors estimated for similar loads.

!
! 220 $
PNorm (t) = # & P(t) (3.2)
" V (t) %

Although the effects of this normalization do not make a notable difference for the

visualization of the signal, they help to clean the signal from noise effects, especially in

moments where there is a large amount of step changes, as in the operation of the

washing machine, which has many cycles due to the irregular operation of the motor.

A linear load, such as a water heater, has a linear relationship between the voltage

and the current, while a nonlinear load, such as a microwave, has a relation that

describes a trajectory particular or specific to that device, and where the current doesn’t

follow the voltage linearly. In the case of the microwave, for instance, the relation

between the voltage and the current describes something similar to an elliptical cycle

around the origin. Our measurements comparing a linear and nonlinear load are shown

in Figure 3-3.
 13

Figure 3-3: Difference between nonlinear and linear load

The most common fundamental frequency signature is the PQ signature, which

corresponds to the variation of real and reactive power in the steady state operation of

the specific appliance. They use a slow sampling frequency and they are relatively

cheap in the use of computational resources. The sampling frequencies can go from 1

PQ sample per second to 1 PQ sample per minute, being 1 second and 16 seconds (used

in (Farinaccio & Zmeureanu, 1999; Marceau & Zmeureanu, 2000)) the most common

choices in the literature. On the other hand, transient signatures are very demanding

regarding computational resources: they usually capture 256 (Chang, Lin, & Lee,

2010) or 128 (Shaw & Laughman, 2007) samples per cycle, which require sampling

frequencies between 6kHz and 13kHz.

 14

3.2.1.2 Harmonic current signatures

Harmonic signatures provide additional information when used together with

fundamental frequency signatures and may be useful for increasing the algorithm’s

effectiveness. Most residential appliances have non-linear characteristics, which

generate significant third, fifth and seventh harmonics. These harmonics can be easily

identified by specialized sensors for particular frequencies and then treated as a steady

state signature. This kind of signature can be very useful for detecting small appliances

such as low power bulbs (Laughman, 2003).

In this area AL-Kandari (Al-Kandari & El-Naggar, 2006) proposes a recursive

algorithm for the identification of harmonic admittance of the load. The advantage of

this method is that it offers a fast estimation of the harmonic admittance even if the

harmonic content varies with time. The disadvantage of this method is that it is based on

complex signal analysis in which the parameters must be calibrated to a particular case.

In Figure 6-1 we show the harmonic content of measurements we obtained from

different appliances. It is clear that it can offer a very good tool to differentiate between

them.

3.2.1.3 DC currents
DC consumption is another non-linear feature that makes an appliance unique. For

example, some blenders have a diode for speed control (only half wave) and they

generate a DC component that can be detected. However, Hart points out that they

haven’t encountered yet an appliance with a significant DC power flow that could be of

interest to utilities.
 15

3.2.1.4 Shape & Others

There are other types of signatures that can provide information for differentiating

between appliances, but they are rather specific. One example could be a 1Hz ripple due

to the cyclic reversal of the tub in washing machines. This would be a steady state shape

signature.

3.2.2 Transient state signatures

In addition to steady state approaches, others (Leeb, 1993; Shaw & Laughman,

2007; Shaw, Leeb, Norford, & Cox, 2008), have used transient approaches to NILM.

These algorithms are more complex and require more computational resources, but they

have demonstrated to be very helpful by providing valuable extra information that can

be used together with steady state signatures to improve the algorithm’s effectiveness.

Transient signatures are more difficult to detect and require more complex and

expensive data acquisition systems. The objective in using transient signatures is to help

differentiate between two very similar loads to each other (Chang, et al., 2010; Leeb,

1993) or identifying low power consumption devices (Leeb, 1993). They may be useful

for differentiating between two appliances that have similar power consumption in a

steady state, but may look quite different during the transient period, for instance having

different in-rush currents. However, a transient signature would not be useful to

determine which of the two appliances turned off, because transients are usually present

when appliances are turned on only. Hart classifies transients in three main categories:

a) Flat character with a sudden step power drop to the steady state operating

level.
 16

b) Large initial increase in the power, followed by an exponential drop lasting

several seconds.

c) Transients lasting a few cycles, which are transients as defined in the linear

circuit theory.

There are more parameters to characterize transients, like time constants, duration,

size or parametric variables in models, which can be fitted to the observed waveform.

Nowadays, a combination mainly composed of shape, size and duration of the transient

is applied to recognize loads. As seen in (Chang, 2010; Leeb, 1993; Shaw, et al., 2008),

where the measured wave transient is fitted to a stored transient wave, and the

waveform that presents the minimum quadratic error with the measured one is assigned

as the most likely one and is chosen.

Chang (Chang, 2010) states that the main problem with power signatures is that

many devices have different consumption levels, like a refrigerator, which has only one

load, a compressor, but it has different physical components for refrigerating and

freezing. He states that the envelopes of transient behavior when appliances are turned

on could offer a better tool for discriminating between similar loads (Chang, 2010;

Chang, et al., 2010; Leeb, 1993; Shaw & Laughman, 2007; Shaw, et al., 2008), which

is very typical in industrial loads. Current in-rush’s when appliances are turned on offer

a very unique and distinguishable parameter to accurately differentiate between two

very similar steady state loads (Chang, 2010).

Some approaches to accommodate the power transient signal to a signal that is

easier to process, obtain the envelope of the signal through signal processing, which is

sometimes complex and requires additional information about the load, like parametric
 17

models of some appliances that are present in it (Shaw & Laughman, 2007; Shaw, et al.,

2008). Shaw (Shaw & Laughman, 2007), proposed a Kalman-filter approach to

compute spectral envelopes of current waveforms, but is complex and uses excessive

computational resources.
 18

4 ALGORITHM, LOAD CHARACTERIZATION AND FIELD TESTS

We developed and applied a steady state fundamental frequency algorithm, being

the most traditional family of algorithms, well documented by Hart and others. The

benefits of this are its low cost, low computational requirements and simplicity. These

are ideal characteristics to combine it with low cost smart meter deployments in

developing countries.

4.1 4 State algorithm

The proposed algorithm builds strongly on Hart and Farinaccio and is divided into

four main stages; these are measurements, normalization, de-noising and appliance-

specific algorithms and rules. After that, individual results are summed up and the total

aggregated consumption is estimated and eventually compared with the original

measurements. Figure 4-1 shows the flow chart of the implemented algorithm.

Figure 4-1: Algorithm's flow diagram

 19

4.1.1 Measurement
The first step of the algorithm is the measurement, which is done at the utility

meter, where the aggregated power demand of the residence is acquired.

4.1.2 Normalization
The following stage is the normalization, where the acquired data is admittance

corrected by means of (3.1). This is well described in our literature review.

4.1.3 De-noising: Border detection

The third stage is a de-nosing stage that is performed in order to clean the data from

its power transients, as they make the signal dirtier, complicating the identification

process; this is done by applying (4.1). This softening of the signal must be performed

before applying appliance-specific rulings onto the normalized aggregated

measurement. This process is realized by an algorithm called “border detection”. A very

useful border detection algorithm can be found in (Farinaccio & Zmeureanu, 1999).

This reference proposed that instead of analyzing the variations of power (!P and !Q),

the sum of the differences between adjacent data should be analyzed, as shown in (4.1):

xi = Pi ! Pi !1
(4.1)
step = xi + xi +1 = Pi +1 ! Pi !1

As seen in (3), the term Pi is eliminated which proves to be very useful for

eliminating steady state in-rushes, which contaminate the signal. This is shown in

Figure 4-2, which corresponds to the comparison between the actual power

consumption signal on the test residence with and without the border detection

algorithm applied.
 20

Figure 4-2: Signal without and with border detection

4.1.4 Rule-based appliance identification

Finally, individual algorithms are applied to each one of the appliances present in

the residence. We developed and applied special or appliance-specific rules and

thresholds that work in different ways for the different appliances. These rules are

derived from careful observation of consumption patterns and appliance usage patterns

and can be further improved on by simple automated learning techniques.

For all appliances there are thresholds for variations of active and reactive power

consumption, !P and !Q. Some of them also have thresholds for the length of

continuous use (e.g. a few minutes for a toaster, water heater, several minutes for a dish

washer, etc) or the timing of its use (toaster typically used in the morning and

afternoon).
 21

The application presented here did not derive timing specific rules for pool motors,

ironing, TV, dishwashers, etc., to avoid achieving a 100% identification based on

studying a specific subject, rather than facing a more general subject. However, it is

possible (and we are working on it) to develop web-based personal and shared

databases on appliance use on order to achieve extremely high identification levels. For

instance, the household of study has lunch at 13:30 hrs and runs the dishwasher at 15:00

hrs after lunch time. We didn’t factor in this valuable information, because other

families could wait until late at night to run the dishwasher.

Other appliances such as clothes washing machine pose larger challenges as they

have multiple customizable complex cycles (wash, rinse, centrifuge, hard centrifuge,

and cycles of different length, power and sequences). In this case, salient features of the

cycle are detected and the signal is reconstructed.

4.2 Dealing with residuals

The consumption not recognized by this algorithm can further be allocated to

specific appliances through a maximum likelihood stage. Others, like Farinaccio, do this

by using a scoring system instead. Here we don’t present this allocation stage because

the proposed algorithm provides good results, thus, making this approximation

unnecessary. We also need a measure of performance for the overall algorithm and the

residual is the best available measure.

4.3 Results: Appliances consumption disaggregation

The application of the proposed algorithm provides good results. While we didn’t

have records, sensors and data registers for about a hundred connected devices to know
 22

the ground truth or actual individual consumption, we did have access to the total

consumption measurement. We found an aggregate error, or unexplained consumption

of approximately 2%.

We classified appliances into 17 different major appliance families which are:

refrigerator, lighting, garage door, TV, dishwasher, sprinkler, Nintendo wii, pool motor,

hair dryer, iron, clothes washer, microwave, toaster, water heater, computer, water and

ice from the refrigerator and clothes dryer. Figure 4-3 shows power consumption from

each family over a sample day.

After the disaggregation among appliance families is done, it is possible to compute

the distribution of energy consumption between them at any point of time or based on

overall energy usage, as shown in Figure 4-3. The two main appliances in power

consumption are the refrigerator and the pool motor.

According to the observers, the algorithm performed very bad on TV detection,

probably because of its low power consumption and irregular usage. For the iron and

computer some observed events were not detected. Again, low power consumption,

irregular usage, together with no additional or specific ruling in the detection algorithm

provide the basis for the errors. However, larger devices are nicely detected as well as

others where specific rules where designed, achieving great results in terms of energy

error (2%).

Further calibration of the algorithm to the specific household would improve this

error, but the algorithm was not conceived only for that specific house, but for a more

general house. It is not meant to show 100% by over-calibrating the algorithm. A

disadvantage with this algorithm is that it is very accurate for the cases for which it is
 23

calibrated, but in other cases it may achieve lower accuracies. For example, an

algorithm calibrated to achieve a high degree of accuracy in one house would provide a

worse result when applied to the neighbor’s house. Keeping high accuracy requires

using the house-specific equipment inventory and building a database to store load

signatures of a variety of equipment from different homes. This means applying the

algorithm to different houses, calibrating models to different appliances, saving the data

into a database and using that data in other houses.

Figure 4-3: Energy disaggregation by appliance families and energy consumption by

appliance family
 24

5 CONCLUSIONS AND DISCUSSIONS

Residential smart meters, especially those used in developing countries, have

limited computational power and communication capabilities. It is customary to obtain

most of them, and techniques such as NILM and NILI have great potential to be

implemented on them, to characterize clients, their appliances use and their

performance. This paper presents a critical review of the different families of methods

used for detecting, identifying and monitoring electric loads; starting from its pioneer

and most cited work, Hart (Hart, 1989, 1992), to the newest approaches using steady

state and transient analysis.

We find that the most common way to detect and identify loads is to use a steady

state approach, but in the most contemporary literature, it is suggested to combine it

with others, such as transient analysis, in order to improve the algorithm’s performance.

We conclude that steady state algorithms must be the base for any NILM or NILI

development, as they are low cost and require minimal computational power and

metering frequency, being compatible with several smart meter deployments. However,

there is still a long way to go to obtain accurate and reliable steady state algorithms.

Transient algorithms are useful and have the potential to improve performance when

used together with steady state techniques, but impose higher costs and measuring and

computational burdens, not always attainable for basic smart meters.

From the literature review we conclude that most of the algorithms have been

applied in very special and favorable environments, either in a bundle of a few loads, in

simple test houses in which no information about the inside appliances is given to the

reader, or in laboratories with too few appliances, which are very easy to differentiate
 25

and therefore to disaggregate. Literature is quite limited (nearly empty) on the

application of algorithms to more challenging setups such as the one proposed here.

In non-ideal environments, like a total aggregated load inside a large house, it is

very difficult to identify small appliances that are very alike between each other, like

illumination devices, such as light bulbs or high efficiency light bulbs. Due to this,

errors in the identification of low consumption appliances are very high, but

surprisingly they aren’t too high in monitoring. This is mainly because the identification

of bulk energy consumption appliances, like the pool motor or the washing machine is

more precise than that of appliances that don’t have a high energy consumption, and

also because the errors in monitoring of low power appliances tend to have a cancelling

tendency, i.e., while some low power appliances are estimated to consume more than

they do, others are estimated to consume less.

Even though it seems very easy to achieve accurate load disaggregation (after all

it’s just adding up and subtracting), parameters such as the nonlinearity of the loads, the

simultaneous operation of appliances, loads with variable cycles, electric noise

produced by the network or the same appliances (that alter the measurements and the

process of disaggregation), and many other issues, complicate the analysis considerably.

This explains why we found no reference on these techniques applied to industrial

loads, where all these issues scale up considerably.

Besides a complete comparative and critical review of the existing literature and

their proposed algorithms, some improvements over these algorithms are proposed and

the new algorithm is implemented in a challenging environment. The approach selected

by us is a nonintrusive steady-state rule-based detection algorithm. In our work we

 26

applied a steady state algorithm enhanced with a system of specific rules to individual

appliances, proving to be much more effective than applying a general algorithm for

detecting all the appliances at once. We achieved a very high accuracy in a house

chosen for its complexity and variety of loads.

5.1 Self-learning limited by lack of intrusive measurements and limitations

Another challenge with these techniques is that in the implementation of a self-

learning network for NILM, obtaining validation data isn’t as simple as it looks. If the

washing machine is identified, there are no individual sensors that could help us

determine if the washing machine really entered in operation. Although we could

determine if it is really the washing machine by waiting for the turn off event, if we

missed the turn on event, we wouldn’t have a teacher for training the network. That

seems to be one of the main challenges in automatic setup of NILM. Invasive

approaches in some test houses could serve as teachers for self-learning networks.
 27

6 FUTURE WORK
The next steps for this project are a) embedding this algorithm into the smart meters

system developed by our group (HappyVolt.Com), b) developing more powerful

algorithms that allow us to scale up into more numerous and complex loads in a single

wire. The proposed algorithm is quite capable of facing complex industrial loads, but

applying it to each individual feeder and adding up the results, instead of applying it to

the connection to the utility. c) Developing a database and multiple houses and trying

the algorithm with houses out of the sample. Our group is applying for funding for this.

A transient state envelope and harmonic distortion index (HDI) detection

algorithms were also explored and evaluated, and although they are not fully described

here due to space limitations. It can be seen from Fig 9 that harmonic content can offer

a useful tool to discriminate between high efficiency and normal light bulbs and several

other different appliances. We propose a simple HDI based on the RMS value of the

normalized harmonic indexes (ai) to take advantage of this.

12

#k "a / a
i i 1
HDI NILM = n=2 ki # 1 "i !{2,...,12} (6.1)
(n ! 1)
 28

Figure 6-1: Current harmonics: (a) Microwave, (b) Vacuum Cleaner, (c) High
Efficiency Light Bulb and (d) Normal Light Bulb

We find preliminarily that envelope and HDI could become a critical tool for future

discrimination between similar loads in a more complex environment, such as the

industrial one. However, in our test bed, results were already very good and assessing

the potential of additional algorithms has limited use. Its serious assessment requires a

more challenging test bed.

A proposed solution for increasing NILM accuracies, by cleaning the signal, is to

measure from the electric panel, separated into the individual circuits instead of

measuring at the main meter. Just isolating the illumination of the residence from the

other appliances would greatly improve the algorithm’s accuracy. With this the

differentiation between different load families could be much more effective. Our recent

meter ekeeper (Happyvolt.Com) is implementing this.

 29

Finally, we see that the weakest link in NILM techniques is training, so future

studies should focus on how to implement systems that require less training or on

systems that are able to be learn by themselves, based on artificial neural networks

(ANN) or on regression models. Having the training and self-learning issues resolved

would help develop viable commercial products for NILM.

 30

REFERENCES

Al-Kandari, A. M., & El-Naggar, K. M. (2006). Recursive identification of harmonic

loads in power systems. International Journal of Electrical Power & Energy Systems,
28(8), 531-536.
Berges, M. E., Matthews, H. S., & Soibelman, L. (2010, 17-19 May 2010). A system
for disaggregating residential electricity consumption by appliance. Paper presented at
the Sustainable Systems and Technology (ISSST), 2010 IEEE International Symposium
on.
Chang, H.-H. (2010). Load identification of non-intrusive load-monitoring system in
smart home. WTOS, 9(5), 498-510.
Chang, H.-H., Lin, C.-L., & Lee, J.-K. (2010). Load identification in nonintrusive load
monitoring using steady-state and turn-on transient energy algorithms. Paper presented
at the Computer Supported Cooperative Work in Design (CSCWD), 2010 14th
International Conference on.
Cole, A. I., & Albicki, A. (1998a). Algorithm for nonintrusive identification of
residential appliances. Paper presented at the Circuits and Systems, 1998. ISCAS '98.
Proceedings of the 1998 IEEE International Symposium on.
Cole, A. I., & Albicki, A. (1998b, 18-21 May 1998). Data extraction for effective non-
intrusive identification of residential power loads. Paper presented at the
Instrumentation and Measurement Technology Conference, 1998. IMTC/98.
Conference Proceedings. IEEE.
El-Ferik, S., Hussain, S. A., & Al-Sunni, F. M. (2006). Identification and weather
sensitivity of physically based model of residential air-conditioners for direct load
control: A case study. Energy and Buildings, 38(8), 997-1005.
Farinaccio, L., & Zmeureanu, R. (1999). Using a pattern recognition approach to
disaggregate the total electricity consumption in a house into the major end-uses.
Energy and Buildings, 30(3), 245-259.
Gavrilas, M., Sfintes, V. C., & Filimon, M. N. (2001, 2001). Identifying typical load
profiles using neural-fuzzy models. Paper presented at the Transmission and
Distribution Conference and Exposition, 2001 IEEE/PES.
Hart, G. W. (1989). Residential energy monitoring and computerized surveillance via
utility power flows. Technology and Society Magazine, IEEE, 8(2), 12-16.
Hart, G. W. (1992). Nonintrusive appliance load monitoring. Proceedings of the IEEE,
80(12), 1870-1891.
Houseman, D. (2005). Smart Metering. Refocus.
Laughman, C. R. L., D; Cox, R.; Shaw, S.; Leeb, S.; Norford, L.; Armstrong, P. (2003).
Advanced Nonintrusive Monitoring of Electric Loads: California Energy Commission.
Leeb, S. B. L., M. S.; Kitley, J. L.; Sweeney, J. P. (1993). Development and Validation
of a Transient Event Detector. AMP Journal of Technology, 3, 69-74.
 31

Marceau, M. L., & Zmeureanu, R. (2000). Nonintrusive load disaggregation computer

program to estimate the energy consumption of major end uses in residential buildings.
Energy Conversion and Management, 41(13), 1389-1403.
Sawyer, R. L., Anderson, J. M., Foulks, E. L., Troxler, J. O., & Cox, R. W. (2009, 20-
24 Sept. 2009). Creating low-cost energy-management systems for homes using non-
intrusive energy monitoring devices. Paper presented at the Energy Conversion
Congress and Exposition, 2009. ECCE 2009. IEEE.
Shaw, S. R., & Laughman, C. R. (2007). A Kalman-Filter Spectral Envelope
Preprocessor. Instrumentation and Measurement, IEEE Transactions on, 56(5), 2010-
2017.
Shaw, S. R., Leeb, S. B., Norford, L. K., & Cox, R. W. (2008). Nonintrusive Load
Monitoring and Diagnostics in Power Systems. Instrumentation and Measurement,
IEEE Transactions on, 57(7), 1445-1454.
van Gerwen, R., Jaarsma, S., & Wilhite, R. (2006). Smart Metering.
APPENDICES
 33

7 APPENDIX A: ALGORITHM RESULTS

We did have a couple of observers that manually registered the events (not the

consumption) and found no error on the pool motor and cloth drier. When higher

frequency and computational power is available, the accuracy for other appliances can

be incremented.

The results of the algorithm were very varied. Fortunately, the identification and

monitoring of appliances that translate in large amounts of energy consumption is far

better and precise than appliances that signify a low power consumption. TABLE 7.1

shows some of our results. It is important to say, that, since the comparison in all the

cases (except for the total) were realized between the algorithm estimated and the

surveillance estimated, we don’t have an actual precise number of the algorithm’s

accuracy by appliance, and in some cases the algorithm was even more precise that the

people surveillance, allowing us to find data that was lost in our surveillance (e.g. Iron).

TABLE 7.1. NILM Algorithm Results

Appliance Manual Error Estimate Real Error*
Dish Washer 0.0943% ?
Refrigerator 0.5358% ?
Washing Machine 14.998% ?
Pool Motor 0% ?
Cloth Dryer 0% ?
Microwave 14.706% ?
Computer 14.636% ?
Iron 216.67% ?
Toaster 28.56% ?
Kettle 33.985% ?
TV 42.723% ?
Total - 2.1%
*Real error is the relation between algorithm estimated and actual appliance measure, which requires
invasive approach, which we didn’t do, we only measured the total power consumption at the main meter,
with an error of only 2.1%. People showed to be bad at performing estimations and surveillance of appliances
usage.
 34

8 APPENDIX B: TYPES OF INTRUSIVENESS

We can classify the intrusiveness and the requirements of a load identification

monitor in four main quadrants, which are delimited by two main limits, if the system is

intrusive or not, which corresponds to the base differentiation, and if the kind of

intrusion is physical or electrical. Physical intrusive signatures are the most intrusive

type of signatures, the consumer notices them, and they can eventually produce

discomfort. They are usually characterized by the presence of sensors, such as

proximity sensors or by devices attached to the power cords of some appliances, called

tags. These tags would allow identifying appliances such as a vacuum cleaner,

regardless the part of the house in which where they are plugged. They could be very

useful to differ between similar appliances or as a teacher for a self-learning algorithm.

Lots of equipment is needed at the consumer premises and an intrusive stage is needed

for the installation of sensors and tags. On the other hand, electrical intrusive monitors,

is a less intrusive approach in which an electric signal is injected at the utility premises

and the variation of that signal at the customers residence is an indicator of the number

and type of appliances in operation at the moment. This electric signal could be a

current harmonic or a transient, however they are a concern to utilities, because they

could compromise the quality of service. In contrast with physically intrusive monitors,

few hardware is needed, in fact, only a smart meter with the ability of generating and

injecting a signal into the residence is needed. In the case of nonintrusive electrical

monitors, there are many approaches, but they all have the same basic characteristics,

which are that a load disaggregation algorithm is applied to the consumption and it

separates it into it’s main appliances, without having much information about what is
 35

happening at the interior of the house. For this, an intelligent meter with advanced

processing and communications capabilities is needed. Finally, there are some unusual

practices that could be applied for making more accurate estimations of what is

happening at the interior of a house. Physically nonintrusive practices could include the

analysis of environmental data, such as ambient temperature for estimating the

feasibility of occurrence of an event, like turning the AC on a snowy day. Also, satellite

images, combined with image processing algorithms could be used for determinate if a

house has a pool, which would explain the operation of a motor in the residence or

thermal cameras could help determine which are the weakest links in a consumption

and what makes it operate do inefficiently. The main disadvantage is that in many cases

interpretation is needed, for example satellite images are some times damaged and a

blind spot could be interpreted as a pool, also, they provide limited information, because

we only realize once that a residence has a pool, and the ability of knowing doesn’t tell

us what is the operating time. Figure 8-1 shows a classification of the types of

intrusiveness in the monitoring.

 36

Figure 8-1: Types of Intrusiveness

 37

9 APPENDIX C: LARGE SCALE APPLICATIONS

Our main concern in this investigation was how to apply this technique in large

scale. Even though our application was a one case, test residence scenario, the further

application to many houses could be achieved by well known and used today’s

practices. Clusterization of different residence types is done by utilities today in order to

characterize and group different client types. This same clusterization could be applied

to help the setup of a NILM system, or to train and teach a self-learning network. For

example, a self-learning network could detect and learn that there are new appliances

present in the residence due to replacement. Even more, the clusterization of clients

could help respond not only to those specific changes in a house, like replacing old

appliances, but also detect structural changes, like, learning from a new inhabitant in the

house, with completely different consumption pattern, by simply comparing his profile

to a general database, obtained from laboratory data of all the appliances (refrigerators,

hair dryers, kettles, etc.) available in the market, or by comparing it to a database

obtained from clustered type clients. In brief, we can characterize the profiles into three

main databases, a general, which contains all the possibilities available in the market, a

clustered, which is obtained from different client types in function of their socio-

economical, geographical, educational characteristics and a local database that has to do

specifically with the consumption of the client and his life habits.

Database information for large-scale implementation can be very useful at the time

of setting rules for appliances algorithms based on the lifestyle and habits of a particular

inhabitant. For example we can characterize habits such as toasting bread all mornings,

so when the client stops toasting his bread we will know that something is happening,
 38

maybe the toaster was changed, the client went on vacation, or the house became

abandoned.

Figure 9-1 shows the way in which the algorithm by being complemented by these

databases could achieve to reduce the error and achieve 100% accuracy for load

monitoring.

Figure 9-1: NILM algorithm with clustering integration

 39

10 APPENDIX D: REFERENCES COMPARISSON

This section organize the references critically, first addressing the methodologies

chosen by the authors, then the sampling period if mentioned and finally summarizing

these two and presenting the scale of the field tests and its accuracy.

10.1 Signatures used

Even though there is a large number of families of load signatures, as defined by

Hart, almost all methodologies followed by authors have been limited mainly to steady

state fundamental frequency analysis, current harmonics, and transient event detection

algorithms. Figure 10-1 shows a classification of the signatures used by the different

references.

Figure 10-1: Classification of Signatures that are actually used for load identification
 40

10.2 Sampling period and frequency

When it comes to sampling periods there doesn’t seem to be a consensus, it

depends mainly in the accuracy being looking for, with respect to the events inside the

residence or load. A high sampling period, e.g. 1 minute, provide a very clean signal,

with relatively low probability of sampling when multiple events occur simultaneously

or when a transient occur, but may miss lots of important events (microwave is often on

for less than 1 minute). On the other hand, a small sampling period (e.g. 1 second),

would avoid missing events, but provides a noisy signal with many transients and

simultaneous events. Figure 10-2 shows a classification of the sampling periods used in

some of the references.

Figure 10-2: Sampling period classification

There is a tradeoff regarding the sampling frequency to use. When using 1 second

or less, there is a huge data density were it could be possible to identify all the

appliances in a house, even those who have a very short time of usage, like a garage

door opener, but it would provide a very noisy signal, with many steady state transients
 41

due to in-rush currents at the turn-on of the appliances. If a long sampling period is

used, e.g. 1 minute, there would be low data density, and a very clean signal, because

the probability of sampling a turn-on event would be very low, but it would be missing

lots of events that last less than a minute, like the microwave, a blender, or the vacuum

cleaner in some cases. The frequency depends also on the algorithm type, i.e. for

algorithms based on transient event detection a higher sampling frequency should be

used, between 7 and 15 kHz, and for steady state algorithms frequencies of less than

one Hz are very common.

Field experience in the existing literature on nonintrusive load identification is

often invalid, because experiments are done in isolated environments (e.g. laboratories

with few loads), seldom field tests are performed with an actual residential load, with

many appliances on it and working at the same time. We organized the main references

in TABLE 17.1 for better and easier visualization of the characteristics of each research.

We emphasize the low accuracy achieved by these approaches.

 42

11 APPENDIX E: HARMONIC DISTORTION INDEX

Although harmonic signatures fit in the definition of steady state signature, they

provide additional information to fundamental frequency algorithms, increasing their

detection accuracy as transient signatures. The main objective of harmonic signatures is

to differentiate between devices with similar P and Q consumption in steady state.

According to (Laughman, 2003), appliances that show a similar behavior in

fundamental frequency power consumption can show differences in higher harmonics

power consumption, which could be decisive for their differentiation. As seen in Figure

11-1 harmonic content in a signal may be a very good parameter for its differentiation,

this figure shows the difference between the harmonic content of two appliances, a

microwave, and a vacuum cleaner. The microwave has a harmonic content more present

in the third and fifth harmonic, while the vacuum cleaner has a harmonic content more

present in the second harmonic.

The differentiation in the harmonic content of a signal is mainly due to its non

linearity, while linear loads have a simple relation between the voltage and the current

in which those two parameters oscillate at the fundamental frequency, in nonlinear

loads, there is a relation in which the current has other components than the

fundamental one, that generates current harmonics, which make a huge difference

between appliances.
 43

Figure 11-1: Current harmonics: (a) Microwave, (b) Vacuum Cleaner, (c) High
Efficiency Light Bulb and (d) Normal Light Bulb

It can be seen in Figure 11-1, that harmonic content can offer a useful tool to

discriminate between high efficiency and normal light bulbs. As part of our research,

we propose a simple harmonic distortion index based on the RMS value of the

normalized harmonic indexes as shown in (12.1).

12

"k ! ai inorm
n=2
HDI NK&DW =
(n #1)
ai
ainorm = $i % {2,...,12} (12.1)
a1

ki & 1 $i % {2,...,12}
 44

ki is a multiplier, which main purpose is to give more importance to the more

significant harmonics, i.e. in the case of the microwave, harmonics indexes 1, 3 and 5,

and in the case of the vacuum cleaner, indexes 1 and 2.

In a preliminary study on harmonic signatures using ki=1 for every i, our distortion

index proved to be very successful. The tests included a comparison between three

different appliances, a refrigerator, a water heater and a microwave. Between these

appliances the differences shown were of nearly a 70% between the heater and the

refrigerator, of a 90% between the heater and the microwave, and the differences

between the refrigerator and the microwave were of 65%. These differences could

certainly be of great value to be used in addition with steady state fundamental

frequency signatures.
 45

12 APPENDIX F: DATA EXTRACTION AND EQUIPMENT

The data extraction for our field test was done on a single and fairly large

residential load, and the instrument we used for our measurements was a Hioki 3197

power analyzer and ekeeper (by Happyvolt.com), a proprietary smart meter we

developed that is capable of measuring mean, maximum and minimum of voltage (rms

value), current (rms), frequency, among other parameters, every two seconds.

12.1 Limited memory

The internal memory of the network analyzer is filled every one and a half hour, so

it required frequent oversight of instrument’s memory and 24 hours almost fully

devoted to this.

Our equipment database was obtained in the field, where we used the same

instrument (Hioki 3197 and our Ekeeper - HappyVolt) to measure every appliance

individually. The true load information was obtained performing a fast equipment

inventory, while for the lighting of the house (mounted and with data not always

available), the information was obtained through an estimate based on the number and

type of bulbs that each illumination electrical circuit of the house.

 46

13 APPENDIX G: GETTING THE GROUND TRUTH

As we didn’t have the possibility of having individual sensors for every appliance

to recover the ground truth, we used a record workbook which was posted in every

circuit in the house (switch and appliance). Its purpose was to register every on and off

event in the testing period. On Figure 13-1 we can see an example of the generic sheet

used for this purposes, this sheet has 60 columns with corresponding to each minute in

an hour, and 24 hours corresponding to a day. Every time an appliance or switch is

turned on, the symbol “/” is placed in the coordinate correspondent to the actual time,

and an “X” is placed when the appliance is turned off.

Figure 13-1: Sample worksheet for registration of events

While the operation of the appliances was determined through the worksheet shown

in Figure 13-1, the real and reactive power database was determined measuring each

appliance independently, except for the lighting, which was determined from the

nominal value of each bulb.

 47

Given our measured data, we wrote an algorithm in Matlab for the comparison

between the measured data and the energy profile estimated from the record workbooks.

Figure 13-2 shows this comparisson.

Figure 13-2: Comparison between estimated power profile (red) and measured power
profile

Before the comparison is performed, a previous softening of the signal must be

realized.

The differences between the estimated energy consumption, and the actual energy

consumption using this practice was of approximately 5%, Which is a number much

bigger than the one that we should obtain placing sensors in each individual appliance.
 48

14 APPENDIX H: PILOT HOUSE STATISTICS

The pilothouse is subscribed to Chilean regulated tariff BT1. The main

characteristics of the residence and the field test day are:

• Total installed capacity: 20.98 kW.

• Point of connection contracted power: 9.9 kW.
• Max. power demand on field test: 4.95 kW.
• Energy consumption on field test: 19.39 kWh.
• Load factor: 18.65%1
• Capacity factor: 4.4%2

1
The load factor corresponds to the energy consumption on the field test divided by the maximum power demand on the field
test multiplied by the number of test hours (21).
2
The capacity factor corresponds to the energy consumption on the field test divided by the total installed capacity multiplied by
the number of test hours (21).
 49

15 APPENDIX I: ENERGY SHARE BY APPLIANCE

The following pie charts show the energy share and which are the appliances that

signify higher energy consumption throughout the day.

Figure 15-1: Energy cake

 50

16 APPENDIX J: INVENTORY AND DATABASE

This is solved entering an equipment inventory to the meter or database system

and/or with a database with consumption features of typical appliances. The pilothouse

inventory consists of 97 devices, which are classified as follows:

16.1 Kitchen

TABLE 16.1. Kitchen Inventory

Quantity Device Aggregated Power (W)
2 Circular fluorescent tube 80
1 Light bulb 75
3 Fluorescent tube 120
1 Electric oven 650
1 Microwave (Daewo) 1,400
1 Kettle (Black & Decker) 2,000
Toaster (Black and
1 1,320
Decker)
1 Television (RCA) 100
1 Refrigerator (GE) 150
1 Diswasher 2,000

16.2 Washing room

TABLE 16.2. Washing Room Inventory

Quantity Device Aggregated Power (W)

2 Circular fluorescent tube 80

1 Halogen Bulb 150
1 Washing Machine 2,200
1 Cloth Dryer 2,300
1 Iron 2,000
 51

16.2.1 Others

TABLE 16.3. Rest of the House Inventory

Device Quantity Aggregated Power (W)
Normal Light Bulbs 15 1,125
Halogen Light Bulbs 2 300
Efficient Light Bulbs 7 140
Candle Light Bulbs 8 320
Dichroic Light Bulbs 27 1,350
Low Power Light Bulbs 5 100
Mini Halogen 2 40
Ventilator 3 150
Hair Dryer 1 2,000
Television 3 450
Nintendo Wii 1 50
Computer 2 300
Printer 1 30
 52

16.3 Total

TABLE 16.4. Total Inventory

Device Total Installed Power
Normal Light Bulbs 1,125
Halogen Light Bulbs 300
Efficient Light Bulbs 140
Candle Light Bulbs 320
Dichroic Light Bulbs 1,350
Low Power Light Bulbs 100
Mini Halogen 40
Ventilator 150
Hair Dryer 2,000
Television 450
Nintendo Wii 50
Computer 300
Printer 30
Circular fluorescent tube 80
Halogen Bulb 75
Washing Machine 120
Cloth Dryer 650
Iron 1,400
Circular fluorescent tube 2,000
Light bulb 1,320
Fluorescent tube 100
Electric oven 150
Microwave (Daewo) 2,000
Kettle (Black & Decker) 80
Toaster (Black and Decker) 150
Television (RCA) 2,200
Refrigerator (GE) 2,300
Diswasher 2,000
Total 20,980
 53

17 APPENDIX K: REFERNCE CLASIFFICATION

TABLE 17.1. Reference Classification
Sampling
Number of
Reference Signature Period or Lab or Field Test Accuracy Observations
Appliances
Frequency
Hart 1989 (Hart, Steady State
75 to Pioneer on the
1989) Fundamental 1 second Field: 3 Houses 25
90% subject
Frequency
Marceau 2000 Steady State
(Marceau & Fundamental 16 seconds Field: Single house > 90%1 72
Zmeureanu, 2000) Frequency
Berges 2010
Steady State
(Berges,
Fundamental - Field: Refrigerator 85% 1
Matthews, &
Frequency
Soibelman, 2010)
Chang 2010a Steady State and
(Chang, et al., Turn on 15 kHz Lab: Between 3 loads 100% 3
2010) Transient
Cole 1998a (Cole Steady State
& Albicki, 1998a) Fundamental 1 second Field: 8 Houses > 99% 7
Frequency
Hart 1992 (Hart, Field: 1 MS-NALM, 3
Steady State Every
1992) first generation AS-
Fundamental 1 second 90% household
NALM and 10 second
Frequency appliance
generation AS-NALM3
Farinaccio 1999 Steady State
Rule Based
(Farinaccio & Fundamental 16 seconds Field: 1 test house 85% 24
Algorithm
Zmeureanu, 1999) Frequency
Cole 1998b (Cole Steady State
2, washer and
& Albicki, 1998b) Fundamental 1 seconds Not specified > 90%5
hp compressor
Frequency
El-Ferik 2006 (El- Steady State Air-
AC6 Physical model
Ferik, Hussain, & Fundamental - Lab: Pilot house 99% conditioning
required
Al-Sunni, 2006) Frequency system
AL-Kandari 2006 Varies between Estimates harmonics
(Al-Kandari & El- 10 and 400 from distorted
Harmonic Lab - -
Naggar, 2006) samples per current and voltage
cycle waves
Leeb 1993 (Leeb,
Transient - Lab - 3
1993)
Shaw 2008 Parametric models
(Shaw, et al., Transient - Lab 99%7 8 are needed, and hand
2008) modification of data
Laughman 2003 Transient and Field: Buildings and
8 kHz - - No results are shown
(Laughman, 2003) Harmonic transportation systems
Chang 2010b Turn-on
15kHz Lab 100% 3
(Chang, 2010) Transient

1
For most evaluation scenarios.
2
7 Major end use appliances with individual clamp-on current sensors.
3
In the first two cases, there was no parallel instrumentation, so the detection of events was using criterion, and in the third case
the field test were not fully analyzed.
4
Rule based algorithms for two appliances, a domiciliary water heater (DWH) and a refrigerator.
5
Comparison made between energy estimation and energy consumption in 15 minute intervals
6
Air-conditioning system
7
For induction motor