Article

Smart Distribution Boards (Smart DB), Non-Intrusive

Load Monitoring (NILM) for Load Device Appliance
Signature Identification and Smart Sockets for Grid
Demand Management
See Gim Kerk 1, Naveed UL Hassan 2 and Chau Yuen 1,*
1 Engineering Product Development Pillar, Singapore University of Technology and Design,
Singapore 487372, Singapore; [email protected]
2 Department of Electrical Engineering, Syed Babar Ali School of Science and Engineering, Lahore University of

Management Sciences (LUMS), Sector U, DHA, Lahore 54792, Pakistan; [email protected]

* Correspondence: [email protected]

Received: 13 April 2020; Accepted: 17 May 2020; Published: 20 May 2020

Abstract: Traditionally, the choices to balance the grid and meet its peaking power needs are by
installing more spinning reserves or perform load shedding when it becomes too much. This problem
becomes worse as more intermittent renewable energy resources are installed, forming a substantial
amount of total capacity. Advancements in Energy Storage System (ESS) provides the utility new
ways to balance the grid and to meet its peak demand by storing un-used off peak energy for peak
usage. Large sized ESS—mega watt (MW) level—are installed by different utilities at their
substations to provide the high speed grid stabilization to balance the grid to avoid installing more
capacity or triggering any current load shedding schemes. However, such large sized ESS systems
and their required inverters are costly to install, require much space and their efficacy could also be
limited due to network fault current limits and impedances. In this paper, we propose a novel
approach and trial for 3000+ homes in Singapore of achieving a large capacity of demand
management by developing a smart distribution board (DB) in each home with the high speed
metering sensors (>6 kHz sampling rate) and non-intrusive load monitoring (NILM) algorithm, that
can assist home users to perform the load/appliance profile identification with daily usage patterns
and allow targeted load interruption using the smart sockets/plugs provided. By allowing load
shedding at device or appliance level, while knowing their usage profile and preferences, this can
allow such an approach to become part of a new voluntary interruptible load management system
(ILMS) that requires little user intervention, while minimizing disruption to them, allowing ease of
mass participation and thus achieving the intended MW demand management capacities for the
grid. This allows for a more cost effective way to better balance the grid without the need for
generation capacity growth, large ESS investment while improving the way to perform load
shedding without disruptions to entire districts. Simply, home users can now know and participate
with the grid in interruptible load (IL) schemes to target specific home appliance, such as water
heaters or air conditioning, allowing interruptions during certain times of the day, instead of the
entire house, albeit with the right incentives. This allows utilities to achieve MW capacity load
shedding with millions of appliances with their preferences, and most importantly, with minimal
disruptions to their consumers quality of life. In our paper, we will also consider coupling a small
sized Home Energy Storage System (HESS) to amplify the demand management capacity. The
proposed approach does not require any infrastructure or wiring changes and is highly scalable.
Simulation results demonstrate the effectiveness of the NILM algorithm and achieving high capacity
grid demand management. This approach of taking user preferences for appliance level load
shedding was developed from the results of a survey of 500 households that indicates >95%
participation if they were able to control their choices, possibly allowing this design to be the most
successful demand management program than any large ESS solution for the utility. The proposed

Sensors 2020, 20, 2900; doi:10.3390/s20102900 www.mdpi.com/journal/sensors

Sensors 2020, 20, 2900 2 of 21

system has the ability to operate in centralized as part of a larger Energy Management System (EMS)
Supervisory Control And Data Acquisition (SCADA) that decide what to dispatch as well as in
autonomous modes making it simpler to manage than any MW level large ESS setup. With the
availability of high-speed sampling at the DB level, it can rely on EMS SCADA dispatch or when
disconnected, rely on the decaying of the grid frequency measured at the metering point in the Smart
DB. Our simulation results demonstrate the effectiveness of our proposed approach for fast grid
balancing.

Keywords: Home-ESS; smart distribution board; smart sockets; grid stability

1. Introduction
In a drive towards enhancing grid stability/optimization and to meet growing energy demand,
many approaches are being adopted from distributed generation, advanced SCADA controls, load
shedding and most recently, demand management and Energy Storage Systems (ESS) [1–13]. For grid
stability, power generation and load demand have to be matched at all the times. As the peak load
demand on the grid is increased, traditional utilities either increase the overall generation capacity or
perform high speed under-frequency load shedding using substation relays [5–9]. Both these
traditional methods have their own disadvantages. Increasing the overall generation capacity to meet
peak load demand which only lasts up to few hours in today’s context, is highly expensive, wasteful,
and environmentally unsustainable [8,9]. On the other hand, load shedding at the substation level can
be hugely disruptive and inconvenient as it affects thousands of homes and businesses and may also
adversely affect the economy if it becomes too frequent [6–9]. Load shedding is not a voluntary choice
and usually adopted as a last resort to avoid a cascading fault and will attract huge penalties by the
energy regulator. Set to shed population centers first, followed by commercial, industrial, financial
institution loads and finally critical loads such as hospitals, military and government facilities, each
outage event will incur millions in penalties, huge losses to the economy which it serves and that does
not include the damages to the end customers, if they are proven negligent. Last but not least, the
reliability and quality of electricity supply can lead to hundreds of millions of losses in the economy
and possibly lives. Hence, any methods that avoid mass load shedding and in enhance the quality of
supply is always preferred.
In recent years, there has been a great interest in renewable energy generation to meet the
growing load demands in an environmentally friendly manner [7,8]. The growth in renewables, while
encouraging, further exacerbate grid stability issues and increase system faults, with many pushing for
more spinning reserve capacities to negate any volatility, which negates the benefits of this type
sustainable energy source. With the advent of low cost communication technologies, high speed
metering with more powerful edge computing capacity and Internet-of-Things (IOT) connectivity
offered by telecoms, demand response management also becomes a tool to manage supply-demand
balance [7–9]. In demand response management, with the help of different pricing signals or frequency
change, loads can be tripped off instantly to meet the generation capacity via voluntary load
interruptions. However, without large scale user participations, demand response management in this
approach has a limited impact in resolving the grid stability issues [10], utilities have to resort to the
unfavorable and traditional method of under-frequency load tripping when it fails to balance at the
substation level to maintain stability, affecting tens of thousands of households with outages [10,11].
The emergence of utility level ESS provides another alternative approach to address the same
concerns on peak demand growth, renewable intermittency and system balancing during generation
or capacity loss [9–13]. In this approach, large sized ESS in MW are installed at the substation levels. A
typical MW level ESS installation at a substation, that can be at the distribution or transmission level
(66 kV and above), consists of the battery, which can be lithium ion, flow batteries and any other High
Power or High Energy types. High Power types are categorized by ability to supply high power in
MW but for very short durations, typically from a few seconds to less than 30 min for frequency
Sensors 2020, 20, 2900 3 of 21

stabilization and High Energy types for less than MW levels and for duration longer than 4 h for
demand management.
Next, it is the Power Conversion System (PCS) inverts the DC power to AC and finally, the
Battery Management System (BMS) that manages the battery to avoid over-discharge or any thermal
runaway. A SCADA EMS system is usually linked to manage the grid and perform load dispatches for
the discharging and charging of the battery, considerations are usually on the network loading for
discharging and for charging, in an open electricity market such as Singapore, spot electricity prices. A
typical load at the 22 KV station in Singapore can vary between 7–12 MVA, about half of a typical 25
MVA transformer. To make any impact, the ESS must be at least 20–30% of the rated load or 1.4 to 3.6
MW power and a backup duration for at least 2 h. Depending on the technology, the space
requirements can vary substantially and cost of setup can be substantial assuming there is availability
of space for installation, a premium in all urban cities such as Singapore. While it is not uncommon to
see many on-going trials with MW batteries [13,14], the widespread adoption of this solution cannot be
expected unless the cost reduces to a level that can justify the investments in capital, land and even
re-laying the cabling infrastructure to manage the higher current system. A typical installation design
is as shown below at Figure 1, where the battery with its attached PCS are connected to the 22 KV bus
and the Supervisory Control And Data Acquisition (SCADA) Remote Terminal Unit (RTU), which
links it to the Remote Energy Management System (EMS).

Figure 1. Typical Setup of a substation level Energy Storage System (ESS).

In this paper, leveraging the advancements and cost reductions in high speed CT metering, the
availability of low cost home gateways with high speed computing power, smart sockets with built-in
breakers and low cost communication technologies, we propose a utility managed Home Energy
Management System (HEMS) that provides appliance level smart load identification, curtailment and
together with a small Home Energy Storage System (HESS), to assist in managing grid supply and
demand mismatch. This approach relies on smart load identification at the house level based on a
Non-Intrusive Load Monitoring (NILM) technique, which is a load signature identification application
that resides in the smart gateway located in the new generation smart distribution box in the house. It
automatically identifies the appliance when it is turned on and where it was turned on at each smart
socket devices. This allows the system to provide users with the information of the usage profile with
respect to Time of Use (TOU) and cost attributed to each house appliance and makes it possible for the
system to control the load.
At the same time, if we combine the HESS, we can discharge it to remove some of the loadings
from the grid without impacting the grid, rotating across the households to maintain the “curtailed”
capacity. The options of demand management, managing user preferences, reducing user
inconveniences, while providing power for critical power to critical applications such as elderly and
home patient care, creates the foundation for this approach to enhance grid stability via home demand
management.
Sensors 2020, 20, 2900 4 of 21

A recent survey, prior to this actual project, was conducted for 500 homes for 2 to 3 bedrooms
public housing flats (between 900–1200 square feet) to measure demand management and IL
opportunities in performing appliance level load shedding. This type of public flat represents 80–90%
of all housing types in Singapore. The measurements were done by only using an incoming single
phase Time of Use (TOU) clamp on meter and smart plugs with built in metering to measure the total
and selected aircon and heater loads for at least 7 days each. The following Figure 2 shows the
breakdown in usage of the aircon and heater loads during the peak and off-peak periods.

Figure 2. Survey summary results of 500 households.

The results show that an average of around 44.65% of total usage of each household is for aircon
and heater loads. It is also found that around 80% of the households in the study group have air
conditioning loads and 100% of them have water heaters, forming a substantial capacity that offers
possible demand management capacity. A sub-breakdown for peak between 10 am to 2 pm was
shown to show the capacity of usage that can offer the most useful interval for peak demand
management.
A post survey shows that substantial percentage of 70.05% is willing to experience some
inconvenience for up to 2 h of interruption a day (of no less than 30 min per interruption) if an
acceptable discount is given off their overall daily usage. The figure increases to 95% if we allow them
to choose what and when the load will be interrupted even for the same discount. From this survey,
we can easily see that we will garner a higher demand management capacity if we have a solution that
is intelligent enough to further identify smaller loads, other than the larger heater and aircon loads, to
allow home user preferences to be further considered for minimizing their inconveniences. This forms
the basis of developing the concept of smart DB to identify more load appliances (>100W) for demand
management purposes. With this design and proof of willingness and availability of excess capacities
in each home, it allows the possibility of trading excess and voluntary load shedding in a secondary
market where industrial and commercial buildings, could possibly buy additional demand capacities
for a short period of time by offering an incentives for home consumers to inhibit or interrupt, without
too many inconveniences to their overall quality of supply. Currently, all building contract capacities
can only be changed annually, a very rigid and ineffective process that does not allow for a more
efficient way to respond to actual production need. This will allow home consumers, with
interruptible loads, who are willing to offset that increase by agreeing to turn off their loads based on a
price that can accept, thus giving building owners a price to pay for taking it, and this is akin to sharing
some of the penalties in certain fractions, which would otherwise be fully incurred to the utility. This
will lead to a more efficient marketplace, an incentive program that encourage demand management
operators to provide the investments of ESS in each home with no space constraints, high setup costs
and when combine with IL, provide a much higher capacity per house for incentives to be shared
between building owners, home users and the operators. This gives the utility more choices to provide
different way to balance the grid, avoid catastrophic load shedding at the substations, allowing for a
more stable grid, without much grid investment, that leads to lower costs and higher reliability to all
stakeholders.
Our contributions in this paper includes:
• Developing a smart distribution board (DB) with high speed CT metering sensor at >100
samples/cycle, NILM algorithm inside the smart DB home gateway can sense and disaggregate
each appliance and smart sockets (SSO) with breaker controls that can turn off the load.
• Proposing a simple methodology of home load shedding system by aggregating a large number
of identifiable and interruptible individual appliance load across households for utility demand
Sensors 2020, 20, 2900 5 of 21

management.
• Developing an HEMS that allows for easy home participation, takes in users preferences for IL
while being part of a larger load shedding system, that can receive direct dispatch for grid
stabilization. potentially can lead to higher IL participation
• Integrating an HESS to the smart DB, to show that by doing so can substantially increase the
distributed ESS capacity throughout the utility, lowering overall cost per MW installed as
compared to utility substation ESS, while working in tandem with the HEMS to minimize load
interruptions, user inconvenience and providing Uninterruptible Power Supply (UPS) supply for
critical applications and services
• Designing this utility managed HEMS as part of an SCADA EMS and ILMS that can also operate
autonomously and will not require real time updates and dispatch simply by sensing the grid
frequency at its point of coupling to decide the amount of load shedding or energy discharge
from the HESS.
The rest of the paper is organized as follows. In Section II we discuss our proposed Smart DB
system model, Home-ESS modeling is discussed in Section III, HEMS including the design
requirements of smart DB, NILM and smart sockets are discussed in Section IV, centralized and
autonomous modes for HEMS aggregation employed for grid stabilization are described in Section V,
a case study is presented in Section VI, while the paper is concluded in Section VII.

2. Smart DB System Model

In this system, we have designed the smart DB system based on a distributed nature, which will
make individual connections to the main SCADA EMS using a device management platform.
Connectivity is provided via the resident home router, which will provide the Internet connection. The
Smart DB will utilize CT sensors to perform the high-speed data capturing or metering to all the
outgoing circuits and perform the NILM algorithm to identify the load/appliance. Using broadband
powerline (BPL) or any other available communications as the sensor network, the Smart DB will then
communicate with the smart sockets with built in breakers that can be used to curtailed the selected
load.
Each HEMS will be connected to the EMS SCADA cloud and users can understand their
individual appliance usage and specify their load shed preferences using a smart phone application.
The Smart DB, with the high speed CT, and a Raspi 3, acting as a smart home gateway, will measure
the following parameters:
• system frequency
• active and reactive power
• Total and individual appliance energy usage including the aircon, heater and other loads (>150
W)
The HEMS, installed inside the Raspi 3 gateway, comes with a quad core CPU and on-board
16GB storage that runs the NILM algorithm for load identification, manage the curtailment of all loads
based on user’s preferences. Users interact with the HEMS and Cloud system via their smart phone
apps.
For grid stabilization, HEMS can operate in two modes (i) centralized mode and (ii) autonomous
mode. In centralized operation, HEMS accepts and executes dispatch commands from the EMS
SCADA and when it is not available, in autonomous mode, based on previously configured limits.
With these limits, the local algorithm makes decisions, to dispatch Home-ESS or trip the appliance
loads, based on the rise or decay of the high-speed sampling of the system frequency measured at the
Smart DB to balance the grid much like an under-frequency relay. To ensure uniformity across all the
HEMS, the EMS SCADA provides frequency thresholds for Home-ESS self-discharge and load
curtailment to each load priorities as set by the users. This approach greatly enhances the stability of
the approach by allowing it to operate with or without constant connection to the EMS SCADA.
Sensors 2020, 20, 2900 6 of 21

The Smart DB, are equipped with different communications options other than BPL to include
Zwave, BLE and WIFI to allow it to integrate with other smart sensors. The connectivity from the
gateway to all the sensors are shown in Figure 3 below.

Figure 3. Connectivity of sensors within the house.

As mentioned, the proposed system design of HEMS with HESS requires little or no
modifications of the home wirings unlike large substation MW ESS installations, i.e., a million homes
with 10 kWh storage with an average of 2 kW of IL at any time is 10,000 MWh with 2000MW of loads
that can tripped for frequency stabilization and a 10,000 MWh high power ESS for demand
management, which is very expensive to implement, with no modifications to the grid.

3. Home-ESS Modeling
Many types of batteries have touted long lifespan and cycle life, but they either suffer from low
round trip efficiency (flow batteries) or have high thermal runaway risks (lithium) [15]. At the grid
transmission and distribution system levels, ESS options are very limited. Large batteries tend to be
more expensive, require much space, require MW level inverters, and also have high maintenance
costs [16–19]. In short, other than battery costs, non-battery cost considerations for large ESS are
• ESS equipment
• Space and network infrastructure
• Engineering and labor
• System setup, management and maintenance
• Part Replacements
On the other hand, Home-ESS batteries are smaller in size, are relatively low costs, with low
power (< 5 kW) inverters that are readily available in the market. However, several different types of
battery technologies (lead-acid, lithium ion or saltwater battery) exist and they have their own
advantages/disadvantages when used as Home-ESS.
A direct cost comparison study for our HESS approach against a typical setup at 22 KV system
must first consider that High Power ESS systems have lower costs per kW and higher costs per KWh
while High Energy Types have lower costs per kWh but higher costs per kW. Hence, the cost of a
battery system can be calculated by looking at the power capacity or energy capacity costs, depending
on the intended application. Consider the reference pricings given by U.S. Energy Information
Administration (EIA) EIA 860, Annual Electric Generator Report for 2013–2016. Prices may have
differed by now but it is still a good reference.
Consider a typical High Power ESS 10 MW 4MWh setup (which is considered short duration
since it is for <30 mins), the power capacity cost is around USD1000/kW while using the energy
capacity cost for short duration will be relatively higher at USD 2000–3400 /kWh. Conversely, a 4 MW,
10 MWh high Energy ESS would have cost between USD 400–800 /kWh for the energy capacity cost
between USD1500–4200 /kW. The estimated costs based on the weighted average cost would be
around > 10 MUSD for the High Power while the High Energy type will cost > 8 MUSD just for the ESS
Sensors 2020, 20, 2900 7 of 21

equipment. The installation and related systems cost can be easily set at USD 3–4 MUSD for either
solution. Land and transmission network cable upgrade can be so prohibitive that it will cost several
times the cost for large ESS setup, even if the space is available. Except for 10+ sites out of 11,000+
substations, rollout of such solutions is not possible due to land constraints in Singapore. This situation
is probably similar in any high population load centers where capacity is most needed.
Consider the HESS alternative where the average power requirement of a typical household of
around 4–6 kW (housing flats in Singapore, where 80% of the people lives in), a cost effective home
ESS can be achieved by installing a 2 kW Home-ESS in our design with 4 h of backup or 8 kWh. It is
important to select a safe battery design that does not create fire risk to the homes. The following
design shows how a typical Home-ESS can be wired and operated in two modes, i.e., Grid Only and
Via Home ESS [5]. This design allows the UPS to be bypassed using the static switch for maintenance as
shown below in Figure 4.

Figure 4. Typical Uninterrupted Power Supply (UPS) setup.

The HESS comes with a built-in battery management system that can monitor and control the
charging/discharging via the HEMS and the upstream EMS SCADA. It also records the number of
cycles that the battery is activated, battery state of charge, charging/discharging status of battery,
operating mode of battery, etc. All these parameters allow us to estimate the remaining battery lifetime
and its continued utility in supporting grid stabilization efforts as with any large ESS.
To compare costs, we set up 1200 homes of 2 kW, 8 kWh HESS to make an equivalent High
Energy type ESS of 4.8 MW, 9.6 MWh when combined, costing around 8 MUSD in total or around
USD5000 each. This is similar in equipment costs except that there is little setup cost, no space needs
for the utility and no other related costs such as cabling upgrades. By combining the IL loads with our
HEMS, it can be used for high-speed frequency stabilization or demand management, similar to a
High Power type ESS at much lower costs with no issue in getting parts for replacements. Simply, it is
more cost effective to setup HEMS with IL, with many distributed storage areas in each home to
achieve the capacity needed for frequency stabilization or demand management in a simple way that
no single or large number of ESS can provide[20,21].

4. HEMS Modeling
In this section, we describe the design requirements of our HEMS. HEMS comprises smart DB
and smart sockets and also includes NILM framework. Safety concerns and minimum changes in
existing wiring infrastructure are our primary design concerns. Moreover, it is also critical that the
smart DB, smart sockets and NILM are collectively able to identify loads that can participate in the
required high speed load shedding needed to balance the grid in safe way.

4.1. Smart DB
Sensors 2020, 20, 2900 8 of 21

Smart DB consists of 16 high-speed individual metering CTs for the incoming and each of the
out-going circuits to record the voltage current, energy, power and frequency of all circuits. Most
importantly, it has a high speed waveform sampling module to record at 6 Khz (6,000 samples per
second) active and reactive power values to form a load signature be used to identify each home
appliance, their on–off timings and with the signature, breakdown of energy usage in terms of each
appliance. The layout of the smart DB, NILM MUX controller and the HEMS Raspi Gateway are
shown below with a schematic of the layout of the CT sensing board and the schematic diagram of the
Current Transformer (CT) board in shown below in Figure 5 below

Figure 5. Smart DB layout and Schematic of 16 CT sensor boards.

From the actual DB layout and the schematic, the CT are spaced to avoid interferences and
examples of load 1, 2 and 3 where different circuits from the main AC bus are wired through each CT
to measure the loads. With the high speed CT sampling, we can then provide the waveform to the
MUX and ARM controllers to handle the waveform processing. This is where the NILM load
desegregation application resides to identify the appliance and load usage before sending it to the
Raspi gateway. The following Figure 6 shows the components of the MUX and ARM controllers.
Sensors 2020, 20, 2900 9 of 21

Figure 6. MUX and ARM Controller and the dataflow diagram from the CT to the ARM Controller and
Raspi.

This is the controller card that interfaces with the CT sensing boards and the Raspi gateway.
Every CT channel correlates with an individual metering chip for individual sensing and waveform
capturing.

4.2. NILM
Using the information captured by smart DB with the help of various CTs, the objective of NILM
algorithms is to identify the type of appliance. We use signature matching technique for appliance
identification. In this technique, the captured profile (current, voltage, active power, reactive power,
etc.) is matched with already stored profiles in a database and then the closest match is found. This
database of signatures can also be gradually built, if not already available.
As new appliances are identified their signatures are added to the database. The user is also made
aware about energy consumption of various appliances thus allowing them to make demand
management choices [22]. The following are the test scenarios for learning and identification process of
the NILM program in our laboratory. The loads tested included typical home appliances such as a
toaster, coffee machines, kettle, vacuum cleaners etc. (loads higher than 150 W).
The following Figure 7 is a typical signature of a high-speed waveform capture of a vacuum
cleaner.

Figure 7. Power (W) signature of a vacuum cleaner.

The high-speed waveform captured of the active power (blue line) and reactive power (orange
line) against the time based sampling interval. As the sampling rate is 6000 samples, the graph is
plotted against the sampling interval number captured. For example: at the 4600 sampling interval, the
actual time is 4600 × 1/6000 = 0.76 s.
Sensors 2020, 20, 2900 10 of 21

To demonstrate the difference in such waveforms captured, the following Figure 8 and Figure 9
are different profiles or signatures of different household loads.

Figure 8. Power (W) signature of an air purifier.

Figure 9. Power (W) signature of a washing machine.

In order to create a library of signatures, we need to understand that the NILM data collected is
based on a time series data for any classification purpose.

where N is the number of types of equipment, Xn,t is the power of equipment n at time t, Xt is the
aggregated signal at time t. Kn is the appliance different state, μk is the power consumed and Zt,k ∈ {0, 1}.
The formula in [23] forms the basis for preparing the input file for its signature profile of each
appliance. Previously, [24] adopted an approach where every household appliance’s change in power
represents a step change in its power values and every appliance stage is a markov state. Using an
iterative Hidden Markov Model (HMM), it attempts to model the household’s specific appliances by
using its known state appliance’s state and its power demand to create an identity matric. Consider the
following Markov Chain Example shown in Figure 10:
Sensors 2020, 20, 2900 11 of 21

Figure 10. A Markov chain example

The Markov chain shown indicates that the output is observable, but its states are hidden and just
a probabilistic function of the output, making the model a non-observable process normally used for
time series processing. However, there are limitations, as complexity increases when there are large
numbers of appliances and it does not consider possible simultaneous transitions from different
individual appliances, thus limiting its adoption.
In order to address such issue, Factorial Hidden Markov Models (FHMM) based on the first order
Markov process was adopted first by [25], that simply considers every observed output such as an
additive function of many hidden states. Simply, this model treats independent HMM that are
produced in parallel for each monitored device. The aggregated data seen as a whole are taken as the
observable output. This is done by combining all the HMM models generated by the appliance in the
hidden states and the appliance identification algorithm models which sequence of Markov hidden
states could have produced that output to identify it [26,27]. It also suffers from
• Time complexity when the number of appliances, hidden states and the length of observation
sequence need to be extended that reduces the classification performance of the model.
• Difficulty in classifying the multistate appliance requiring a long-range pattern resulting the
existing models that are based on the first-order Markov process could not be trained efficiently.
• Error in classifying the appliances having similar power consumption.
In this paper, we just adopt the FHMM method and a high-speed 6 kHz sampling rate CT and a
separate MUX and Arm controller to address the known issues.
First, the increased computing power provided by the separate controller allows for it to handle
more appliances without affecting its classification performances. Secondly, by being able to capture
highly accurate (Class 0.5) accuracy values at 6000 samples in 1 s, it is able to produce a distinct power
change signature of each load, avoiding a need for long range pattern, However, it does have limited
accuracy when the appliances have similar power consumptions and types of loads such as motor
loads that are found in many appliances.
The system can trigger a learning mode where the appliance is explicitly trained by getting the
end user to identify the load for the first time by triggering the training mode using the phone app to
the controller. Once the training mode is on, the user will be asked to turn on and off the load at least 3
times to capture multiple signatures of the same appliance for identification purposes, based on the
FHMM. During the normal operations, the signature will be used to compare against the absolute
output previously recorded. This method of direct comparison is accurate within a tolerance of 5% of
error is easy, fast and makes the adoption simpler for each household at the start. However, it does not
cater for long term appliance pattern change due to appliance degradation, which some adopted using
recurrent neural network as part of a deep learning process adopted to address this issue [28]. As of
now, the system can trigger the learning mode to update the outputs and adjust to the widening error
if needed. More research is needed to know the frequency and period needed for re-training. The data
flow diagram is shown in Figure 11.
There is a self-learning program that will learn any new appliance or signature that was
previously not known. At each of the number shown in Figure 11 are the set points used by the NILM
Sensors 2020, 20, 2900 12 of 21

algorithm to form the profile of each appliance uses changes in active and reactive power. Repeated
profiles of the same appliance for up to 3 times within 30 days will trigger to be presented to the user
to ask for the specific appliance name.

Figure 11. Data Flow of Non-Intrusive Load Monitoring (NILM) Program.

In both processes, the identification process is always run first and checked against a known
database stored in the local raspi and then the cloud, which is regularly updated. If it is unknown, it
goes back to the learning process as a new load. With the appliance identified and their on/off status,
the program will add the usage from the moment the appliance starts and stop to the same device
appliance. Over time, the accumulated and interval energy usage of each appliance is computed.
The assumption we have is that the load signature is generally constant and continuously
triggering the training mode allows it to update the program to update as the devices aged.
Test Scenarios:
1. Learning and Detection of Toaster (800 watt), Kettle (2300 watt), and Vacuum Cleaner (1200
watt).
2. (Sequence 1) Detection of Toaster, Kettle, and Vacuum Cleaner during Subsequent ON/OFF.
3. (Sequence 2) Detection of Toaster, Kettle, and Vacuum Cleaner during Subsequent ON/OFF.
In Figure 12, the NILM is in training mode and attempts to learn a new appliance signature as
shown in item 3 through 11, while at item 12, it fetches a new Appliance ID from the cloud central
database for the load learned. At item 7, it learns after at least 3 previous signatures to learn a New
Appliance and at item 16, it updates the local DB and at item 18 to 21, it is able to identify the new
Appliance ID 666 being switch on and off. The following table shows the simulations of appliances
against pretrained modules of NILM of each load and their error levels.
Sensors 2020, 20, 2900 13 of 21

Figure 12. NILM learns new signatures of load appliances.

Many types of typical home appliances are tested and generally, appliances with heater or
resistive types of loads are more accurate as compared to those that uses power electronics or switch
mode power supply. For loads that are less than 150 W, there are usually difficult to identify as it does
not produce significant signature against the background noises generated by all the appliances in
operations. The testing was done by turning one appliance on/off at each time.
For loads exceeding 150 W, the accuracy of 1000 tests are known to be almost 95%. It must be
qualified that the aircon and water heater are not included because a dedicated circuit was designed to
handle it. With NILM, users can understand their usage behavior and set their preferences for load
tripping. For example, a certain user may record the following priority for the shedding of loads in his
home.
1. Air conditioner—Priority 1
2. Water Heater—Priority 2
3. Kettle—Priority 3
4. Fan—Priority 4
5. Lights—Priority 5
Generally, incentives are higher for higher priority loads. For example, by setting the air con at
priority 1, the discounts are higher for all consumption by the air con and the breakdown can be
measured by the dedicated circuit and verified by the NILM. This makes it attractive for voluntary
participation of interruptible loads (ILs) in our proposed scheme. However, at the same time, utility
retain the right to shed specific types of loads, if policy allows, and more than the preferred load list of
a user to avoid resorting to the under-frequency load shedding affecting whole regions. Nonetheless,
by knowing what is critical and what is not as deemed by the consumers with the help of NILM, the
depth of load shedding can be done with minimum inconveniences and with wider choices [20,21].
In Figure 13, it shows the high speed identification offered by the approach where all the different
appliances are already pretrained and from item 60 to 69, it is able to identify the operations of the
appliances being on and off. This is critical as any IL will only make sense to the system if it is being
turned on during the load shed operations. Note that the identification is done using the power at the
x-axis and the sampling interval where every interval is 1/6000 s.
Sensors 2020, 20, 2900 14 of 21

Figure 13. The load identification of different loads learned by the NILM program.

Mass quantities of appliances can aggregate using this approach by summing all the appliances
under different priorities controlled by the EMS SCADA by issuing load shed commands to the
HEMS. A successful Interruptible Load Management System (ILMS) can only be possible if it allows
easy participation and when activated, minimize any inconvenience to the end users.

4.3. Smart Sockets (SSO):

Smart sockets are used to provide breaker control to individual loads when the demand
management decision is made. With this smart SSO, it makes it possible to curtail specific appliances.
By using BPL, the breaking of all socket loads can be done within 1–2 s. A typical smart socket with
built-in relay and BPL is shown below in Figure 14.

Figure 14. Smart Sockets with built-in temperature sensors. Relay and broadband powerline (BPL)
communications.

With the help of a smartphone app, the user can specify their load priorities and tripping intervals
to allow the EMS SCADA to compute the shedding capacity across the grid at different times.
However, actual monitoring through NILM is important to ensure that the load indicated by the user
is actually available in his home and operating behind a smart socket at that interval to allow it to
actually participate. Therefore, accurate estimation of how much load is offered by the user and how
Sensors 2020, 20, 2900 15 of 21

much is actually online when grid stabilization is required are important to know the actual load
shedding amount.
The appliance load shedding commands from EMS SCADA to HEMS can be communicated
using internet or loaded into the smart DB against the system frequency [9,12] for under-frequency
tripping. These modes are discussed in the next section.

5. HEMS Aggregation Modes for Grid Stabilization

We propose a two-pronged control approach for energy management.

5.1. Centralized Approach

EMS SCADA lists and aggregates all the Home-ESS capacity depending on the current state of
charge of each Home-ESS. It also aggregates the load of all the appliances, which are indicated by
participating HEMS for load shedding according to the appliance priorities and the indicated duration.
This aggregation allows the EMS SCADA to know the total available load shedding capacity. Further
information about local weather conditions, total load demand, SCADA generation data are also
collected. With this information, ESS SCADA can easily know the difference between the generation
capacity and forecasted load demand and the requirement of additional support from aggregated
Home-ESS. To minimize user inconvenience, EMS SCADA first discharges enough Home-ESS to
achieve stability. If further support is required, it would then start tripping appliances according to the
indicated priority list.

5.2. Autonomous Approach

Autonomous approach further simplifies the grid stabilization process. This novel approach,
when there is loss of communications, operates based on the decay detection of local supply frequency
and user settings of the load priorities for demand management simplifies the management of large
number of distributed ESS. In this mode, the frequency threshold values and the action list is
published by the EMS SCADA and is broadcast to all the HEMS. Once the HEMS have obtained this
list, it uses the home network as the primary network to allow the EMS SCADA to control load
curtailment at each home. This autonomous approach allows certain robustness in case of
communication loss or for simplicity in management by having a self-autonomous mode of operation.
The settings provided by EMS SCADA are locally adjusted by each HEMS according to the state of
charge of its Home-ESS and its own appliance priorities.
An example setting list could be the following:
49.5 Hz—Discharge of the Home-ESS
49.4 Hz—Tripping of Priority 1 loads
49.3 Hz—Tripping of the Priority 2 loads
49.2 Hz—until 48.4 Hz—Tripping of the Priority 10 loads
48.0 Hz—Substation Relay.
Note that priority level 1 means the first load to be shed and many levels can be set.
Our two-pronged planning and control (centralized or autonomous modes) strategy allows the
utility a novel way to avoid substation level load shedding that indiscriminately affects every
households and offers a better way to manage demand growth, improve network stability and offer a
better way to affect user usage behavior. Currently in Singapore, the Interruptible Load Scheme
available is only offered to industrial customers with each interruptible load coming with a min of 100
kW. Their incentives are peg to the prices offered for the spinning reserve markets based on primary
reserve (8 s to trip), secondary reserves (10 min to trip) and contingency reserves (30 min to trip). The
reasonably fast response time of our HEMS makes our approach suitable for the more demanding
primary, secondary spinning reserve markets, and for those with high-speed communications or
utilizing our Smart DB under-frequency tripping, as part of the overall load shedding scheme for
frequency stabilization, which currently is not priced as this strategy has never been adopted. This
potentially create a more vibrant interruptible load schemes by allowing end consumers to participate
Sensors 2020, 20, 2900 16 of 21

as interruptible loads, by aggregating many different home loads with their preferences, as part of the
spinning reserve markets [22].
The inclusion of Home-ESS in our proposed scheme also provides a huge capacity already readily
available for demand management with several added benefits, e.g., absorbing excess and buffer solar
intermittency and allowing a more widespread solar adoption with minimal impact to the grid. The
demand response potential of our scheme is also huge. This is shown in our earlier survey for 500
homes that consumers are willing to participate as long as they can set their preferences of when and
what they want to participate, although the capacity at home during the industrial peak timing
between 10 am to 2 pm is lower compared to the other hours. A simple calculation can be easily shown
by adding 2 kW Home-ESS capacity with possible load interruption of up to 1 kW of loads in each
home giving a total capacity of 3 kW of demand management for up to 2 h. With 1 million homes, a
distributed ESS capacity of 3000 MW with 6000 MWh for instantaneous dispatch can easily reduce up
to 45% of Singapore peak usage of 6600 MW for 2 h or staggered across all users reducing the impact
and average duration of appliance disruption to each household. In many countries with insufficient
capacity, it can assist in reducing investments in generation capacities leadings to huge economic and
environment benefits by delaying any such investments, if required at all.

5.3. Security of This System

The system adopted are highly secure and adopts an end to end security design, as shown in
Figure 15, by adopting:
• PKI Infrastructure CA and PKI servers based on RSA2048 encryption on the data and Transport
Layer Security TLS communication between the gateway and device data collection layer.
• Each gateway consists of a secure element (eSE) to store the signed SSL certificates in each
gateway, which is tamperproof
• Use of session keys for every transaction to upload data
• SHA-256 hashing for all information.
• Reviewed by Cybersecurity Agency of Singapore (CSA) and the customer internal security
department.
• All sensors also adopt the same TLS architecture

Figure 15. Schematic of the security system.

From Figure 16 shown, all communications are secured using TLS communication and all the
devices are provisioned via an internal process where the SSL are signed and loaded into the eSE
before it is issued to the device gateway to be assembled. All the keys are stored at the HSM provided
at the backend. We have provided a security controller to manage all the keys with the PKI server,
which is used to manage the expiration and revocation of the keys, when needed. Through this design,
we only allow devices that have our signed certificates to connect, encrypt all data that passes between
Sensors 2020, 20, 2900 17 of 21

the front gateway and backend based on the session key generated. For ease of management, we set
the expiry of the keys with a 10-year limit in time for the hardware renewal.

Figure 16. TLS Architecture of the smart DB.

For user privacy, the signature profile of the appliance is stored locally and in the cloud. We only
collect the energy data of each household, subject to their agreement for using the system. No users’
personal information was collected and only the unit address against the data was collected.

6. Case Study
A trial is done based on the potential of setting up smart home for 3000+ HDB households with
Home-ESS, smart DB, NILM and 10 smart sockets/smart plugs in each home. Please note that there are
1.38 million households in Singapore and 3096 households only represent 0.21% of total households.
The purpose of this trial is to demonstrate that even with a tiny participation of households, we can
easily aggregate MW capacity and with the two pronged frequency stabilization mechanism, we can
easily cut off selective load according to user preferences, provide a huge capacity for demand
management and avoid load shedding at the substation levels.
By considering a typical loads and capacity, the following approach can easily obtain thousands
of MW of ESS capacity with a much larger capacity that we can obtain for demand management via
our HEMS system. In Table 1 we give a typical system size and the amount of load that can be cut
according to user priority, assuming 100% Home-ESS state of charge.

Table 1. Simulations of NILM against known appliances.

Types of loads No. of NILM Simulations No. Error cases % of Error

Above 150W

Central AC for large hooms 1,000 298 29.8%

Room AC/Split Type 1,000 321 32.1%
AC Heat Pump 1,000 13 1.3%
Kettle/Electric heater 1,000 35 3.5%
Dishwasher 1,000 328 32.8%
Freezer 1,000 16 1.6%
Refrigerator 1,000 12 1.2%
Second refrigerator 1,000 15 1.5%
Hot water heater 1,000 18 1.8%
Tankless Hot Water Heater 1,000 30 3.0%
Heat Pump Water heater 1,000 3 0.3%
Washing Machine 1,000 86 8.6%
Electric Dryer/Hair dryer 1,000 68 6.8%
EV Charger 1,000 56 5.6%
LED TV 1,000 389 38.9%
Sensors 2020, 20, 2900 18 of 21

Toaster 1,000 11 1.1%

Microwave Oven 1,000 38 3.8%

Below 150 W

Fan 1,000 530 53.0%

Dehumidifier 1,000 543 54.3%
Lights 1,000 679 67.9%
Aquarium Water Pump 1,000 235 23.5%

In the Tables 1 and 2 above, the EMS SCADA can aggregate 6–10 MW or more power across 3000
homes, which would have required substantial system wiring and land space to achieve similar ESS
capacity at the substation level. By using a Smart DB with high speed CTs and NILM, the information
makes it possible for households to participate, allow the grid to verify the availability of each
appliance and simply varying the incentives for participation to attract more participants. This study
demonstrates that the proposed approach provides a novel alternative design approach for utilities to
achieve demand management schemes easily as compared to the approach for large scale ESS
currently deployed, costing less while making the grid smarter and more user friendly.

Table 2. Aggregation of HESS and Voluntary Demand Management capacity.

Number of homes 3000

Available power Home-ESS 2 kW
Available energy Home-ESS 5 kWh
Peak power (3000 homes) 6 MW
Home-ESS State of charge 100%
Priority load 1-Aircon 2 kW
Priority load 2-Water Heater 1 kW
Priority Load 3-Kettle 1.2 kW
Priority load 4-Vacuum Cleaner 0.8 kW

For every 1 million households equipped with our system, it would translate to 4000 MW (8000
MWh based on 2 h of discharge), representing around 65–70% of Singapore’s peak power demand of
6600–6800 MW. This would save huge investments in power generation or large scale ESS to balance
an ever-growing demand and expectations of a more stable grid, notwithstanding the value add for
offering critical services with an UPS installed in the home.
With this usage information on each key load appliances, the system will lead to even more users
to make informed choices with their preferences, changing their usage behavior to participate in
voluntary demand management participation as our survey suggested. Any methods that can increase
user participation either via NILM to understand their own usage, indicate their preferences thus
allowing a secondary market aggregation of all interruptible loads for trading of incentives with
buildings to avoid paying penalties for exceeding their contract capacities, offers a real solution against
any known methods of demand management that are currently deployed. As of now, the Singapore
ILMS market has less than 10 participants of facility owners after 14 years of deployment and not open
to end home consumers. With our proposed approach, it opens up many possibilities for different
Demand Management operators to target small consumers with different possibilities across different
geographical locations by aggregating them to meet the minimum requirement of at least 100 kW
before they can participate in ILMS offering.
In this approach, a utility managed HEMS have the following cost factors and consideration:
• Cost of setting up the smart DB and gateway hardware in every home
• Cost in managing and maintaining a utility wide smart home gateway and system that is “after
the meter”, which is traditionally not the utility space and cost benefits must be recovered in
terms of savings from having to pay for higher spinning reserve capacity to ensure the stability of
the network
Sensors 2020, 20, 2900 19 of 21

Applicability when scaling up:

• The system is highly scalable as utility revenue metering, currently recorded every 30 min or
even bi-monthly for mechanical meters can now be broken down into higher resolution or even
real time streaming when at home without loading on the utility server as each user will be
polling its own gateway
• Granularity of load usage profile of every appliance will lead to higher customer trusts and
satisfaction on their monthly bill as it leads to higher transparency.
• Condition monitoring of appliances when compared to similar usage periods for possible
degradation
• Ability to provide additional services for new revenue stream for the utility
• Payback model for setting up Home ESS can include other services such as providing a better
quality of supply for home critical care services
Some of the challenges are:
• Need to develop a custom made smart DB to house all the required components and to get the
type testing approvals are not easy
• Use of different communication protocols that must allow the offering of additional sensors and
services not part of the utility space
• Accuracy of the NILM for smaller loads (<100 W) which can be easily mixed up with the
background noise
• Designing a unified security plan for all sub-components
• Selection of the recording and uploading interval, implementing the schema on the amount of
data to upload to avoid using too much bandwidth to the end user. Hence the 6 kHz sampling
interval is only locally captured and used to identified the load signature and discarded to avoid
collecting too much data. V, I, kW, kVar, kWh and frequency appliances are recorded and
uploaded only every 15 min to limit the amount of data. Each file is less than 1 kbyte each and
only takes less than 1–2 s to upload on a typical home WIFI of around 100–200 MBits. However,
when loads are turn on/off, which affects the ILMS capacity available, these status changes are
uploaded immediately. On a grand scale, load changes across thousands of households changes
gradually as appliances are being on/off only a few times a day. Even turning on/off the TV load
100 times a day is considered nothing to the system over 24 h.

7. Conclusions
In this paper, we proposed a new system design that comprises the use of high-speed sensors
such as CT and sockets to make the grid smarter. By aggregating them and integrating them with
small capacity Home-ESS coupled with selective load tripping in homes while taking in user priorities,
it allows for a more vibrant and intuitive way for grid stabilization that is likely to attract a higher user
participation rate for demand response as part of a larger ILMS. We outlined the key design
requirements of Home-ESS, HEMS, and the EMS SCADA. The proposed system has the ability to
operate in centralized as well as in autonomous modes where it can discharge Home-ESS and shed
enough load depending on grid frequency. The fast response time of the proposed scheme makes it a
low cost viable alternative in spinning reserve markets. This approach, while providing huge savings
in grid investments and achieving huge aggregation of demand capacity at different locations, also
makes it more effective than any existing load shedding schemes to balance the grid, while minimizing
inconveniences by taking their load priorities into consideration.

Author Contributions: G.K. was the lead author in developing the systems, test programs and applications as
described in the main paper. His contributions in providing the background on the current grid system, impacts of
current under-frequency load shedding schemes, challenges in encouraging Interruptible Load participation, and
most importantly, developing the approach to solve these issues form the basis for this paper. The selection of the
NILM methods, use of right sensors to provide a more definitive solution allows the solution to make a practical
impact for possible implementation. N.UL H. was co-authoring on the approach and assist in advising how the
approach could be refined and assisted in the testing of the solution. His biggest contribution comes in refining the
Sensors 2020, 20, 2900 20 of 21

testing techniques to better improve the accuracy of the NILM.C.Y. is the supervising professor that provide the
guidance on the works to ensure the proper execution of the experiments. His contribution comes in his
knowledge on advising the authors on targeting user participation and the use of energy storage systems. All
authors have read and agreed to the published version of the manuscript.

Acknowledgments: We would like to acknowledge the assistance provided by partners who are involved in the
project as such HDB, Bosch Germany and Singapore Power.

Conflicts of Interest: We declare that we have no conflict of interest.

References
1. Chang, L.; Zhang, W.; Xu, S.; Spence, K. Review on Distributed Energy Storage Systems for Utility
Applications. CPSS Trans. Power Electron. Appl. 2017, 2, 267–276.
2. Kim, J.S.; Yang, H.; Choi, S.G. Distributed real-time stochastic optimization based ESS management strategy
for residential customers. In Proceedings of the 19th International Conference on Advanced Communication
Technology, PyeongChang, Korea, 19–22 February 2017.
3. Petrichenko, R.; Chuvychin, V. Smart load shedding system. In Proceedings of the 3rd IEEE International
Symposium on Power Electronics for Distributed Generation Systems (PEDG), Aalborg, Denmark, 25–28
June 2012.
4. Syadli, H.; Abdullaha, M.P.; Hassana, I.F.M.Y.; Hussina, F. An improved load shedding scheduling strategy
for solving power supply defisit. J. Tecknologi (Sci. Eng.) 2016, 78, 61–66.
5. Tang, J.; Liu, J.; Ponci, F.; Monti, A. Adaptive Load Shedding Based On Combined Frequency And Voltage
Stability Assessment Using Synchrophasor Measurements. IEEE Trans. Power Syst. 2013, 28, 2035–2047.
6. Quyen, H.A.; Bui, Q.A.; Le, T.N.; Le, T.T.T. Optimal Load Shedding Based on Frequency, Voltage
Sensitivities and AHP Algorithm. In AETA 2013: Recent Advances in Electrical Engineering and Related Sciences;
Springer: Berlin/Heidelberg, Germany, 2014; pp. 75–84.
7. Atteya1, I.; Fahmi1, N.; Strickland, D.; Ashour, H. “Utilization of Battery Energy Storage Systems (BESS) in
Smart Grid: A Review”. In Proceedings of the International Conference on Renewable Energies and Power
Quality (ICREPQ’16), Madrid, Spain, 4–6 May 2016.
8. Logenthiran, T.; Srinivasan, D. Intelligent Management of Distributed Storage Elements in a Smart Grid. In
Proceedings of the 2011 IEEE Ninth International Conference on Power Electronics and Drive Systems, IEEE
PEDS, Singapore, 5–8 December 2011.
9. Tushar, W.; Chai, B.; Yuen, C.; Smith, D.B.; Wood, K.L.; Yang, Z.; Poor, H.V. Three-Party Energy
Management With DistributedEnergy Resources in Smart Grid. IEEE Trans. Ind. Electron. 2015, 62, 2487–
2498.
10. Lindahl, S.; Runvik, G.; Stranne, G. Operational experience of load shedding and new requirements on
frequency relays. In Proceedings of the Sixth International Conference on Developments in Power System
Protection, Nottingham, UK, 25–27 March 1997.
11. Perumal, N.; Amran, A.C. Automatic load shedding in power system. In Proceedings of the National Power
Engineering Conference, PECon 2003, Bangi, Malaysia, 15–16 December 2003.
12. Chen, C.-S.; Ke, Y.-L.; Hsu, C.-T. Protective Relay Setting of the Tie Line Tripping and Load Shedding for the
Industrial Power System. IEEE Trans. Ind. Appl. 2000, 36, 1226–1234.
13. Eller, A.; Gauntlett, D. Energy Storage Trends and Opportunities in Emerging Markets. Available online:
https://www.ifc.org/wps/wcm/connect/ed6f9f7f-f197-4915-8ab6-56b92d50865d/7151-IFC-EnergyStorage-rep
ort.pdf?MOD=AJPERES (accessed on 13 March 2019).
14. Hart, D.M.; Mason, G.; Bonvillian, B.; Nathaniel, M. Energy Storage for the Grid: An MIT Energy Initiative
Working Paper; MIT Energy Initiative: Cambridge, MA, USA, 2018.
15. Hesse, H.C.; Schimpe, M.; Kucevic, D.; Jossen, A. Lithium-Ion Battery Storage for the Grid—A Review of
Stationary Battery Storage System Design Tailored for Applications in Modern Power Grids. Energies 2017,
10, 2107.
16. Schneider, M.; Biel, K.; Pfaller, S.; Schaede, H.; Rinderknecht, S.; Glock, C.H. Using inventory models for
sizing energy storage systems: An interdisciplinary approach. J. Energy Storage 2016, 8, 339–348.
17. Borsche, T.; Ulbig, A.; Koller, M.; Andersson, G. Power and energy capacity requirements of storages
providing frequency control reserve. In Proceedings of the IEEE Power and Energy Society General Meeting
(PES), Vancouver, BC, Canada, 21–25 July 2013; pp. 1–5.
Sensors 2020, 20, 2900 21 of 21

18. Buckley, B. Putting More Energy into Peak Savings: Integrating Demand Response and Energy Efficiency
Programs in the Northeast and Mid-Atlantic. In Proceedings of the 2016 ACEEE Summer Study on Energy
Efficiency in Buildings, Pacific Grove, CA, USA, 21–26 August 2016.
19. Thirugnanam, K.; Kerk, S.G.; Yuen, C.; Thirunavukarasu, B. Battery Integrated Solar Photovoltaic Energy
Management System for Micro-Grid. In Proceedings of the IEEE ISGT-Asia 2015, Bangkok, Thailand, 3–6
November 2015.
20. Hassan, N.U.; Khalid, Y.I.; Yuen, C.; Huang, S.; Pasha, M.A.; Wood, K.L.; Kerk, S.G. Framework For
Minimum User Participation Rate Determination To Achieve Specific Demand Response Management
Objectives In Residential Smart Grids. Int. J. Electr. Power Energy Syst. (IJEPES) 2016, 74, 91–103.
21. Tushar, W.; Yuen, C.; Chai, B.; Huang, S.; Wood, K.L.; Kerk, S.G.; Yang, Z. Smart Grid Testbed for Demand
Focused Energy Management in End User Environments. IEEE Wirel. Commun. 2016, 23, 70–80.
22. Rules Modification for EMA’s “Implementing Demand Response in the National Electricity Market of Singapore”;
Publisher: Energy Market Authority , Singapore, 24 June 2011.
23. de Mello, P.B.; Pinto, R.G.D.; Silva, P.B. Load Disaggregation of Industrial Machinery Power Consumption
Monitoring Using Factorial Hidden Markov Models. In Proceedings of the NILM Workshop 2018, Duisburg,
Germany, 1–2 October 2018.
24. Parson, O.; Ghosh, S.; Weal, M.J.; Rogers, A. Non-intrusive load monitoring using prior models of general
appliance types. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, Toronto,
ON, Canada, 22–26 July 2012.
25. Kolter, J.Z.; Jaakkola, T. Approximate inference in additive factorial hmms with application to energy
disaggregation. In Proceedings of the Fifteenth International Conference on Artificial Intelligence and
Statistics, Ser. Proceedings of Machine Learning Research, Palma, Canary Islands, 21–23 April 2012;
Lawrence, N.D., Girolami, M., Eds.; PMLR: Palma, Canary Islands, 2012; Volume 22, pp. 1472–1482.
26. Zoha, A.; Gluhak, A.; Nati, M.; Imran, M.A. Low-power appliance monitoring using Factorial Hidden
Markov Models. In Proceedings of the 2013 IEEE 8th International Conference on Intelligent Sensors, Sensor
Networks and Information Processing: Sensing the Future (ISSNIP ’13), Melbourne, VIC, Australia, 2–5
April 2013; pp. 527–532.
27. Rabiner, L.R. A tutorial on hidden markov models and selected applications in speech recognition. Proc.
IEEE 1989, 7, 257–286.
28. Kim, J.; Le, T.-T.-H.; Kim, H. Non Intrusive Load Monitoring based on advanced deep learning and Novel
Signature. Comput. Intell. Neurosci. 2017, doi:10.1155/2017/4216281.

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).