NON-INTRUSIVE LOAD MANAGEMENT SYSTEM FOR RESIDENTIAL

LOADS USING ARTIFICIAL NEURAL NETWORK BASED ARDUINO

MICROCONTROLLER

ISIYAKU ABUBAKAR

A thesis submitted in fulfilment of the

requirements for the award of the degree of
Doctor of Philosophy

Faculty of Electrical Engineering

Universiti Teknologi Malaysia

APRIL 2018
 iii

DEDICATION

I dedicated this thesis to my mother Hajiya Aishatu Ishaq and my father

Alhaji Abubakar Ibrahim for their tireless efforts in ensuring my good upbringing
through support, guidance and unlimited prayers. May their soul rest in perfect peace
and may Allah grant them Jannatul Firdausi. This dedication is also extended to my
late wife Asiya Kabir Suyudi, may her soul rest in perfect peace and may she be
granted the Jannatul Firdausi by the almighty Allah. My entire family will also be
covered in the reward of this work specifically my wife Binta Kabir, my children
Asiya Ishaq, Ibrahim Ishaq, Abubakar Ishaq and Ishaq Ishaq. May Allah assists us in
giving you a very good upbringing and may all of us be among those that the prophet
peace be upon him will be happy with in the day of judgment.
 iv

ACKNOWLEDGEMENT

All praise be to almighty Allah, the most merciful for his guidance in successful
completion of my PhD research work. My sincere thanks and appreciations to my
supervisor Dr. Saifulnizam Bin Abd. Khalid for his support, guidance, encouragement
and patient throughout the research activities. His dedication, support, advice and
technical expertise are the key elements to the succesful completion of my doctorial
research work. Moreover, I would like to extend my gratitudes to my co-supervisor
Prof. Ir. Dr. Mohd Wazir bin Mustafa for his motivations, guidance, encouragement
and patience to attend my numerous research challenges. Furthermore I would like to
express my appreciation to Associate Prof. Dr. Hussain Shareef for his advices and
contributions in the research work. I thank you all for your valuable contributions and
may almighty Allah rewards you abundantly.

My appreciation also goes to the technical staff of Power System Lab for
observing their various duties in the coordination of the Laboratory activities. I am
also grateful to all my research colleagues, especially Mamunu Mustapha, Saleh Y.
Abujarad, Mudasiru Mustapha, Jalal, Ameer, Sani Salisu, Raja Masood, Mohammed
Olatunji, Abdurrahman Otoze and many others, for their valuable advices and
encouragements. Same appreciation goes to all my my friends and colleagues who
have been here during our program. The support by Bayero University, Kano and
TETFUND is also very much appreciated. Finally, I wish to express honest gratitueds
to my parents and my entire relatives for their various supports and prayers. The
support and encouragements by my Family is also appreciated specifically my wife
Binta Kabir and my children Asiya, Ibrahim, Abubakar and Ishaq. The scrafice made
to me by Asiya Ishaq and Ibrahim Ishaq is highly appreciated, may Allah make them
great in their future endavourss and the entire family, ameen.
 v

ABSTRACT

The energy monitoring is one of the most important aspects of energy

management. In fact there is a need to monitor the power consumption of a building
or premises before planning technical actions to minimize the energy consumption. In
traditional load monitoring method, a sensor or a group of sensors attached to every
load of interest to monitor the system, which makes the system costly and complex.
On the other hand, by Non-Intrusive Load Monitoring (NILM) the aggregated
measurement of the building’s appliances can be used to identify and/or disaggregate
the connected appliances in the building. Therefore, the method provides a simple,
reliable and cost effective monitoring since it uses only one set of measuring sensors
at the service entry. This thesis aims at finding a solution in the residential electrical
energy management through the development of Artificial Neural Network Arduino
(ANN-Arduino) NILM system for monitoring and controlling the energy consumption
of the home appliances. The major goal of this research work is the development of a
simplified ANN-based non-intrusive residential appliances identifier. It is a real-time
ANN-Arduino NILM system for residential energy management with its performance
evaluation and the calibration of the ZMPT101B voltage sensor module for accurate
measurement, by using polynomial regression method. Using the sensor algorithm
obtained, an error of 0.9% in the root mean square (rms) measurement of the voltage
is obtained using peak-peak measurement method, in comparison to 2.5% when using
instantaneous measurement method. Secondly, a residential energy consumption
measurement and control system is developed using Arduino microcontroller, which
accurately control the home appliances within the threshold power consumption level.
The energy consumption measurement prototype has an accurate power and current
measurement with error of 3.88% in current measurement when compared with the
standard Fluke meter. An ANN-Arduino NILM system is also developed using steady-
state signatures, which uses the feedforward ANN to identify the loads when it
received the aggregated real power, rms current and power factor from the Arduino.
Finally, the ANN-Arduino NILM based appliances’ management and control system
is developed for keeping track of the appliances and managing their energy usage. The
system accurately recognizes all the load combinations and the load controlling works
within 2% time error. The overall system resulted into a new home appliances’ energy
management system based on ANN-Arduino NILM that can be applied into smart
electricity system at a reduced cost, reduced complexity and non-intrusively..
 vi

ABSTRAK

Pemantauan tenaga adalah salah satu aspek yang paling penting dalam
pengurusan tenaga. Malah, terdapat keperluan untuk memantau penggunaan kuasa
sebuah bangunan atau premis sebelum merancang tindakan teknikal untuk
meminimumkan penggunaan tenaga. Dalam kaedah pemantauan beban tradisional, satu
pengesan atau sekumpulan pengesan dipasangkan kepada setiap beban untuk
pemantauan sistem, yang menjadikan sistem itu mahal dan kompleks. Sebaliknya,
dengan Pemantauan Bebanan Tidak Robos (NILM) pengukuran teragregat peralatan
bangunan boleh digunakan untuk mengenal pasti dan/atau mengasingkan peralatan
yang bersambung di bangunan. Oleh itu, kaedah ini menyediakan pemantauan yang
mudah, boleh dipercayai dan kos efektif kerana hanya menggunakan satu set pengesan
pengukur pada kemasukan perkhidmatan. Tesis ini bertujuan mencari penyelesaian
dalam pengurusan tenaga elektrik kediaman melalui pembangunan sistem yang tidak
robos untuk memantau dan mengawal penggunaan tenaga peralatan rumah. Matlamat
utama penyelidikan ini adalah pembangunan pengenal pasti peralatan kediaman
berasaskan Rangkaian Neural Tidak Robos dipermudah. Sistem Rangkaian Neural
Buatan (ANN)-Arduino (NILM) adalah masa nyata dalam pengurusan tenaga kediaman
dengan penilaian prestasi dan penentukuran modul pengesan voltan ZMPT101B untuk
pengukuran yang tepat, dengan menggunakan kaedah regresi polinomial. Dengan
menggunakan algoritma pengesan yang diperolehi, ralat 0.9% dalam pengukuran voltan
punca min kuasa dua (rms) diperolehi menggunakan kaedah pengukuran puncak ke
puncak, berbanding 2.5% dengan kaedah pengukuran ketika. Keduanya, sistem
pengukuran dan kawalan penggunaan kediaman dibangunkan menggunakan
mikropengawal Arduino, yang mengendalikan peralatan rumah secara tepat di dalam
tahap penggunaan kuasa ambang. Prototaip pengukuran penggunaan tenaga
mempunyai pengukuran kuasa dan arus yang tepat dengan ralat 3.88% dalam
pengukuran arus apabila dibandingkan dengan meter piawai Fluke. Sistem ANN-
Arduino NILM juga dibangunkan menggunakan signatur keadaan mantap, yang
menggunakan ANN suap hadapan untuk mengenalpasti beban apabila menerima kuasa
sebenar teragregat, arus rms dan faktor kuasa daripada pada Arduino. Akhir sekali,
ANN-Arduino NILM berasaskan sistem pengurusan dan kawalan peralatan
dibangunkan untuk mengesan peralatan dan menguruskan penggunaan tenaga mereka.
Sistem ini mengenal sepenuhnya dengan tepat semua kombinasi beban dan kawalan
kerja beban dalam ralat masa 2%. Sistem keseluruhannya menghasilkan sistem
pengurusan tenaga elektrik rumah yang baru berdasarkan ANN-Arduino NILM yang
boleh diaplikasikan untuk sistem elektrik pintar dengan kos yang dikurangkan,
mengurangkan kerumitan dan tidak robos.

.
 vii

TABLE OF CONTENTS

CHAPTER TITLE PAGE

DECLARATION ii
DEDICATION iii
ACKNOWLEDGEMENT iv
ABSTRACT v
ABSTRAK vi
TABLE OF CONTENTS vii
LIST OF TABLES xii
LIST OF FIGURES xiv
LIST OF ABBREVIATIONS xvi
LIST OF SYMBOLS xviii
LIST OF APPENDICES xix

1 INTRODUCTION 1
1.1 Background 1
1.2 Problem Statement 3
1.3 Objectives of the Research 5
1.4 Scopes of the Research 5
1.5 Significance of the Research 6
1.6 Thesis Organization 7

2 LITERATURE REVIEW 8
2.1 Introduction 8
2.2 The Load Monitoring Concept 9
2.2.1 Traditional Load Monitoring 10
2.2.2 The Emergence of Non-Intrusive Load Monitoring 10
2.3 The Electrical Load Monitoring of Appliances 11
 viii

2.4 The Load Signature 11

2.4.1 Classification of Load Signatures 13
2.4.1.1 Intrusive Signatures 13
2.4.1.2 Non-Intrusive Signatures 14
2.5 Load Monitoring Devices 15
2.5.1 Measuring Devices 16
2.5.2 Communication Devices 16
2.5.3 Soft Computing Devices 18
2.5.4 Control Device 18
2.5.5 Recognition Devices 19
2.5.6 Display Devices 19
2.6 Intrusive Load Monitoring 20
2.6.1 The ILM Framework 22
2.6.1.1 Activity of Daily Living Recognition 22
2.6.1.2 Occupancy Detection 23
2.6.1.3 User-Appliance Interaction 23
2.6.2 The ILM Structure 23
2.6.3 An Overview of Some ILM Research Works 25
2.7 Non-Intrusive Load Monitoring 28
2.7.1 NILM Framework 32
2.7.1.1 Data Acquisition 33
2.7.1.2 Feature Extraction 34
2.7.1.3 Appliance Classification 34
2.8 NILM Disaggregation and Identification Approaches 34
2.8.1 Steady-State Approaches 35
2.8.2 Transient-State Approaches 37
2.8.3 Approaches Based on Hybrid Analysis 39
2.8.4 Approaches Using Non-traditional Signature
Analysis 40
2.9 Event based and Non-event based Detection in NILM 40
2.10 Artificial Intelligent Learning in NILM 41
2.10.1 Supervised Learning 42
2.10.2 Unsupervised Learning 42
2.11 NILM Reviews 43
 ix

2.12 General NILM Applications and Future Research Directions 44

2.13 Application of Load Monitoring in Appliances’ Energy
Management and Control 46
2.13.1 Energy Management Approaches Based on ILM 47
2.13.2 Energy Management Approaches Based on NILM 51
2.14 Energy Savings Techniques for Minimizing Power Consumption. 55
2.15 AC Voltage Measurement Methods 57
2.15.1 Peak Value Method 58
2.15.2 Peak to Peak Value Method 58
2.15.3 Instantaneous Calculation Method 58
2.16 The Regression Analysis 59
2.16.1 Linear Regression 59
2.16.2 Polynomial Regression 60
2.17 Summary 61

3 RESEARCH METHODOLOGY 63
3.1 Introduction 63
3.2 Load Selection and Profiling of the Home Appliances 64
3.3 Sensors Calibration 66
3.3.2 Voltage Sensor Calibration 67
3.3.2.1 Peak to Peak Voltage Measurement
Method 68
3.3.2.2 Instantaneous Voltage Measurement
Method 69
3.3.2.3 Data Collection and Regression
Procedure 69
3.3.2.4 The Programing Code for the Voltage
Sensor Calibration 71
3.4 Residential Appliances consumption Measurement and control 71
3.4.1 Appliances Energy Consumption Measurement and
Control System Design 73
3.4.2 System Development 74
3.4.3 Energy Measurement and Control Programming
Code 75
 x

3.5 Design of the Non-Intrusive Appliance Load Identification System 76

3.5.1 Energy Consumption Data Collection for the NILM
System 77
3.5.2 The ANN Training for the NILM System 79
3.5.3 The Arduino Programming for the ANN-NILM
System 80
3.6 Development of ANN-Arduino NILM based Residential Appliances
Management and Control System 82
3.6.1 Data Collection for the Appliances Management and
Contol System 84
3.6.2 ANN Training for the Appliances Management and
Control System 86
3.6.3 Arduino Programming for the Appliances Management
and Contol System 86
3.7 Summary 88

4 RESULTS AND DISCUSSIONS 89

4.1 Introduction 89
4.2 The Sensors Calibration 89
4.2.1 The Current sensor Results 90
4.2.2 Calibration of the Voltage Sensor 90
4.2.1 Resulted Polynomial Equations 91
4.2.2.1 The Instantaneous Voltage Method
Analysis 92
4.2.2.2 The Peak to peak Voltage Method
Analysis 94
4.2.2.3 Comparison between the Instantaneous
Method and Peak to Peak Voltage
Method and General Analysis 94
4.3 Residential Electrical Energy Consumption Measurement and control
system 99
4.3.1 Measurement Analysis of the System 99
4.3.1.1 Voltage Measurement 100
4.3.1.2 Current Measurement 102
 xi

4.3.1.3 Power Measurement 102

4.3.3 The Loads Control of the System 103
4.3.4 The System Control when the Highest Priority Load is
Greater than the Threshold 104
4.3.5 General Analysis 108
4.4 Development of the Non-Intrusive Appliances Identification
System 108
4.4.1 The ANN Training Results of the NILM System 108
4.4.2 The Network Simulation Results 110
4.4.3 System Performance 112
4.4.4 General Analysis 112
4.5 The ANN-Arduino NILM based Residential Appliances
Management and Control System 114
4.5.1 ANN Training Results of the Appliances Management
and Control System 114
4.5.2 Network Simulation and Testing 116
4.5.3 Comparison with Previous Work 118
4.5.4 Load Monitoring Performance of the Appliances
Management and Control System 119
4.5.5 Appliances’ Timing and Load Control Performance of
the System 119
4.5.6 System Analysis 120
4.6 Comparison of the ANN-Arduino NILM with Support Vector
Machine (SVM) 121
4.7 Summary 121

5 CONCLUSIONS AND RECOMMENDATIONS FOR FUTURE

WORK 124
5.1 Conclusions 124
5.2 Recommendations for Future Work 125
5.3 Research Contributions 126
REFERENCES 128
Appendices A – G 148 – 155
 xii

LIST OF TABLES

TABLE NO. TITLE PAGE

2.1 Comparison of Communication Methods 17

2.2 Some Key Studies of Energy Management in ILM 49
2.3 Some Key Studies of Energy Management in NILM 54
3.1 The home appliances and their power consumption 65
3.2 The Sample of the Data for one Combination of loads 79
3.3 The Data Distribution for the ANN-NILM system 80
3.4 Pseudo code for the ANN-Arduino NILM 82
3.5 Sample of the Data for a combination of Loads 85
3.6 Data breakdown for the ANN-Arduino appliances management 86
3.7 Pseudo code for the ANN-Arduino NILM appliances management
system 88
4.1 The Current Sensor Parameters 90
4.2 Equations from the Polynomial Regression Analysis 91
4.3 Arduino Voltage Measurement Using Instantaneous Method 93
4.4 Arduino Voltage Measurement Using Peak to peak Method 95
4.5 Comparison of the Peak to peak and Instantaneous Methods 96
4.6 Comparison of the Voltage Measurement with Miron Alexe Method
[197] 98
4.7 System Voltage Measurement and its Comparison with Commercial
Meter 101
4.8 Current measurement of the system 102
4.9 Power measurement of the system 103
4.10 The Loads Control of the System 1055
4.11 The Summary of the ANN for Non-Intrusive Appliances
Identification 109
4.12 The System performance 113
4.13 Summary of the ANN for ANN-Arduino NILM based Appliances
Management and Control System 115
 xiii

4.14 The Appliances Timing by the System 120

 xiv

LIST OF FIGURES

FIGURE NO. TITLE PAGE

2.1 Appliance Signature Classification 14

2.2 Energy Monitoring Devices 20
2.3 Intrusive Load Monitoring 21
2.4 The ILM Approach 24
2.5 Overlapping Cluster in ∆P - ∆Q plane 31
2.6 Configuration of NILM System 32
2.7 The NILM Approach 33
2.8 V-I Trajectory of some group of loads 36
2.9 Transient Waveform 38
2.10 Global GHG Emission 53
2.11 AC Supply quantities changing with time 57
3.1 The flowchart for implementation of the research work 64
3.2 The block diagram of the ANN-Arduino system 66
3.3 Analog and Digital Voltages 68
3.4 Flow Chart of the rms Voltage Calculation 72
3.5 The Proposed Block Diagram of the Energy Consumption
Measurement and Control System 74
3.6 Flow chart of the Energy Consumption Measurement and Control
System 76
3.7 The Block Diagram of the ANN-Arduino NILM Configuration 78
3.8 Flow chart of the ANN-Arduino NILM system 81
3.9 Block diagram of ANN-Arduino appliances management system 85
3.10 The ANN-Arduino NILM Network Training Flowchart 87
 xv

4.1 The response of the instantaneous and peak to peak voltage methods
when compared with FLUKE meter readings (A) The whole voltage
range (B) Section of the graph 97
4.2 Comparison of Voltage Measurement 100
4.3 The Trained ANN for the Non-Intrusive Load Identification
system 110
4.4 The Network Simulation Result 111
4.5 The Result of Testing the Network 112
4.6 The Trained ANN for the Load Controlling System 115
4.7 The Network Simulation Result 1177
4.8 The Result of Testing the Network with Different Data 118
 xvi

LIST OF ABBREVIATIONS

AC - Alternating Current
ADC - Analog to Digital Conversion
ADL - Activity of Daily Living
AFAMAP - Additional Factorial Approximate MAP
AIA - Artificial Immune Algorithm
ANN - Artificial Neural Network
BEMS - Buildings Energy Management System
CCTV - Closed Circuit Television
CFL - Compact Fluorescent Lamp
CT - Current Transformer
CVD - Continuously Varying Devices
CW - Current Waveform
DC - Direct Current
DENs - Distributed Electricity Networks
DMPR - Disjoint Multi Path base Routine
DR - Demand Response
DWT - Discrete Wavelet Transform
ECM - Energy Control Manager
EIG - Eigen Values
EM - Energy Management
EMCUs - Energy Management Control Units
EMS - Energy Management System
ESD - Edge Symbol Detector
EMPT - Electromagnetic Transient Program
FHMM - Factorial Hidden Markov’s Model
FPGA - Field Programmable Gate Array
GHG - Greenhouse Gas
GMM - Gaussian Mixture Model
 xvii

HAR - Harmonics
HEMS - Home Energy Management System
HMM - Hidden Markov’s Model
IAW - Instantaneous Admittance Waveform
iHMES - In-Home Energy Management System
ILM - Intrusive Load Monitoring
IPW - Instantaneous Power Waveform
IR - Infrared
k-NN - k – Nearest Neighbour rule
LED - Light Emitting Diode
LDR - Light dependent Resistor
MAP - Maximum A posteriori Probability
NBC - Naïve Bayes Classifier
NILM - Non-Intrusive Load Monitoring
OREM - Optimization-based Residential Energy Management
PC - Personal Computer
PCAT - Power Conservation Analysis Tool
PLC - Power Line Communication
PSO - Particle Swarm Optimization
PT - Potential Transformer
rms - root mean square
RMSE - Root Mean Square Error
ROC - Receiver Operating Characteristic
SDM - Supply Demand Matching
SoC - System on Chip
STFT - Short Term Fourier Transform
STW - Switching Transient Waveform
SVM - Support Vector Machine
TPW - Transient Power Waveform
USD - United States Dollar
USEPA - United State Environmental Protection Agency
WLAN - Wireless Area Networks
WMRA - Wavelet Multi-Resolution Analysis
ZEM - Zigbee Energy Meter
 xviii

LIST OF SYMBOLS

I - Current
IADC - ADC current
IADCref Reference ADC current value
Ii - Sampled Current
Ipeak - Peak Current
Ipeak-peak - Peak to Peak Current
Irms - Root Mean Square Voltage
j - Order of a Polynomial
L1 to L4 - House Appliances
n - Total number of samples
P - Power
pf - Power Factor
S - Apparent Power
Q - Reactive Power
UT - Total Transient Energy
V - Voltage
Vi - Sampled Voltage
Vpeak - Peak Voltage
Vpeak-peak - Peak to Peak Voltage
Vrms - Root Mean Square Voltage
x - Independent variable
y - Dependent variable
δ - Sensitivity of current sensor
 xix

LIST OF APPENDICES

APPENDIX TITLE PAGE

A List of Publications 148

B Table of Equations for Electrical Quantities 150
C The System of Residential Appliances 151
D The Voltage Measurement Experimental Set Up 152
E The Monitoring and Management System 153
F The Load Identification for ANN-Arduino NILM 154
based Appliances Management and Control
G The Appliances Control of the ANN-Arduino NILM 155
System
 CHAPTER 1

1INTRODUCTION

1.1 Background

The usage of electricity at home has since become an essential activity in both
rural and urban settings throughout the world. Every now and then the demand of
electricity is increasing. New electrical devices and appliances are always emerging at
the expense of the users, for they have to be paying the utility companies for operating
those devices. However through proper usage, a lot of benefits can be achieved by both
the consumers and electrical utility companies [1, 2]. Those benefits, which include
energy saving and reduction in greenhouse gas emission will surely come as the results
of proper monitoring and management of the power usage.

The appliance load monitoring is the prerequisite activity to every load

management, in fact it is through proper monitoring that a management decision can
be taken on the connected loads to make the best utilization of the energy. This
therefore calls for a system that can monitor the appliances and takes a real time action
to manage the energy usage [3-6]. The largest amount of electrical energy sources at
global level is from oil and gas sector, which are non-renewable in nature and are
estimated to exhausted in the near future. Hence the load monitoring with the view of
managing and optimizing the sources is essential in order to reduce the risk of energy
crises in the world, by taking energy management actions on the monitored loads. Over
the last three decades, the non-intrusive load monitoring has witnessed so many
contributions by the researchers in all angles of the research area [7-12]. Some of the
works attempted to come out with systems for load controls [13, 14]. Several ways of
 2

identifying the connected loads non-intrusively, have existed which include steady-
state and transient-state methods of signature extraction [15-17]. Using the steady-state
signatures for identifying the home electrical appliances in real time will be more
suitable because most of the appliances’ operation is in steady-state; therefore a steady-
state system can go with the real time load identification of the loads [18]. Previous
studies in Non-Intrusive Load Monitoring (NILM) have also indicated the need to use
it in energy management of the appliances [2, 19].

The NILM system can be applied in many energy consuming systems to

facilitate energy management. For instance, the NILM system can determine the load
schedule in an aircraft or motor vehicle from the measurement at the generator level
[20-23]. With proper NILM system which is powered by energy consumption data
feedback, a very significant saving can be achieved, in fact energy minimization of up
to 15% can be achieved from the world wide perspective through changing daily
routine of the appliances, low cost stake holding and acquiring more efficient and
energy saving electrical appliances [24, 25]. Researchers in the area of energy
management pointed out that due to globalization and urbanization of the rural areas,
the electricity demand can only be contained by appropriate and efficient energy usage
and management [26-28]. Moreover, another analysis found that the residential
settings can waste one-third of their energy consumption due to unmonitored
operations of various appliances [26].

The home automation through NILM has also been found very useful in
helping elderly people to sustain their lifestyle with respect to home daily activities
[19]. Also the advances in computer technology, sensor technology and
communication facilities will give opportunities for researchers and energy
management equipment manufacturers to device many ways of monitoring and
managing the load energy consumption [29, 30].

Most of the residential electrical energy management research works are

intrusive in nature, which suffers from the weakness of multiple point sensing. The
NILM researchers have done a lot in appliances disaggregation and signature
extraction, but the NILM systems with real-time energy management capability have
 3

not been treated in the research area. The issue of measuring the performance of NILM
is another challenging area of NILM, that is why most of the NILM works use
distributive metering to check the performance of the systems [2, 31]. This research
work focuses on the development of ANN-Arduino based NILM system in order to
develop a single point sensing energy management system to replace the existing
multiple points sensing systems. The research will also bridge the gap of non-
application of the NILM in real-time residential electrical energy management.
Moreover, the evaluation of the ANN-NILM in residential appliances control will be
a novel method of measuring the performance of NILM,

1.2 Problem Statement

There is need for accurate, simple and cost effective monitoring and
management system in all aspects of electrical power consumption. The electrical
power consuming system suffers a lot of energy wastage whether at commercial,
industrial or residential settings. The residential settings are prominent in energy
wastage due to mismanagement, energy unwanted activities and lack of proper load
control. The existing researches on home energy management proposed many works
on distributed or intrusive load monitoring for home appliances’ management [32-34].
However intrusive load monitoring requires sensors or meters at every load of interest,
which makes it complex and costly. On the other hand there are many research works
regarding the non-intrusive load monitoring which uses one set of measuring devices
at the service entry. Though some of these researches are related to the energy
management [35-37], but the studies have reputed the need to have an improved real-
time NILM for home appliances management at reduced cost and complexity. Various
techniques of NILM had been brought up by the researchers that discovered a lot
regarding the load identification accuracy, signature extraction and different
identification algorithm [38-41], however there is hardly in the literature the use of
steady-state NILM technique for residential electrical energy management especially
using ANN-Arduino microcontroller based. The use of ANN-Arduino based NILM
system for residential energy management will systematically solve the problem of
multiple point sensing in ILM and eventually resulted into a cost effective and simple
 4

residential energy management using NILM, which will promote the present smart
electricity system. The following summarises the problems in the efforts to promote
the residential appliances energy management using ANN-Arduino NILM.

 The present residential energy management systems all concentrated on the

distributed sensing for controlling the loads. This not only make the system
complex and costly, but also affects the reliability of the appliances
management as the results of system dependent on multiple sensing. The
introduction of single point ANN-Arduino based NILM has vital role to
play in ensuring a reliable energy management system.

 Using the steady-state load signatures to identify the appliances will

provide more identification accuracy because the steady-state condition is
the dominant stage of the load operation [31]. Also the ANN has a great
capability of mapping the the load signatures into the load identification
even with non-linear relationship between them.

 The Arduino microcontroller takes samples from voltage and current

sensors to calculate the steady-state quantities of the appliances
consumption. Hence there is need toensure accurate calibration of the
sensors, especially the ZMPT101B voltage sensor module which does not
have a single sensitivity like the ACS716 current sensor. The calibration of
ZMPT101B voltage sensor will provide a universal transfer equation for
the sensor.

 There is need to evaluate the ANN-NILM system to see its capability in

controlling the home appliances for energy management.

The system can be integrated into smart electricity system to serve as smart
home energy monitoring and management system for efficient power consumption.
The non-intrusiveness of the NILM system ensures the privacy of the consumers
because it can be applied outside the building at meter level.
 5

1.3 Objectives of the Research

The objectives of the research are:

(i) To develop a real-time ANN- Arduino microcontroller based NILM
system for home appliances energy consumption management with
energy savings capability.
(ii) To develop a simplified ANN-based non-intrusive load identifier for
home appliances’ identification using steady-state load signatures.
(iii) To develop the voltage sensor algorithm, using polynomial regression
analysis for direct and accurate application in the Arduino
microcontroller

(iv) To evaluate the performance of the ANN-NILM system in load

controlling applications.

1.4 Scopes of the Research

The scope of the research work are bounded as follows:

(i) The research work targeted monitoring and management of residential
electrical appliances for energy savings.

(ii) The research work will involve the use of single point measurement, in
NILM, to identify the appliances instead of multipoint sensing.

(iii) Steady-state signatures of the loads, which comprises of the real power,
rms current and power factor, with acceptable accuracy, are used in the
load identification.

(iv) ANN model of the load identifier will be developed using MATLAB.

(v) Arduino microcontroller, current and voltage sensors will be used to

develop the system and the ANN model will be implemented in the
Arduino.

(vi) The voltage sensor (ZMPT101B module) is calibrated while a factory

settings of the current sensor (ACS716) is used.
 6

The ANN-Arduino NILM presented in this research work is limited to some

home appliances consisting of Electric bulbs, Standing fans, Electric toaster, Rice
cooker, Electric kettle and Blender. Some other appliances like the continuously
varying loads are not considered. Also all the experiments are carried out under the
standard voltage supply of Malaysia (240V 50Hz), therefore it may have different
performance in different voltage supply scenario like 120V or 110V systems.

1.5 Significance of the Research

Non-Intrusive Load Monitoring is a developing research area with ongoing

research activities taking place in the area. The application of modern technology in
electrical appliance load monitoring and management will revolutionize the energy
consumption so that the consumers will make the best use of the power supply. The
significance of this research include but not limited to the following:

(i) Improvement in load monitoring, control and load management will make
significant improvement and advancement in smart electricity system.
(ii) The sub metering energy monitoring, which involves the monitoring of
the appliances in sub groups will benefit immensely from the positive
contribution of this research work.
(iii) The research work will provide an alternative way of ascertaining the
home appliances operating conditions non-intrusively.
(iv) Researches in load identification using NILM contribute immensely in the
load monitoring area, which will subsequently facilitate the best
economical and efficient electricity consumption.
(v) The method of residential electrical energy management emanated from
this research will be beneficial to the consumers, utility companies,
regulatory bodies and all the stakeholders in the nation. As such, the
electrical energy monitoring and management is the key factor in
conserving the energy resources.
 7

(vi) The ANN-Arduino NILM system also has the advantage that it can be
applied outside the building where the appliances are working. This will
allow the system to be accessed without disturbing the customer privacy.

1.6 Thesis Organization

The remaining parts of the thesis structure are organized as follows;

Chapter 2 presents an extensive literature review of techniques and methods of

NILM which includes signature extraction, disaggregation and identification
techniques. It also includes a general literature about ILM techniques and their
applications. The chapter also highlights the application of load monitoring in
appliances energy management including some of the future research directions.

In Chapter 3 the research methodology is explained, which provides the step

by step methods of conducting the research work. It includes the sensor calibration,
load monitoring and load controlling systems design.

Chapter 4 provides the results and discussion of the research work. The results
of ZMPT101B voltage sensor calibration, the residential energy consumption
management using Arduino microcontroller and the results of the non-intrusive load
identification based on ANN-Arduino are given in the chapter. It finally presents the
results of ANN-NILM evaluation.

The conclusion of the various aspects of the research and recommendations are
given in Chapter 5 which also includes the recommendations for future works.
 CHAPTER 2

2LITERATURE REVIEW

2.1 Introduction

This chapter gives a general overview of load monitoring and its applications,
especially in the appliances’ power consumption management. It described the state
of art of both Intrusive Load Monitoring (ILM) and Non-Intrusive Load Monitoring
(NILM) and their approaches. It also analyzed different methods of load monitoring
and how different researchers in the area give various contributions in developing
various load monitoring and management techniques. Those techniques include but
not limited to signature extraction and disaggregation procedures. In this research the
NILM system is adopted in order to achieve the appliances’ energy management by
considering a single point sensing. The load identification using steady-state signatures
facilitated the appliances’ energy management in the research. The chapter also
highlighted some of the techniques for minimizing power consumptions of home
appliances and it also gives the review of greenhouse gas (GHG) emission at global
level.

The appliance load monitoring concept can be traced back to the time when
utilization of electricity started, that is a particular electric appliance can only work
when the supply has enough power to operate it. In fact even the earliest types of
electrical loads have their energy capacity to be considered before they can be
powered. The analog electrical energy meter is the earliest equipment used in
determining the energy consumption of the appliances, which is subsequent to the
usage of number of lightings in a house to bill the consumers, which was the tradition
 9

in the earliest history of electricity billing. Over the last century there has been
advancements in electrical measurement, especially with the advances in digital
electronics, sensor technology and communication technology. These advancements
lead to the invention of smart meter, a digital electrical energy meter with
communication capability [2, 42]. Energy management and hence the conservation of
power is the main purpose of every electrical load monitoring, which are achieved by
taking energy efficiency measures like the use of energy saving appliances, eliminating
unnecessary power hungry activities and use of appropriate timing of the appliances
loading [26, 43].

In the times the usage of electricity was commercialized, the working condition
of the appliances could be observed physically by the operation of the appliances.
However, the energy consumption of the appliances could also be observed from the
readings of the old energy meter, even though there is no any automation in the system,
but the consumer will at least notice the increase in the meter level when there is
additional energy consumption in the premises. In any case, the energy consumption
pattern of the electrical appliances can accurately be known by a proper monitoring of
the appliances’ operation, in which the actual energy consumption and state of
operation of the appliances can be ascertained. Looking at the recent development in
the area of load monitoring, the monitoring system can be classified into Intrusive
Load Monitoring (ILM) or Non-Intrusive Load Monitoring (NILM) depending on
whether the load monitoring system provides a measuring facility for every appliance
or the system used only one set of measuring facilities for all the appliances
respectively [19]. The researches in load monitoring especially the NILM is highly
needed in upgrading the performance of smart grid, in order to achieve the sustainable,
secured and economic electricity supply in the future [43, 44]. Also the records of the
consumer activities in the smart grid will be simplified using NILM.

2.2 The Load Monitoring Concept

The appliance load monitoring generally involves one of the two activities on
the monitored appliances, the load disaggregation and load identification [1, 2]. The
 10

load disaggregation in the power utility system is the disintegration of the aggregated
measurements of the system into its constituent individual loads. In other words the
disaggregation means knowing the individual load consumption from the total or
aggregated measurement. On the other hand the load identification is the knowledge
of the operating conditions of every individual loads also from the aggregated
measurement as the case maybe depending on the load monitoring type. The load
identification determines which loads are ON and which ones are OFF [3, 45]. The
load monitoring will create awareness of the level of energy consumption and energy
mismanagement if any, among the utility companies which will motivate efficient
energy consumption [46].

2.2.1 Traditional Load Monitoring

Traditional load monitoring referred to as the simplest way of monitoring the

energy consumption of the appliances. Normally a meter or sensor is attached to the
load to be measured in order to have its energy consumption measurement. To get the
measurement of another load, another measuring device has to be provided or the
measuring device will be connected from one load to another. The ILM is mostly
described as the traditional load monitoring which is characterized by distributive
sensing [47].

2.2.2 The Emergence of Non-Intrusive Load Monitoring

The non-intrusive load monitoring was pioneered by G. W. Heart from

Massachusetts Institute of Technology. Hart was able to make this discovery when he
was taking measurements on PV modules, where he discovered that it is possible to
use the aggregated measurement of the energy consumption of a system to know the
energy consumption of individual loads [7]. In fact Hart and his teammates concluded
that this method can be beneficial to the utilities and later works in the area of NILM
showed that it can also be useful to the consumers when applied in load monitoring,
 11

control and management, especially with the smart meter becoming more affordable.
In the area of energy management, NILM can play important role in the development
of simple, efficient and cost effective energy monitoring and management systems that
can help consumers to become energy cautious in the utility.

2.3 The Electrical Load Monitoring of Appliances

The load monitoring stated earlier can be in the form of disaggregating the
energy consumptions of the appliances into their constituent individual loads
consumption, or it can be the identification of the appliances’ operating conditions in
the house or building. Hence the primary purpose of every appliances’ energy
monitoring is to create awareness of the appliances operating condition and/or their
individual energy consumption. When the operating conditions of the appliances is
known, whether intrusively or non-intrusively, the energy consumption can easily be
managed and controlled for better usage. The efficient and optimized energy usage
will promote the energy savings and GHG reduction. Researchers in the field of load
monitoring performed various works in both NILM and ILM, but they focused more
on NILM because of its effectiveness and less complexity. The NILM method
generally involved the use of algorithm, with optimization in most cases, to do the
disaggregation and/or identification. On the other hand the ILM is considered to be the
traditional method of load monitoring, yet researches are still being carried out in the
area with the view of improving the method and applying it in different energy
monitoring scenarios as in [48]. Moreover the NILM comes with the advantage of its
non-intrusive nature, therefore both the consumers and utility companies prefer to use
NILM in their monitoring system for more privacy [49, 50].

2.4 The Load Signature

The load signatures are the most important parameters in every load
monitoring, especially the NILM. They are normally run in an algorithm to
 12

disaggregate/identify the appliances. The NILM system is a single point monitoring in

which the measurement obtained at the service drop are utilized in extracting the load
signatures. A load signature is a specific and unique feature that characterizes the
behavior of the load when the load is in operation. It can also be defined as the
measurable parameter of the load that gives information about the operating condition
of the load and sometimes level of its energy consumption [20, 51, 52]. An example
of appliance load signature can be the steady-state real or reactive power it consumed
when it is turned on. Moreover the emergence of the new load monitoring techniques
made some power quality measurements like voltage and current sags, spikes,
surge/swell, harmonic distortion and unbalance, useful in load monitoring application.
These power quality signatures are mostly found in NILM applications. The appliance
load signature can be divided into intrusive and non-intrusive depending on whether it
involved individual appliances (intrusive) or it is taken from the service entry point
[1]. When analyzed based on the sampling rate the load signature can be classified as
micro and macro level signatures. The macro signatures are signatures with sampling
interval longer than 1 second, whereas micro level signature require a sampling
interval of less than 1 second [53, 54]. The load signatures in NILM are sometimes
classified into steady-state and transient state signatures. The steady-state signatures
are extracted when the appliances are in steady-state operation, whereas the transient
signatures are taken when the loads are in transient stage, either during on or off
transition [55].

In a real monitoring system, especially NILM, it is difficult to identify load

with a single signature, this is because some loads may have similar measurements.
Therefore it is easy to find an overlapping among the appliances’ signatures, especially
in a system of large number of appliances. Therefore a combination of more than one
signatures is used in most cases. Some cases may even require more complicated
signatures like harmonic and transient in order to identify the loads, especially in cases
where different loads have the same real and reactive power consumption [56].
Furthermore the signature extraction and the load disaggregation/identification mostly
involved the use of algorithms and pattern recognition networking, therefore, the
general performance of the load monitoring system will be greatly influenced by
proper selection of appropriate appliance features.
 13

2.4.1 Classification of Load Signatures

Electrical voltage and current are the basic quantities in signature extraction or
measurement. All other quantities can be derived from these two major parameters.
The voltage and current samples are used in an algorithm to determine the various
signatures of the appliances. There are two basic characteristic to be considered when
choosing the load signatures. These characteristics are as follows:

(i) The signature has to (as much as possible) presents distinct values for different
loads. This will help in avoiding overlapping of signatures.
(ii) The signature extraction has to require fewer computational works, which will
make it easy to be implemented algorithm for real time applications [57, 58].
In real sense, the use of fewer signatures to disaggregate/identify the appliances may
results into a simpler or less complicated monitoring system, but sometimes it may
decreases the overall accuracy of the load monitoring system.

As stated earlier the load signatures are classified in different forms such as
micro and macro or intrusive and non-intrusive, which are the main classification
methods. The intrusive signatures are the signatures extracted at the device level
whereas the non-intrusive signatures are extracted at the service drop. The intrusive
signatures are of two forms the electrical and physical, and on the other hand the non-
intrusive signatures can be steady-state, transient-state or non-traditional signatures.
The load disaggregation in NILM sometimes requires more distinct signatures,
especially when there is overlapping signatures among the appliances, that is why
some researchers in NILM make use of the non-traditional signatures to identify the
loads. Figure 2.1 shows the taxonomy of the classification of the signatures.

2.4.1.1 Intrusive Signatures

The intrusive signatures are the signatures taken when the load monitoring
involved taking the signatures at appliance level and it is mostly applied in intrusive
 14

load monitoring. The intrusive signatures can be taken electrically or physically

depending on the nature of signatures to be taken. Electrical signatures can be real
power consumption, current, voltage, power factor or any measurement that can be
derived from these quantities. The physical signatures are the signatures extracted from
the physical operation of the appliances such as be temperature, light intensity, sound,
speed et cetera, depending on the operation of the appliances.

Appliance
Signatures

Non-Intrusive Intrusive

Steady-state Transient State Non-traditional Electrical Physical

Current Temperature
Transient shape
Power Light sensing
Transient duration
Harmonic Distortion Time of the day
Transient wavelet
Admittance et cetera
Transient size

Figure 2.1: Appliance Signature Classification

2.4.1.2 Non-Intrusive Signatures

The signatures captured non-intrusively, without contact with the appliances

are known as the non-intrusive signatures. They are the signatures used for NILM
systems. One unique characteristic of non-intrusive signature is that it is extracted from
the aggregated quantities, unlike intrusive signatures that may come from a single
appliances. The steady state signatures, under non-intrusive signatures includes,
current, power, harmonic distortion, voltage and admittance. These signatures are
obtained when the appliances are in steady state operation. These signatures are
 15

normally stable and they represent the normal operating characteristics of the
appliances.

The transient state of the appliances is the state preceding the steady-state in
the case of turn-on transient and it comes after the steady-state in the case of turn-off
transient. The transient state is characterized by overshooting of the current and power
consumption. The signatures of this behaviors of the appliances such as the transient
shape, transient duration and transient size are used in the load identification. The
transient state signatures are more unique than its steady state counterpart, which
makes its load recognition to have higher accuracy than the steady state [2, 3].

Although the transient signatures are very good in appliance recognition due
to their unique characteristics, but in some situations the signatures can misbehave due
to noise or other power quality problems. For this reason some load monitoring
researches used the non-traditional signatures to identify the loads. Non-traditional
signatures are similar to the physical quantities in intrusive signatures. Signatures like
temperature, time of the day and light sensing are sometimes combined with the
traditional (electrical) signatures to identify/disaggregate the appliances [2, 59, 60].

2.5 Load Monitoring Devices

The only way to ascertain the appliances’ energy consumption of a utility is

through proper monitoring systems. The use of traditional energy meter to know the
energy consumption of a utility had been used before the advent of modern appliance
monitoring systems. These energy meters could only give the total energy consumed
in the building within a certain period of time. The energy consumption of the
individual appliances or what cause the total energy consumption cannot be known
using the meter. However modern electrical load monitoring combines recent
innovations in technology such as digital electronics, sensor technology, computer
technology and telecommunication to form a smart monitoring system for the utility.
The load monitoring devices are the devices that make the building blocks of every
 16

load monitoring which can be categorized into the following; Measuring Devices,
Communication Devices, Optimization Tools, Recognition Devices, Display Devices
and Control Devices. Each of the monitoring devices perform different operation in
the system with some of the devices, like computer that can serve more than one
purpose.

2.5.1 Measuring Devices

The measuring devices, referred to as devices used in acquiring the

measurements of the electrical quantities consumed by the loads, which include any
measurement that can be used in monitoring the system. The choice of measuring
devices depends on the type and method of the load monitoring. These devices
includes voltage sensor, current sensor, energy meter, smart meter, current transformer
(CT), potential transformer (PT), smart plug, zigbee energy meter (ZEM-30), fluke
meter, energy detective (TED-pro) and power quality transducers. These equipment
measure the various quantities in the monitoring systems apart from the smart
appliances that come with their built in meters. In any case the measurement is required
for the data acquisition of the monitoring system.

2.5.2 Communication Devices

Communication is a vital issue in load monitoring systems. It performs the

function of conveying the data from one point to another in the system depending on
the load monitoring framework. Before the advancement in telecommunication, the
traditional communication networks uses wires to transfer the data from one point to
another, though some of the modern appliances monitoring systems still utilize the
wired communication network. Wireless communication has helped immensely in the
development of modern load monitoring and management systems.
 17

The wireless protocol widely used in load monitoring include: Bluetooth, WI-
FI, Infrared, Zigbee, Zigbee pro and other wireless area networks (WLAN). The
wireless communication techniques had reached the extent that some measuring
devices such as smart plug, TED-pro, smart meter and ZEM-30 have communication
capability. These devices can send the data they measured over the network of the
monitoring system. The wireless communication is flexible, cheap and less complex
than the wired communication, therefore when used in a load monitoring or
automation system an improved operation is obtained.

A comparison of different techniques in communication, that are applicable to

load monitoring in terms of operating frequency, data rate, power consumption and
application are shown in Table 2.1. When designing a load monitoring or management
system, the choice of the communication system can be made to suit the system
requirement. Zigbee communication is preferred by many researchers because of its
very low power consumption and simplicity in installation among others [61].

Table 2.1: Comparison of Communication Methods

Function Communication Method

Zigbee WIFI Bluetooth IR wireless
Data Rate 20, 40 & 50 54Mbit/s 1 Mbit/s 20 – 140 Kbit/s
Kbit/s 115Kbit/s 4 &
16 Mbit/s
Range 10 – 100m 50 – 100m 10 m <10m (line of
sight)
Operating 900 – 928 2.4 & 5 GHz 2.4 GHz 300Ghz –
Frequency MHz 430THz
Power Very low High High Low
Consumption
Typical Industrial Wireless LAN Wireless Remote control
Application control Communication
 18

2.5.3 Soft Computing Devices

When the measurements of the appliances’ electrical consumption (like current

and voltage) is taken, the next step in the appliance monitoring is the extraction of the
load signatures from the measured quantities [1]. This signature extraction normally
includes optimization so that the best features of the loads are extracted. This
optimization will match the load signatures to a possible combinations in the data base
with lowest possible error.

The optimization techniques in the load monitoring system can be either

deterministic or stochastic depending on the monitoring technique. The deterministic
optimization focuses on finding the global solution, whereas the stochastic
optimization uses randomness to maximize the process.

The computers play important role in the optimization process, as such the
algorithms are run using computer programs like matlab, C++ language and other
programming software. Also the microcontrollers available as the stand alone device
can serve as optimization devices in some monitoring systems.

2.5.4 Control Device

The appliance energy monitoring is conducted in order to control the load or

be aware of the energy consumption to facilitate energy saving measures. In any case
actions may be needed to control the loads, especially when the monitoring and
controlling of the system is automatic. Therefore when talking about the control of the
loads, there are devices like relay, contactors, breakers and timers that play the role.
Normally the control devices can switch on or off the load as the case may be and
when required by the monitoring system. Recently there are infrared switching devices
that control loads when they recognize movements, they are in the form of plugs,
switches and lamp holders.
 19

2.5.5 Recognition Devices

Appliance recognition is the key work of the load monitoring system. The
system uses the load signatures to recognize the status of the appliances in the system.
When the signatures are extracted and optimized they are then used to recognize the
appliances using appropriate algorithm. The recognition devices are mostly computers
and microcontrollers which may sometimes use data base to select appliance pattern
by optimization. When the appliances are recognized they can then be controlled for
energy management purpose

2.5.6 Display Devices

When the system appliance energy consumption is monitored, there is need to

see the monitored quantities whether in aggregated or disaggregated form depending
on the requirement of the monitoring system. The working conditions of some
electrical loads can be seen physically without coming closer to them, whereas on the
other hand there are appliances whose operation can hardly be observed from a far
distance. In any case, the display devices in the form of LCDs, computer monitors,
digital meters, I-pad and I-phones are used in so many monitoring systems to display
the required quantities.

Some load monitoring research works involves the transmission of data

through internet, zigbee or any communication facility, which may need the presence
of a display devices at the receiving end to indicate the coming of the data [62].
Sometimes the monitoring device system can be controlled from a remote places,
therefore there is a great need of the display devices for the system to work properly.
Every monitoring system should have measuring, recognition and communication
devices and there may be systems that do not have display. Figure 2.2 shows some of
the typical energy monitoring devices.
 20

Figure 2.2: Energy Monitoring Devices

2.6 Intrusive Load Monitoring

The intrusive load monitoring is the appliance load monitoring in which a

measuring device or sensor is installed on every load of interest. The ILM system is
meant to be systematic because the monitoring of every load is independent on the
other and it doesn’t require the aggregated quantities of the connected loads. Though
this kind of monitoring is systematic, it imposes high cost and complex installation
because of so many measuring sensors.

The ILM is the traditional load monitoring system which started since the
earliest time of electricity usage, and which advances with the advancement of sensor
and communication technologies. Both the researchers and manufacturers of load
monitoring earlier on decided to use multiple sensors for precision and simplicity. This
doesn’t lead to only high cost of installation but may also face reliability problem since
all sensors must work before an accurate monitoring is achieved. That is, should any
of the sensors fail, the appliances to be monitored by that sensor cannot be seen by the
 21

monitoring system. Nevertheless there are ILM systems all over the world and
researches are still being carried out in ILM [63].

Though the ILM systems are characterized by having a measuring sensor at

every load, but some monitoring systems can have a sensor at every plug level, in
which the plugs have multiple loads connected to them. Nevertheless, every
monitoring system with multiple number of sensing positions can be referred to as
ILM. Figure 2.3 gives an illustration of ILM system indicating the sub metering for
the appliances.

Figure 2.3: Intrusive Load Monitoring

ILM is the traditional method of load monitoring before the advent of NILM.
A monitoring sensor for every appliance makes it systematic eventhough with some
disadvantages, for long it has numerous applications [63]. The real applications of ILM
include; Understanding Local and Global energy consumption, Load Monitoring,
NILM simulation and evaluation, Recognition of human activities and Appliance
localization. By understanding the energy consumption ILM can be used to know the
energy consumption per appliance or it can be used to know the energy consumption
of the whole appliaces in the case of global energy consumption. The load monitoring
is the primary application of ILM and it can also be applied in simulation and
evaluation of NILM, which will leads to the accurate calibration of the NILM.
 22

Moreover the ILM system can also be applied to recognize human activities in
a building, which can be applied in the building’s energy management. The position
of an appliance can also be located using ILM, which can play role in fault location
and clearance [64]. Even in sub metering monitoring cases the appliance zone can
easily be detected for appropriate action on the appliance.

2.6.1 The ILM Framework

The intrusive load monitoring as stated earlier is characterized by having

multiple number of sensing positions in the monitoring system. It can have a group of
appliances assigned to a single set of sensors while having many of such sets of sensors
in the monitoring system. A. Ridi et al [63] has classified the ILM into three
subdomain. The first one is the sub metering system which is characterized by having
a meter or a set of sensors assigned to a zone of appliances. This means each meter or
set of sensors will give the aggregated measurement of that particular appliance zone.
Secondly we can have a smart plug classification in which each plug will measure the
aggregated quantities of the appliances connected to it and all together the plugs make
the ILM system. The third category is the smart appliance, comprising of a meter or a
set of sensors attached to each and every appliance in the system, which is the most
systematic of all. All the three classifications are ILM because they involved multiple
number of sensors. A. Ridi et al in [65] describes the activity recognition in appliances
monitoring as having three classes depending on load monitoring application. The
classifications are Activity of Daily Living (ADL) Recognition, Occupancy Detection
and User Appliance Interaction as illustrated below:

2.6.1.1 Activity of Daily Living Recognition

Some activities in home like eating, cooking, sleeping and bathing usually
involve the use of electrical appliances. ADL is the recognition of such activities using
 23

Electrical signatures in load monitoring systems and it makes things easier with elderly
and disabled/sick people.

2.6.1.2 Occupancy Detection

Occupancy detection is the recognition of the presence of activity in a building.

With occupancy detection it is very possible to take energy saving measures. Recently
there are switches that recognize the presence of people in a place and they can
automatically control the appliances with this functionality. They are really useful in
monitoring systems this days.

2.6.1.3 User-Appliance Interaction

The use of some appliances in a building may be time specific. Therefore it is

possible to know what time the user operates an appliances and use it to facilitate load
monitoring. The weather condition for example, can also give a clue about what
appliances is being used in a house.

2.6.2 The ILM Structure

The general frame work of the ILM can be subdivided into three; Detecting
Single Appliances, the Middleware Stage and Appliance Status Stage as illustrated in
Figure 2.4.
 24

Detecting Single
Appliance

Middleware

Appliance Status

Figure 2.4: The ILM Approach

This structure describes the stages of appliances recognition in ILM as

described below:

(iii) Detecting Single Appliances: The detecting single appliance stage is a stage
where sensor technology is used to detect the appliance status (ON or OFF) as
well as its location. Sensors like smart plug, light and temperature sensors can be
found in this stage
(iv) Middleware Phase: When the detecting side senses the appliance’s
measurement, the middleware, which is a software interprets status of the
appliances from measurements obtained in the detecting side
(v) Appliance Status Phase: This is the final stage of the ILM which shows the
status of the appliances, from which monitoring and control can be achieved.

Smart appliances plays important role in ILM systems. They mostly provides
the measurements in the detecting stage of the ILM, hence they help in providing a
fine-grained energy decomposition at appliance level, which facilitate the optimization
of the energy consumption.
 25

2.6.3 An Overview of Some ILM Research Works

The load monitoring researchers are still patronizing the ILM despite being
traditional load monitoring system. The ILM consists of a multiple number of sensing
points, as described in previous sections, therefore any research work in load
monitoring that adopts multiple sensing can be categorized as ILM. The ILM frame
work described ILM as having three classification, Sub metering, smart plug and Smart
appliances [63]. Majority of the ILM researchers used smart plugs as the energy
sensing devices. These smart plugs can display the measured power on LCD and can
transmit the measurements to the monitoring or computer system through USB,
wireless connection or via internet. The wireless connection normally use WI-FI,
Bluetooth, VLAN and Zigbee protocol.

A. Ridi et al [63] provided a survey on current research works in ILM and its
methodologies. It presented the potential applications of ILM and compare them with
the NILM counterpart. Though the work discovered that there are much more research
articles in NILM than in ILM, but an increasing number of published researches in
ILM were seen in the last few years. Lastly the survey demonstrated some key issues
to be focused by researchers, such as the energy consumption of ILM system as the
result of multiple components. In another work by A. Ridi and J. Hennebert [48]
presented an application of Hidden Markov’s Model (HMM) for appliance load
monitoring using ILM. The state-based model of the HMM is compared with Gaussian
Mixture Model (GMM) based on criteria of log-likelihood maximization and heuristic
approaches. The comparison done between GMM, log-likelihood and heuristic
methods when applied to intersection and unseen protocol shows that generally the
GMM outperform the log-likelihood method. However, the models in the system
require a compromise between the accuracy rate and complexity of the system. Also
imposing fixed parameters on the HMM and GMM as the total number of Gaussian,
as adopted in the system, can lead to a suboptimal solution.

Villanueva Marcel L. G. et al [66] developed an ILM system that can

automatically recognize home appliances based on an electrical consumption data set.
The system used k-NN for both the intersection and unseen protocols with accuracy of
 26

up to 99% in both protocols. Interestingly, a web application was developed for the
system, adapting web of thing architecture that presents the recognized appliances and
their corresponding power consumption. However, its cost will go up with the number
of sensors and it remain complex to configure D. Han and J. Lim [67] had developed
a smart home energy management system using zigbee sensor network. A new routine
protocol was developed known as the Disjoint Multi Path base Routine (DMPR) that
improved the performance of the zigbee sensor network. Both the hardware and
software of the system were developed and using active sensor network and actuators.
The control of various home appliances was achieved, even though their system lack
wide coverage, which makes it not applicable in a big premises.

S. Hussain et al [33] developed a cost effective and efficient home energy

management system with demand response capability. The system uses zigbee
communication and Power Conservation Analysis Tool (PCAT) and is found suitable
for real-time home energy management and smart grid application. Their system also
suffered the problem of having meters at every node to measure the power
consumption of the appliances. Also the cost of the home area network system should
be further reduced by employing cheaper communications like Zigbee, Zigbee pro and
RF technology. Several other research works in ILM have been done and researchers
are still exploring the area for more opportunities especially in building’s energy
management, control and automation.

Although the ILM is systematic by having a sensing device for every load of
interest, it suffers a lot of draw backs. In the first place its intrusive nature disturb the
privacy of the appliances and even the occupants of the building. One other drawback
of the system is that the more the number of appliances to be monitored, the more the
number of sensing devices to be used. This therefore means that if any of the sensing
devices fails, it may affect the whole monitoring system. That is in an event the system
is using an algorithm, the absence of the functionality of a sensor can lead the
algorithm to misidentification of the appliances [47].

The issue of system reliability can also be affected because there is higher
chance of failure when you have so many components performing different tasks in a
 27

system [68]. Moreover, the presence of so many sensing devices in the ILM makes it
more tedious to repair and maintain, and in case of failure of the system, it is likely to
incur more cost of reconstruction. Also an ILM system will consume more energy than
its NILM counterpart as the result of so many components, which in one way or the
other consume different amounts of energy. An ILM system will also require higher
initial cost than the NILM counterpart and may have higher computational
requirements [63].

The ILM is more of traditional method of load monitoring, but there are
researches that are still being carried out using the method with the intention of finding
better ways of load monitoring and management. Different ways forward have been
pointed out by different researchers in the field especially in reducing the power
consumption of the ILM [63]. Some of the issues to be handled by the researchers in
ILM include but not limited to the following:-

(i) One of the major reasons of load monitoring is to facilitate energy

management, therefore more efficient and cost effective applications of
ILM in energy management will be much welcomed in the area. For
example, a more reliable home energy management system using ILM
will make much impact in this research area.
(ii) The number of sensing positions in ILM can be reduced by adopting the
sub metering system, in which a set of sensing devices will monitor a zone
of appliances. This will not only reduce number of components in the
system, but will also makes the system cost effective and less energy
consumption (by the system components). Researches can be done to have
more exploration of this method especially when the appliances are
appropriately zoned. Appropriate zoning of the appliances in the sub
metering system will also solve the problem of similarity of the
appliances.
(iii) Unlike the NILM the ILM has very limited data base that is publically
available for researchers. Almost all research works in ILM use personal
data base, therefore the researchers can go toward forming a public data
 28

base to serve as reference for other ILM works, the work in [66] serves as
an example.
(iv) Though there are different types of appliance categories which makes the
electrical appliances data set somehow unpredictable, yet a more reliable
machine learning technique in ILM is in view.
(v) The ILM system usually consume more energy than the NILM because of
the numerous sensing devices. Hence a research with the aim of finding a
less energy consumption ILM, through reduced energy consumption
techniques, is appropriate. Reduced energy consumption can be achieved
by proper design and careful selection of the monitoring devices especially
control devices, communication devices and sensing devices

2.7 Non-Intrusive Load Monitoring

The Non-Intrusive Load Monitoring (NILM) system is a monitoring system

that used only one set of monitoring devices at the service entry to
identify/disaggregate the appliances in the building. This method provides a simple,
intelligent and cost effective methodology of ascertaining the status of the appliances
without necessarily getting contact with the appliances [69-71]. The energy
consumption patterns of any utility can best be known when an appropriate load
monitoring is observed, which will give the state of operation of each appliance and
its energy consumption. Providing energy consumption awareness is the primary work
of every load monitoring system and the knowledge of the energy consumption pattern
is what motivate the customers to take energy saving measures. The energy saving
measures may include home energy management and control which can lead to great
savings in the energy and ofcouse the savings in electricity bill. The global energy
demand is always increasing due to civilization, urbanization and increase in world
population, therefore load monitoring especially NILM need to be implemented to
device energy management systems for home, industries, offices and other commercial
buildings [14, 72].
 29

According to some studies in NILM, if the consumers are promptly given a

feedback about their energy consumption at appliance level, that will greatly motivate
them to take appropriate energy saving measures, which will as well help the electricity
utility company and other energy stake holders [1, 73]. As the result of excessive
carbon dioxide emission due to global energy related activities, there are efforts by
different organizations and authorities in different part of the world to reduce the Green
House Gas (GHG) emission. Fossil fuel source provides the majority of the electrical
energy at global level, these sources are non-renewable which are estimated to decline
in about 50 years to come. It is therefore necessary to plan ahead of the time, an action
that will cater for the problem. One of the quickest solution to this is to device a means
of energy management and optimization through load monitoring in order to help in
avoiding unnecessary and illegal power usage.

Researchers discovered that a maximum energy saving can be achieved if the

consumers are given direct feedback about their energy consumption. That is rather
than monthly readings if a real time consumption can be available at smallest possible
interval of time this will help the consumers greatly [2, 25]. The NILM is the cost
effective method of load monitoring, therefore if the majority of the utility consumers
are equipped with proper NILM systems, they will be quite aware of energy
consumption of every individual appliances and this can help them to identify energy
saving appliances for patronage. With this, even the manufacturers of the appliances
will compete in producing more energy efficient appliances such as lightings and
kitchen appliances.

One of the useful concepts that highly promotes energy management especially
in the smart grid system is the Demand Response (DR). The DR system operates in
such a way that some loads will be cutout during peak periods to allow the power
system to contain the peak demand. Normally the consumers take the action of
switching off those appliances in response to the DR and the utility offer them a
reduction of tariff in return or sometimes for the customers to participate in greenhouse
gas reduction [74]. In traditional DR system the consumers may not reliably follow
the DR instruction due to some circumstances. In that case a NILM system can be
applied in the DR program to have an overall system that can follow the load shedding
 30

instructions [38, 75]. The electronics and communication are also playing role in load
monitoring and management. Their recent advancement especially communication
contribute immensely in load monitoring and management aspects through a system
of interconnected home area network. Numerous electronics components have
emerged over the last few decades that promotes the load monitoring research works.
A research estimated that through a good demand management and tariff innovation,
a utility company with 5 million customers can save roughly $10 million USD per
annum plus a possible addition of $6 million USD through customer usage information
[76].

The NILM system perform the load monitoring without intrusion into the
consumer premises. The load disaggregation/identification uses the voltage and
current waveforms obtained at the service entry to extract the signatures of the loads
and use them for the load identification. Hart used the steady-state variations of the
real power (P) and reactive power (Q) consumption of the appliances to determine the
status of the appliances in ∆P - ∆Q plane. Hart’s work was based on the assumption
that the power consumption (both real and reactive) of each and every load in the
building is unique. Hence a change in real power and reactive power should correspond
to a unique load in the system. Even though Hart and his team mates were the first to
discover the usefulness of the NILM in appliance recognition, they still discovered that
appliances with similar real and reactive power consumption can suffer the problem
of overlapping cluster in the ∆P - ∆Q plane as shown in Figure 2.5 [7].

The appliance identification/disaggregation also depends on the type of

appliances, which are classified according to functionality of the appliance and its
energy consumption pattern. There are four such classifications and the more
complicated the energy consumption pattern of the load the more the complexity of
the identification/disaggregation procedure. These classifications are divided into
Type 1, 2, 3 and 4 [7].
 31

(i) Type 1:
This refers to the appliances with only two states of operation, they can
only be ON or OFF. Example of such loads include; normal Electric lamp,
Electric kettle and Toaster.
(ii) Type 2:
The type 2 loads are the multi-state appliances that have finite number of
operating states. The ON state of the appliances can have different states
like in Table fan, Blender and Washing machine.
(iii) Type 3:
These are the continuously varying devices (CVD) having variable power
consumption with infinite number of states. Example of such loads are
dimmer light and power drill. More advanced load monitoring system is
required to disaggregate these kind of appliances.
(iv) Type 4:
The Type 4 devices are “permanent consumer devices”. They don’t
necessarily change power consumption, but they are active throughout
certain period of time. The devices like Telephone set, Television receiver,
CCTV camera, and Smoke detector are some of the appliances in type 4.
The power consumption of type 4 appliances are normally subtracted from
the aggregated value before disaggregation.

Figure 2.5: Overlapping Cluster in ∆P - ∆Q plane

 32

2.7.1 NILM Framework

The NILM system uses voltage and current waveform of the aggregated load
consumption at the service entry to perform the load identification. Normally analysis
is carried out on the voltage and current waveforms to extract the features of the
connected loads. This is what makes the NILM system to have only one set of
measuring sensors at the service drop as shown in Figure 2.6.

Figure 2.6: Configuration of NILM System

The appliance disaggregation/identification in NILM basically involves three

stages: these are Data Acquisition, Feature Extraction and Appliance Classification [2,
77]. The major difference between ILM and NILM is that the appliance recognition in
ILM will be much simpler since there is a meter connected to each appliance, whose
reading can be accessed to ascertain the appliances operating condition. Therefore the
issue of appliances’ operating condition can easily be determined. On the other hand
the NILM has only one set of measurement to be analyzed and determine the status of
many connected loads in the building. Thus the measurements in NILM have to
undergo the three stages in order for the loads to be recognized. Figure 2.7 shows the
stages of NILM which are described in the following sections.
 33

Figure 2.7: The NILM Approach

2.7.1.1 Data Acquisition

The data acquisition in NILM is the measurement or sampling of the

aggregated load consumption (primarily current and voltage) at appropriate rate for
use in the appliance identification. The data can be sampled at appropriate rate
according to the nature of the signatures to be acquired. Basic electric metrics like real
power, reactive power, root mean square voltage, root mean square current, apparent
power and power factor can be calculated or captured using the voltage and current
samples. The need of higher harmonics in the load signatures may arise in some load
monitoring systems under the steady state analysis, in that case the sampling rate has
to follow Nyquist criterion. Some of the disaggregation procedures in NILM uses noise
or transients signatures to identify the loads, in this case the sampling rate must be high
enough to capture the waveform quality [15, 78, 79]. The work developed by T. M.
Chung and H. Daniyal [80] is capable of accurately measuring rms voltage, rms
current, real power and power factor which is possible to be modified for NILM
application. The data acquisition can also include the non-electrical signatures like
motion detection, light sensing and timing, which can also be used to identify the
appliances.
 34

2.7.1.2 Feature Extraction

After the data acquisition, the next step is the feature extraction, which can also
be called signature extraction. It is the collection or calculation of the appliances’
signature from the acquired data which use to be current and voltage samples. The
signatures are the results of manipulation of the aggregated measurement, therefore in
most cases of NILM the signature extraction involves data processing and data
optimization in order to extract the most accurate signatures of the appliances [81, 82].
The extracted features can be steady-state, transient or non-traditional, and depending
on the NILM type, the steady-state signatures are usually simpler and cheaper to use.

2.7.1.3 Appliance Classification

The signatures obtained from the feature extraction are used to run algorithms
for the load identification which is called appliance classification. In most research
works a supervised method is used in the disaggregation, which require an inference
data for training the classification algorithm. The unsupervised method has also been
tried by some researchers to eliminate the need for the training data [83-86]. When the
appliances energy consumption is disaggregated the result of the monitoring system
can be made available to the consumers or it can be used by the system for further
control measures.

2.8 NILM Disaggregation and Identification Approaches

The signature used in the load identification in NILM can be used to classify
the disaggregation methods. This is so because the signature used in NILM system
normally determine its complexity. There are two major signature classification in
NILM namely the traditional (electrical) and non-traditional (non-electrical)
signatures. The electrical signatures can be steady-state or transient whereas the non-
traditional signatures are non-electrical quantities that can be captured from the
 35

appliances operation, such as temperature and light. The disaggregation approaches

may use steady-state, transient-state, hybrid or non-traditional signatures [87, 88].
Every research work in NILM has to select appropriate features that can best classify
the loads with as high accuracy as possible.

2.8.1 Steady-State Approaches

This method uses the signatures of the loads in steady state operation to
perform the load identification. These steady state signatures include: Real power (P),
Reactive Power (Q), Apparent Power (S), root mean square Voltage (V), root mean
square Current (I) and the steady state harmonics. All the measurements in the case
of steady state analysis are taken when the loads are steadily in operation [17, 89]. The
real power and reactive power are the most common features used in steady-state
analysis though it is difficult to identify Type 3 load with power consumption. Another
problem with P and Q signatures is the possibility of overlapping clusters, which is
why some researchers combine many steady-state signatures for their load
identification [90].

Some researchers used the steady-state real power alone to disaggregate the
loads [9, 91], where it was found to be very accurate when disaggregating high power
appliances and not very good in disaggregating low power consumption loads and
those with similar power consumption. H. H. Ming [90] tried to increase the reliability
of the steady-state signatures by combining many steady-state metrics as the load
signatures. He used real power, reactive power, apparent power, rms current and power
factor of the aggregated load as signatures for the identification. A load profile of 33
lightings consisting of three different type of lighting is used. The load profile was
constructed using a sub metering and it is used as inference for the load identification.
The Microsoft Visual studio is used to analyze the system which was found to be
97.37% accurate when tested with arbitrary load combination. Though, the system has
an improved load recognition, but it will be more complex to implement the algorithm
as the result of too many signatures. In another related work N. F. Esa et al [41]
analyzed the impact of different sampling rate of the aggregated data in determining
 36

the energy consumption pattern of various appliances. The system is analyzed using
one second, one minute and five minutes sampling intervals. And it has been found
out that the smaller the sampling time the better the estimation, even though the smaller
the sampling time, the more the system becomes impracticable as the result of memory
requirement for storing the data, which render the system to uncertain appliance
detection and low accuracy.

T. Hassan et al [92] and Nur Iksan et al [93] developed a method of extracting

load signatures using V-I trajectory. The V-I trajectory is a plot of voltage against
current over one cycle of the sinusoidal voltage. The voltage and currents are
normalized before plotting the trajectory. The nature of the trajectory is used to identify
the loads and the researchers used properties like asymmetry, area, looping direction,
curvature of mean line, self-intersection, slop of middle segment, area of left and right
segment and peak of middle segment to classify the appliances. The V-I trajectory
really explore an important method of acquiring load signatures which can help in
realizing a more accurate load monitoring systems. Figure 2.8 shows the V-I trajectory
of some group of loads [92]. This works also suffered the problem of computational
complexity as the result of the formation of the complex load taxonomy.

Figure 2.8: V-I Trajectory of some group of loads

 37

Some researchers uses the steady-state harmonics of the connected loads as

signatures for the identification [16, 94, 95]. Harmonics occur when there is non-linear
load in the system which as the result produces non sinusoidal waveform with multiple
frequencies. This frequencies are the odd multiples of the supply frequency. S-G Jesus
Ivan [91] presented a method of detecting and classifying a steady-state events using
fuzzy c-means clustering algorithm. The steady-state analysis uses only the
fundamental and first harmonic, which improved the computational cost of the
algorithm. However the c-means clustering algorithm needs proper validation. C. Y.
Feng et al [16] created a load profile containing 3rd to 15th harmonics of the current to
be used in identifying the load combinations. The system was successful using three
kinds of load; PC, compact fluorescent lamp (CFL) and T5 fluorescent. The
disaggregation using harmonics will be much efficient if the loads in the monitoring
system exhibit harmonic behavior, though the system will also suffer from the
computational complexity as the result of many harmonics computation. The
introduction of the odd harmonic signature in [94] made a valuable contribution of
improving the efficiency of the identification algorithm. To improve the accuracy of
the steady-state signatures in NILM M. Temneanu and A. Ardeleanu [96] combined
the conventional changes in real and reactive power (∆P and ∆Q) with steady-state
harmonics. The detection accuracy of the system is about 100% when applied on four
working machines. Even though the work focused only on harmonic loads.

2.8.2 Transient-State Approaches

The transient usually occurred at the time of switching a load in a power

supply system, and it is associated with rise in current consumption for a very short
period of time. The transient behavior of electrical appliances is distinct, that is why it
is used for load identification in some NILM systems [2, 60]. The transient time differ
from one appliance to the other and the transient has many characteristics such as
transient time, transient shape and total energy that can be used as load signatures.
When an appliance with transient behavior is turn ON, the momentarily transient
happen which precedes the steady-state operation. Figure 2.9 shows a typical transient
waveform, which normalizes after the transient period.
 38

Figure 2.9: Transient Waveform

The work of H. H. Chang [15] uses transient time and transient energy with
short term Fourier transform (STFT) and discrete wavelet transform (DWT) to identify
the loads. In this case the transient analysis was compared with the steady-state
analysis comprising real power, reactive power and harmonics signatures where the
transient analysis was observed to have highest accuracy when recognizing three major
appliances. The computational requirement was also better when the analysis is carried
out using the transient features, though the system will face more challenges when
applied on more than three loads and it could solve the real-time NILM problem when
applied in sub-metering NILM. Liu B. et al [97] uses transient power waveform (TPW)
sample and template time series in dynamic time algorithm to identify the appliances.
Through a comparison test with the existing methods, the proposed method improved
the accuracy of load transient identification and is cost effective for implementation,
even though the frequency of 10Hz used for the data sampling is low compared to the
standard. Another transient analysis method was proposed by C. Duarte et al [98]
where the turn on transient is analyzed using filtering method. A clean transient
features were obtained using the “envelop trigger”, even though the process introduces
temporary selectivity reduction and larger computation time.
 39

The transient analysis gives a very good method of load disaggregation in

NILM, even though the monitoring system may face challenges in differentiating the
turn on and turn off transients, which is why sometimes the hybrid method is adopted.

2.8.3 Approaches Based on Hybrid Analysis

The hybrid signature analysis is the combination of transient and steady-state

signatures in a NILM system in order to increase the recognition accuracy. This has
been adopted by some researchers because of the diversity of the electrical appliances.
Some of the electrical loads are better recognized using steady-state features while
some are better recognized using the transient features. Therefore combining the two
types of signatures will make use of the advantages of both methods. L. K. Norford
and S. B. Leeb [99] had earlier on worked on a NILM system that analyzed turn-on
transient and steady-state algorithms for NILM. H. H. Chang et al [55] provided an
analysis with a combination of real power (P), reactive power (Q) and total transient
energy (UT). The work was able to disaggregate different appliances with the same real
and reactive power consumption, where all the load combinations are accurately
recognized using the combined signatures. However differentiating the turn on and
turn off transient remain the limitation of the work.

In another hybrid based method, Liang et al [51, 52] extracted multiple

signatures from both steady state and transient operating states of the loads. The
appliance disaggregation uses “committee decision mechanism” to produce final load
estimation from the disaggregation using individual signatures. The separate
disaggregation used the following signatures: Active and reactive power (PQ), Current
waveform (CW), Eigen values (EIG), Harmonics (HAR), Instantaneous admittance
waveform (IAW), Switching transient waveform (STW) and Instantaneous power
wave (IPW). Thus after disaggregation with these signatures the committee decision
will decide the final recognition. Though the recognition efficiency of the method is
very high, its high computational requirement may make it impracticable in a real
situation.
 40

2.8.4 Approaches Using Non-traditional Signature Analysis

Non-traditional signatures are the non-electrical behaviors that can be

exhibited by different electrical appliances. For example the temperature inside a
refrigerator can give a valuable information for load monitoring and management.
Some natural or habitual behaviors of the loads can also be used as non-traditional
signatures. This non-traditional signatures include time of the day, frequency of
appliance usage, light sensing, sound made by the appliances and on/off duration of
the appliances, which can be added to the normal electrical signature to identify the
loads.

H. Kim et al [100] used the time of use and how the appliances are used with
the normal electrical signatures to disaggregate the loads. Their work demonstrated
that the unsupervised techniques can provide more accurate appliance disaggregation
with the additional features of when and how the appliances are used. However
determining the state of individual loads is the limitation of the work. A non-traditional
signature of the appliance can be used to distinguish it from another appliance with
similar electrical properties [101].

2.9 Event based and Non-event based Detection in NILM

The load identification methods in NILM can be classified into event-based

detection and non-event based detection methods. In the event-based approach, the
aggregated measurement of the signal is used in the appliance detection. Normally an
edge detection algorithm is used on the power consumption curve of the system to
trace the changes in the curve, which will indicate state changes of the monitored loads.
The edge features of the power curve are normally classified using a set of rule by
machine learning approaches. In a research work related to event-based detection
Anderson et al [102] used the event-based method to highlight the importance of using
power metrics in selecting an event detection algorithm. Their system needs the
incorporation of the power change information metrics to determine the total energy
 41

disaggregation. The work of L. Jiang et al [103] utilizes edge symbol detector (ESD)
and transient detection approach based on SVM to precisely locate the loads event.
However, the work can only accommodate limited number of appliances and no
transient event is involved. Most recent event-based detection was demonstrated by
[104] and [105]. In the research work by L. D. Baets et al [105], a new event detection
method was proposed. The system works in frequency domain and uses Cepstrum
smoothing method to eliminate the noise from the signals. Although the frequency
components can be used for both event detection and appliance recognition, but the
method is more opaque than the traditional chi-square method. C. C. Yang et al [106]
presented a systematic approach to ON and OFF event detection algorithm for NILM.
The algorithm is capable of filtering out the noise and detecting the events
successfully. But the use of only apparent power (VA) as a detection signature can
affect the accuracy of the recognition.

Non-event based detection approaches on the other hand do not relay on the
edge detection to classify the appliances. Instead it takes every sampled data as
inference to be utilized for the classification. In other words non-event based detection
is a method that combines event detection and classification in one process. The non-
event based methods are applied in [100] and [107] . J. Z. Kolter and T. Jaakkola [107]
achieved a very good and accurate disaggregation when they used Additional Factorial
Approximate MAP (AFAMAP) inference. The method have the potential to
disaggregate virtually all appliances in home, but its applicability is greatly reduced
by the substantial number of HMMs.

2.10 Artificial Intelligent Learning in NILM

The NILM is a system that uses one set of measurement to

identify/disaggregate the connected loads in the whole building. The load classifier
may have to learn the system using artificial learning techniques that can be solved
using different algorithms. The learning techniques can be broadly classified into two
groups: the supervised learning and unsupervised learning techniques. In artificial
 42

intelligent learning, a matching is performed between data base created and extracted
features from the metering point to estimate the disaggregated appliances [108].

2.10.1 Supervised Learning

The supervised machine learning techniques is a system that requires initial

data for training the classifier. The data is needed to perform an offline training of the
classifier so that it can recognize the appliances when the signatures of the loads are
delivered to it. The supervised learning uses either optimization or pattern recognition
to perform the classification. In optimization method the algorithm compares the load
signatures with the data base and try to minimize the error between the two. The output
in the data base whose input is the closest to the input data at hand will be presented
as the output. The pattern recognition uses pattern matching of the appliances in the
classification. There are numerous researches under the supervised learning
approaches, some of them uses ANN, some used Support Vector Machine (SVM) and
some used k-Nearest Neighbor rule (k-NN) [109], [110], [111] & [112]. H. T. Yang et
al [111] used Electromagnetic Transient Program (EMTP) to analyze NILM system
realized using ANN. The experimental results shows that the backpropagation
classifier outperform the learning quantization classifier with up to 95.6% accuracy.
However, the number of hidden neurons in the back propagation network will affect
the computing time of the NILM.

2.10.2 Unsupervised Learning

As mentioned in the previous section, the supervised learning requires training

data before the appliance classifier is developed and this training data may be tedious,
time consuming and expensive to obtained. The unsupervised learning on the other
hand is a learning method in which no training data is required and hence the data set
will only have inputs without outputs. The work presented in [85] is an unsupervised
learning approach which used Gaussian Mixture Model (GMM), Sequential
 43

Expectation Maximization and Adaptive Fine-turning to detect the unknown states of

the loads. B. Wild et al [113] presented an unsupervised NILM based on Kernel Fisher
Discriminant Analysis (KFDA). The system shows an excellent detection performance
when experimented with BLUED dataset, although the analysis reveal some short
peaks in the system that affects the detection performance.

2.11 NILM Reviews

The NILM method forms the majority of the research works in load
monitoring, in fact, more than 90% of load monitoring articles are in NILM areas. It
is recommended to study review articles in an area of research in order to know the
current state of art in the area. Several reviewers presented review articles promoting
the state of the art of NILM. General issues like feature extraction, optimization,
disaggregation algorithms, appliances classification and NILM applications have been
discussed and analyzed by so many reviewers [2, 10, 114]. [2] recommended that there
is need for general NILM evaluation method. Faustine A. et al [114] discovered that
there is lack of proper NILM research works in the developing countries and the need
of real-time application of NILM in homes is still anticipated. Regarding the signature
extraction and disaggregation algorithm [60] and [115] provide the relevant literatre.
[115] recommended that a load feature that provides little value should be removed
from the load signatures in order not complicate the NILM model. However, not all
the review articles give considerable support to its application in appliances’ energy
management, which is one of the most important in power sector [2].

NILM methodologies have also been studied by [116], [117] and [118] in an
effort to assess the NILM methodologies applied by various researchhers. Klemenjak
C. and Goldborough P. suggested that NILM is an ongoing research area and is still
open for researchers as we are yet to find a perfect NILM. A general review of NILM
context and solutions for implimentation of the NILM systemis provided by [119]. The
work also found that the continuous development of NILM system will help in
developing energy efficient programs and reduce energy usage. However, the NILM
 44

methodologies are still emergining and therefore the NILM reviews will have to be
updated in the near future.

2.12 General NILM Applications and Future Research Directions

The major purpose of load monitoring is to keep track of the energy

consumption in the building, as the importance of knowing the energy consumption
per appliance cannot be overemphasized. The NILM provides the appliances
disaggregation/identification while protecting the consumer privacy, as the result of
single point sensing at the service entry. The monitoring of the connected loads is what
motivate and facilitate energy saving measures and even energy safety measures in the
building. Problems like fault detection and clearance can easily be solved using proper
NILM system with a proper algorithm. Beside the one point monitoring system offered
by the NILM, the complexity of its circuitry is also less compared to ILM where the
monitoring system has to have direct access to each load. This will also help in making
the NILM system, simple, cost effective and reliable.

Energy management is one of the major concerns of load monitoring

researchers and that is why NILM researchers have concern over energy savings
related works. Ninad K. et al [13] developed a non-intrusive load monitoring that
overlooks the maximum demand of the system to make sure that the system does not
consume more than the threshold power. S. Barker et al [120] made a recommendation
that the research works in NILM should now concentrate on application rather than
spending all our research efforts in accuracy improvement and searching for new
disaggregation algorithms. In another work by R. W. Cox [21] the application of NILM
in shipboard supervisory control was demonstrated and the results showed that NILM
can be applied in shipboard management and control both efficiently and cost
effectively. [121] did similar work in shipboard fault detection, although the system
needs proper validation procedure.
 45

Application of NILM is also found in other professions like fluid/water

mechanics where the flow rate of the fluid is observed by some sensing devices non-
intrusively. The research work by M. M. Campagna et al [122] uses the linear
relationship between the fluid flow in the pipeline and the amplitude of its transversal
vibration oscillation to develop a way of measuring the flow rate non-intrusively.
Other similar applications of NILM in water/fluid monitoring can be found in [123]
and [124]. Moreover, the NILM is a potential application tool for future smart grid and
smart home energy management systems. T. K. Nguyen et al [14] developed the real
time, low cost and high performance NILM system based on System on Chip (SoC)
with field-Programmable Gate Array (FPGA) for smart system application. The
system has the capacity of solving the real-time feedback feature in a low cost NILM
system, although the algorithm cannot detect and classify two simultaneous events.

Researchers have worked tirelessly in the area of NILM mostly trying to

improve the recognition efficiency and exploring more load signatures and recognition
algorithms. However more efforts are needed from the researchers in order to realize
a real time applications of NILM with as high accuracy as possible. Though every
research effort and finding will be appreciated in the area, but issues that should be
focused by the researchers include the following:

(vi) So many researchers in NILM have addressed the issue of recognition accuracy,
but efforts to come out with better recognition algorithms that will yield better
accuracy are still needed in the area.
(vii) The application of NILM in appliance energy management, control and
automation also needs the attention of the researchers. This is together with the
need for a method of measuring the performance of NILM [125].
(viii) The supervised learning approach can be very accurate in monitoring some fixed
appliances in the NILM system. It can achieve very high accurate monitoring
when proper network is obtained. Hence, the researchers can explore more in the
application of supervised method for NILM incorporating home appliance
energy management.
 46

(ix) The NILM process that can recognize so many loads is also needed. This is for
the fact that the electrical appliances are numerous in number and it is easy to
find many appliances in residential or home settings.
(x) There are many researches in NILM incorporating the use of smart meter, which
has sampling and communication capability. Therefore, energy management and
automation researchers can make use of the opportunity of smart meter to explore
more applications of NILM.
(xi) There are limited research works in NILM involving the use of non-traditional
signatures. Yet non-traditional signatures can increase the disaggregation
accuracy [101]. Hence, the non-traditional signatures should be explored more.
(xii) Also the NILM research works has limited data-sets from the developing
countries [114]. Hence there is need to develop a standard NILM data-sets from
different geographical regions of the world.

2.13 Application of Load Monitoring in Appliances’ Energy Management and

Control

Energy monitoring and management are very important in every electrical

power system. The fundamental action in energy management and automation systems
is the proper monitoring of the appliances’ energy consumption. By default, majority
of the electrical energy generations worldwide comes from oil and gas resources which
are decreasing in reserve and perishable over time. The load monitoring system can be
used to eliminate the critical inefficiencies in the power system, which will lead to the
solution of the problem of energy crises and global warming [126, 127]. Advances in
measurements and instrumentations, computer hardware and software technology, and
communication technology make it possible for a load monitoring system of a large
number of utility consumers to be realized accurately and economically [29, 128].
Researchers adopted different methods of appliance monitoring that facilitate energy
management and optimization. With the invention of an energy management system
that can be used globally, the energy crises and even the global warming can be
reduced to minimal [129, 130].
 47

Majority of the energy management researchers are targeting home and offices
because there are a lot of mismanagement of the energy happening in the sectors and
they form the majority of energy consumption in most countries of the world,
especially the developing countries. It is estimated that about one-third of the energy
consumed in residential settings is wasted as the result of unaware activities of some
of the appliances. To have a quick view of what the researchers have done in energy
management, a lot of review articles have been written in the area [129, 131-140]. The
study of the literature articles will give the researchers a highlight of the energy
management researches in the current state of art. Lee D. and Cheng C. C. [131]
analyzed the management performance of 305 energy management systems from 1976
to 2014. The report discovered that the building energy management is increasing by
11.39% to 16.22% annually.[132] provided a systematic review of academic journals
on EMS that covered about 44 selected journals. Lopez R. I. et al [135] suggested that
the supply demand matching (SDM) model is a valid mechanism for managing
distributed electricity networks (DENs).

In [136] various methods of controllable loads management have been

presented and the work discovered that the controllable loads management approaches
are effective in providing balancing service in distribution systems. Energy
management cannot be achieved without proper controls of the loads. Nanda A. K. and
Panigrahi C. [137] presented a literature review of energy management system
scheduling, control, real-time automation and communication. The study suggested
some better approaches for smart home energy management in future energy efficient
environment. The appliance energy management area is also a growing area, therefore
more recent reviews will reveal the new idears.

2.13.1 Energy Management Approaches Based on ILM

The ILM is a multiple sensing monitoring system as described in the previous

sections. The classification of ILM in terms of appliance intrusiveness are sub
metering systems, smart plugs systems and smart appliances systems. Therefore any
load monitoring or energy management system that encompasses the use of
 48

distributive sensing method is considered to be under ILM and on the other hand the
NILM is solely for researches that have one set of sensing devices.
Researchers in home energy management and automation have made various
contributions regarding the multi sensing approach. W. H. Kim et al [141] developed
a system of energy monitoring and control with postgre SQL using operating system
independent. The system can provide real-time energy consumption of the appliances,
which can be displayed on smart phones, although the system cannot automatically
detect the appliances. J. Han et al [34] used zigbee and PLC in home energy
management system with renewable energy alternative. The system used Energy
Management Control Units (EMCUs) as the smart plugs to monitor the power system.
The system is expected to enhance the home energy management and energy saving,
even though the use of Zigbee pro instead of Zigbee would improved the system
communication. Table 2.2 summarizes some of the recent ILM works which are
applicable in appliances’ energy management. The ILM is very systematic in energy
management application because each appliance has its own meter or sensor for
measurement. The first drawback of the method is its multiple sensing that makes it
complex. Another shortcoming of the ILM system in energy management is its high
cost due to so many sensors and also high energy consumption. The ILM system is
also susceptible to reliability problem, because failure of one of the sensing devices
can affect the ILM algorithm which can lead to misdetection or misclassification of
the appliances.
 49

Table 2.2: Some Key Studies of Energy Management in ILM

Study Methods Observations

[48] Ergodic Hidden Markov GMM outperform the HMM
Models using log-likelihood in the intersection protocol
maximization and heuristic and HMMH outperform
methods and using GMM on both GMM and HMM in the
the other hand. It is carried unseen protocol. The
out using intersection and accuracy of the methods
unseen protocols. makes them suitable for
energy management system.
[142] Multilayer decision The system monitoring has
architecture 10% error which is good for
energy management
especially when only
important loads are
considered. Also the
processing of the sensor data
at the node limit the data
transmission.
[143] A green building with sensor The developed green
based intelligent system that building prototype is capable
monitors and control the of saving energy and
appliances used in the adapting the behavior of the
building. appliances to the real
environmental conditions.
Thus the system formed a
very good energy
management system.
[144] Involved using HTML and PLC allow the system to be
Java scripts. used without extra
connection. It can form a
very reliable HEMS.
 50

[66] The study uses AGILASx The k-NN recognition

Dataset consisting of real experiment with k=10 for
power, reactive power, intersession protocol got
apparent power, rms voltage, highest accuracy of 99.17%
rms current and power at a data frequency of 0.1Hz.
factor. k-Nearest Neighbor
(kNN) is used as the
recognition algorithm for
both intersession and unseen
pfotocols
[145] End device, router and The work achieved energy
control Algorithms using saving of 69% by reducing
microcontroller. the light intensity of the LED
lighting system. The energy
saving of more than 55%
observed makes the system a
promising HEMS.

[146], [147] DR Algorithm using The system can effectively

simulated using C++ keep the total power
consumption below a
specific limit and The
controller response time is in
milliseconds therefore it can
serve as a real time HEMS.
[33] Power consumption analysis The system has the
tool (PCAT) using moving capability of remote
average algorithm. controlling the power supply
to the appliances hence is
good for HEMS.
[148] Load tree using functional The system can recognize
decomposition and load with similar and
individual decomposition. overlapping signatures.
Hence good for HEMS.
 51

[149] GMMs and k-NN The GMMs outperformed

the k-NN with up to 93.6%
accuracy which is also very
good for energy monitoring
and management.
[32] Naive Bayes classifier The work presented a
(NBC) and HMM hardware design of HEMS, a
good effort in energy
management using non-
traditional signatures.
[150], [151] & [152] The system uses a The in-Home Energy
scheduling algorithm at the Management (iHEM)
EMU implemented using developed reduces the
C++. energy expenses,
contribution of the customer
to the peak load as well as
reduces carbon dioxide
emission. The iHEM also
provides savings close to
that of optimization-based
residential energy
management (OREM),
hence the iHEM can form a
very good HEMS.

2.13.2 Energy Management Approaches Based on NILM

The NILM methods use algorithms on the collected single point measurements to
recognize/disaggregate the loads. Therefore NILM works that are related to energy
management are classified under this category. Men-Shen Tai and Yui-Hsiu Lin [153]
produced an adaptive NILM system for energy conservation. The system make use of
the turn-on and turn-off transient signatures of the appliances to identify them using k-
Nearest Neighbors (k-NN) and Back Propagation ANN. The system also used
Artificial Immune Algorithm (AIA) with fisher criterion for adjusting the loads feature
 52

parameters when a new load is introduced into the system, thus it forms an extendable
NILM system. The recognition accuracy of the system is 95% and above. This is an
indication that the system can be utilized as a good monitoring and management
system. Although the system identifies the operating status of the appliances, but the
k-NNR recognizer can only recognize the de-energizing operation status of the
appliances.

The work proposed by N. J. Nunes et al [154] demonstrated a low cost eco-

feedback load monitoring system that uses audio input and single point current and
voltage measurements to identify the real-time energy usage of the appliances. The
study on 21 houses over a period of 9 weeks achieved an overall of 9% decrease in the
total energy consumption. Although the work demonstrated that the user events are
not directly related to the aggregated consumption, but further investigate is needed to
understand the correlation between user event and energy consumption. Table 2.3
shows some of the key research works in NILM applicable in energy management
especially HEM.

Due to the continuous increase in population, urbanization of rural areas and

advancement in civilization all over the world, there is continuous increase in energy
demand from all aspects of life. Industries, transportation and electricity generation
are some of the major energy consumption sectors of the world. According to a recent
report of the United States Environmental Protection Agency (USEPA) [155], the
major greenhouse gases emitted by human activities include Carbon dioxide, Methane,
Nitrous oxide and Fluoride gases with Carbon dioxide having the highest percentage
(Fig. 2.10). This therefore indicates a very serious issue regarding global warming and
a way to decrease the gases should be devised [156, 157]. The USEPA report also
shows that Industries, Electricity generation and Agricultural activities are the major
contributors of GHG emission. However when classified by countries China is the top
in carbon dioxide emission followed by the United States. Fig. 2.10 illustrates the
charts of the global GHG emission by Gas, by economic sector and through fossil fuel
combustion by countries [155].
 53

Figure 2.10: Global GHG Emission

According to a research published by PBL Netherlands Environmental

Assessment Agency [158] a reduction of 40% GHG emission can be achieved if
countries reduce their emission by some percentage including 25 to 31% for US, 38 to
48% for Japan, 19 to 34% for China, 26 to 41% for Canada and 32 to 49% for Brazil.
No doubt the process of achieving these targets must involve the load management
[131]. From the global perspective, the GHG emission can be reduced by taking
actions to see that the use of these energy sources is minimized and optimized [159],
hence the energy management should be applied to contribute its quota in reducing the
world GHG emission.
 54

Table 2.3: Some Key Studies of Energy Management in NILM

Study Methods Observations

[153] k-NNR and back Recognition percentage of
propagation ANN. more than 95%.
[160] The system used NILM to Simulation results of the
Monitor the appliances framework shows that the
and used decentralize system is efficient and
system to control the accurate.
loads.
[161] Hybrid SVM/GMM. Recognition rate of more
than 90%.
[36] The method used Over 90% recognition
quadratic programming efficiency is achieved and
followed by simulations the disaggregated power
and laboratory of the individual
experiments appliances is within 1%
error.
[162] The study used Stacked The results indicated that
Denoising Autoencoders the model could
to provide feedback for disaggregate the
the customers and the appliances correctly with
results of the analysis scores above 0.9.
were evaluated using
absolute mean loss. The
energy proportion is
assigned for the signal
disaggregation.
[39] Multilayer Feed forward Load identification of
Neural Network. 100% in lighting, 94.35%
for PC and 89.57% for
Rack.
 55

[3] Neural Network and The transient energy

Genetic programming. signature gives 100%
recognition of the loads.
[35] A prototype capable of The system formed a cost
providing a platform effective disaggregation
supporting 65KHz tool that can enable
sampling rate and 24bits energy management
amplitude resolution is system.
developed.
[163] Unsupervised load The error between
disaggregation algorithm estimated and consumed
employing Factorial energy is around 1%.
Hidden Markov Model
(FHMM).

2.14 Energy Savings Techniques for Minimizing Power Consumption.

Minimizing and optimizing energy consumption is one of the first steps of

reducing the greenhouse gas (GHG) emission, this is also in line with the appropriate
actions to cater for the continuous global increase in energy demand [164]. Recently
there are advancements in electronics, communication and sensor technologies, which
can help in devising automatic monitoring and control systems that can turn the
consumers into energy cautious society. Smart meter, smart plug, different types of
sensors and even smart appliances are at the discretion of the researchers and energy
management system manufacturers to use and come out with universal energy saving
systems. The smart metering system is one of the ways to make consumers closer to
the appliances by having feedback from the meter. The smart meter can send its
measurements through wireless communication which can help the consumers to
manage their energy consumption with their existing appliances [42]. Efforts are still
being made by utility companies and authorities concerned to upgrade the power
system to a smart system, even though some of the customers are still lacking
enlightenment about the smart system [165].
 56

A very important step in optimizing the energy consumption of a consumer is

to create awareness about energy consumed by individual appliances in the building.
Hence monitoring the building appliances will be the first step of energy management.
Also the enlightenment of the consumers about saving the energy for greener
environment will also have to take effect in appropriate measures to minimize energy
consumption [166, 167]. The home energy management researches conducted by
different researchers can also help a lot in containing the problems of power shortage.
The work developed by Ninad K. et al [13] can be modified to be used as a home
energy management system.

The use of highly efficient appliances is also a very good measure of energy
minimization, therefore replacing the appliances with the ones that can serve the same
purpose with less energy consumption is a very good way forward. Also the utility
company gives low tariff during off-peak period, therefore if the consumers can
change the time of use of their non-essential appliances, they will make a great saving
at the end of the month. The heating and air conditioning systems in homes and offices
are some of the high energy consuming appliances, therefore the consumers need to
monitor such appliances very well. Energy saving lights (like LED lamps) should
replace the high energy consuming lightings in the building. The computer loads in
the building should be shut down if applicable or they should be hibernated instead of
using screen saver. Other appliances should be switched off instead of standby mode
and the timing system for lighting loads should be considered to avoid unnecessary
lightings during the day [167].

Using natural lighting in a building can also save some amount in the bill, this
is especially during the day when a transparent window can be enough for lighting
some rooms [168]. Proper insulation of building walls is also a promising way of
saving energy, where less heating or cooling will be required by the insulated building
as the case may be. There are some low energy consuming applications like phone and
computer charging that can be done at very low cost using solar photovoltaic
applications, such renewable energy sources can supplement for the house and result
into cost saving. Recently there are motion detector and LDR lightings that can switch
off in the absence of motion and/or in the absence of another lightings for the purpose
 57

of energy management. Such facilities, which are also applicable in other appliances,
can save the energy greatly. The electrical heating is relatively more expansive than
some other ways of heating such as gas and solar thermal, therefore replacing electrical
heating with some other means can save a lot in the monthly billing.

2.15 AC Voltage Measurement Methods

The nature of the voltage supply in DC power system is characterized by

having a constant voltage and current, unless the supply voltage or the load is changed.
But when it comes to an AC supply, the voltage and current are sinusoidal in nature,
changing direction in every half of the cycles. Hence in an AC system the current and
voltage sensors can only read the value at a particular time. This therefore makes the
AC supply to be measured by some characteristic such as root mean square (rms)
value, peak value, peak to peak value and average value. The rms value is the most
common measure used to describe the magnitude of AC voltage and current. Figure
2.11 shows the waveform of an AC supply indicating the AC voltage, current, power
and total energy consumption changing with time.

Figure 2.11: AC Supply quantities changing with time

 58

As seen from the Figure 2.11, normally the current and voltage are out of phase
with voltage leading the current in most AC supply systems. The AC voltage
measurement methods used in this reseach are the instantaneous voltage and peak to
peak voltage methods. They are the voltage measurement methods that can best be
used with microcontroller [169]. Comparison of the voltage measurement by the two
methods will provide a better method for the voltage measurement. Appendix B
provides the table for the AC metrics calculation.

2.15.1 Peak Value Method

The peak voltage or peak current method is a method of calculating rms value
using the peak value of the voltage or current as the case may be. With microcontroller
system or any sampler, the voltage or current is sampled several times in a cycle and
the maximum sampled value is taken as the peak value.

2.15.2 Peak to Peak Value Method

The peak to peak value method uses the peak to peak value of the sampled
voltage (maximum and minimum values) to calculate the rms values. Normally the
difference between the maximum and minimum sampled values is taken as the peak
to peak value and equations in Appendix B are used in calculation of the rms values.

2.15.3 Instantaneous Calculation Method

The instantaneous calculation method uses the samples of the voltage and
current in an algorithm to determine the rms values as well as the average power, which
is the same as the real or active power. As the microcontroller samples the voltage and
current through the sensors the rms value is the square root of the averaged square
 59

samples as illustrated in Appendix B. One important thing about the instantaneous

calculation method is that the sampling rate of the voltage and current should be at
least twice the frequency of the supply as required by Nyquist sampling theorem. Also
the voltage and current must be sampled at the same time and at the same sampling
rate by the microprocessor in order to facilitate the real power calculation. The
instantaneous power calculation requires the product of voltage and current (instant
power) at each sampling interval and it can be calculated using maximum instant
power or average power as shown in the equations in Appendix B [80, 170].

2.16 The Regression Analysis

The regression analysis is a way of predictive modelling technique in which

the relationship between a dependent variable and independent variables is
investigated and estimated. With the data of the dependent variable and that of
corresponding independent variables a model of the relationship between the two can
be found which can predict the dependent variable in the presence of only the
independent variables. Most of the times the regression analysis can use Linear
Regression, Stepwise Regression, Logistic Regression, Polynomial Regression, Lasso
Regression, Ridge Regression or Elastic Net Regression. Linear Regression and
Polynomial Regression are considered in framing out the sensors input and output in
this research.

2.16.1 Linear Regression

Many natural phenomena can be predicted using straight line or linear

regression method. The linear regression determines the relationship between the
dependent and independents variable using best fit straight line. The analysis is termed
simple linear regression if there is only one independent variable and it is multiple
linear regression when there is multiple independent variable. The relationship
between the two variables in linear regression resulted into equation (2.1).
 60

Y  a b X e (2.1)

Where a is the intercept of the line on Y axis, b is the slope of the line and e is the
error term. To determine the values of a and b, the least square method is used, which
is obtained by minimizing the sum of the squares of the error between the line and the
real data [171].

2.16.2 Polynomial Regression

The polynomial regression uses polynomial of certain order (j) to realize the
fitting between the dependent variable and independent variable (s). In a typical
application the polynomial regression can be used to fit the relationship between the
dependent variable (y) and independent variable (x) with a polynomial of order j. The
nature of the data and its quantity influences the accuracy of the polynomial [171].
Accordingly the solution of the general equation (2.2) will give the coefficient of the
polynomial which is shown in equation (2.3).

  yi   n x i x 2
i x 3
i ...... x i
  a0 
j

    
  yi xi    xi x x x x
2 3 4 j 1
i i i ...... i   a1 
 :    xi2 : : : : :  : 
    (2.2)
 :    xi
3
: : : : :  : 
 :   : : : : : :  : 
    
  yi xi    xi  xij 1  xij 2  xij  j  a j 
j j
...... ......

y  a0  a1 x  a2 x 2  ................  a j x j
(2.3)

Where (xi , yi) are the data points, n is the number of data and ajs are the
coefficients of the polynomial of order j.
 61

2.17 Summary

The general concepts of load monitoring have been discussed in this chapter.
Different approaches of Intrusive Load Monitoring and Non-Intrusive Load
Monitoring have been discussed including the home energy management research
works. Voltage measurement methods and regression analysis procedures have also
been highlighted. The literature study helped in identifying the research gaps and the
research direction.

From the literature it is realized that most of the existing research works in
residential electrical energy management are intrusive in nature and the ones involving
the non-intrusive methods have very little applicability in home energy management.
Therefore the non-intrusive load monitoring in residential appliances management
with real-time energy saving capability has not been sighted in the literature. When
NILM for residential electrical energy management is developed, all the advantages
of NILM such as the consumer privacy, the reliability of the system due to only one
set of measuring sensor, the cost effectiveness, less complexity and less energy
consumption will be at our disposal.

On the other hand, the use of microcontroller to develop NILM system for
energy management, especially the Arduino microcontroller, has not been addressed
by the researchers. This is more especially when it comes to combining the residential
energy appliance monitoring, control and energy savings. The Arduino
microcontroller board is in expensive, capable of handling a number of intelligent
projects with extendable hardware and software that can work across many platforms.
Therefore the use of Arduino microcontroller in solving residential energy
management problem will surely provide a lasting solution.

Moreover, the research work uses ANN identification for the non-intrusive
load identification. The ANN has the capacity to detect all possible interaction and
relationship between the variable [3], hence ANN can also help in providing a lasting
solution to the NILM. The research work involving ANN that focuses on residential
energy management with energy saving capability has not been sighted in the
 62

literature. This is especially the researches that combines microcontroller and ANN
for NILM. The combination will surely be highly promising in solving the residential
energy mangement problem.

The issue of using the steady-state signatures in the appliances identification

has been used by other researchers, but the combination of real power, rms current and
power factor as the load signatures has not been utilized. The real power, rms current
and power factor are among the metrics that have distinct values for most of the home
appliances [90, 172].

The ILM or distributive monitoring system has higher energy consumption as

the result of too many sensing devices. Researchers in the area pointed out the
importance of minimizing the system energy consumption by reducing the number of
sensors or by adopting NILM in the appliances energy consumption management
instead of using ILM with higher energy consumption [63, 173]. Hence developing
the ANN-Arduino NILM system will quite address the issue.
 CHAPTER 3

3RESEARCH METHODOLOGY

3.1 Introduction

The task of developing real-time energy management system must involve

series of steps. This chapter provides the detailed steps to achieve the residential
electrical energy management through the ANN-Arduino NILM system. The load
identification uses the steady-state aggregated measurements of the appliances to
identify the connected loads for proper loads control and management. The chapter
consists of the load selection of the home appliances, sensors calibration and algorithm
development, appliances energy consumption measurement and control system
design, the design of the ANN-Arduino NILM identification system and finally the
development of non-intrusive load management and controlling system using the
ANN-Arduino.

The use of the steady-state quantities to identify the home appliances is a good
choice of the appliances features because the steady-state conditions of the loads is the
dominant operating condition of the appliances [31, 174]. The functionalities of the
ANN-Arduino NILM system involve the measurement of the aggregated consumption
of the appliances and the utilization of those aggregated quantities to identify the
connected appliances. Therefore the choice of Arduino microcontroller together with
the current and voltage sensors in the system will facilitate the accurate measurements
[175, 176]. The flowchart showing the overall methodology for the implementation of
the research work is illustrated in Figure 3.1.
 64

Start

The home appliances selection and profiling.

Application of polynomial regression method to

develop the sensors ADC algorithm

Measurement of Aggregated consumption of the home

appliances and Laboratory Experimental validation
using Fluke 115 meter

Design of Non-intrusive load identification using the

steady-state measurements of the appliances, Arduino
microcontroller and ANN

Development of Residential appliances energy control

and management system through the ANN-Arduino
NILM.

End

Figure 3.1: The flowchart for implementation of the research work

3.2 Load Selection and Profiling of the Home Appliances

The research is aimed at developing a non-intrusive energy management

system that can be applied in real situation to monitor and manage the home electrical
 65

devices for the purpose of efficient and economic energy consumption. Some typical
home appliances are sampled and profiled, in terms of power consumption, to serve as
the smart home appliances. The selection of the appliances followed the convenient
sampling method, which provide the researcher the chance to select samples that are
directly linked to the research and relatively easy to access among the ones
available.[177, 178]. The appliances selected are among the most common appliances
found in typical residential house in Malaysia [179] and they are among the available
home appliances in the laboratory. Moreover the steady-state behaviour of those
appliances is quite similar to most of the other home appliances. Television, music
player and dimmer lights are more or less continuously varying devices which are out
of scope of this research. Each of the appliances is supplied via a single relay from the
multichannel relay in the system that is controlled by the Arduino microcontroller. The
voltage and current sensors enable the microcontroller to sample the real-time voltage
and current for measurements of the load consumption. Also the 20 x 4 LCD display
is used for displaying the energy monitoring information. The home appliances and
their power rating profiles are shown in Table 3.1 and Figure 3.2 represent the block
diagram of the system. Matlab and Arduino Programming software are used as the
software tools for the research. Appendix C shows the picture of the set up.

The modelling of the appliances will enable the profiling of the loads to
investigate the aggregate consumption of different load combinations.

Table 3.1: The home appliances and their power consumption

Appliance Rated Power

Consumption (W)

Electric Bulb 100

Standing Fan 50

Standing Fan 60
 66

Electric Toaster 700

Rice Cooker 700

Electric Pressing Iron 1,300

Electric Kettle 1,700

Blender 350

Figure 3.2: The block diagram of the ANN-Arduino system

3.3 Sensors Calibration

The microcontroller program is responsible for measuring the real time energy
consumption of the appliances. It samples the voltage and current of the supply within
certain interval of time and use them to determine the energy consumption metrics of
the appliances like power consumption and rms values of the voltage and current. This
makes it necessary for the system to have voltage and current sensors connected to the
Arduino for proper sampling of the voltage and current. The voltage and current
 67

sensors are therefore calibrated so that an accurate algorithm will translate the ADC
value read by the microcontroller into the real analog value (voltage or current).
Therefore the current and voltage sensors are experimented to find the accurate transfer
characteristics between the input analog and digital values.

The current sensor adopted for this research work is ACS712 current sensor. It
is an integrated hall effect based linear current sensor that comes with factory
sensitivity which depends on the current rating of the sensor module. The sensor has
high temperature tolerance and it can withstand higher current than the rated current
[180]. The value of the analog input current to the sensor is related to the ADC output
value through Equation (3.1) [181].

I
I ADC  I ADCref   5
(3.1)
1023   

Where I is the analog value of the current, IADC is the analog to digital
conversion value of the current, IADCref is the reference ADC value of the sensor (ADC
value at 0 input current) and δ is the sensitivity of the sensor. The sensitivities of 30A,
20A and 10A sensor modules are 66mV/A, 100mV/A and 185mV/A respectively. The
theoretical value of the reference ADC value (IADCref) is 512 but sensor should be
experimented to find the experimental value which always differ from the theoretical
value. The sensor is therefore experimented by sampling the ADC at open circuit
condition and the ADC output is observed.

3.3.2 Voltage Sensor Calibration

The ZMPT101B voltage sensor module is the voltage sensor adopted for this
research. It is a voltage sensor board made from ZMPT101B potential transformer
which is capable of measuring up to 250V AC [182]. The sensor has to be trimmed
using the trimmer potentiometer and it has to be calibrated for accurate measurement.
However, the ZMPT101B sensor does not have a fixed sensitivity like the current
 68

sensor, therefore the accurate transfer characteristic can only be found using
appropriate calibration method.

In this case the peak to peak input voltage (y) to the sensor (measured using
the standard FLUKE 115 meter) is correlated with the peak to peak ADC output (x) of
the sensor using 1st to 5th order polynomial regression, in order to determine the best
fitting relationship between them. The peak to peak ADC value is determine by
sampling the supply voltage in the Arduino program and the best polynomial is
ascertained using absolute error analysis. Figure 3.3 shows an illustration of the analog
input and the digital output after sampling. The advantages of calibrating the
ZMPT101B voltage sensor module using polynomial regression includes finding a
simple, accurate and straight forward relationship between the analog input and the
ADC output. In addition the use of ZMPT101 voltage sensor does not require an
external driver circuit to be used in the microcontroller, hence it reduces the
complexity of the system. After the derivation of the polynomials, the peak to peak
and instantaneous methods of voltage calculations are used to test and bring out the
best algorithm.

Figure 3.3: Analog and Digital Voltages

3.3.2.1 Peak to Peak Voltage Measurement Method

To determine the rms value of the voltage using peak to peak method there is
need for the maximum and minimum values of the sampled voltage, at sampling rate
 69

that satisfy the Nyquist criteria. The Nyquist Shannon criteria theoretically maintained
that the sampling rate should be more than twice the maximum frequency of the signal.
However in practical application the sampling frequency should be more than 10 times
the maximum frequency of the signal [169]. To that effect the voltage is sampled 1000
times at an interval of 1ms to make 20 sampling points per period which satisfy the
requirement. The maximum and minimum are observed by the microcontroller
program. The rms voltage is therefore expressed in Equation 3.2.

Vmax  Vmin
Vrms  (3.2)
2 2
Where Vmax and Vmin are the maximum and minimum voltages respectively.

3.3.2.2 Instantaneous Voltage Measurement Method

The use of instantaneous method requires the whole samples in order to

calculate the square average. The Arduino also takes 1000 samples at time interval of
1ms for accuracy. The samples are averaged to get the rms value using Equation (3.3).
The instantaneous calculation method is also useful in finding the real power
consumption when the voltage and current are sampled together.

1 n 2
Vrms  Vi
n 1
(3.3)

Where Vi is the ith sample voltage and n is the total number of samples.

3.3.2.3 Data Collection and Regression Procedure

The FLUKE 115 meter is used to measure the input voltage which is varied
from 0 to 250V AC using the variable transformer. The ADC output of the sensor at 0
input voltage is 512, therefore 0 to 512 will represent the negative part of the AC
voltage signal and 512 to 1024 will represent the positive part of the signal. Hence if
 70

512 to 1024 is allowed to scale 0 to 1000V range (the maximum theoretical value of
the voltage sensor [182]), then 512 to 640 will be suitable for scaling 250V range. In
this regard with 250V connected to the sensor, the maximum ADC is adjusted to be
640 using the sensor trimmer (this serves as the reference for any measurement to be
done using the results of this work). For 120V system 512 to 574 can scale the voltage.
A variable voltage from variable transformer (measured using FLUKE 115 meter) is
applied to the sensor and through the Arduino program maximum and minimum values
of the ADC are recorded against each input voltage. The peak to peak input voltage
and peak to peak ADC output are tabulated for the analysis. The Arduino code sampled
the voltage 1000 times, at an interval of 1 millisecond and obtain the minimum and
maximum ADC values from the samples. Appendix D shows the experimental set up.

For the regression up to the 5th order, the solution to equation (3.4) and its
lower orders is needed for the regression analysis which will results into polynomials
from order 1 to 5. The maximum of 5th order is selected to limit the sensor calibration
analysis to 5th order polynomial in order to avoid higher computational requirement in
the sensor algorithm [183].

  yi   5 xi x
2
i x
3
i x 4
i  x  a 
i
5
0
  
  yi xi    xi x
2
i x
3
i x
4
i x 5
i  x   a 
i
6
1
 yi xi2   xi2 x
3
x
4
x
5
x 6
 x  a 
7

 3
 3 i i i i i 2

  yi xi    xi x x x x  x  a  (3.4)
4 5 6 7 8
i i i i i 3
 y x4   x4
 i i5   i5 x
5
i x
6
i x
7
i x 8
i  x  a 
i
9
4

  yi xi    xi x x x x  x  a 
6 7 8 9 10
i i i i i 5

Where n is the order of the polynomial, xi are the samples of the ADC voltage,
yi are the samples of the analog voltage and ai are the coefficients of the polynomial.
 71

3.3.2.4 The Programing Code for the Voltage Sensor Calibration

After forming the polynomials using the regression analysis in section 3.3.2.3,
a programming code is written to calculate the rms value of the input voltage using
each of the 1st to 5th order polynomial as the transfer function between the analog
input to the sensor and the ADC value. The accuracy of each polynomial is ascertained
by observing the absolute errors between the real analog value and the calculated
value, which is done using instantaneous and peak to peak method. The Arduino is
programmed provides the code for instantaneous voltage method and peak to peak
voltage method which works in line with the flowchart in Figure 3.4.

3.4 Residential Appliances consumption Measurement and control

The advancement in computer electronics, communication and sensor

technologies, provides a good opportunity in designing systems for automatic
appliances’ energy consumption management. This section described the steps taken
in the development of a system that can measure the energy consumption of home
appliances and use the measurement to control the loads. The system uses the
instantaneous power calculation method to calculate the total power, voltage, current
and power factor consumption of the house using Arduino microcontroller and other
accessories. It displays the results and through the actions of the relays that supply the
house, the loads will be reduced based on the priority to ensure that the total kW
consumption of the house is not more than the threshold value. The selected home
appliances like fans, electric kettle, rice cooker and electric iron are used in the system.
The system realized can be used as home automation system and it has the advantage
of working in real-time.
 72

Start

Select the voltage

calculation Method (Peak
to Peak or Instantaneous
method)

No Yes
Inst. Method ?

Initiate max. and min value Initiate the counter to 1 and

of the Samples. Also the Sum of square voltage
initiate counter to 1 to 0.

Sample the voltage to Sample the voltage to

obtain the ADC value. obtain the ADC value.

Use the ADC value in the

Compare the sampled value
selected polynomial to
with the previous Max. and
compute the analog value.
Min. and replace
accordingly.

Square the voltage and add

to previous sum.
No Increment N
Is N =1,000?
by 1

Yes
Increment N No
Is N =1,000?
by 1
Obtain the rms value of the
voltage using the maximum
value, minimum value and
Yes
the selected polynomial.

Obtain the rms value of the

voltage using the sum.
Display the rms value

Display the rms value

End

Figure 3.4: Flow Chart of the rms Voltage Calculation

 73

3.4.1 Appliances Energy Consumption Measurement and Control System

Design

The system is designed to monitor the energy consumption of the appliance

and notify the occupants of the residence through the display system. The system, also
automatically control the energy consumption of the appliances by reducing the load
in excess of 2kW, which serves as the threshold value. It also encompasses automatic
reconnection of the loads for automation purpose. Figure 3.5 illustrate the block
diagram of the system and the following explain the step by step of the system
operation.

(i) The Arduino microcontroller will be sampling the ADC values of the voltage
and current through the voltage and current sensors. This follows the
initialization of some parameters (reset time, V2, I2, VI and N) for the subroutine
of the system program.
(ii) The system will be using the ADC samples to obtain the corresponding values
of the voltage and current. These values will be used in the program subroutine
to calculate the rms voltage, rms current and the real power using instantaneous
calculation method. The apparent power, reactive power and power factor are
also calculated and all the calculated parameters are displayed on the LCD for
monitoring purpose.
(iii) The system will check if all the relays are off. All the relays are off means, the
highest priority relay (L8 in Figure 3.6) is connected to a load more than the
threshold value. This will be followed by a warning on the display and 3 minute
alarm (unless the microcontroller is manually reset) to indicate that the system
that is meant to operate within 2kW, now has a single load of more than 2kW
connected to the highest priority outlet. The system will off the alarm, clear the
display and reset all the relays after 3 minute. By manually resetting the Arduino
microcontroller (through pressing a button) the system will be reset and start
from the beginning.
(iv) The system will also check the real-time aggregated power consumption of the
appliances, in which if it is greater than 2kW it will reduce the loads based on
priority (from L1 to L8) until the total load is within the threshold value. A short
 74

alarm will be sound to indicate that an over load has occurred and a load has been
reduced.
(v) Lastly, the system programming will always be automatically resetting all the
relays after every three minute for automation purpose. The program is based on
the consideration that the overload will be removed within that 3 minutes and
therefore this will avoid manual resetting always a load reduction occurred. In
any case, the time is adjustable in the program.

Figure 3.5: The Proposed Block Diagram of the Energy Consumption Measurement
and Control System

3.4.2 System Development

The components used in the development of the system include, the Arduino
MEGA 2560 development board, eight channel relay, the ACS716 (30A) current
 75

sensor, ZMPT101B AC voltage sensor module, 20x4 LCD display, buzzer alarm,
power supply for the Arduino, bread board and socket outlet for connecting the loads.
Figure 3.6 shows the complete flowchart that make up the system. The result for the
sensor calibration in section 3.5 is used the ZMPT101B voltage sensor module.

3.4.3 Energy Measurement and Control Programming Code

An Arduino program run the system which includes the various measurements
through the sensor reads, alarming, LCD display and controls of the loads. By
experimental observations, the current sensor gives an output of 528 ADC value at
zero input current whereas the voltage sensor gives 511 ADC at zero input voltage.
These values are subtracted as appropriate in the programming code in order to get the
instant values of the voltage and current.
 76

Start

Initiate Reset time to

current time

Initiate V2, I2, VI

and N to 0

Get the ADC values of the current and voltage

from the sensors; x and y respectively

f(x) = 0.00000412x3 – 0.000857x2 + 2.675x – 3.198

g(y) = 5*y/(1023*0.066)

V2 = f(x)*f(x) + V2

I2 = g(y)*g(y) + I2

VI = f(x)*g(y) + VI

Increment Yes is N<1000 No

N by 1 ?

Display P, Q, S, Vrms, Irms and pf

Reset all the relays and Sound alarm for 3 min and Yes Are all the relays
clear the display display warning off
?
No

Yes
Sound a short alarm and reduce is P > Threshold
the load based on priority ?

No

Yes No
Reset all the is the time up to 3 min from the
relays reset time ?

Figure 3.6: Flow chart of the Energy Consumption Measurement and Control System

3.5 Design of the Non-Intrusive Appliance Load Identification System

 77

The system of identifying connected loads in a utility using the measurement

at the service entry (NILM) will help greatly in promoting the smart electricity system.
This section describes the design and development of non-intrusive appliance
identification using ANN. The work involves the measurement of the aggregated real
power, rms current and power factor of the connected loads using Arduino micro
controller, voltage and current sensors. The ZMPT101B voltage sensor and ACS712
current sensor are used. The feedforward ANN use the aggregated quantities (real
power, rms current and power factor) as the input variables and the load connection
pattern as the output variables. The feedforward ANN is selected to be the identifier
because it has the capacity of yielding highest accuracy for train cases [184, 185]. The
function of the ANN model is to identify the connected loads when it receives the
aggregated quantities as the input. Figure 3.7 shows the configuration of the ANN
connected to the Arduino in the system. The appliances used for the ANN-NILM
system are 120W light bulb, an energy saving bulb, a table fan and rice cooker.
Because of the home appliances are numerous in number, it is impractical to device a
NILM identification system for all the existing home appliances, but the NILM system
can be put into practice with smaller NILMs with each NILM monitoring some few
appliances forming a sub-metering system [63, 72] .

3.5.1 Energy Consumption Data Collection for the NILM System

Instantaneous calculation method is used in calculating the total power

consumption of the system from the samples of voltage and current collected from the
sensors. This instantaneous calculation makes the total power and other quantities to
have a little variations over the cycles of the operations of the microcontroller. Ideally
the microcontroller supposed to produce the same measurement for each load
combination in each circle of operation, but the variation of the measured quantities,
as the results of the instantaneous computation is used as the data for the analysis. With
the four appliances experimented, there are 16 possibilities of load combination, from
“0000” representing all the appliances are “OFF” to “1111” representing all appliances
are “ON”. Now regarding the variation of the measurements, within the same load
combination (e.g. 1010), a total of eight different measurement is taken, this is in
 78

conformity with the instantaneous measurement that can give up to eight distinct
measurements for the same loads [186]. Therefore in each of the 16 possible load
combinations the eight different measurements are taken in which 6 are assigned for
training and 2 are assigned for testing. This makes a total of 96 training data and 32
testing data. Table 3.2 is a sample of the data for a combination of load (i.e. “1010”)
and when the data of the table is combined for the whole 16 load combinations it will
make the 96 training data. The quantity of the data is enough for training the identifier
with three steady-state signatures and four loads [15, 55].

Figure 3.7: The Block Diagram of the ANN-Arduino NILM Configuration

 79

3.5.2 The ANN Training for the NILM System

The feedforward ANN, which has been selected for its best performance [184]
is trained using the training data generated. The load pattern is the output of the
network whereas the Power, rms Current and Power factor are the inputs to the
network. Hence there are three input neurons and four output neurons. Only one hidden
layer is used for this work to enable easier implementation of the network model in the
Arduino program. Out of the 128 set of data 96 data (6 from each load combination)
is used in training the network and the remaining 32 is used in simulating the network
to observe how it fit the input to the output. Table 3.3 shows the distribution of the
data. In each training and retraining, which involved changing the number of hidden
neurons and activation functions, the performance of the network is observed. Apart
from the identification accuracy the feedforward ANN identifier is also expected to
have least number of neuron in the hidden layer to avoid complexity in the Arduino
programming. The more the number of hidden neuron, the more the computation
requirement.

Table 3.2: The Sample of the Data for one Combination of loads

Rice Energy Table Electric Power, P Current, Power

Cooker Saving Fan Bulb (W) Irms (A) Factor,
Lamp pf
559 2.75 0.87
561 2.78 0.87
ON OFF ON OFF 562 2.72 0.89
562 2.75 0.88
564 2.70 0.90
565 2.78 0.88
 80

Table 3.3 The Data Distribution for the ANN-NILM system

Input Output Generated Training Testing

Variables Variables Data Data Data

The input The 7 data Training Testing

variables are output variables per Data is 96 x 7 Data is 32 x7
the variables are data sample
aggregated the load Making a total
real power P, connection of (16 x 8) by 7
rms current I status of the 4 = 128 x 7 data
and power loads 0 for (each row data
factor pf (3 OFF and 1for 3 inputs and 4
inputs) ON (4 outputs).
outputs)

3.5.3 The Arduino Programming for the ANN-NILM System

The program will first capture the aggregated measurements of the Power,
Current, and Power factor using the samples from the voltage and current sensors.
These aggregated values are then passed into the network model to determine the
connected loads which will be displayed on the LCD. The flowchart of the operation
of the ANN-Arduino NILM system is illustrated in Figure 3.8. Whereas the Pseudo
code describing the steps is shown in Table 3.4.
 81

Start

Get the ADC values from the Voltage and

Current sensors

Use the sensor equations to obtain the

corresponding analog values.

Use the analog samples to obtain the

aggregated P, Irms and pf.

Use the aggregated values in the ANN to

identify the connected loads.

Display the status of the Loads.

End

Figure 3.8: Flow chart of the ANN-Arduino NILM system

 82

Table 3.4: Pseudo code for the ANN-Arduino NILM

1: Begin
2: Initialize the Sum of Instant power and Sum of Square current to 0.
3: For n from 1 to 1,000
4: Sample the ADC values of the voltage and current from the sensors
5: Use the sensor equations to calculate the analog values from the ADC samples
6: Calculate the Instant power and Square of current from the analog values
7: Add to the previous Sum of Instant power and Sum of Square current respectively
8: End
9: Calculate P, Irms and pf from the Sums
10: Use the P, Irms and pf in the ANN to identify the connected loads
11: Display the status of the connected loads
12: End

3.6 Development of ANN-Arduino NILM based Residential Appliances

Management and Control System

The ANN-Arduino NILM based measurement and control system is a modified

system of the ANN-Arduino NILM, which is modified to include the appliances
management and control for home automation. The work is conducted to evaluate the
performance of the ANN-Arduino NILM in electrical energy management of the
residential appliances.

With a monitoring system that can accurately recognize the connected loads
using measurements at the service entry, it is possible to develop appliances’ energy
management system that can keep track of the connected loads, control and manage
them. This section provides the detail design of the ANN-Arduino NILM for the
appliances control and management. The appliances control and management involve
both the monitoring and controlling of the appliances so that they will work for the
specific time they are designed to work within a specific period (say 24 hours). The
advantage of the system is that it allows the appliances to use their time of work at
 83

their discretion within the 24 hours. Therefore the appliances can use their time of
work at once or they can divide it as long as the cumulative time does not exceed the
appliances’ time of work. The automatic appliances control can play important role in
dynamic electric power scheduling [187].

The work involves the measurement of the aggregated real power, rms current
and power factor of the connected appliances using Arduino microcontroller, current
and voltage sensors. The aggregated measurements are then passed on to a feedforward
ANN model embedded into the Arduino code for identifying the connected loads. The
output of the ANN is 1 if the corresponding load is ON and is 0 if the load is OFF,
hence the ANN will always give the load pattern of the four loads whenever it receives
the aggregated quantities. The NILM real-time system can best be achieved with sub-
metering system in which each NILM will monitor some number of appliances [2].
The NILM has demonstrated 100% identification in [55] and the 100% load
recognition is achieved with ANN-Arduino NILM in this work. Initially data of the
four house appliances used in the analysis is generated using the Arduino micro
controller and sensors set. The appliances are 300W Filament Lamp, 100W Electric
Bulb, 50W Standing Fan (on the high speed mode) and 24W Energy Savings Lamp,
whereas the data generated is used in training and simulating the ANN. Moreover,
when the connected appliances are identified by the ANN-Arduino NILM, the Arduino
program will use the results of the appliances identification to determine their
cumulative time of work. This cumulative time will be used by the system to
disconnect the loads that exhausted their time limit until next operating cycle (probably
the next day). Figure 3.9 gives the block diagram of the system configuration.

The advantages of the ANN-Arduino NILM based appliances management system

cannot be overemphasized. With the control and management among the house
appliances, the power supply system will not be over used. Each appliance will work
according to the scheduled time for it, which will ensure reliability and efficiency of
the system. A typical example is the UPS or solar supply system, which as the result
of storage system will work perfectly if the appliances did not work beyond their
designated times [188-194].
 84

3.6.1 Data Collection for the Appliances Management and Contol System

Instantaneous calculation method is used in calculating the aggregated power

consumption of the system from the voltage and current samples from the sensors. The
power calculated using the instantaneous method have little variations over the cycles
of operations of the microcontroller, this is because of the averaging of the sampled
power. There are 16 possible load connection pattern or combination with the four
loads.in each load combination (e.g. “0101”) 8 different record of the aggregated data
is taken, of which 6 data are used in training the ANN and the remaining 2 is used in
testing the ANN (after the network is formed). The data collection is done at an interval
of 1 minutes after the system setup and after every 6 minutes the load connection
pattern is advanced until the training data is taken for all the load combinations. This
makes the training data to be 96. Also the testing data is taken at an interval of 1 minute
makes the 32 simulating data. Table 3.5 shows the training data for one of the loads
combinations i.e. “0101” and the data breakdown is shown in Table 3.6.
 85

Figure 3.9: Block diagram of ANN-Arduino appliances management system

Table 3.5: Sample of the Data for a combination of Loads

Filament Electric Table Energy Power, P Current, Power

Lamp Bulb Fan Saving (W) Irms (A) Factor,
Lamp pf
127 0.67 0.84
125 0.68 0.81
OFF ON OFF ON 122 0.68 0.78
121 0.69 0.83
126 0.68 0.83
120 0.65 0.80
 86

Table 3.6: Data breakdown for the ANN-Arduino appliances management

Data Type Sampling Samples Taken Total Load Total Data

Interval per Load combinations Taken
(minutes) combination (“0000” to
“1111”)
Training 1 6 16 96
Data
Testing Data 1 2 16 32

3.6.2 ANN Training for the Appliances Management and Control System

The ANN training is similar to the previous case of the NILM system. The data
generated is used in training the feedforward ANN for loads recognition. The power,
rms current and power factor of the aggregated loads are the inputs of the ANN and
the load pattern is the output. Since the ANN model is going to be implemented in the
Arduino program only one hidden layer is considered to avoid complex computations
in the Arduino coding [195]. The 96x7 training data is used for training the ANN and
the network obtained is simulated using the 32x7 simulating data, to see the
performance of the network outside the training data. In each training and retraining
which involves changing the number of neurons and changing the activation functions,
the network performance is observed for the best monitoring results. The ANN training
and retraining is illustrated by the flowchart in Figure 3.10.

3.6.3 Arduino Programming for the Appliances Management and Contol

System

The program will first calculate the aggregated power, current and power factor using
the current and voltage samples from the sensors (ZMPT101B voltage sensor module
and ACS712). The aggregated values will pass through the ANN model to determine
the connected loads and display it on the LCD. At last the program will use the
 87

monitoring results and determine the cumulative working time of each load and
disconnect the loads exhausting their time as appropriate. Moreover, the pseudo code
for the appliances management and control system is a modification of the pseudo
code for ANN-Arduino NILM code as shown in Table 3.7.

Start

Input the Training Set

Adjust the weight and biases to

minimize the error to global minimum
Retrain the ANN

Input the Validation set

Change Other
Training Parameters
Is MSE between the output of the ANN (No. of Hidden
and target of Validation set acceptable ? Layers, Activation
Function, Number of
Hidden Neurons and/
or Learning Rate)

Test the ANN with the Testing Data

Is the ANN performance OK

with the Testing Data?

Accept the ANN to be used for the

Non-Intrusive Load Identification

End

Figure 3.10: The ANN-Arduino NILM Network Training Flowchart

 88

Table 3.7: Pseudo code for the ANN-Arduino NILM appliances management system

1: Begin
2: Initialize the Sum of Instant power and Sum of Square current to 0.
3: For n from 1 to 1,000
4: Sample the ADC values of the voltage and current from the sensors
5: Use the sensor equations to calculate the analog values from the ADC samples
6: Calculate the Instant power and Square of current from the analog values
7: Add to the previous Sum of Instant power and Sum of Square current respectively
8: End
9: Calculate P, Irms and pf from the Sums
10: Use the P, Irms and pf in the ANN to identify the connected loads
11: Display the status of the connected loads
12: Determine the Cumulative operation time for each appliance using the load status
13: Display the cumulative ON time for each appliance
14: Disconnect the loads that have exhausted their time
15: End

3.7 Summary

This chapter has presented the methodology for conducting this research work.
The chapter started with the calibration of the sensors for accurate measurement. The
ZMPT101B voltage sensor calibration using polynomial regression is a novel and first
work of its kind in the voltage and current sensor area. There is need for more accurate
sensing in load monitoring and management. The appliances energy consumption
management presented in the chapter also provide a novel method of home automation
for energy savings. The ANN-Arduino NILM and its design has also been presented
in the chapter which provides a major contribution in the non-intrusive load
identification. Finally the load management and control of the residential appliances
is realized using the ANN-Arduino NILM system which gives an advanced method of
the appliances energy consumption automation and it is a novel. The appliances
choices has been done based on convenient sampling as described in the chapter.
 89

CHAPTER 4

4RESULTS AND DISCUSSIONS

4.1 Introduction

This chapter provides the general results and discussions of the research work. The chapter
started presenting the presenting the results and discussion of the sensors calibration for the
purpose of getting accurate measurement from the microcontroller. This is followed by the results
of the developed energy consumption measurement and control using Arduino microcontroller.
The energy consumption measurement and control system has shown the readyness of the Arduino
microcontroller to develop the ANN-NILM with energy savings capability. The results and
analysis of the developed ANN-Arduino non-intrusive load identification is also presented, which
resulted from the combination of ANN and Arduino microcontroller. Finally the results and
analysis of the ANN-Arduino NILM based appliances management and control gives the
modification of the ANN-NILM which also evaluate the system in residential energy management.

4.2 The Sensors Calibration

The voltage and current sensors are the facilitators of all the measurements in the system.
The microcontroller uses the real-time current and voltage samples to calculate all the electrical
quantities in the system. A transfer characteristic model that relate the ADC output of the sensor
to the analog input is found for both the current and voltage sensors.
 90

4.2.1 The Current sensor Results

The ACS712 experimentation gives the average reference ADC value (IADCref) of 519. This
will be the amount to be subtracted from the ADC value read (IADC) to provide the instant sampled
value. The 30A module is used throughout the research works, therefore the transfer characteristics
of the current sensor which is found through the experimentation is illustrated in Table 4.1. The
resulting equation is used throughout the research work in representing the sensor current.

Table 4.1 The Current Sensor Parameters

General Equation Sensitivity Reference Resulted Equation

Value (δ) ADC
(IADC-ref )
Value

I
I ADC  I ADCref   5 0.066 519
I
 I ADC  519   5
1023    1023  0.066 

4.2.2 Calibration of the Voltage Sensor

This section discusses the results of the calibration of ZMPT101B voltage sensor module
using polynomial regression for the purpose of accurate measurement and load monitoring. When
the voltage measurement is accurate the load monitoring will also be accurate. The load monitoring
of this research work also involve the measurement of the appliances consumption, therefore it is
essential to have more accurate measurement through calibration. The calibration of the
ZMPT101B sensor has resulted into a more accurate representation of the sensor as shown in the
subsequent sub-sections of this chapter. ZMPT101B is a linear sensor module but polynomial
regression analysis had been carried out on it to check for better modelling.
 91

In this work, the polynomial regression analysis of 1st to 5th order is carried out on the
sensor, which resulted into 5 model equations. The accuracy of the resulting polynomials is tested
in measuring the rms values of the supply voltage using instantaneous and peak voltage methods
of calculation. It has been shown that the 3rd order polynomial provides the best relationship in
terms of error in the measurement in both the instantaneous and peak voltage methods, when
compared with the measurement of standard FLUKE 115 meter.

4.2.1 Resulted Polynomial Equations

The steps taken in the calibration of the ZMPT101B sensor is described in section 3.3.2.
The work intended to investigate the performance of the polynomials from order 1 to 5, therefore
solutions of the polynomial regression of up to 5th order is required. Hence the 5th order equation
and the lower orders are needed in order to perform the analysis. The polynomial regression of
the sensor data using Equation (3.4) in chapter 3 resulted into the polynomial equations in Table
4.2 which represent the first order to fifth order polynomials respectively.

Table 4.2: Equations from the Polynomial Regression Analysis

Order of The Polynomial Equation Equation

the Number
Polynomial
1 y  2.709 x  8.35 (4.1)

2 y  0.0007711x 2  2.506 x  0.2662 (4.2)

3 y  0.00000412 x3  0.000857 x 2  2.675 x  3.198 (4.3)

4 y  0.00000004888x 4  0.00002986 x3  0.005183x 2  2.922 x  6.085 (4.4)

5 y  0.0000000001278x5  0.00000003529x 4  0.00001023x3  0.003271x 2  2.853x  5.57 (4.5)

 92

4.2.2 Voltage Measurements Analysis

Each of the equations found in the analysis gives the transfer characteristic between the
ADC output of the sensor and actual analog input voltage to the sensor. Generally and
approximately all the five polynomials supported the linearity of the sensor as can be seen in Table
4.2. The table indicated that there are small coefficients for the higher order. To ascertain which
of the five polynomial equation has highest accuracy, the equations are run on the arduino program
to calculate the rms values of different input voltage using instantaneous calculation and peak to
peak voltage methods. The results are illustrated in the sections that follows

4.2.2.1 The Instantaneous Voltage Method Analysis

The instantaneous calculation method uses the appropriate voltage samples to estimate the
root mean square value using the Arduino program. This number of samples should at least satisfy
the Nyquist chariteria.The input voltage which is varied through an autotransformer is calculated
in the arduino using the programming code and is validated with the measurement of the FLUKE
115 meter for the analysis. The response of each polynomial equation including the errors in the
measurement are shown in Table 4.3, which indicated that there are higher errors when the
computations are done using equations (4.2), (4.4) and (4.5). Generally the result shows that the
third order polynomial, equation (4.3) is the best because the percentage errors are the lowest,
followed by the first order equation (4.1).
 93

Table 4.3: Arduino Voltage Measurement Using Instantaneous Method

Input
voltage
Voltage measurement using the instantaneous calculation method.
FLUKE
Meter
(V) Equation (4.1) Equation (4.2) Equation (4.3) Equation (4.4) Equation (4.5)
Voltage Error Voltage Error Voltage Error Voltage Error Voltage Error
(V) (%) (V) (%) (V) (%) (V) (%) (V) (%)
250 246.48 1.41 229.15 8.340 248.5 0.60 305.39 22.16 269.55 7.82
248 247.17 0.33 228.54 7.847 247.67 0.13 301.14 21.43 268.96 8.45
246 244.92 0.44 226.14 8.073 245.91 0.04 301.46 22.54 226.17 8.06
244 244.33 0.14 234.17 4.029 242.11 0.77 297.93 22.10 264.68 8.48
242 241.76 0.10 222.46 8.074 242.11 0.05 297.93 23.11 264.68 9.37
240 240.53 0.22 220.61 8.079 240.31 0.13 295.03 22.93 263.53 9.80
238 237.38 0.26 217.48 8.622 238.42 0.18 292.03 22.70 260.74 9.55
236 236.80 0.34 216.69 8.182 236.61 0.26 289.06 22.48 257.69 9.19
234 235.26 0.54 215.44 7.932 234.76 0.32 287.44 22.84 255.41 9.15
232 231.56 0.19 214.47 7.556 232.65 0.28 284.41 22.59 245.48 5.81
230 230.97 0.42 213.74 7.070 230.62 0.27 282.62 22.88 250.80 9.04
228 229.34 0.59 212.44 6.825 228.68 0.30 278.48 22.14 247.93 8.74
226 227.25 0.55 211.60 6.372 226.58 0.26 276.78 22.47 246.40 9.03
224 225.78 0.79 209.38 6.527 224.95 0.42 274.27 22.44 244.95 9.35
222 222.91 0.41 207.76 6.414 222.79 0.36 270.44 21.82 243.60 9.73
220 222.19 1.00 205.98 6.373 221.42 0.65 268.41 22.00 242.51 10.23
218 220.39 1.10 202.82 6.963 218.66 0.30 265.55 21.81 239.42 9.83
216 217.29 0.60 199.68 7.556 217.12 0.52 262.33 21.45 235.33 8.95
214 216.22 1.04 197.53 7.696 215.83 0.86 260.25 21.61 233.99 9.34
212 213.99 0.94 196.24 7.434 213.76 0.83 257.19 21.32 231.73 9.31
210 212.10 1.00 194.42 7.419 211.84 0.88 255.47 21.65 230.07 9.56
208 210.40 1.15 192.78 7.317 210.33 1.12 252.35 -21.32 228.32 9.77
 94

206 208.71 1.32 191.31 7.131 208.51 1.22 249.72 21.22 226.76 10.08
204 206.42 1.19 190.42 6.657 206.38 1.17 247.53 21.34 224.89 10.24
202 205.18 1.57 188.42 6.723 204.47 1.22 244.42 21.00 222.58 10.19
200 204.04 2.02 186.25 6.875 203.84 1.92 241.26 20.63 220.53 10.27

4.2.2.2 The Peak to peak Voltage Method Analysis

The peak to peak voltage method samples the voltage appropriately and take the difference
between the maximum and minimum samples and perform the rms calculations. The
autotransformer and FLUKE 115 meter facilitate the variation of the input voltage and voltage
measurement in the system. The response of each polynomial is taken and the percentage errors
are tabulated in Table 4.4. This also shows the superiority of the third order polynomial because it
has the lowest error when compared with the FLUKE meter measurement.

4.2.2.3 Comparison between the Instantaneous Method and Peak to Peak Voltage Method
and General Analysis

The instantaneous and peak to peak methods are tested on all the polynomials in order to
ascertain the best in terms of minimum error. Consequently, the analysis on both methods shows
that the third order is the most accurate among them. Moreover, the performance of the selected
polynomial (3rd order or equation (4.3)) using the instantaneous and peak to peak method is tested
for the whole voltage range of 0 to 250V. The result is shown in Table 4.5 which indicated that
the peak to peak method is more accurate than the instantaneous voltage method with less absolute
percentage errors.
 95

Table 4.4: Arduinno Voltage Measurement Using Peak to peak Method

Input
voltage
FLUKE
Voltage measurement using the peak to peak voltage method
Meter
(V) Equation (4.1) Equation (4.2) Equation (4.3) Equation (4.4) Equation (4.5)
Voltage Error Voltage Error Voltag Error Voltage Error Voltage Error
(V) (%) (V) (%) (V) (%) (V) (%) (V) (%)
250 246.07 1.57 248.88 0.45 249.89 0.04 249.15 0.34 249.01 0.40
245 242.24 1.13 233.75 4.59 244.49 0.21 244.11 0.36 244.06 0.38
240 237.54 1.03 239.66 0.14 240.2 0.08 240.05 0.02 240.05 0.02
235 233.62 0.59 235.57 0.24 234.87 0.06 234.97 0.01 236.02 0.43
230 227.87 0.93 231.49 0.65 230.64 0.28 230.89 0.39 229.94 0.03
225 225.00 0.00 226.41 0.63 225.39 0.17 225.78 0.35 226.88 0.84
220 220.21 0.10 221.33 0.60 221.21 0.55 221.68 0.76 221.77 0.80
215 215.42 0.20 217.29 1.07 216.02 0.47 216.56 0.73 217.67 1.24
210 210.63 0.30 211.23 0.59 210.86 0.41 211.44 0.69 211.50 0.71
205 205.84 0.41 206.20 0.59 204.71 0.14 206.32 0.64 206.37 0.67
200 201.05 0.53 201.18 0.59 199.63 0.19 202.21 1.11 202.26 1.13

As shown by the results of Table 4.3 and 4.4,both the instantaneous and peak voltage
methods indicated that Equation (4.3) (the third order polynomial) is the best among the five
equations, based on the analysis of the error and in terms of representing the sensor response as
far the determination of the rms value of the voltage. The peak voltage method according to the
tables has less absolute error compared to the instantaneous method, even though it gives higher
error when there is spike problem in the system, which occurs very rarely. The instantaneous
voltage method has higher error, though it is needed in calculating power consumption of the
supply, which employs the instantaneous power measurement method [186].
 96

Table 4.5: Comparison of the Peak to peak and Instantaneous Methods

FLUKE Meter Peak to peak Standard
Input Voltage Instantaneous Error Voltage method Deviation
(V) Calculation (V) (%) (V) Error (%) SD (V)
0.08
250 250.05 0.02 249.89 0.04
0.13
240 240.24 0.10 240.2 0.08
0.48
230 230.93 0.40 230.64 0.28
0.93
220 221.82 0.83 221.21 0.55
1.19
210 212.36 1.12 210.86 0.41
1.32
200 202.54 1.27 200.65 0.33
1.28
190 192.44 1.28 190.56 0.29
1.63
180 183.26 1.81 181.58 0.88
1.38
170 172.67 1.57 170.72 0.42
1.59
160 163.1 1.94 160.96 0.60
1.92
150 153.48 2.32 150.32 0.21
1.77
140 143.36 2.40 140.74 0.53
1.42
130 132.83 2.18 131.23 0.95
1.39
120 122.71 2.26 120.84 0.70
1.12
110 112.14 1.95 110.51 0.46
1.01
100 101.86 1.86 100.24 0.24
0.85
90 91.48 1.64 90.01 0.01
0.53
80 80.81 1.01 79.82 0.23
0.36
70 70.36 0.51 69.64 0.51
0.60
60 60.65 1.08 59.46 0.90
0.51
50 50.27 0.54 49.29 1.42
0.18
40 40.32 0.80 40.02 0.05
0.24
30 30.28 0.93 29.8 0.67
0.38
20 20.76 3.80 20.48 2.40
0.62
10 11.16 11.60 10.18 1.80
 97

The performance of the two method tested using the selected polynomial, over the range
of 0 to 250 V rms value, indicated that the two methods are accurate as can be seen in Figure 4.1.
The line of the peak to peak voltage method is closer to the FLUKE meter line as shown. The
index in the graph represent a point where the same measurement is taken using the FLUE meter,
instantaneous method and peak to peak method for the purpose of analysis. The research works
performed in the area of electrical consumption measurements using microcontrollers all used
errors to analyze the measurements [80, 196]. As can be seen in the Figure 4.1 (B), which is a
section of the Figure 4.1(A) the output using the peak to peak method is more along the FLUKE
meter line than the instantaneous method. Here the laboratory FLUKE meter used is the benchmark
of that ascertain the performance of the peak to peak and instantaneous measurements methods.
The standard deviation (SD) analysis also indicated the validity of both the instantaneous and peak
to peak methods since the values of the SD are always less than 2V.

Figure 4.1 The response of the instantaneous and peak to peak voltage methods when compared
with FLUKE meter readings (A) The whole voltage range (B) Section of the graph

The ZMPT101B voltage sensor module has been experimented and/or tested by some
researchers for the purpose of calibration or scaling. What has been done in one way or the other,
is finding the transfer model of the sensor so that an accurate interpretation of the input voltage is
 98

realized from the ADC value. The ZMPT101B voltage sensor module has been calibrated by
Miron Alexe V. [197] using peak to peak voltage method. The standard voltage used is 220V
system and the measurement indicated an error of up to 0.97%. Comparison with the peak to peak
method in this research shows that it outperform their method with a maximum error of 0.55% as
illustrated in Table 4.6. The ZMPT101B sensor module is also used by Muhannad J. B. et al [198]
Arduino NANO is programmed to calculate the rms values of the currents voltage of a three phase
system using instantaneous calculation method.

Obtaining accurate rms value of the input voltage is the primary purpose of using the
ZMPT101B sensor. The work of Drang M. et al [199] is also obtained using ZMPT101B voltage
sensor module. The work monitor the line currents and phase voltage of a three phase system.
Analysis on the resulting phase voltage measurement shows a maximum error of 1.33% when
compared with the conventional meter. The error obtained in their analysis is higher than the error
indicated in Table 4.5 for the same voltage range, which is less than 1%. Moreover, the calibration
in this research work is accurately achieved using a third order polynomial, hence it provides a
better option in terms of computation requirement than the calibration in [183], in which the
sensor’s response is represented by 9th order polynomial. As such the 9th order polynomial will
require more computational time and space than the 3rd order obtained in this research.

Table 4.6: Comparison of the Voltage Measurement with Miron Alexe Method [197]

Peak to peak Miron Alex Method

Voltage method [197]
Reading Voltage Error Voltage Error
No. (V) (%) (V) (%)
1 219.98 0.01 221.42 0.65
2 221.21 0.55 220.70 0.32
3 220.55 0.25 221.42 0.65
4 221.16 0.53 220.70 0.32
5 220.48 0.22 220.70 0.32
6 220.03 0.01 222.14 0.97
7 220.52 0.24 219.97 0.01
8 220.36 0.16 222.14 0.97
 99

4.3 Residential Electrical Energy Consumption Measurement and control system

This section provides the results and discussion of the residential electrical energy
consumption measurement and control (section 3.5.1). The automatic energy consumption
management of home appliances will play an important role in smart electricity system, whose
purpose is to ensure proper usage of the electricity in an economic and optimize manner. In a
system of Arduino microcontroller, relays, voltage and current sensors and other accessories, this
work provides a system that uses instantaneous calculation method to calculate and display the
rms voltage, current, power factor and aggregated real power consumption of the loads. Moreover,
the work is designed to control the appliances so that the total power consumption does not exceed
the threshold value of 2000W. This 2,000W maximum threshold value will be appropriate for
electricity consumers between 1-200kWh per month and those that consume between 201-300kWh
per month [200].

The results of the energy consumption measurement and control will show the system
accuracy in measuring the aggregated quantities and will also demonstrate the capability of the
system to maintain the total load below the threshold value. These will be illustrated by the results
in the subsequent sub-sections. The research work will significantly reduce the total energy
consumption of a house or building, which will promote green energy consumption.

4.3.1 Measurement Analysis of the System

The measuring capability of the work is tested using some measurements of voltage,
current and aggregated power. The standard FLUKE 115 meter is used as a reference and all the
measurements are evaluated with the corresponding measurements from the FLUKE meter.
 100

4.3.1.1 Voltage Measurement

The voltage measurement is analyzed using the whole voltage range of 250V AC, which
is varied using an autotransformer. Taking the FLUKE 115 meter measurement as the standard.
The system voltage measurement is compared with that of a commercial multi meter. Table 4.7
gives the voltage measurements with the corresponding percentage error of each measurement.
The graphical representation of the measurement in Figure 4.2 shows that the system measurement
is more accurate than that of the commercial multi meter especially when measuring voltage above
220V. The index of the measurement as shown in the graphs represents the point at which same
measurement is taken using the three measuring devices.

Figure 4.2: Comparison of Voltage Measurement

 101

Table 4.7: System Voltage Measurement and its Comparison with Commercial Meter

Commercial Multimeter The System

Voltage Percentage Voltage Percentage
Fluke (V) Error (%) (V) Error (%)
Meter (V)
250 251 0.40 250.13 0.05
240 241 0.42 239.14 0.36
230 231 0.43 229.26 0.32
220 221 0.45 218.51 0.68
210 211 0.48 208.48 0.72
200 201 0.50 198.49 0.75
190 191 0.53 187.82 1.15
180 181 0.56 176.88 1.73
170 171 0.59 166.94 1.80
160 161 0.63 157.78 1.39
150 151 0.67 147.33 1.78
140 141 0.71 137.14 2.04
130 131 0.77 127.21 2.15
120 121 0.83 117.30 2.25
110 111 0.91 107.55 2.23
100 101 1.00 97.53 2.47
90 91 1.11 87.52 2.76
80 81 1.25 78.23 2.21
70 70 0.00 68.57 2.04
60 59 1.67 58.70 2.17
50 48 4.00 49.06 1.88
40 38 5.00 39.53 1.18
30 28 6.67 29.74 0.87
20 19 5.00 20.44 2.20
10 9 10.00 10.19 1.90
 102

4.3.1.2 Current Measurement

The current measurement capacity of the system is observed by connecting some of the appliances
to the system and comparing the system measurement with that of FLUKE meter. Table 4.8 shows
the comparison of the measurement showing an error of 3.88% in measuring rice cooker current.

Table 4.8: Current measurement of the system

Measuring Devices

Appliance(s) Fluke Meter (A) The Measurement and Error (%)

Control System (A)
Toaster 3.327 3.39 1.89
Rice cooker 2.705 2.81 3.88
Electric Kettle 8.300 8.22 0.96
Pressing Iron 5.220 5.24 0.38
Rice cooker +
Toaster 5.836 5.85 0.24

4.3.1.3 Power Measurement

As mentioned earlier, the system uses the instantaneous power measurement method to
measure the power consumption of the connected appliances. Therefore, the power measurements
of some of the loads are compared with the measurement of a wattmeter and the errors are shown
in Table 4.9. The power measurement has less accuracy than the voltage and current measurements
because the instantaneous power calculation comes with some errors as the result of cumulative
errors in the samples.
 103

Table 4.9: Power measurement of the system

Measuring Devices

Wattmeter The Measurement Percentage Error

Appliance(s) (W) and Control System (%)
(W)
Toaster 679.23 716.96 5.55
Rice cooker 533.21 579.34 8.65
Electric Kettle 1654.15 1769.78 6.99
Pressing Iron 1029.39 1073.18 4.25
Rice cooker +
Toaster 1135.95 1219.5 7.36

4.3.3 The Loads Control of the System

The major purpose of the system, apart from the measurement, is to manage the electricity
consumption of the loads so that the total power consumption does not exceed the threshold value, which
can also be the maximum demand of the system. The load reduction is based on priority from Load1 (L1)
to Load8 (L8) with L8 having the highest priority. The system is found to be accurately reducing the loads
when the total power consumption is above the threshold value (2,000W in this case), thus it will reduce
the load to the power consumption less than or equals the threshold value. Each time a load is to be reduced
the system will sound a brief alarm, indicating that it has detected an overload and then a load will be cutoff.
The system performance is tested using 32 arbitrary load combinations, with each condition changed after
every 3 minutes, in which the system accurately maintain the total load below the threshold level in all the
overload conditions, while maintaining the loads where there is no overload.

The home energy management system forms a simple and effective methods of home automation
for electrical energy consumption. Secondly, the system will contribute immensely in the current trend of
of smart electricity by upgrading the smart system. The system will also make a significant contribution in
 104

the field of UPS and Solar supply systems, where this system can be used to protect them automatically
without stopping their operation.

The results of the load control analysis of the system is shown in Table 4.10 in which several load
combinations are tested with the system. Some of the load combinations as can be seen in the table have
total power consumption less than the threshold value (2,000W) and some have the total load greater than
the threshold. As shown in the table when the total load is less than 2,000W the system will allow all the
loads to remain connected, which can be seen in cases (serial number) 1, 2, 6, 8 etc. On the other hand when
the total load is greater than 2,000W the system will turn off the least priority loads (from L1 to L8) until
the total load becomes less than the threshold value. In some cases like 3, 5 and 13, the system reduced only
one load to maintain the load within the threshold, but in other cases like 11, 16 and 31 multiple loads are
turn off to maintain the total load less than the threshold level.

Whenever an action is taken on a load the system will automatically reset all the loads after 3
minutes from the last load reduction. This is in anticipation that the unnecessary load will have been
disconnected by the user. In the case when all the loads are switch on, the system will turn off the loads
according to priority until the total load is not more than the threshold.

4.3.4 The System Control when the Highest Priority Load is Greater than the Threshold

The system is designed to operate within the threshold power consumption. Therefore under normal
circumstances a load more than the threshold value is not expected to be connected in the house, unless if
there is a fault in the system. When the highest priority load L8, is replaced by a load greater than the
threshold, the system will sound a 3 minutes alarm showing that there is an overload on the highest priority
load alone. The alarm will take 3 minute and then reset the system, by assuming that the overload has been
corrected by the users. The system can also be reset from the Arduino board using the microcontroller reset
button. The 3 minutes alarm is set up in the Arduino program which sounds through the buzzer.
 105

Table 4.10: The Loads Control of the System

Before load control After load control

S/N Time Total Total
(min) Power Power Remark
L8 L7 L6 L5 L4 L3 L2 L1 (W) L8 L7 L6 L5 L4 L3 L2 L1 (W)
1 0 0 0 1 0 1 1 1 1 919 0 0 1 0 1 1 1 1 919 No action
2 3 0 1 0 1 0 0 1 0 1911 0 1 0 1 0 0 1 0 1911 No action
3 6 0 1 0 1 1 1 1 0 2018 0 1 0 1 1 1 0 0 1926 L2 turn off
4 9 0 0 1 1 1 1 1 1 1539 0 0 1 1 1 1 1 1 1539 No action
5 12 0 1 0 1 0 1 1 1 2033 0 1 0 1 0 1 1 0 1944 L1 turn off
6 15 1 0 0 0 0 1 1 1 1758 1 0 0 0 0 1 1 1 1758 No action
7 L3 & L7 turn
18 1 1 0 0 0 1 0 0 2786 1 0 0 0 0 0 0 0 1578 off
8 21 0 0 1 1 1 0 1 0 1409 0 0 1 1 1 0 1 0 1409 No action
9 24 0 0 1 1 0 0 1 1 1451 0 0 1 1 0 0 1 1 1451 No action
10 27 1 0 1 0 0 0 0 0 2182 1 0 0 0 0 0 0 0 1563 L6 turn off
11 L1, L2, L4 &
30 1 0 0 1 1 0 1 1 2430 1 0 0 0 0 0 0 0 1557 L5 turn off
 106

12 33 0 1 0 0 1 1 1 1 1479 0 1 0 0 1 1 1 1 1479 No action

13 36 0 1 0 1 0 1 0 1 2008 0 1 0 1 0 1 0 0 1889 L1 turn off
14 39 0 1 1 0 1 0 0 1 1985 0 1 1 0 1 0 0 1 1985 No action
15 L5 & L6 turn
42 1 0 1 1 0 0 0 0 2812 1 0 0 0 0 0 0 0 1549 off
16 L2, L3, L4 &
45 1 1 0 0 1 1 1 0 2948 1 0 0 0 0 0 0 0 1535 L7 turn off
17 48 0 0 1 1 0 0 1 0 1378 0 0 1 1 0 0 1 0 1378 No Action
18 L1, L2, L3,L4
51 1 0 0 1 1 1 1 1 2448 1 0 0 0 0 0 0 0 1567 & L5 turn off
19 L1,L4 &L7
54 1 0 1 0 1 0 0 1 2352 1 0 0 0 0 0 0 0 1563 turn off
20 57 0 1 0 0 0 0 1 0 1276 0 1 0 0 0 0 1 0 1276 No action
21 60 0 1 1 0 1 0 0 0 1976 0 1 1 0 1 0 0 0 1976 No action
22 63 0 0 0 1 1 1 1 1 1893 0 0 0 1 1 1 1 1 1893 No action
23 L1, L2 & L5
66 0 1 1 1 0 0 1 1 2663 0 1 1 0 0 0 0 0 1838 turn off
24 69 1 0 0 0 1 1 0 0 1659 1 0 0 0 1 1 0 0 1659 No action
25 72 0 0 1 0 0 1 1 1 867 0 0 1 0 0 1 1 1 867 No action
26 L2, L4, L6 &
75 1 1 1 0 1 0 1 0 3548 1 0 0 0 0 0 0 0 1562 L7 turn off
 107

27 L5 & L7 turn
78 1 1 0 1 0 0 0 0 3401 1 0 0 0 0 0 0 0 1561 off
28 81 0 0 0 1 1 0 0 0 690 0 0 0 1 1 0 0 0 690 No action
29 84 0 1 0 0 0 1 1 1 1412 0 1 0 0 0 1 1 1 1412 No action
30 L2 & L5 turn
87 1 0 0 1 0 0 1 0 2272 1 0 0 0 0 0 0 0 1558 off
31 L3, L5 & L6
90 1 0 1 1 0 1 0 0 2891 1 0 0 0 0 0 0 0 1561 turn off
32 93 0 0 1 0 1 0 0 0 690 0 0 1 0 1 0 0 0 690 No action
4.3.5 General Analysis

First and foremost the system is very good in managing the load consumption
of the house because it does not allow the loads to be greater than the threshold value
and has automatic reconnection of the relays after three minutes of action. Though
there are higher errors in power measurement due to instantaneous calculation method,
the accuracy the power measurement is enough to perform the load control. The
voltage and current measurement are very good with acceptable errors though the
errors in power measurement are a bit higher than that of the voltage and current.
Another little constrain with the system is that sometimes the measured values use to
fluctuate a little, but it is still readable as most of the times it keeps on repeating one
value. Moreover, the system can be applied on other systems with different threshold
value which can be accommodated by changing the value in the Arduino program.

4.4 Development of the Non-Intrusive Appliances Identification System

As described in section 3.6, the monitoring system here uses the aggregated
real power (P), rms current (Irms) and power factor (pf) of the connected loads to
identify the loads using feedforward ANN as the identifier. The identification system
perform this task non-intrusively, which means it can be applied outside the residence
and be operated without disturbing the occupants. All the times the monitoring system
will take the aggregated quantities and use them to identify the connected loads. This
work explore more opportunity of developing home NILM using microcontroller.

4.4.1 The ANN Training Results of the NILM System

Training and retraining of the ANN network involved the adjustment of the
number of neurons in the hidden layer, change of activation function and adjustment
 109

of the bias, which is performed automatically in the MATLAB program. After each
training of the network, the network overall regression and root means square error
(RMSE) are observed until the best network is found with overall regression of
0.98068 and an average RMSE of 0.1006, the complete summary of the ANN is given
in Table 4.11. The network has 12 neurons in the hidden layer, ‘tansig’ activation
function in both the hidden and output layers as shown in Figure 4.3. The ‘tansig’
activation function is very fast in delivering the desired output in training the network.

Table 4.11: The Summary of the ANN for Non-Intrusive Appliances Identification

Parameter Specification

Network Feedforward
Configuration

Training Technique Levenberg - Marquardt

Performance Mean Square Error (0.009552)

No. of Input 3

Number of Outputs 4

No. of Input Layer 3

No. of Hidden Layer 12

No. of Output Layer 4

Training Function ‘trainlim’

No. of training 64
Pattern

No. of Testing 28
Pattern

No. of Epochs 42

Initial Learning rate 0.01

Momentum rate 0.9

Correlation Training Validation Testing All 0.98322

Coefficient Values 0.98286 0.96663 0.98068
 110

Hidden Layer ‘tansig’

Activation Function

Output Layer ‘tansig’

Activation Function

Figure 4.3: The Trained ANN for the Non-Intrusive Load Identification system

4.4.2 The Network Simulation Results

After the network is formed, its performance is observed. The output of the
network is compared with the target output of the training data through graphical
representation as shown in Figure 4.4. The predicted and targeted outputs are
 111

accurately matched and since the output desired is 0 or 1, approximating the predicted
value to the nearest whole number will make it exactly the targeted output. However
what is desired from the ANN is the load recognition with “1” representing ON and
“0” representing OFF.

Figure 4.4: The Network Simulation Result

Similarly the network is tested with the 32 testing data (which is not part of the
training data), and the mapping of the simulated value against the actual value is shown
in Figure 5.3. Also the approximation of the simulated output to the nearest whole
number will fit exactly on the actual value. This will give the desired ON and OFF of
the appliances. This has shown that the network can accurately identify the connected
loads. The issue of memorizing phenomina can arise when the ANN just memorize
the input and output without being able to predict the outputs for new inputs [201,
202]. This phenomina is avoided in this work by simulating the network in a different
data scenario as illustrated in Figure 4.5.
 112

Figure 4.5: The Result of Testing the Network

4.4.3 System Performance

Looking very well at Figures 4.4 and 4.5, it can be seen that the output of the
network (the simulated) is not in the form of 0 and 1 as in actual or desired output.
This output is converted to binary be the Arduino program and that will make it the
same as the desired values. All the possible load combinations are tested and the
corresponding load pattern are displayed on the LCD as shown in Table 4.12. The table
presented the conditions of the appliances and the corresponding load identification by
the ANN.

4.4.4 General Analysis

The power calculation using instantaneous method is always fluctuating due to

the averaging of the instantaneous power. Hence we may not have exactly the same
value of the power in two consecutive displays of the system. Nevertheless, the system
 113

accurately identifies the connected loads despite the little fluctuations of the power
measurement as the result of the instantaneous method. Another effect that leads to
incorrect identification of the loads is the drop in the supply voltage.

When the voltage in the laboratory drops from the normal 240V to around
230V, it is observed that there is a drop in the power consumption measurement and
that can affect the load identification. This anomaly can be addressed by including the
voltage in the load signatures, which may also increase the complexity of the ANN.
On the other hand, ensuring a stabilized voltage supply in the system will solve the
problem of the drop in power consumption. In any case, without the problem of supply
voltage drop, the system is said to identify the connected loads with 100% accuracy.

Table 4.12: The System performance

Appliances Condition ANN Output

(the Load Identification)
Load 4 Load 3 Load 2 Load 1 L4 L3 L2 L1
OFF OFF OFF OFF 0 0 0 0

OFF OFF OFF ON 0 0 0 1

OFF OFF ON OFF 0 0 1 0
OFF OFF ON ON 0 0 1 1

OFF ON OFF OFF 0 1 0 0

OFF ON OFF ON 0 1 0 1

OFF ON ON OFF 0 1 1 0

OFF ON ON ON 0 1 1 1

ON OFF OFF OFF 1 0 0 0

ON OFF OFF ON 1 0 0 1

ON OFF ON OFF 1 0 1 0

ON OFF ON ON 1 0 1 1

ON ON OFF OFF 1 1 0 0

ON ON OFF ON 1 1 0 1
 114

ON ON ON OFF 1 1 1 0

ON ON ON ON 1 1 1 1

4.5 The ANN-Arduino NILM based Residential Appliances Management and

Control System

The previous section described the results of the non-intrusive appliances

identification using microcontroller. Similar method has been adopted here to monitor
and manage the home appliances with load controlling capability. In the same way the
system here measures and use the aggregated real power (P), rms current (Irms) and
power factor (pf) of the connected appliances and run the ANN model to recognize the
connected loads. The monitoring result is then used by the microcontroller to
determine the cumulative time of work of each appliances and switch off the
appliances that exhausted their hours of work for the purpose of the appliances’
management. The block diagram of the system can be seen in Figure 3.9 in chapter 3.
The system is a real-time monitoring system and non-intrusive in nature. The results
of the monitoring, appliances’ timing and control all demonstrated a very good
performance, which make the system applicable in real life.

4.5.1 ANN Training Results of the Appliances Management and Control System

As in the previous case training the ANN involves the adjustment of neurons
in the hidden layer (also only one hidden layer is used here) and change in activation
function to observe the performance of the network. In each training and retraining,
which is performed automatically in the MATLAB, the mapping of the actual output
against the simulated output using the whole training data is observed. Also the root
mean square error (RMSE) is observed until the best preforming network is found with
least number of hidden neurons possible. The minimum number of hidden neurons
obtained after several retraining is 9, with overall regression of 0.98528 “tansig” and
 115

“purelin” as the activation functions in the hidden and output layers respectively, as
shown in Figure 4.6. The complete summary of the ANN is shown in Table 4.13.

Figure 4.6: The Trained ANN for the Load Controlling System

Table 4.13: Summary of the ANN for ANN-Arduino NILM based Appliances
Management and Control System

Parameter Specification

Network Feedforward
Configuration
 116

Training Technique Levenberg - Marquardt

Performance Mean Square Error (0.009552)

No. of Input 3

Number of Outputs 4

No. of Input Layer 3

No. of Hidden Layer 9

No. of Output Layer 4

Training Function ‘trainlim’

No. of training 67
Pattern

No. of Testing 26
Pattern

No. of Epochs 52

Initial Learning rate 0.01

Momentum rate 0.9

Correlation Training Validation Testing All

Coefficient Values 0.98726 0.97506 0.98713 0.98528

Hidden Layer ‘tansig’

Activation Function

Output Layer ‘purelin’

Activation Function

4.5.2 Network Simulation and Testing

The network obtained is simulated with the training data where an average
RMSE of 0.0872 is obtained, this is the average RMSE before the output is
approximated to nearest whole numbe (0 and 1). The mapping of the simulated output
against the actual output is shown in Figure 4.7 which shows that the simulated output
will map exactly on the actual output when approximated to the nearest 0 and 1. This
has shown that the ANN provide the correct load identification of the appliances.
 117

Figure 4.7: The Network Simulation Result

Another data which that not part of the training data are also used to test the
performance of the network. The mapping of the actual value and simulated value, in
Figure 4.8 also shows that the simulated value is exactly the same as the actual value
when approximated. This of course shows that the network can perform very well in
the load cotrolling system. Also the ANN-Arduino NILM is free of the memorising
phenomina as it predicted the results very well in Figure 4.8. The memorising
phenomina normally cause the network to providing in correct predictions with other
test values [202].
 118

Figure 4.8: The Result of Testing the Network with Different Data

4.5.3 Comparison with Previous Work

The performance of the network obtained here is higher than that of the
monitoring system in the previous section. Previously, the monitoring network was
realized using 12 hidden neurons as against the 9 hidden neurons in the management
and control system. The reduction in the number of hidden neurons will ease the
computational requirement of the microcontroller, as the results of reduced length of
the network equations and hence there will be quicker response in managing the
appliances. Also the overall regression of 0.98528 obtained in this network is higher
than the 0.98068 obtained in the previous case. This can be seen in Table 4.13 which
gives the general summary of the ANN.
 119

4.5.4 Load Monitoring Performance of the Appliances Management and

Control System

The ANN output, which is in decimal number is approximated by the Arduino

microcontroller to the nearest whole number, so that the desired output in the form of
0 and 1 are obtained. All the possible load combinations of the system are tested and
the corresponding load patterns are accurately displayed on the LCD. The block
diagram of the system can be seen in Figure 3.9 of chapter 3 and Appendix E shows
the diagram of the system with some of the displayed load connection patterns. Also
the complete results of the load monitoring aspect is shown in Appendix F.

Apart from displaying the load connection pattern, the system also displays the
cumulative ON time of each of the appliances. The cumulative ON tine is determined
by the Arduino microcontroller when it process the load monitoring results. Both the
monitoring result and cumulative ON time are displayed on the LCD as illustrated in
Appendix G.

4.5.5 Appliances’ Timing and Load Control Performance of the System

The appliances timing and control of the system can be determined by setting
a time of work for each of the appliances and observing the system control. These
working hours depend on the desired application of the appliances based on its
operation duration which has to be included in the Arduino programing code, but for
demonstration purpose an ON time of 5 minutes, 15 minutes, 30 minutes and 1 hour
are assigned for loads 1, 2, 3 and 4 respectively. The system is run continuously for 1
hour, with all the loads switched ON and the loads are turn off by the system according
to their ON time, with small time errors as shown in Table 4.14.
 120

Table 4.14: The Appliances Timing by the System

S/N Appliance Power Allowed Disconnection Time

Rating Duration time Error
(W) (ON time) (Seconds) (%)
(Minutes)
1 Energy Saving 24 5 317 5.67
Lamp
2 Standing Fan 50 15 916 1.78
3 Electric Bulb 100 30 1823 1.28
4 Filament Lamp 300 60 3638 1.06

When the assigned time of an appliance is over, the system will turn off the
supply from the appliances, even though there is a little error between the allocated
duration and disconnecting time as shown in Table 4.14. Figure 5.10 shows the
pictures of the appliances after each disconnection time. The system operation is
shown in Appendix L which illustrates how the system control the appliances
throughout the 1 hour. As observed during the system operation each of the loads is
turn off after its designated ON time, with little errors as shown in the figure and as
presented in Table 4.14.

4.5.6 System Analysis

The monitoring system in this case is achieved with 9 number of hidden

neurons as compared to the 12 hidden neurons in the Non-Intrusive identification
system. Moreover the feedforward ANN correctly identifies the load patterns of the
system. The operating time of each load is observed by the system and the loads are
disconnected when their operation time is over, this is illustrated by the results in Table
4.14 where the allowed ON time against the disconnection time is shown. Also the
performance of the load identifier is higher than in previous case and the timing system
has an error of 1.78% for 15 minutes timing (as in Table 4.14) and the error keeps
reducing as the timing goes longer for 30 minutes and 1 hour. This therefore shows
that the system can accurately be applied in real situation and for real-time operation.
 121

4.6 Comparison of the ANN-Arduino NILM with Support Vector Machine

(SVM)

Other methods of machine learning have been used in other research works to develop
the non-intrusive load identification that can be compared with the ANN-Arduino
NILM in this research work. The SVM which is also a supervised learning algorithm
has been implemented into NILM by some researchers [110, 112]. Marisa B. F. et al
[112] uses SVM and k-NN classification algorithm for NILM system. Their NILM
system is very accurate in identifying the residential appliances, but both the scenarios
of Linear and RBF analyzed in the work, none of them give 100% recognition like the
ANN-Arduino NILM.

Similarly Y, H. Lin and M. S. Tsai [110] applied the hierarchical SVM for load
identification using NILM systems. Although the system identifies the operation status
of each load, but the hierarchical nature of the system allows does not allow multiple
loads to be identified at the same time. This is unlike the ANN-Arduino NILM in this
research work where multiple operations of the appliances are identified. In general
the ANN-Arduino NILM is more capable of being applied in real-time residential load
monitoring than the former SVM methods. Also when it comes to computational
requirement the SVM will be more complex than the ANN-Arduino NILM.

4.7 Summary

This chapter contains the general results and discussion of the research work.
It contains the result and analysis of the ZMPT101B sensor calibration, Energy
consumption measurement and control using Arduino microcontroller, the ANN-
Arduino non-intrusive appliances identification and the results of Ann-Arduino NILM
based appliances management and control system. The non-intrusive load
identification method is a quite systematic way of load monitoring using single point
sensing. The four loads identification system can be made to monitor as many loads
as possible using sub metering NILM system [48], where the appliances are divided
into sub groups.
 122

In the ZMPT101B sensor calibration, an accurate relationship between the

input voltage to the sensor and the output ADC is determined. The most important rms
voltage measurement in low voltage AC supply are the 220 – 240V range and 110 -
120V range depending on the power supply of a particular country. The calibration
results obtained in this research showed that for rms measurement of 220 – 240V range
(which fall within the range of Malaysia) the maximum error, as observed in Table 4.5
is 0.83% for instantaneous method and 0.41% for the peak to peak method. On the
other hand the measurement of 110 – 120V has 2.26% and 0.70% errors in the
instantaneous and peak to peak method respectively. The voltage measurement below
50V is associated with higher errors especially in the instantaneous method. However,
less than 50V AC voltage is referred to as extra-low voltage, which is less relevant in
power monitoring systems. To use the calibration in this research work the trimmer of
the ZMPT101B sensor has to be adjusted to a maximum of 640 ADC when 250V AC
is connected to the input of the sensor. Similarly the ADC output can be trimmed to a
maximum of 575 ADC at 120V input voltage. This will serve as the reference before
Equation (4.3) can be used as the transfer model of the sensor. The Arduino code is
used to determine the maximum ADC when the voltage is sampled 1000 times.

The second part of the chapter provides the results of energy consumption
measurement and control system using Arduino microcontroller. The system measures
and display the aggregated real power consumption, rms current and rms voltage of
the connected loads. It also manages the total power consumption of the loads so that
the threshold value is not exceeded. The voltage measurement using the system has a
maximum error of 2.47% as against some higher values using a commercial multi
meter and taking FLUKE meter as the standard. The current measurement as compared
to the FLUKE meter has a maximum error of 3.88%, whereas an error of up to 8.65%
is obtained in power measurement. The load control is accurate as presented in Table
4.10.

This chapter provides the results and discussions of ANN-Arduino non-

intrusive appliances identification. The major advantage of non-intrusive load
monitoring is its ability to monitor or identify the appliances without having to enter
the building and with only the measurements at the service entry.
 123

The results and discussion of the ANN-Arduino NILM based appliances

management and control is also presented in the chapter. The performance of the
systems are promising and all the results indicated the system’s readiness to be
implemented in a real situation and for real-time application. Moreover, the
performance of the ANN-Arduino NILM in home appliances management and control
is quite demonstrated by the work.
 124

CHAPTER 5

5CONCLUSIONS AND RECOMMENDATIONS FOR FUTURE WORK

5.1 Conclusions

This thesis presented a non-intrusive load management system for residential

loads using ANN based Arduino microcontroller. The work involves the calibration of
ZMPT101B voltage sensor module, which developed a more accurate sensor equation.
It also involves the development of energy consumption measurement and control
using Arduino microcontroller, the system that achieved the home automatic energy
consumption management. ANN-Arduino non-intrusive load identification system has
also been developed, which monitors the residential appliances non-intrusively and
lastly an ANN-Arduino NILM based residential appliances management and control
is developed. The ANN-Arduino non-intrusive appliances identification has achieved
the steady-state identification objective and the ANN-Arduino NILM based residential
appliances management has achieved the performance of the ANN-Arduino NILM.
The overall developed systems show the readiness to be implemented in real situation
and real-time applications. This proposed work can be implemented in smart electricity
systems to promote the energy management and automation of the system.

The need to optimize and automatize the usage of electricity is always

becoming essential as the result of increase in energy demand and the way electricity
is becoming more and more necessary. Both the optimization and automation of
energy consumption cannot be achieved without proper monitoring of the appliances
that consume the power. Therefore the ANN-Arduino NILM based system has been
implemented using Arduino microcontroller and ANN identifier. The following
conclusions can be drawn from the achievements of the research works.
 125

(i) It can be concluded that the new approach of ANN-Arduino based non-intrusive
identification of home appliances using steady –state real power, rms current and
power factor as load signatures is capable of correctly identifying all the loads
combinations in the system. The identification accuracy is retained even with
some fluctuations in the input voltage. Hence the method has presented an
improved accuracy in identifying home appliances using aforementioned
signatures.
(ii) As the result of the calibration work of the ZMPT101B voltage sensor module it
can be concluded that a more accurate voltage measurement is obtained when
the sensor is modeled using a third order polynomial than the linear system as
shown in the results and analysis.
(iii) The developed residential energy consumption measurement and control system
will give an improvement in home appliance management sector. The problem
of energy unaware activities and the limitation of user intervention in the energy
management will be overcome. The system will ensure energy savings among
the home appliances.
(iv) From the experimental findings on the ANN-Arduino NILM based reisidential
appliances management it can be concluded that an improved system of home
appliances’ control is obtained.
(v) As per the NILM contribution in providing simple and cost effective method of
monitoring the appliances, it can be concluded from the works in this research
that the NILM has something to offer beyond just monitoring the appliances. In
fact its greatest contribution will come from the residential management side of
the appliances as demonstrated by the research work.

5.2 Recommendations for Future Work

Based on the outcomes of the research work, some recommendations

regarding future work in the study are highlighted below:
(i) Further analysis on the NILM system is required so that more numbers of
appliances can be identified by the ANN. This will help in the home
appliances monitoring of more loads using a single NILM system.
 126

(ii) It is recommended to carry out further research on this NILM method so

that sub metering system of load monitoring is formed, with each NILM
system monitoring a group of appliances. This will solve the problem of
high number of appliances in the home.

(iii) The energy consumption measurement and control system can be modified
to work entirely on wireless basis. This will include the use of wireless
relay system to control the loads, in order to further reduce the complexity
of the network.

(iv) The non-traditional signatures like temperature, light and sound and their
sensing devices can help in improving the monitoring system. Therefore
there is need to incorporate such signatures to simplify and improve the
appliances identification.

(v) In cases where the home appliances are distinct in power consumption, a
smart meter reading can recognize the loads in a similar microcontroller
programming. Therefore systems utilizing the smart meter reading for load
identification and management can further be investigated using
microcontroller and related projects.

(vi) This system will have a very good application in solar inverter system and
UPS, where the appliances are designated to work for certain operating
hours. Therefore it is recommended to incorporate this ANN-Arduino
NILM in solar inverters and UPS systems for automation and load
management. This work will also help greatly in preventing the energy
world from serious energy crises. Moreover the combination of the system
with communication facilities like WIFI and other wireless system will
results into a more desired monitoring and management system.

5.3 Research Contributions

The following list gives the contributions of the research in the residential
energy monitoring and management.
 127

(i) Residential Appliances’ Identification using Steady-State Load Signatures.

The ANN-Arduino non-intrusive load Identification system has shown an

improved method of load identification with higher accuracy compared to previous
works.

(ii) Residential Energy Consumption Measurement and Control.

The residential energy consumption measurement and control system, using

Arduino microcontroller, from the research work will form a basis for implementing
house energy management especially in smart housing system.

(iii) ANN-Arduino NILM based residential energy management and control.

The load ANN-Arduino NILM based residential management and control system
will help the smart housing researchers in developing a more efficient and energy
savings smart system.

(iv) ZMPT101B Voltage Sensor Module Algorithm

The research work has also achieved accurate ZMPT101B sensor algorithm using
polynomial regression. This work is the first of its kind and it gives more accurate
representation of the sensor ADC conversion than the linear algorithm.
 128

REFERENCES

[1] Yung Fei, W., Ahmet Sekercioglu, Y., Drummond, T. and Voon Siong, W. (2013).
Recent approaches to non-intrusive load monitoring techniques in residential
settings. IEEE 2013 IEEE Symposium on Computational Intelligence
Applications In Smart Grid (CIASG).IEEE: 73-79.
[2] Zoha, A., Gluhak, A., Imran, M. A. and Rajasegarar, S. Non-intrusive load
monitoring approaches for disaggregated energy sensing: A survey. Sensors.
2012. 12(12): 16838-16866.
[3] Chang, H.-H. Load identification of non-intrusive load-monitoring system in smart
home. WSEAS Transactions on Systems. 2010. 9(5): 498-510.
[4] Das, S., Srikrishna, S., Shukla, A., Harsha, G. and Deb, S. (2013). A low-cost non-
intrusive appliance load monitoring system. IEEE 2013 3rd International
Advance Computing Conference (IACC). IEEE: 1641-1644.
[5] Ali, M., Yousaf, A. and Usman, F. (2017). Designing and simulation of load control
& monitoring system through Demand Side Management technique. IEEE
2017 8th International Renewable Energy Congress (IREC). IEEE: 1-4.
[6] Nielsen, J. J., Ganem, H., Jorguseski, L., Alic, K., Smolnikar, M., Zhu, Z., Pratas,
N. K., Golinski, M., Zhang, H. and Kuhar, U. Secure Real-Time Monitoring
and Management of Smart Distribution Grid using Shared Cellular Networks.
IEEE Wireless Communications. 2017. 24(2): 10-17.
[7] Hart, G. W. Nonintrusive appliance load monitoring. Proceedings of the IEEE.
1992. 80(12): 1870-1891.
[8] Marceau, M. L. and Zmeureanu, R. Nonintrusive load disaggregation computer
program to estimate the energy consumption of major end uses in residential
buildings. Energy Conversion and Management. 2000. 41(13): 1389-1403.
[9] Bijker, A. J., Xia, X. and Zhang, J. (2009). Active power residential non-intrusive
appliance load monitoring system. IEEE 2009 AFRICON. IEEE: 1-6.
[10] Hosseini, S. S., Agbossou, K., Kelouwani, S. and Cardenas, A. Non-intrusive load
monitoring through home energy management systems: A comprehensive
review. Renewable and Sustainable Energy Reviews. 2017. 79: 1266-1274.
 129

[11] Yang, C. C., Soh, C. S. and Yap, V. V. A non-intrusive appliance load monitoring
for efficient energy consumption based on Naive Bayes classifier. Sustainable
Computing: Informatics and Systems. 2017. 14: 34-42.
[12] Wu, X., Han, L., Wang, Z. and Qi, B. A nonintrusive fast residential load
identification algorithm based on frequency‐domain template filtering. IEEJ
Transactions on Electrical and Electronic Engineering. 2017. 12(S1).
[13] Ninad Khandekar, N. P., Kalpak Thube and Dr. P. B. Mane. Non-Intrusive
Appliance Load Monitoring System Using Zigbee Protocol. International
Journal of Engineering Research & Technology (IJERT). 2014. 3(4): 2415-
2417.
[14] Nguyen, T. K., Dekneuvel, E., Jacquemod, G., Nicolle, B., Zammit, O. and
Nguyen, V. C. Development of a real-time non-intrusive appliance load
monitoring system: An application level model. International Journal of
Electrical Power & Energy Systems. 2017. 90: 168-180.
[15] Chang, H.-H. Non-intrusive demand monitoring and load identification for energy
management systems based on transient feature analyses. Energies. 2012.
5(11): 4569-4589.
[16] Feng, C. Y., Hoe, H. M., Abdullah, M. P., Hassan, M. Y. and Hussin, F. (2013).
Tracing of energy consumption by using harmonic current. IEEE 2013 Student
Conference on Research and Development (SCOReD), IEEE: 444-449.
[17] Jimenez, Y., Duarte, C., Petit, J., Meyer, J., Schegner, P. and Carrillo, G. Steady
state signatures in the time domain for nonintrusive appliance identification.
Ingeniería e Investigación. 2015. 35(1): 58-64.
[18] Wichakool, W (2011). Advanced nonintrusive load monitoring system. PhD
Thesis. Massachusetts Institute of Technology.
[19] Valero Pérez, M. N. (2011). A non-intrusive appliance load monitoring system
for identifying kitchen activities. PhD Thesis. Aalto University - School of
Electrical Engineering, Espoo, Finland.
[20] Christopher Laughman, K. L., Robert Cox, Steven Shaw, Steven Leeb, Les
Norford, and Peter Armstrong. Power Signature Analysis. IEEE Power &
Energy. 2003. 99(2): 56-63.
[21] Cox, R. W., Bennett, P., McKay, T., Paris, J. and Leeb, S. B. (2007). Using the
non-intrusive load monitor for shipboard supervisory control. IEEE 2007
Electric Ship Technologies Symposium (ESTS'07). IEEE: 523-530.
 130

[22] Shrestha, A., Foulks, E. L. and Cox, R. W. (2009). Dynamic load shedding for
shipboard power systems using the non-intrusive load monitor. IEEE 2009
Electric Ship Technologies Symposium (ESTS). IEEE: 412-419.
[23] DeNucci, T., Cox, R., Leeb, S. B., Paris, J., McCoy, T. J., Laughman, C. and
Greene, W. C. (2005). Diagnostic indicators for shipboard systems using non-
intrusive load monitoring. IEEE 2005 Electric Ship Technologies Symposium
(ESTS) IEEE: 413-420.
[24] Suzdalenko, A. and Galkin, I. (2013). Case study on using non-intrusive load
monitoring system with renewable energy sources in intelligent grid
applications. 2013 International Conference on Compatibility and Power
Electronics (CPE). 115-119.
[25] Berges, M. E., Goldman, E., Matthews, H. S. and Soibelman, L. Enhancing
electricity audits in residential buildings with nonintrusive load monitoring.
Journal of industrial ecology. 2010. 14(5): 844-858.
[26] Shaikh, P. H., Nor, N. B. M., Nallagownden, P., Elamvazuthi, I. and Ibrahim, T.
A review on optimized control systems for building energy and comfort
management of smart sustainable buildings. Renewable and Sustainable
Energy Reviews. 2014. 34: 409-429.
[27] Kailas, A., Cecchi, V. and Mukherjee, A. (2012). A survey of contemporary
technologies for Smart Home energy management. Handbook of green
information and communication systems. Waltham, USA. Elsevier: 35-56.
[28] Fernandes, F., Morais, H., Vale, Z. and Ramos, C. Dynamic load management in
a smart home to participate in demand response events. Energy and Buildings.
2014. 82: 592-606.
[29] Westergren, K.-E., Högberg, H. and Norlén, U. Monitoring energy consumption
in single-family houses. Energy and buildings. 1999. 29(3): 247-257.
[30] Wang, J.-j. and Wang, S. (2010). Wireless sensor networks for home appliance
energy management based on ZigBee technology. IEEE 2010 International
Conference on Machine Learning and Cybernetics (ICMLC). IEEE: 1041-
1046.
[31] Bonfigli, R., Principi, E., Fagiani, M., Severini, M., Squartini, S. and Piazza, F.
Non-intrusive load monitoring by using active and reactive power in additive
Factorial Hidden Markov Models. Applied Energy. 2017. 28: 1590-1607.
 131

[32] Hu, Q. and Li, F. Hardware design of smart home energy management system
with dynamic price response. IEEE Transactions on Smart Grid. 2013. 4(4):
1878-1887.
[33] Hussain, S., Ikram, M. J. and Arshad, N. (2014). A Low Cost Implementation of
Home Area Networks for Home Energy Management Systems. IEEE 2014
Fourth International Conference on Big Data and Cloud Computing
(BdCloud). IEEE: 688-695.
[34] Han, J., Choi, C.-S., Park, W.-K., Lee, I. and Kim, S.-H. Smart home energy
management system including renewable energy based on ZigBee and PLC.
Consumer Electronics, IEEE Transactions on. 2014. 60(2): 198-202.
[35] Adabi, A., Manovi, P. and Mantey, P. (2016). Cost-effective instrumentation via
NILM to support a residential energy management system. IEEE 2016
International Conference on Consumer Electronics (ICCE). IEEE: 107-110.
[36] Lin, S., Zhao, L., Li, F., Liu, Q., Li, D. and Fu, Y. A nonintrusive load
identification method for residential applications based on quadratic
programming. Electric Power Systems Research. 2016. 133: 241-248.
[37] Lee, S., Song, B., Kwon, Y. and Kim, J.-h. Non-intrusive Load Monitoring for
Home Energy Usage with Multiple Power States Recognition. Advanced
Science and Technology Letters. 2015. 111: 282-289.
[38] Dawei, H., Weixuan, L., Nan, L., Harley, R. G. and Habetler, T. G. Incorporating
Non-Intrusive Load Monitoring Into Building Level Demand Response. IEEE
Transactions on Smart Grid. 2013. 4(4): 1870-1877.
[39] Racines, D. L. and Candelo, J. E. Non Intrusive Load Identification with Power
and Impedance obtained from Smart Meters. International Journal of
Engineering & Technology. 2014. 6(4): 1867-1876.
[40] Giri, S. and Bergés, M. An energy estimation framework for event-based methods
in Non-Intrusive Load Monitoring. Energy Conversion and Management.
2015. 90: 488-498.
[41] Esaa, A., Farahin, N., Abdullah, M. P., Hassan, M. Y. and Hussin, F. Electricity
consumption pattern disaggregation using non-intrusive appliance load
monitoring method. Jurnal Teknologi. 2016. 78(5-7): 29-35.
[42] Aliyu, G., Khalid, S. B. A., Mustafa, M. W. and Shareef, H. An improved three‐
phase reactive power measurement algorithm using walsh functions transform.
IEEJ Transactions on Electrical and Electronic Engineering. 2014. 9(1): 7-14.
 132

[43] Jazizadeh, F. and Becerik-Gerber, B. A Novel Method for Non Intrusive Load
Monitoring of Lighting Systems in Commercial Buildings. Bridges. 2014. 10:
523-530.
[44] Sianaki, O. A., Hussain, O., Dillon, T. and Tabesh, A. R. (2010). Intelligent
decision support system for including consumers' preferences in residential
energy consumption in smart grid. IEEE 2010 Second International
Conference on Computational Intelligence, Modelling and Simulation
(CIMSiM). IEEE: 154-159.
[45] Batra, N., Dutta, H. and Singh, A. (2013). INDiC: Improved Non-intrusive Load
Monitoring Using Load Division and Calibration 2013 12th International
Conference on Machine Learning and Applications (ICMLA). 79-84.
[46] Wichakool, W. (2011). Advanced Non-Intrusive Load Monitoring System. PhD
Thesis. Massachusetts Institute of Technology.
[47] Burbano Acuña, M. D. Intrusive and Non-Intrusive Load Monitoring (A Survey).
Latin American Journal of Computing LAJC. 2015. 2(1): 1-9.
[48] Ridi, A. and Hennebert, J. Hidden Markov Models for ILM appliance
identification. Procedia Computer Science. 2014. 32: 1010-1015.
[49] McLaughlin, S., McDaniel, P. and Aiello, W. (2011). Protecting consumer
privacy from electric load monitoring. ACM 2011 18th ACM conference on
Computer and communications security: ACM: 87-98.
[50] Yu-Hsiu, L. and Men-Shen, T. (2012). Application of neuro-fuzzy pattern
recognition for Non-intrusive Appliance Load Monitoring in electricity energy
conservation. IEEE 2012 International Conference on Fuzzy Systems (FUZZ-
IEEE). IEEE: 1-7.
[51] Liang, J., Ng, S. K., Kendall, G. and Cheng, J. W. Load signature study—Part I:
Basic concept, structure, and methodology. IEEE Transactions on Power
Delivery. 2010. 25(2): 551-560.
[52] Liang, J., Ng, S. K., Kendall, G. and Cheng, J. W. Load signature study—Part II:
Disaggregation framework, simulation, and applications. Power Delivery,
IEEE Transactions on. 2010. 25(2): 561-569.
[53] Cheng, J. W., Ng, S. K., Liang, J. and Zhong, J. (2012). An alternative use of
power quality information-Load signature studies & applications. IEEE 2012
15th International Conference on Harmonics and Quality of Power (ICHQP),
IEEE: 150-155.
 133

[54] Po-An, C., Chi-Cheng, C. and Ray, I. C. (2012). Automatic appliance

classification for non-intrusive load monitoring IEEE 2012 International
Conference on Power System Technology (POWERCON). IEEE: 1-6.
[55] Chang, H.-H., Lin, C.-L. and Lee, J.-K. (2010). Load identification in
nonintrusive load monitoring using steady-state and turn-on transient energy
algorithms. IEEE 2010 International Conference on Computer Supported
Cooperative Work in Design (CSCWD). IEEE: 27-32.
[56] Hsueh-Hsien, C., Ching-Lung, L. and Hong-Tzer, Y. (2008). Load recognition
for different loads with the same real power and reactive power in a non-
intrusive load-monitoring system CSCWD 2008 12th International
Conference on Computer Supported Cooperative Work in Design, (CSCWD
2008). CSCWD: 1122-1127.
[57] Bacurau, R., Dante, A., Duarte, L. and Ferreira, E. Experimental Investigation on
the Load Signature Parameters for Non-Intrusive Load Monitoring. Przegląd
Elektrotechniczny. 2015. 91(8): 86--90.
[58] Kahl, M., Ul Haq, A., Kriechbaumer, T. and Jacobsen, H.-A. (2017). A
Comprehensive Feature Study for Appliance Recognition on High Frequency
Energy Data. ACM 2017 Proceedings of the Eighth International Conference
on Future Energy Systems: ACM: 121-131.
[59] Tabatabaei, S. M., Dick, S. and Xu, W. Toward non-intrusive load monitoring via
multi-label classification. IEEE Transactions on Smart Grid. 2017. 8(1): 26-
40.
[60] Esa, N. F., Abdullah, M. P. and Hassan, M. Y. A review disaggregation method
in Non-intrusive Appliance Load Monitoring. Renewable and Sustainable
Energy Reviews. 2016. 66: 163-173.
[61] Somani, N. A. and Patel, Y. Zigbee: A low power wireless technology for
industrial applications. International Journal of Control Theory and Computer
Modelling (IJCTCM). 2012. 2: 27-33.
[62] Batista, N., Melicio, R., Matias, J. and Catalao, J. P. (2012). Zigbee wireless area
network for home automation and energy management: Field trials and
installation approaches. IEEE 2012 3rd International Conference and
Exhibition on Innovative Smart Grid Technologies (ISGT Europe). IEEE: 1-5.
 134

[63] Ridi, A., Gisler, C. and Hennebert, J. (2014). A survey on intrusive load
monitoring for appliance recognition. IEEE 2014 22nd International
Conference on Pattern Recognition (ICPR), IEEE: 3702-3707.
[64] Cottone, P., Gaglio, S., Re, G. L. and Ortolani, M. User activity recognition for
energy saving in smart homes. Pervasive and Mobile Computing. 2015. 16:
156-170.
[65] Ridi, A., Gisler, C. and Hennebert, J. (2015). User interaction event detection in
the context of appliance monitoring. IEEE 2015 International Conference on
Pervasive Computing and Communication Workshops (PerCom Workshops).
IEEE: 323-328.
[66] Villanueva, M. L. G., Dumlao, S. M. G. and Reyes, R. S. (2016). Appliance
recognition system for ILM using AGILASx—Dataset of common appliances
in the Philippines. IEEE 2016 Cloudification of the Internet of Things (CIoT):
IEEE: 1-5.
[67] Han, D.-M. and Lim, J.-H. Design and implementation of smart home energy
management systems based on zigbee. IEEE Transactions on Consumer
Electronics,. 2010. 56(3): 1417-1425.
[68] Song, S., Coit, D. W., Feng, Q. and Peng, H. Reliability analysis for multi-
component systems subject to multiple dependent competing failure processes.
IEEE Transactions on Reliability. 2014. 63(1): 331-345.
[69] Makonin, S. (2012) Approaches to Non-Intrusive Load Monitoring (NILM) in the
Home. PhD Thesis. Simon Fraser University.
[70] Aiad, M. and Lee, P. H. (2017). Non-intrusive monitoring of overlapping home
appliances using smart meter measurements. IEEE 2017 Power and Energy
Conference at Illinois (PECI) IEEE: 1-5.
[71] Srinivasan, G., Anandan, C., Jain, S. A. K., Ahmed, S. S. and Vijayaraghavan, V.
(2016). Low-cost non-intrusive device identification system. IEEE 2016
Dallas Circuits and Systems Conference (DCAS), IEEE: 1-4.
[72] Bouhouras, A. S., Milioudis, A. N. and Labridis, D. P. Development of distinct
load signatures for higher efficiency of NILM algorithms. Electric Power
Systems Research. 2014. 117: 163-171.
[73] Alcalá, J. M., Ureña, J., Hernández, Á. and Gualda, D. Assessing Human Activity
in Elderly People Using Non-Intrusive Load Monitoring. Sensors. 2017. 17(2):
351.
 135

[74] Alahmad, M., Zulfiqar, M. F., Hasna, H., Sharif, H., Sordiashie, E. and Aljuhaishi,
N. A. (2011). Green and Sustainable Technologies for the Built Environment
2011 Developments in E-systems Engineering (DeSE), DeSE: 521-526.
[75] Bergman, D., Jin, D., Juen, J., Tanaka, N., Gunter, C. and Wright, A. Nonintrusive
load-shed verification. IEEE Pervasive Computing. 2011. 10(1): 49-57.
[76] Anna Bertona, J.-P. V. t., Alexander Tamminga. (2014). Energy Management:
can Utilities seize the Opportunity ? A. T. Kearney.
[77] Aiad, M. and Lee, P. H. Non-intrusive load disaggregation with adaptive
estimations of devices main power effects and two-way interactions. Energy
and Buildings. 2016. 130: 131-139.
[78] Sudiharto, I., Anggriawan, D. O. and Tjahjono, A. Harmonic Load Identification
Based on Fast Fourier Transform and Levenberg Marquardt Backpropagation.
Journal of Theoretical and Applied Information Technology. 2017. 95(5):
1080.
[79] Wójcik, A., Winiecki, W. and Kowalik, R. Characterization of electrical
appliances based on their immitance. International Society for Optics and
Photonics. 2016 Photonics Applications in Astronomy, Communications,
Industry, and High-Energy Physics Experiments 2016: International Society
for Optics and Photonics. 10031: 1003121-1003128.
[80] Chung, T. M. and Daniyal, H. Arduino Based Power Meter Using Instantaneous
Power Calculation Method. ARPN Journal of Engineering and Applied
Sciences. 2006. 10(21): 9791-9795.
[81] Lin, Y.-H. and Tsai, M.-S. The Integration of a Genetic Programming-Based
Feature Optimizer With Fisher Criterion and Pattern Recognition Techniques
to Non-Intrusive Load Monitoring for Load Identification. International
Journal of Green Energy. 2015. 12(3): 279-290.
[82] Lam, H., Fung, G. and Lee, W. A novel method to construct taxonomy electrical
appliances based on load signatures. IEEE Transactions on Consumer
Electronics,. 2007. 53(2): 653-660.
[83] Pattem, S. (2012). Unsupervised Disaggregation for Non-intrusive Load
Monitoring 2012 11th International Conference on.Machine Learning and
Applications (ICMLA). 515-520.
 136

[84] Parson, O., Ghosh, S., Weal, M. and Rogers, A. An unsupervised training method
for non-intrusive appliance load monitoring. Artificial Intelligence. 2014. 217:
1-19.
[85] Po-An, C. and Ray, I. C. (2013). Unsupervised Adaptive Non-intrusive Load
Monitoring System. IEEE 2013 International Conference on Systems, Man,
and Cybernetics (SMC), IEEE: 3180-3185.
[86] Jia, R., Gao, Y. and Spanos, C. J. (2015). A fully unsupervised non-intrusive load
monitoring framework. IEEE 2015 International Conference on Smart Grid
Communications (SmartGridComm), 2015 IEEE: 872-878.
[87] Rahimi, S (2012). Usage Monitoring of Electrical Devices in a Smart Home.
Master Thesis. Carleton University Ottawa, Ontario, Canada, K1S 5B6.
[88] Ming, D., Meira, P. C. M., Wilsun, X. and Chung, C. Y. Non-Intrusive Signature
Extraction for Major Residential Loads. IEEE Transactions on Smart Grid.
2013. 4(3): 1421-1430.
[89] Figueiredo, M. B., de Almeida, A., Ribeiro, B. and Martins, A. Extracting features
from an electrical signal of a non-intrusive load monitoring system. Intelligent
Data Engineering and Automated Learning–IDEAL 2010. Springer. 2010:
210-217.
[90] Ming, H. H. (2013). Tracing of Lighting Electricity Consumption Using Non-
Intrusive Load Monitoring Method. Master Thesis. Universiti Teknologi
Malaysia, Skudai.
[91] Sanchez-Gomez, J. I., Morales-Velazquez, L., Osornio-Rios, R. A. and Guillen-
Garcia, E. (2016). Fuzzy c-means clustering for steady state events
classification of electrical signals. IEEE 2016 12th Congreso Internacional de
Ingeniería (CONIIN). IEEE: 1-5.
[92] Hassan, T., Javed, F. and Arshad, N. An Empirical Investigation of VI Trajectory
based Load Signatures for Non-Intrusive Load Monitoring. IEEE Transaction
on Smart Grid. 2014. 2( 2): 870-878.
[93] Iksan, N., Sembiring, J., Haryanto, N. and Supangkat, S. H. (2015). Appliances
identification method of non-intrusive load monitoring based on load signature
of VI trajectory. IEEE 2015 International Conference on Information
Technology Systems and Innovation (ICITSI). IEEE: 1-6.
[94] Bouhouras, A. S., Vaggos, M., Poulakis, N. and Christoforidis, G. C. (2016). Load
signatures enhancement via odd-order harmonic currents. IEEE 2016 16th
 137

International Conference on Environment and Electrical Engineering

(EEEIC), 2016 IEEE 16th International Conference on: IEEE: 1-6.
[95] Jung, D., Nguyen, H. H. and Yau, D. K. (2015). Tracking appliance usage
information using harmonic signature sensing. IEEE 2015 International
Conference on Smart Grid Communications (SmartGridComm), 2015 IEEE:
459-465.
[96] Temneanu, M. and Ardeleanu, A. Non-intrusive hybrid energy monitoring
system. Advanced Materials Research: 2014. 837: 495-499.
[97] Liu, B., Luan, W. and Yu, Y. Dynamic time warping based non-intrusive load
transient identification. Applied Energy. 2017. 195: 634-645.
[98] Duarte, C., Delmar, P., Barner, K. and Goossen, K. (2015). A signal acquisition
system for non-intrusive load monitoring of residential electrical loads based
on switching transient voltages. IEEE 2015 Power Systems Conference (PSC)
Clemson University. IEEE: 1-6.
[99] Norford, L. K. and Leeb, S. B. Non-intrusive electrical load monitoring in
commercial buildings based on steady-state and transient load-detection
algorithms. Energy and Buildings. 1996. 24(1): 51-64.
[100] Kim, H., Marwah, M., Arlitt, M. F., Lyon, G. and Han, J. (2011). Unsupervised
Disaggregation of Low Frequency Power Measurements. SIAM 2011 SDM:
SIAM: 747-758.
[101] Bouloutas, A., Hart, G. W. and Schwartz, M. Two extensions of the Viterbi
algorithm. IEEE Transactions on Information Theory,. 1991. 37(2): 430-436.
[102] Anderson, K. D., Berges, M. E., Ocneanu, A., Benitez, D. and Moura, J. M. F.
(2012). Event detection for Non Intrusive load monitoring 2012 IEEE 38th
Annual Conference on IEEE Industrial Electronics Society (IECON). IEEE:
3312-3317.
[103] Jiang, L., Luo, S. and Li, J. (2013). Automatic power load event detection and
appliance classification based on power harmonic features in nonintrusive
appliance load monitoring. IEEE 2013 8th IEEE Conference on Industrial
Electronics and Applications (ICIEA). IEEE: 1083-1088.
[104] Barsim, K. S. and Yang, B. Sequential Clustering-Based Event Detection for
Non-Intrusive Load Monitoring. Computer Science & Information
Technology. 2016. 10: 5121.
 138

[105] De Baets, L., Ruyssinck, J., Deschrijver, D. and Dhaene, T. (2016). Event
Detection in NILM using Cepstrum smoothing. 2016 3rd International
Workshop on Non-Intrusive Load Monitoring, 1-4.
[106] Yang, C. C., Soh, C. S. and Yap, V. V. A systematic approach to ON-OFF event
detection and clustering analysis of non-intrusive appliance load monitoring.
Frontiers in Energy. 2015. 9(2): 231-237.
[107] Kolter, J. Z. and Jaakkola, T. (2012). Approximate inference in additive factorial
hmms with application to energy disaggregation. 2012 International
conference on artificial intelligence and statistics, 1472-1482.
[108] Anderson, K. D. (2014). Non-Intrusive Load Monitoring: Disaggregation of
Energy by Unsupervised Power Consumption Clustering. PhD Thesis.
Carnegie Mellon University Pittsburgh, PA.
[109] Roos, J. G., Lane, I. E., Botha, E. C. and Hancke, G. P. (1994). Using neural
networks for non-intrusive monitoring of industrial electrical loads. IEEE 1994
Instrumentation and Measurement Technology Conference, (IMTC/94).
IEEE:. 1115-1118.
[110] Yu-Hsiu, L. and Men-Shen, T. (2011). Applications of hierarchical support
vector machines for identifying load operation in nonintrusive load monitoring
systems 2011 Congress on Intelligent Control and Automation (WCICA).
WCICA: 688-693.
[111] Hong-Tzer, Y., Hsueh-Hsien, C. and Ching-Lung, L. (2007). Design a Neural
Network for Features Selection in Non-intrusive Monitoring of Industrial
Electrical Loads 2007 11th International Conference on Computer Supported
Cooperative Work in Design, (CSCWD). 1022-1027.
[112] Figueiredo, M. B., De Almeida, A. and Ribeiro, B. An experimental study on
electrical signature identification of non-intrusive load monitoring (nilm)
systems. Adaptive and Natural Computing Algorithms: Springer. 2011. 6594:
31-40.
[113] Wild, B., Barsim, K. S. and Yang, B. (2015). A new unsupervised event detector
for non-intrusive load monitoring. IEEE 2015 Signal and Information
Processing (GlobalSIP), 2015 IEEE Global Conference on: IEEE, 73-77.
[114] Faustine, A., Mvungi, N. H., Kaijage, S. and Michael, K. A Survey on Non-
Intrusive Load Monitoring Methodies and Techniques for Energy
Disaggregation Problem. arXiv preprint arXiv. 2017. 1703: 00785.
 139

[115] Sadeghianpourhamami, N., Ruyssinck, J., Deschrijver, D., Dhaene, T. and

Develder, C. Comprehensive Feature Selection for Appliance Classification in
NILM. Energy and Buildings. 2017. 1703: 00785.
[116] Zeifman, M. and Roth, K. Nonintrusive appliance load monitoring: Review and
outlook. IEEE Transactions on Consumer Electronics, 2011. 57(1): 76-84.
[117] Liu, Y. and Chen, M. (2014). A review of nonintrusive load monitoring and its
application in commercial building. IEEE 2014 4th Annual International
Conference on Cyber Technology in Automation, Control, and Intelligent
Systems (CYBER), IEEE: 623-629.
[118] Klemenjak, C. and Goldsborough, P. Non-Intrusive Load Monitoring: A Review
and Outlook. arXiv preprint arXiv. 2016. 1610: 01191.
[119] Hoyo-Montaýo, J. A., Pereyda-Pierre, C. A., Tarýn-Fontes, J. M. and Leon-
Ortega, J. N. (2016). Overview of Non-Intrusive Load Monitoring: A way to
energy wise consumption. IEEE 2016 13th International Conference on Power
Electronics (CIEP). IEEE: 221-226.
[120] Barker, S., Kalra, S., Irwin, D. and Shenoy, P. (2014). NILM redux: The case
for emphasizing applications over accuracy NILM-2014 Workshop. NILM:1-4,
[121] Nation, J. C., Aboulian, A., Green, D., Lindahl, P., Donnal, J., Leeb, S. B.,
Bredariol, G. and Stevens, K. (2017). Nonintrusive monitoring for shipboard
fault detection. IEEE 2017 Sensors Applications Symposium (SAS). IEEE: 1-5.
[122] Campagna, M. M., Dinardo, G., Fabbiano, L. and Vacca, G. Fluid flow
measurements by means of vibration monitoring. Measurement Science and
Technology. 2015. 26(11): 115306.
[123] Schantz, C., Donnal, J., Sennett, B., Gillman, M., Muller, S. and Leeb, S. Water
Nonintrusive Load Monitoring. Sensors Journal, IEEE. 2015. 15(4): 2177-
2185.
[124] Borin, V. P., Barriquello, C. H. and Pinto, R. A. (2013). An improved technique
for load identification in residential buildings 2013 International Conference
on New Concepts in Smart Cities: Fostering Public and Private Alliances
(SmartMILE). 1-5.
[125] Muthuselvi, G. and Saravanan, B. Real Time Speech Recognition Based
Building Automation System. ARPN Journal of Engineering and Applied
Sciences. 2014. 9(12): 2831-2839.
 140

[126] Sawyer, R. L., Anderson, J. M., Foulks, E. L., Troxler, J. O. and Cox, R. W.
(2009). Creating low-cost energy-management systems for homes using non-
intrusive energy monitoring devices. IEEE 2009 Energy Conversion Congress
and Exposition (ECCE). IEEE: 3239-3246.
[127] Kucuk, S., Arslan, F., Bayrak, M. and Contreras, G. (2016). Load management
of industrial facilities electrical system using intelligent supervision, control
and monitoring systems. IEEE 2016 International Symposium on Networks,
Computers and Communications (ISNCC). IEEE: 1-6.
[128] Donnal, J. S. (2013). Home NILM: A Comprehensive Energy Monitoring
Toolkit. Master Thesis. Massachusetts Institute of Technology.
[129] Rasool, G., Ehsan, F. and Shahbaz, M. A systematic literature review on
electricity management systems. Renewable and Sustainable Energy Reviews.
2015. 49: 975-989.
[130] Landry, J. S. and Matthews, H. D. The global pyrogenic carbon cycle and its
impact on the level of atmospheric CO2 over past and future centuries. Global
change biology. 2017. 23(8):3205-3218.
[131] Lee, D. and Cheng, C. C. Energy savings by energy management systems: A
review. Renewable and Sustainable Energy Reviews. 2016. 56: 760-777.
[132] Schulze, M., Nehler, H., Ottosson, M. and Thollander, P. Energy management
in industry–a systematic review of previous findings and an integrative
conceptual framework. Journal of Cleaner Production. 2016. 112: 3692-3708.
[133] Rafsanjani, H. N., Ahn, C. R. and Alahmad, M. A review of approaches for
sensing, understanding, and improving occupancy-related energy-use
behaviors in commercial buildings. Energies. 2015. 8(10): 10996-11029.
[134] Lobaccaro, G., Carlucci, S. and Löfström, E. A Review of Systemsand
Technologies for Smart Homes and Smart Grids. Energies. 2016. 9(5): 1290-
1314.
[135] Lopez-Rodriguez, I., Hernandez-Tejera, M. and Lopez, A. L. Methods for the
management of distributed electricity networks using software agents and
market mechanisms: A survey. Electric Power Systems Research. 2016. 136:
362-369.
[136] Shen, J., Jiang, C. and Li, B. Controllable load management approaches in smart
grids. Energies. 2015. 8(10): 11187-11202.
 141

[137] Nanda, A. K. and Panigrahi, C. Review on smart home energy management.

International Journal of Ambient Energy. 2016. 37(5): 1-6.
[138] Du, Y., Du, L., Lu, B., Harley, R. and Habetler, T. (2010). A review of
identification and monitoring methods for electric loads in commercial and
residential buildings. IEEE 2010 Energy Conversion Congress and Exposition
(ECCE) IEEE: 4527-4533.
[139] Vega, A., Santamaria, F. and Rivas, E. Modeling for home electric energy
management: A review. Renewable and Sustainable Energy Reviews. 2015.
52: 948-959.
[140] Abedin, Z. U., Shahid, U., Mahmood, A., Qasim, U., Khan, Z. A. and Javaid, N.
(2015). Application of PSO for HEMS and ED in Smart Grid. IEEE 2015 Ninth
International Conference on Complex, Intelligent, and Software Intensive
Systems (CISIS). IEEE: 260-266.
[141] Kim, W. H., Lee, S. and Hwang, J. Real-time energy monitoring and controlling
system based on Zigbee sensor networks. Procedia Computer Science. 2011.
5: 794-797.
[142] Guvensan, M. A., Taysi, Z. C. and Melodia, T. Energy monitoring in residential
spaces with audio sensor nodes: TinyEARS. Ad Hoc Networks. 2013. 11(5):
1539-1555.
[143] Anastasi, G., Corucci, F. and Marcelloni, F. (2011). An intelligent system for
electrical energy management in buildings. IEEE 2011 11th International
Conference on Intelligent Systems Design and Applications (ISDA), 2011:
IEEE, 702-707.
[144] Lien, C.-H., Chen, H.-C., Bai, Y.-W. and Lin, M.-B. (2008). Power monitoring
and control for electric home appliances based on power line communication.
IEEE 2008 Instrumentation and Measurement Technology Conference
Proceedings. (IMTC). IEEE: 2179-2184.
[145] Magno, M., Polonelli, T., Benini, L. and Popovici, E. A Low Cost, Highly
Scalable Wireless Sensor Network Solution to Achieve Smart LED Light
Control for Green Buildings. Sensors Journal, IEEE. 2015. 15(5): 2963-2973.
[146] Pipattanasomporn, M., Kuzlu, M. and Rahman, S. An algorithm for intelligent
home energy management and demand response analysis. IEEE Transactions
on Smart Grid. 2012. 3(4): 2166-2173.
 142

[147] Kuzlu, M., Pipattanasomporn, M. and Rahman, S. Hardware demonstration of a

home energy management system for demand response applications. IEEE
Transactions on Smart Grid. 2012. 3(4): 1704-1711.
[148] Jiang, X., Van Ly, M., Taneja, J., Dutta, P. and Culler, D. (2009). Experiences
with a high-fidelity wireless building energy auditing network. ACM 2009 7th
ACM Conference on Embedded Networked Sensor Systems: ACM: 113-126.
[149] Ridi, A., Gisler, C. and Hennebert, J. (2013). Automatic identification of
electrical appliances using smart plugs. IEEE 2013 8th International Workshop
on Systems, Signal Processing and their Applications (WoSSPA). IEEE: 301-
305.
[150] Erol-Kantarci, M. and Mouftah, H. T. Wireless sensor networks for cost-efficient
residential energy management in the smart grid. IEEE Transactions on Smart
Grid,. 2011. 2(2): 314-325.
[151] Erol-Kantarci, M. and Mouftah, H. T. (2010). Using wireless sensor networks
for energy-aware homes in smart grids. IEEE 2010 Symposium on Computers
and Communications (ISCC). IEEE: 456-458.
[152] Erol-Kantarci, M. and Mouftah, H. T. (2010). TOU-aware energy management
and wireless sensor networks for reducing peak load in smart grids. IEEE 2010
72nd Vehicular Technology Conference Fall (VTC 2010-Fall). IEEE: 1-5.
[153] Tsai, M.-S. and Lin, Y.-H. Modern development of an adaptive non-intrusive
appliance load monitoring system in electricity energy conservation. Applied
Energy. 2012. 96: 55-73.
[154] Nunes, N. J., Pereira, L., Quintal, F. and Berges, M. (2011). Deploying and
evaluating the effectiveness of energy eco-feedback through a low-cost NILM
solution. 2011 6th International Conference on Persuasive Technology. 2-5.
[155] United States Environmental Protection Agency. Global Greenhouse Gas
Emissions Data. https://www3.epa.gov/climatechange/: Available 10th
December, 2014.
[156] Babu, A. V., Rao, C. U. and Tirupataiah, L. Energy Conservation, Green House
Gas Emission Reduction and Management Strategies of VFSTR University: A
Case Study. Journal of Advanced Research and Dynamical Conrol System
(JARDCS). 2017. 9(4): 21-27.
 143

[157] Sreekanth, K. Review on integrated strategies for energy policy planning and
evaluation of GHG mitigation alternatives. Renewable and Sustainable Energy
Reviews. 2016. 64: 837-850.
[158] Andries Hof, C. B., Angelica Mendoza Beltran, Michel den Elzen.
GREENHOUSE GAS EMISSION REDUCTION TARGETS FOR 2030
CONDITIONS FOR AN EU TARGET OF 40%.
[159] Calvillo, C., Sánchez-Miralles, A. and Villar, J. Energy management and
planning in smart cities. Renewable and Sustainable Energy Reviews. 2016.
55: 273-287.
[160] Li, W. J., Tan, X. and Tsang, D. H. (2015). Smart home energy management
systems based on non-intrusive load monitoring. IEEE 2015 2015 IEEE
International Conference on Smart Grid Communications (SmartGridComm):
IEEE: 885-890.
[161] Lai, Y.-X., Lai, C.-F., Huang, Y.-M. and Chao, H.-C. Multi-appliance
recognition system with hybrid SVM/GMM classifier in ubiquitous smart
home. Information Sciences. 2013. 230: 39-55.
[162] Garcia, F. C. C., Creayla, C. M. C. and Macabebe, E. Q. B. Development of an
Intelligent System for Smart Home Energy Disaggregation Using Stacked
Denoising Autoencoders. Procedia Computer Science. 2017. 105: 248-255.
6[163] Egarter, D. and Elmenreich, W. (2015). Autonomous Load Disaggregation
Approach based on Active Power Measurements. IEEE 2015 International
Conference on Pervasive Computing and Communication Workshops
(PerCom Workshops). IEEE: 417-422.
[164] Environment, U. S. Available and Emerging Technologies for Reducing
Greenhouse Gas Emissions from Coal-Fired Electric Generating Units. Sector
Policies and Programs Division Office of Air Quality Planning and Standards
U.S. Environmental Protection Agency Research Triangle Park, North
Carolina 27711. 2010.
[165] Niesten, E. and Alkemade, F. How is value created and captured in smart grids?
A review of the literature and an analysis of pilot projects. Renewable and
Sustainable Energy Reviews. 2016. 53: 629-638.
[166] Ruth, M., Pratt, A., Lunacek, M., Mittal, S., Wu, H. and Jones, W. (2015).
Effects of Home Energy Management Systems on Distribution Utilities and
 144

Feeders Under Various Market Structures. 2015 23rd International Conference

on Electricity Distribution (ICED).1-6 .
[167] Rodrigues, E., Godina, R., Pouresmaeil, E., Ferreira, J. and Catalão, J. Domestic
appliances energy optimization with model predictive control. Energy
Conversion and Management. 2017. 142: 402-413.
[168] Tauqeer, T., Ansari, M. A. and Hasan, A. Realization for low cost and energy
efficient ceiling fans in the developing countries. Renewable and Sustainable
Energy Reviews. 2017. 76: 193-201.
[169] Shen, H.-Y., Chen, Y.-C. and Hsu, C.-H. A Power Frequency Sensing Device
Using an Arduino Device and Zero-Crossing Algorithm and Its
Implementation on Android App. Sensors and Materials. 2017. 29(6): 741-
756.
[170] Jao, C. and Guo, X. PowerBox: The Safe AC Power Meter. Cit. Cornell.. 2013.
1(04): 1-22.
[171] Synergy Software. The KaleidaGraph Guide to Curve Fitting
www.synergy.com/Tools/curvefitting.pdf. Available 16th March 2016.
[172] Guo, C., Wang, M., Yang, L., Sun, Z., Zhang, Y. and Xu, J. A review of energy
consumption and saving in extra-long tunnel operation ventilation in China.
Renewable and Sustainable Energy Reviews. 2016. 53: 1558-1569.
[173] Youngwook, K., Seongbae, K., Rakkyung, K. and Sung-Kwan, J. (2014).
Electrical event identification technique for monitoring home appliance load
using load signatures. 2014 International Conference on Consumer Electronics
(ICCE). IEEE: 296-297.
[174] Klein, P., Merckle, J., Benyoucef, D. and Bier, T. (2013). Test bench and quality
measures for non-intrusive load monitoring algorithms. IEEE 2013 39th
Annual Conference of the Industrial Electronics Society (IECON 2013). IEEE:
5006-5011.
[175] Baraka, K., Ghobril, M., Malek, S., Kanj, R. and Kayssi, A. (2013). Low cost
arduino/android-based energy-efficient home automation system with smart
task scheduling. IEEE 2013 Fifth International Conference on Computational
Intelligence, Communication Systems and Networks (CICSyN). IEEE: 296-
301.
 145

[176] Adriansyah, A. and Dani, A. W. (2014). Design of small smart home system
based on Arduino. IEEE 2014 Electrical Power, Electronics, Communications,
Controls and Informatics Seminar (EECCIS). IEEE: 121-125.
[177] Marshall, M. N. Sampling for qualitative research. Family practice. 1996. 13(6):
522-526.
[178] Etikan, I., Musa, S. A. and Alkassim, R. S. Comparison of convenience sampling
and purposive sampling. American Journal of Theoretical and Applied
Statistics. 2016. 5(1): 1-4.
[179] Mekhilef, S., Saidur, R., Said, S., Hong, P. and Islam, M. Techno-economic
evaluation of energy efficiency measures in high rise residential buildings in
Malaysia. Clean Technologies and Environmental Policy. 2014. 16(1): 23-35.
[180] LLC, A. M. ACS712: Fully Integrated, Hall-Effect-Based Linear Current Sensor
IC with 2.1 kVRMS Voltage Isolation and a Low-Resistance Current
Conductor. http://www.allegromicro.com/.
[181] Srividyadevi, P., Pusphalatha, D. V. and Sharma, P. M. Measurement of Power
and Energy Using Arduino. Research Journal of Engineering Sciences. 2013.
2(10): 10-15.
[182] Global Logica Software Technologies. AC Voltage Sensor (ZMPT101B).
www.makemyproduct.com.: Available 22nd December, 2016.

[183] Ramos, P. M., Brás, N. B. and Serra, A. C. (2006). A new calibration method
for current and voltage sensors used in power quality measurements. IEEE
2006 Instrumentation and Measurement Technology Conference Proceedings
(IMTC): IEEE: 2283-2288.
[184] Chang, H.-H., Yang, H.-T. and Lin, C.-L. Load identification in neural networks
for a non-intrusive monitoring of industrial electrical loads. Computer
Supported Cooperative Work in Design IV: Springer. 664-674; 2008.
[185] Thacker, M. S. and Jhala, M. B. Domestic Load Identification with Single CT
Using Artificial Neural Network. Paripex - Indian Journal of Research. 2013.
2(4): 90-93.
[186] Fiorucci, E. The measurement of actual apparent power and actual reactive
power from the instantaneous power signals in single-phase and three-phase
systems. Electric Power Systems Research. 2015. 121: 227-242.
 146

[187] Bilil, H., Aniba, G. and Gharavi, H. Dynamic Appliances Scheduling in

Collaborative MicroGrids System. IEEE Transactions on Power Systems.
2017. 32(3): 2276-2287.
[188] Addabbo, T., Fort, A., Mugnaini, M. and Vignoli, V. Distributed UPS Control
Systems Reliability Analysis. Measurement. 2017. 110: 275-283.
[189] Aamir, M., Kalwar, K. A. and Mekhilef, S. Review: Uninterruptible Power
Supply (UPS) system. Renewable and Sustainable Energy Reviews. 2016. 58:
1395-1410.
[190] Bagheri, E., Tashakor, N., Farjah, E. and Ghanbari, T. (2017). Single-phase
multi-source on-line UPS with isolated battery charger. IEEE 2017 Power
Electronics, Drive Systems & Technologies Conference (PEDSTC). IEEE:
223-228.
[191] Patil, V. R., Biradar, V. I., Shreyas, R., Garg, P., Orosz, M. S. and Thirumalai,
N. Techno-economic comparison of solar organic Rankine cycle (ORC) and
photovoltaic (PV) systems with energy storage. Renewable Energy. 2017. 113:
1250-1260.
[192] Goel, S. and Sharma, R. Performance evaluation of stand alone, grid connected
and hybrid renewable energy systems for rural application: A comparative
review. Renewable and Sustainable Energy Reviews. 2017. 78: 1378-1389.
[193] Al-Shamani, A. N., Othman, M. Y. H., Mat, S., Ruslan, M., Abed, A. M. and
Sopian, K. Design & Sizing of Stand-alone Solar Power Systems A house Iraq.
Recent Advances in Renewable Energy Sources. 2013: 145-150.
[194] Charron, R. and Athienitis, A. Design and Optimization of Net Zero Energy
Solar Homes. ASHRAE transactions. 2006. 112(2).
[195] Svozil, D., Kvasnicka, V. and Pospichal, J. Introduction to multi-layer feed-
forward neural networks. Chemometrics and intelligent laboratory systems.
1997. 39(1): 43-62.
[196] David, N., Anozie, F. N., Ebuka, F. O. and Nzenweaku, S. A. Dessign of an
Arduino Based Wireless Power Meter. International Journal of Scientific &
Engineering Research. 2016. 7(9): 446-449.
[197] Miron-Alexe, V. Iot Power Consumption Monitoring System for Off-Grid
Households. University of Pitesti Sciencetific Bulletin : Electronics and
Computers Science. 2017. 17(2): 64-71.
 147

[198] Mnati, M. J., Van den Bossche, A. and Chisab, R. F. A smart voltage and current
monitoring system for three phase inverters using an android smartphone
application. Sensors. 2017. 17(4): 872.
[199] Mlakić, D., Nikolovski, S. N. and Alibašić, E. Designing Automatic Meter
Reading System Using Open Source Hardware and Software. International
Journal of Electrical and Computer Engineering (IJECE). 2017. 7(6): 3282-
3291.
[200] Rahman, K., Leman, A., Mubin, M. F., Yusof, M., Hariri, A. and Salleh, M.
(2017). Energy Consumption Analysis Based on Energy Efficiency Approach:
A Case of Suburban Area. EDP Sciences 2017 MATEC Web of Conferences:
MTEC:87.
[201] Winther, O., Lautrup, B. and Zhang, J.-B. Optimal learning in multilayer neural
networks. Physical Review E. 1997. 55(1): 836.
[202] Hansel, D., Mato, G. and Meunier, C. Memorization without generalization in a
multilayered neural network. EPL (Europhysics Letters). 1992. 20(5): 471.
 148

Appendix A

List of Journal Publications

1. I. Abubakar, S. N. Khalid, M. W. Mustafa, Hussain Shareef and M. Mustapha

Application of Load Monitoring in Appliances’ Energy Management – A Review.
Renewable and Sustainable Energy Reviews. 2017. 67: 235-245. ISI Journal IF
8.050.

2. I. Abubakar, S. N. Khalid, M. W. Mustafa, Hussain Shareef and M. Mustapha

Recent Approaches and Applications of Non-Intrusive Load Monitoring. ARPN
Journal of Engineering and Applied Sciences. 2016. 11(7): 4609-4618. SCOPUS.

3. Abubakar, I., Khalid, S., Mustafa, M., Shareef, H. and Mustapha, M. Calibration
of ZMPT101B Voltage Sensor Module Using Polynomial Regression for
Accurate Load Monitoring. ARPN Journal of Engineering and Applied Sciences.
2017. 12(4): 1076 – 1084. SCOPUS.

4. I. Abubakar, S. N Khalid, M. W. Mustafa, Hussain Shareef and M. Mustapha.

Residential Energy Consumption Management using Arduino Microcontroller.
Advanced Science Letters. SCOPUS.

5. I. Abubakar, S. N. Khalid, M. W. Mustafa, Hussain Shareef and M. Mustapha

Development of Arduino Microcontroller Based Non-Intrusive Appliances
Monitoring System using Artificial Neural Network. Advanced Science Letters.
SCOPUS.

List of Book
1.. Isiyaku Abubakar, Saifulnizam Bin Abd. Khalid, Mohd Wazir Bin Mustafa
“Electrical Appliance Load Monitoring and Its Applications”, 2016, LAP LAMBERT
Academic Publishing, ISBN: 978-3-330-00190-9.
 149

List of Conference Proceedings

(1) I. Abubakar, S. N. Khalid, M. W. Mustafa, Hussain Shareef and Mamunu

Mustapha. An Overview of Non-Intrusive Load Monitoring Methodologies.
Proceedings of the 2015 IEEE Conference on Energy Conversion
(CENCON). October 19 - 20: IEEE. 54-59. SCOPUS.

(2) I. Abubakar, S. N. Khalid, M. W. Mustafa, M. Mustapha and Shareef, H. A

New Calibration Method For ZMPT101B Voltage Sensor Module for
Accurate Load Monitoring. Proceedings of the 6th International Graduate
Conference on Engineering, Science & Humanities (IGCESH 2016).
Universiti Teknologi Malaysia 2016. P 112-115.

(3) I. Abubakar, S. N. Khalid, M. W. Mustafa, M. Mustapha and Shareef, H.

Energy Consumption Management of a Smart House using Arduino
Microcontroller. Proceedings of the 2nd International Conference on Science,
Engineering and Social Sciences 2016 (ICSESS2016) Universiti Teknologi
Malaysia, 2016. p 288 – 289.

(4) I. Abubakar, S. N. Khalid, M. W. Mustafa, M. Mustapha and Shareef, H.

Development of Arduino Microcontroller Based Non-Intrusive Appliances
Monitoring System using Artificial Neural Network. Proceedings of the
International Conference on Science, Engineering, Management and Social
Sciences 2016 (ICSEMSS2016) Universiti Teknologi Malaysia, 2016 p 108.
 150

Appendix B

Table of Equations for Electrical Quantities

S/N Method Metrics

1 AC Measurements 1 T 2
V  t  dt
T 0
Vrms 

1 T 2
I  t  dt
T 0
I rms 
t
V 2 t  t
Pave   dt   R  I 2  t  dt
0
R 0
t
ET   Pave dt Pave  t
0

2 Peak Method V peak

Vrms 
2
I peak
I rms 
2

3 Peak to Peak Method V peak  peak

Vrms 
2 2
I peak  peak
I rms 
2 2

4 Instantaneous Method 1 n 2
Vrms  Vi
n i 1
1 n 2
I rms   Ii
n i 1

P
Vi  Ii max
2
n
1
P Vi Ii
n i 1
Where P = real power, Vrms = rms Voltage, Irms = rms Current, ET = Total Energy,
Vpeak = Peak Voltage, Ipeak =Peak Current, Vpeak-peal = Peak to Peak Valtage and Ipeak-
peak = Peak to Peak Current.
 151

Appendix C

The System of Residential Appliances

 152

Appendix D

The Voltage Measurement Experimental Set Up

 Appendix E

The Monitoring and Management System

 154

Appendix F

The Load Identification for ANN-Arduino NILM based Appliances

Management and Control

Appliances Condition The Load Identification

Load 4 Load 3 Load 2 Load 1 L4 L3 L2 L1
OFF OFF OFF OFF 0 0 0 0
OFF OFF OFF ON 0 0 0 1
OFF OFF ON OFF 0 0 1 0
OFF OFF ON ON 0 0 1 1
OFF ON OFF OFF 0 1 0 0
OFF ON OFF ON 0 1 0 1
OFF ON ON OFF 0 1 1 0
OFF ON ON ON 0 1 1 1
ON OFF OFF OFF 1 0 0 0
ON OFF OFF ON 1 0 0 1
ON OFF ON OFF 1 0 1 0
ON OFF ON ON 1 0 1 1
ON ON OFF OFF 1 1 0 0
ON ON OFF ON 1 1 0 1
ON ON ON OFF 1 1 1 0
ON ON ON ON 1 1 1 1
 155

Appendix G

The Appliances Control of the ANN-Arduino NILM System