Light weight AI/ML algorithms for edge based load

disaggregation

A Stage II Dissertation submitted in partial fulfillment of

the requirements for the degree of

Master of Technology

by

Aashish Pawankumar Pnadia

(Roll No. 203196001)

Under the guidance of

Prof. Anupama Kowli

Electrical Engineering
INDIAN INSTITUTE OF TECHNOLOGY BOMBAY
July 2023
 Dissertation Approval

This dissertation entitled

Low Power Programmable CMOS Buck-Boost Regulator for Energy Harvesting Systems

by

Aashish Pandia
Roll No. 203196001

is approved for the degree of

Master of Technology in Electrical Engineering

..................................................... .....................................................
Prof. Anupama Kowli Prof. Mukul Chandorkar
(Supervisor) (Examiner)

..................................................... .....................................................
Prof. . Prof. .
(Chairman) (Examiner)

Date: July 14, 2023

Place: IIT Bombay

i
 Declaration

I declare that this written submission represents my ideas in my own words and where others’
ideas or words have been included, I have adequately cited and referenced the original sources.
I also declare that I have adhered to all principles of academic honesty and integrity and have
not misrepresented or fabricated or falsified any idea/data/fact/source in my submission. I un-
derstand that any violation of the above will be cause for disciplinary action by the Institute and
can also evoke penal action from the sources which have thus not been properly cited or from
whom proper permission has not been taken when needed.

..................................
Aashish Pandia
Roll No. 203196001

Date : July 14, 2023

ii
Acknowledgements

I would like to express my sincere gratitude to Prof. Anupama Kowli for her valuable guidance,
constant support and kindness, & motivation at each and every step of this work. I would like to
thank her for her insightful suggestions which helped me in smooth and successful completion
of the project.
I would like to thank Sust Labs. for the field data and technical insights they provided
during the course of the thesis.
I would also like to thank Technocraft center for Applied Artificial Intelligence (TCAAI)
for funding this research.
Lastly, I thank all the wonderful people who kept making me learn new things everyday.

Aashish Pawankumar Pandia

iii
Abstract

Domestic electricity usage accounts for around a quarter of the total electricity consumption
in the country. A detailed breakdown of energy use by different household appliances has
many applications – it can be used to design better tariffs, and/or identify opportunities for
demand response. It can also help consumers better understand their own consumption so as
to better manage the end-use loads and leverage available incentives such as time-of-use tariffs.
Appliance level monitoring can provide such breakdowns but it is expensive and increases the
sensor footprint. Prior work has established that the energy, current, and voltage signatures of
typical home use appliances exhibit distinct patterns. This thesis investigates the use of AI and
ML to leverage these distinctions and disaggregate appliance power signals at a single point-
of-connection. The majority of industry based solutions work on the cloud and dis-aggregates
energy demand on monthly timescales. This thesis will try to extend the product capabilities
to facilitate online disaggregation for better demand-side management, a first proof-of-concept
algorithm is provided here which is lightweight and allows the possiblity of AI/ML applications
on the Edge, this architecture tries to detect On-off and Intermittent loads (FSM type) so that
clean data can be made available for cloud inference as well as provide better system-wide
decision-making.

iv
Contents

Abstract iv

List of Figures vii

1 Introduction ix
1.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1.2 Proposed improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

2 House PoC Simulator xiv

2.1 Need for a simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
2.2 The Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
2.2.1 Model of different appliances . . . . . . . . . . . . . . . . . . . . . . xv
2.2.2 Inputs and outputs of simulator . . . . . . . . . . . . . . . . . . . . . xvi
2.2.3 Functional verification of the simulator . . . . . . . . . . . . . . . . . xvii
2.2.4 Adding noise to the simulator . . . . . . . . . . . . . . . . . . . . . . xvii

3 Load Disaggregation xx
3.1 Motivation for a house wise model . . . . . . . . . . . . . . . . . . . . . . . . xx
3.2 NILM: a state based approach . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii

4 Methodology xxvii
4.1 Data cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvii
4.2 Unsupervised Learning for FSM type Loads . . . . . . . . . . . . . . . . . . . xxix

5 Experiment & Results xxxiv

v
6 Future work xl
6.1 Limitations of current work . . . . . . . . . . . . . . . . . . . . . . . . . . . . xl
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xli

References xlii

vi
List of Figures

1.1 Appliances in typical household . . . . . . . . . . . . . . . . . . . . . . . . . x

1.2 Relation between appliance disaggregation possibilty and data frequency . . . xi
1.3 Architecture of edge NILM . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

2.1 Modelling of appliances in simulator . . . . . . . . . . . . . . . . . . . . . . . xv

2.2 Block Diagram of Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi
2.3 Appliance data recorded in a dataset . . . . . . . . . . . . . . . . . . . . . . . xviii
2.4 Appliance data simulated for validation . . . . . . . . . . . . . . . . . . . . . xix

3.1 PoC collected readings of real house . . . . . . . . . . . . . . . . . . . . . . . xxi

3.2 collected data of 3 appliances(Fridges) from different manufacturer . . . . . . . xxii
3.3 Subset sum algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv
3.4 dealing in ∆S . Courtesy: G.W. Hart, NALM, proceedings of IEEE 1992 . . . . xxiv
3.5 FSM loads as Time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv
3.6 State machine representation of loads . . . . . . . . . . . . . . . . . . . . . . xxv

4.1 decision threshold for piecewise linearisation . . . . . . . . . . . . . . . . . . xxviii

4.2 Raw data vs Clean data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxviii
4.3 State and transition in PQ-domain of a house . . . . . . . . . . . . . . . . . . xxix
4.4 Connected graph in limited time . . . . . . . . . . . . . . . . . . . . . . . . . xxx
4.5 Time domain Waveform subtraction . . . . . . . . . . . . . . . . . . . . . . . xxxi
4.6 Time between transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxi
4.7 More data more discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxii
4.8 Flowchart of Complete NILM . . . . . . . . . . . . . . . . . . . . . . . . . . xxxiii

5.1 simulated data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxv

vii
5.2 FSM loads in different scenarios example . . . . . . . . . . . . . . . . . . . . xxxvi
5.3 FSM extracted by NILM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxvii
5.4 Discovered FSM faulty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxvii
5.5 Discovered FSM correct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxviii

6.1 Variation in Power of an inverter AC . . . . . . . . . . . . . . . . . . . . . . . xli

viii
Chapter 1

Introduction

In India, domestic power use makes up about 24% of the nation’s overall electricity consump-
tion. Studies that break down the energy use of various home appliances offer significant in-
sights. For instance, while lighting accounts for around 20% of the demand, air conditioners,
water heaters, fans, and air coolers account for more than half of the overall electricity use. [1]
The inhabitants of an energy-aware house can benefit from these insights if they have proper
feedback such as the consumption of each appliance in the sum of the total billing.
The operating condition of appliances used in a home cannot be truly determined without
the proper monitoring system. In general terms, load monitoring is the process of identifying
and acquiring the load measurement in a power system. This load monitoring will determine
the consumption and appliances’ status, in order to comprehend the behavior of individual loads
in the whole system. [2]
Depending on the approach used for monitoring appliances, the load monitoring can be
classified into

Intrusive Load Monitoring

• A measurement device (sub-meter) for each appliance of interest.

• Expensive and hard to maintain, install and expand.

• Gives complete and accurate description of power uses by each appliance.

ix
Non - Intrusive Load Monitoring

• Data polled at a single point, generally at the meter connection.

• Cheaper and easier installation, since typically only one metering device is used.

• Computationally impossible to get the exact true operating condition.

Energy disaggregation, a quantitative method for calculating the power demand of individ-
ual appliances from a single meter that measures the total demand across numerous appliances,
is another name for NILM.
Introduced by Hart et al. [3] the basic problem statement is, given the aggregate power
consumption can we find the power of the individual appliance.
N
Pt = ∑ Pti + Et
i=1

where,
Pt = Total power consumed at tth time instant
Pt i = Power consumed by ith appliance at tth time instant
Et = measurement noise at tth time instant

Figure 1.1: Appliances in typical household

x
1.1 Literature review
Each appliance has steady-state features as well as transient features, depending on sampling
frequency these features can be used to distinguish them, transient features need high sampling
rate of the meter while steady-state features can be observed after one-tenth to one-third of a
second depending on the appliance, a low sampling rate of about 1Hz is sufficient to capture
these. Some non traditional features(like time of day, ambient temperature etc.) are also used
in current state-of-art techniques. [4] [5]
In current literature various techniques based on high frequency sampling are used which
are able to identify unique appliance loads using their transient waveforms[6]. Use of high
frequency sampling gives better resolution when similar types of loads are present because of
involvement of higher harmonics and use of Fourier analysis like techniques but both acquir-
ing the data and the space and computation time needed for analysis of data sampled at such
a high rate are difficult task in themselves. Low frequency sampling is enough to disaggregate
appliances in ∆I-∆φ domain the data acquisition is much simpler and cheaper compared to high
frequency sampling and for these reasons low frequency sampling is preferred.

Figure 1.2: Relation between appliance disaggregation possibilty and data frequency

Many Machine learning based techniques are in practice to get the task of load disaggrega-
tion done, the NILM can be proposed as a classification problem where we have to detect which

xi
 Sampling
Ref Fre- Features
ML Method Appliances of interest Remarks
No. quency Used
Used

LSTM, Denois-
kettle fridge washing
1
[7] 6 Hz P, Q ing Autoencoder, only major loads
machine dishwasher
Specific deep NN
training on PC
supervised dryer washing ma-
implementation
1
[8] 60 Hz P regression, mo- chine dishwasher
on Edge mobile
bileNet microwave cooker
device
1 Hz microwave coffee ma- classification of
[9] and 15 P, Q, φ KNN, SVM chines toaster incan- light load at high
KHz descent lamp LCD TV freq
use of real world
fan, lamp, hair dryer,
[10] 10 Hz VI k-NN Classifier voltage varia-
LED light
tions
hair dryer fridge
K-NN, Gaussian washing machine variety of loads
30 I, P, Q, V-I
[11] Naive Bayes, microwave AC CFL at very high fre-
KHz trajectory
SVM fan heater bulb laptop quency sampling
vacuum

appliances are active at a time t, classify ON and OFF, or it can be represented as a regression
problem in which one estimates how much each appliance contributes to the total percentage of
consumption. [12] Techniques based on supervised methods (like pattern recognition), unsu-
pervised clustering, [13] HMM, [14] CNNs, LSTMs and Deep learning are already in practice
to solve these problems of regression and classification. [15]
There are a few edge based solutions in the literature also, these edge devices use similar
ML techniques but detects a reduced number of appliances or their overall disaggregation ac-
curacy suffers either because of computational power limitation or lack of storage space. Event
based detection is commonly seen in such methods where inferencing only happens if a large

xii
change is detected in the features. Edge based NILM techniques use pruning methods to re-
duce the overall storage needed for rule set[16]. Moreover, powerful devices like RaspberryPi
or android mobile device are used or in-house developed complete system are used for load
inferencing[16][8][17].

Figure 1.3: Architecture of edge NILM

1.2 Proposed improvements

In the reviewed literature it is observed that small load appliances like CFLs, fans etc. whose
power ratings are not significant compared to larger loads like heating loads, fridges and ACs
are ignored from account, such loads are often treated as noise in the system when performing
the load disaggregation. While as mentioned before the lighting loads comprised almost 20%
of the household total power demand which is a substantial part hence it should not be ignored
as noise.
The state-of-art Machine learning models used for load disaggregation typically require a
lot of memory and processing power. As a result, the inference (disaggregation for a test house)
is carried out majorly on the cloud utilising powerful and effective GPUs. Such an architecture
where the smart meter data is sent from home to the cloud for inference has two disadvantages
one is higher data transmission and the other is the privacy of the user.[18] These problems
can be tackled by using an inference / dis-aggregator device situated at the end user location,
preferably on a smart meter itself. Such an edge device will cut both the need for higher data
transmission and will protect the privacy of the user [19] by sending selective data to the cloud
which is already processed and de-noised.
In the following sections, a light algorithm is proposed which separates the different
type(On-off, intermittent) of load appliances from the meter data and denoises the data for the
further inference of smaller loads at cloud, this light algorithm has potential to be implemented
on an edge device located at end user’s house.

xiii
Chapter 2

House PoC Simulator

2.1 Need for a simulator

For the purpose of the load disaggregation we need to have some kind of rule set which can be
used to differentiate the different appliances, we can use ∆P- ∆Q domain or the I - φ domain.
In order to generate this rule set we need to acquire labeled data of different appliances
from an actual house. This data not only should contain enough information about different
types of commonly used household loads such as fans, lights (tube or CFL), TV, fridge etc.,
but it should also contain the complete cycle of operation of an appliance in different use cases.
The quantities like active power, reactive power, magnitude of the current, phase of the current
(relative to voltage) must be known along with the operational state of appliance (ON/OFF) for
all the appliances separately in order to generate a disaggregation algorithm working to a fair
degree of correctness.
While the above task is doable, the amount of complications that are present in obtaining
such a data set having sufficient reliability is very tough as the measurements of each and every
appliance have to be recorded at all times, furthermore, it is not possible to remove all the
measurement noise. Even if when we have all the data from one house by some golden method
a different house having a different set of appliances would need us to repeat the whole process
of data acquisition.
In contrast to this we can use a simulator that can simulate a household system having an
arbitrary number of appliances of any kind. We will not be needing the setup of hardware and
can skip the data acquisition part altogether.

xiv
2.2 The Simulator
Motivated by the need, a simulator is designed to simulate the behaviour of a house as seen
from a single PoC, like a real-world house’s energy meter. In the model of this simulator,
just like a real house, the loads are connected parallel to each other and between the Mains
and the neutral terminal. When the switch corresponding to a load is connected the load is
considered in operating state (ON) and it will draw power according to its rating, else the load
is in disconnected state(OFF) no power is drawn and that branch is open circuited.

2.2.1 Model of different appliances

For simulations, different appliances need to be modeled, each appliance is modeled as a series
combination of resistance R and inductance X, further, the appliances having multiple operating
states in themselves (like a Fridge) are modeled as having separate R-X combinations for each
state which can be switched internally by a selection switch.

Figure 2.1: Modelling of appliances in simulator

xv
 The R and X values of different loads are obtained from ACS data set via the use of data-
driven methods. The ACS Data set contains 1-hour monitored electrical recording of appliances.
The data set contains recordings of the values of voltage, current, phase, active power, reactive
power, and frequency for a type of appliance. The R and X values are extracted from this data
set using formulas
V×V
S=
Z̄
Z̄ = R − ιX

where Z̄ is the conjugate of the impedance of appliance.

2.2.2 Inputs and outputs of simulator

Figure 2.2: Block Diagram of Simulator

The simulator takes the number of branches in the house connected in parallel between
Mains and the neutral lines, the type of load i.e. the R and X values of each branch and the time
series of the state of operation of each branch for a total time duration of ’T’ as inputs.
The simulator then simulates the electrical appliances in the house at a duration of 1 sec
interval, only the steady state behaviour is of importance, the transient behavior is ignored as
this behaviour of appliances settles well within one sec. and provides the total AC current
magnitude, phase AC current, total active power, and total reactive power consumed at t th time
instance, as observed at the PoC. The mathematical equations being simulated are as follows

S = V I∗

xvi
 N
I = ∑ Ii
i=1
where Ii is the complex current in branch i of the system.
The simulator also gives the value of branch currents,and power for all branches at all times for
further analysis.

2.2.3 Functional verification of the simulator

Intending to verify the correctness of the simulator, some nominal values for R and X are chosen
(for some type of household appliance) and the simulation are run, the resulting magnitude
of current, phase, active power, and reactive power values are checked against the available
datasets. The different combinations of appliances are tested and verified using the rigor based
analytical methods.

2.2.4 Adding noise to the simulator

The simulator works very close to ideality, but in real-world scenarios, there are inherent uncer-
tainties and variability in the data and measurements, incorporating a noise model can replicate
these uncertainties, making it more realistic and representative of the actual house being sim-
ulated. A noise model allows to compare the output of the simulator with real-world data or
measurements. The ruleset developed on such simulator are expected to work well on real
houses. As the algorithm thus generated have enhanced generalizability and robustness.

xvii
Figure 2.3: Appliance data recorded in a dataset

xviii
Figure 2.4: Appliance data simulated for validation

xix
Chapter 3

Load Disaggregation

3.1 Motivation for a house wise model

Same kind of appliances from different manufacturer exhibit separate electrical characteristic
and temporal changes. (e.g., fridge of different makes). Every house has its own set of appli-
ances and hence every house is unique. A general model of appliances will not fit in all the
houses. But the same house will have a highly corelated behaviour in terms of power consump-
tion from one day to other. The set of appliances do not change quite often, in fact most of the
significant appliances takes years to replace. Having model of one house on an edge device is
that sits on the target house gives better accuracy in terms of appliance disaggregation. We can
also try and make the device light as the scope of its job only includes one house.

xx
Figure 3.1: PoC collected readings of real house

xxi
Figure 3.2: collected data of 3 appliances(Fridges) from different manufacturer

xxii
3.2 NILM: a state based approach
The operational electrical condition of a house containing different appliances can be repre-
sented as a system which takes a finite state depending on the combination of appliances which
are ON in one of their own valid operating states. In essence, we transformed the ON combi-
nation of set of appliances to a unique number, the total number of states in this system will be
equal to 2N (considering only on-off conditions and ignoring the rest of the states in multi-state
appliances)
where N is the total number of appliances in the house.
A case where nothing is switched on is mapped to a state 0 while each appliance being on is
mapped to a state 2N -1. The I, φ , P, Q recorded at PoC of such a house on any given time instant
is representative of the combination of appliances which are operating at that particular time.
As the total electrical power is equal to the sum of the individual power components by each
appliance.
N
St = ∑ Xi Si + εt
i=1
where Si is the power contributed by ith appliance
Xi is the operating state of appliance i
Going from one state to another we can observe changes in one or more quantities, this means
that appliances operation of a house is now completely described by a system having some
states and the transition in between those states, if at any time we know the state of the system,
i.e., the appliance combination ON at that time, we can deduce the subsequent time states just
by observing the change in the elec. quantity’s values. Such knowledge of states and their
transition with time is equivalent to knowing which load(s) are on at that particular time instant
owing to the mapping of load status to state number.

St = ∆St + St−

where ∆St is the change observed at time = t

We can solve this by using the sub-set sum method, the problem with this method is that
either it gives one correct solution and terminates resulting in same answer each time even for
different combinations or it searches the whole subset space to get all the possible solutions,
which is very time consuming due to the fact that it searches the all possible subsets (powerset).

xxiii
 Figure 3.3: Subset sum algorithm

Some heuristics are needed to decreases the available search space.

We can make use of certain facts like between 2 consecutive time instances there can occur
only maximum one switching event, if a subset is satisfying the total power, then it persists in
that state until next change comes i.e., only one of the sub-set sum is correct.

Figure 3.4: dealing in ∆S . Courtesy: G.W. Hart, NALM, proceedings of IEEE 1992

If we match the up and down transactions in power and bind them together, we form the
model of an appliance.

Now the system can have multiple appliances and there are some which have more than
one internal state such appliances are termed as Finite state machine (FSM) type appliances,

xxiv
such appliances have more than one transition attributed to them. If we associate each observed
transition correctly to appliances, we will have a model for each appliance and consequently for
the house.

Figure 3.5: FSM loads as Time series

Figure 3.6: State machine representation of loads

Since each appliance consumes no power in off state, we can club the positive changes
with their equivalent negative counter parts and in so we will label the on off appliances. If those

xxv
changes are observed for more than a finite time and time between the matching transactions is
in some finite range.
Because of the nature of the problem, if more than one appliance of the same kind (i.e.
having same R and X value) are present in the house, each will behave exactly similarly when
observed from the PoC hence those 2 states are indistinguishable and it is impossible to de-
termine which one of the two appliances are exactly operating, this phenomenon gives rise to
classes of equivalent states.

xxvi
Chapter 4

Methodology

4.1 Data cleaning

The recorded data from a real house contains a lot of noise some of it is induced by the voltage
variation from the grid, some is introduced by the instruments and the limitations of them. In
order to get consistent and meaningful results from them we need to pre-process/clean them.
Non-ideal grid voltage and voltage swings results in power consumption as the Impedance of
an appliance is fixed, thus a voltage swing of 10% value will cause 2 times a change in power.
To mitigate the effects of this non idealilty we use voltage normalisation.

2202
Snormt = St
Vt2

Some errors in readings are introduced by the recording device itself maybe because of poor
quality instrument or maybe the hardware and software techniques employed to collect data
causes the introduction to such noise. In order to mitigate the effect of this noise we use a piece-
wise linearisation of the observed quantities. (On real house data it was found by experiments
that a boundary of decision (∆ P = 20 and ∆ Q = 15) is sufficient to remove further noise and
make data linear.

xxvii
Figure 4.1: decision threshold for piecewise linearisation

Figure 4.2: Raw data vs Clean data

xxviii
4.2 Unsupervised Learning for FSM type Loads
In order to do the unsupervised clustering, we need some features and metrics of similarity so
that the computing machine can learn, there are 2 orthogonal features that the smart meters
record namely active power(W) and reactive power (VAr) other than that higher features can be
derived from data set which include (but not limited to) the initial position and the final position
of the house after the completion of transaction, number of time a transaction is observed ,
time spend in a particular state etc. The 2 such metric that are used in this thesis report are the
connected graph of house in a limited time frame and the exact wave shape as a time series.

Figure 4.3: State and transition in PQ-domain of a house

The connected graph of a house in a limited time window represents the transactions and
the states encountered in that time interval. The FSM transactions can start from any initial state
but will follow the same pattern of the transitions(changes). In order to reduce the search space,
we use the time between transitions. The transition delta and counts are the metric of similarity
in this domain. As can be seen in two cases the operation starts from s3 but in other it starts
from s22, the transactions are similar but an additional change was observed in 1 case due to
another appliance.
The other metric of similarity that we use is the time domain wave form subtraction. As

xxix
 Figure 4.4: Connected graph in limited time

stated earlier any FSM load in house can start operating at electrical state ‘S’ of the house but
it will operate through its cycles, these cycles can vary in time a little. The subtraction of 2
such time-series instances give the measure of similarity between them. A zero-resulting wave
after wave subtraction means both time slice instances are equal in time. A constant difference
indicates that the waves are similar in shape but shifted with a constant bias, this constant bias
corresponds to the switching of one of the appliances or switching of one of the states of any
FSM appliance.Thus, the piecewise flatness is the metric of similarity between 2 time series in
this domain.

An FSM based load is going to give transition much more frequently and consistently
than a human can. So, when an FSM load is operating the features like time between events
provide us with the time windows where frequent transitions can be seen. The length of these
windows will vary, now we group the obtained slices of time with similar time duration together.

xxx
 Figure 4.5: Time domain Waveform subtraction

Figure 4.6: Time between transitions

We then proceed to make the connected graph and do the time domain wave subtraction of
each of the members of one group and calculate the similarity between them. To aid to this the

xxxi
connected graph will give us the information about duration of state and the transition seen in
that time slice while amount of piecewise flatness of the waveform subtraction will give us the
metric of similarity in time domain between any two given time slices, if the two-time domain
slices are similar to each other, we assign the differential changes observed in those windows to
one appliance regardless of the state position at the start of the time slice window.

Figure 4.7: More data more discovery

In this manner the more data we have the more similarities can be found and hence more
appliances can be discovered with high accuracy. We gradually assign the observed deltas to
appliances, once all FSM are done, we assign remaining delta to on off devices.

Single connected graph is a graph where the edges are only present between those vertex
which transited to and from a state to another, such types of graphs differentiates between FSM
and On-off kind of loads.

xxxii
Figure 4.8: Flowchart of Complete NILM

xxxiii
Chapter 5

Experiment & Results

For the purpose of the proof of this thesis a house with 3 FSM loads (one slow changing and
other rapidly changing), 1 cyclic always on, and 4 on-off is simulated and the data is clustered
unsupervised to find the FSMs. The proposed method was shown the simulated data of a house
and with training on 3 days and asked to find FSM loads. Once completed the disaggregation
is done on newly generated schedule of the house from simulator having same appliances.
Different kinds of loads are included to that the algorithm can be tested thoroughly.
Moreover, for the first implementation only on off type of loads are used. The value of
impedance for each type of appliance is provided to the simulator from the data-set.

xxxiv
Figure 5.1: simulated data

xxxv
 As there are 8 appliances there will be 256 states, not all of them are distinct, moreover,
some transitions are not possible because of the constrain that any load can operate in only one
of its internal states. Once the states and transitions are known, we bind them to n number of
apps (to be decided by user), then we can perform the load disaggregation of appliance on any
running schedule. To test this stratergy a random schedule of operation of appliances of this
house is generated in which each appliance is turned on and off accordingly. Since we need
to know the initial state of the system for convenience and easier debugging it is set to state 0
where all appliances are in OFF state.

Figure 5.2: FSM loads in different scenarios example

The results shows that a highly active FSM load is easily disaggregated because the pres-
ence of more transition of same type makes it easy to recognise similarities in the connected
graph domain. Similarly, an FSM with a few dilation and expansion in time of one of the run-
ning states of an appliance is also recognised easily because of the presence of the long sections
of piece wise linearity in the time waveform subtracted plot domain.
On the other hand, a slowly transiting FSM load whose time to cycle between states is
near to some other cyclic always on type of appliance is difficult to find, typically because the
ontime of always on appliance almost always overlaps with the slow changing FSM, in such
cases more than one FSM type behaviour is registered and attributed to same appliance. The
value of this threshold decides the final operating state of the system, as we can see in Figure
4.2 a low value of threshold results in correct classification of none of the operating states of the
system and the accuracy of the algorithm is very poor(nearly 0), as the threshold is increased the
accuracy improves and at some threshold the algorithm becomes 100 % accurate, as also can be

xxxvi
 Figure 5.3: FSM extracted by NILM

Figure 5.4: Discovered FSM faulty

seen from the figure the value of threshold for 100 % correct classification is highly dependent
on the presence of load, this is because the ∆values of electrical quantities are dependent on
the initial state of the system, turning a small inductive load (say a fan) on or off will produce
a different set of ∆values when a large resistive only load was already operating than only
inductive load was operating. Thus changing the operating condition of a particular load will

xxxvii
 Figure 5.5: Discovered FSM correct

result in separate ∆values each time depending on the initial state.

Choosing the right value of threshold for both connected graph and time series wave sub-
traction is crucial as a value too small will defeat the purpose of having the threshold in the first
place, whereas choosing a value too large will result in incorrect classification of states.
Applying the above steps to the entirety of the simulation gives the time series of operating
states, which when mapped back to the appliance level information gives us the information of
the operating schedule of loads of the house, which was previously unknown and randomly
generated.
The energy disaggregation can now be done to see how much energy each appliance is
consuming as the power consumed in ON state is known to us along with the time of duration
for which each appliance is ON.

xxxviii
 Table 5.1: Golden running schedule

Golden Truth Energy Consumed (W.s) ∆S associated (P,Q)

appliance 1 1521750 +/-(50,95)
appliance 2 48300 +/-(120,-300)
appliance 3 135000 +/-(170,-260), +/-(350,-105), +/-(160,155)
appliance 4 276000 +/-(80,-75)
appliance 5 695160 +/-(60,-20)
appliance 6 843750 +/-(1250,0)
appliance 7 150000 +/-(250,-75), +/-(150,220)
appliance 8 7200 +/-(60,20)

Total 3677160 –

Table 5.2: NILM results schedule

Pridicted/clustered Energy Consumed (W.s) ∆S associated (P,Q)

appliance a 1500690 +/-(50,95)
appliance b 47655 +/-(120,-300)
appliance c 112950 +/-(170,-260), +/-(350,-105), +/-(160,155)
appliance d 272800 +/-(80,-75)
appliance e 694980 +/-(60,-20)
appliance f 843750 +/-(1250,0)
appliance g 150000 +/-(250,-75), +/-(150,220)
appliance h 2340 +/-(60,20)
appliance i 37770 +/-(50,95), (350,-105)
appliance j 7200 +/-(60,20) , +/-(50,95)

Total 3667805 –
Error 9355 –

xxxix
Chapter 6

Future work

6.1 Limitations of current work

1. Scalability:

The present work is a state based approach in which the total number of states are a
function of the number of appliances present in the house(ON or OFF), addition of one
appliance double the max number of states, this exponential increase in the state blows
out of proportion very very quickly as appliances increase. Also, a multi-state appliance
having ’M’ distinct states increases the number of states by ’M’ fold instead of doubling
them.

2. Effect of ambient conditions:

The variable type of loads whose R and X values are a function of external factors (am-
bient temperature etc.) can have a range of R and X values and it may happen that one
of the appliance have the R and X values comparable to the change in R and X of other
appliance in that case the small load will be interpreted as operational while it was never
turned ON.

3. Merging of similar FSM loads:

The method used is of unsupervised learning and it may happen that 2 or more appliances
are mixed into one appliance because they both were operated simultaneously and have
similar duration of operation.

xl
 Figure 6.1: Variation in Power of an inverter AC

6.2 Future work

1. Better Heuristics: field trials on edge devices, monitor how the environmental noise
affects the algorithm, augmentation steps to remedy the effects.

2. Field trials: field trials on edge devices, monitor how the environmental noise affects
the algorithm, augmentation steps to remedy the effects and complete implementation on
edge device.

3. Others: Development of methodology to disaggregate variable power appliance.

xli
References
[1] C. Aditya and A. Sreenivas, “Towards an understanding of residential electricity consump-
tion in india,” Building Research & Information, vol. 47, no. 1, pp. 75–90, 2019.

[2] e. a. bubakar, I., “Application of load monitoring in appliances energy management a

review.” Renew. Sustain. Energy Rev, vol. 67, no. 1, pp. 235—-245, 2017.

[3] G. W. Hart, “Nonintrusive appliance load monitoring,” Proceedings of IEEE, vol. 80,
no. 12, pp. 1870—-1891, 1992.

[4] D. V. Christoforos Nalmpantis, “Machine learning approaches for non-intrusive load mon-
itoring: from qualitative to quantitative comparation,” 2018.

[5] M. Zeifman, “Disaggregation of home energy display data using probabilistic approach,”
IEEE Trans Consum Electron, vol. 58, no. 1, pp. 23–31, 2012.

[6] e. a. Laughman, C., “Power signature analysis,” IEEE Power Energy Mag, vol. 1, no. 2,
pp. 56—-63, 2003.

[7] S. M. Jack Kelly and I. V. Bajic´, “Neural nilm: Deep neural networks applied to energy
disaggregation,” Conference on Acoustics, Speech, and Signal Processing, vol. 44, 2019.

[8] M. B. Shamim AHMED, “Edge computed nilm: a phone-based implementation using

mobilenet compressed by tensorflow lite,” Proceedings of NILM, 2020.

[9] M. B. Figueiredo, “Contributions to electrical energy disaggregation in a smart home,”

PhD thesis, 2013.

[10] Y.-H. L. S.-K. H. M.-S. Tsai, “Study on the influence of voltage variations for non-
intrusive load identifications,” International Power Electronics Conference, vol. 44, 2018.

[11] J. G. E. C. K. S. G. M. Bergés, “A feasibility study of automated plug-load identification

from high-frequency measurements,” IEEE Global Conference on Signal and Information
Processing, vol. 44, 2015.

xlii
[12] D. G.-U. Daniel Precioso, “Nilm as a regression versus classification problem: The im-
portance of thresholding,” 2020.

[13] e. a. Gon¸calves, H., “Unsupervised disaggregation of appliances using aggregated con-

sumption data.” Environmental Engineering, 2011.

[14] M. M.-A. M. L. G. H. J. Kim, H., “Unsupervised disaggregation of low frequency power

measurements,” roceedings of the 2011 SIAM International Conference on Data Mining,
pp. 747 –– 758, 2011.

[15] A. L. M. e. a. Jorge Revuelta Herrero, “Non intrusive load monitoring (nilm): A state of
the art,” Advances in Intelligent Systems and Computing, 2018.

[16] D. B. Enrico Tabanelli and L. Benini, “A feature reduction strategy for enabling
lightweight non-intrusive load monitoring on edge devices,” IEEE, pp. 805 – 810, 2020.

[17] S. M. Alon Harell and I. V. Bajic´, “Wavenilm: A causal neural network for power disag-
gregation from the complex power signal,” Conference on Acoustics, Speech, and Signal
Processing, vol. 44, 2019.

[18] L. W. Z. G. H. Cao, S. Liu and X. Du, “Achieving differential privacy against non-intrusive
load monitoring in smart grid: A fog computing approach,” Concurrency and Computa-
tion: Practice and Experience, vol. 31, no. 22, 2019.

[19] Q. Z. Y. L. W. Shi, J. Cao and L. Xu, “Edge computing: Vision and challenges,” IEEE
Internet of Things, vol. 3, no. 5, pp. 637–646, 2016.

xliii