>> Tanuja is a co-founder of DataGlen,
which is a early,
around startup, in this space of solar energy.
Tanuja was
earlier a technical staff member at IBM Research,
She finished a masters from MIT.
It's a pleasure to have you, and she's also part of
the summer workshop that we have here,
so she'll be around for next couple of weeks.
>> Thanks Patrick. Good morning, everyone.
It's great to present some of the work that we're doing,
and I think it's very relevant to
Azure Analytics part that
we are talking as part of the workshop.
So, as the name suggests,
I'm going to talk about Bits and Joules.
So, some of the use cases in energy
vertical and what are the challenges?
How the paradigm is shifting?
How the modern day's technologies
or digital technologies of IoT and
AI are relevant for some of the problems
that you're trying to solve in energy particle?
So, as all of us would agree that energy is
a fundamental necessity of modern society,
and as we have the access to clean and cheap energy,
we can do a lot in terms of the other sectors,
including the healthcare and
education and various other factors.
So, it's important to
basically look at the problem of energy,
which is the driving factor for
various other industries and technologies,
in terms of the innovations.
Energy vertical right now is
going through a paradigm shift,
and we call that paradigm shift as Democratization.
So, on the left-hand side,
you would see the democratization of media.
So, few years back or let say 20,
30 years back, the media was pretty centralized,
so there was content which was
created centralized way and
the content was basically
disseminated in a distributed manner.
So, it's like through national BBC
who's creating the content,
and the mass population
is basically just absorbing the content.
But now in last 10 years,
there is with the social media like Facebook and YouTube,
everyone can create the content
and everyone can consume that content as well,
so it's completely changing from
the centralized to decentralized manner,
very similar thing which is happening
in the energy vertical right now,
which is democratization of energy.
So, previously, there was
a central generation which was
happening at the form of the hydro or thermal energy.
It was, the distribution and transmission was
happening through central entities.
Again, in India, most of the utilities are doing
the transmission and distribution right now,
and the consumers who are
basically the residential or industrial consumers,
are just consuming the energy without
participating in the entire model,
but now we are shifting,
very similar to social media,
the change is happening in the energy vertical as well.
So, there is distributed
solar and batteries which is coming into picture,
there are electric vehicles which
are basically getting commercialized right now.
So, it's completely changing
the way the energy was generated and consumed.
So now, the commercial or residential consumers
are able to generate their own energy,
part of that they might be consuming,
part of that they can basically
inject back into the grid as well.
So, it's not a one way flow but now
there is a two-way flow of energy which is happening.
Obviously, the traditional power plants
and traditional transmission distribution network
is not designed to handle
this two-way type of communication
or two-way type of energy flow,
and hence there are a lot of
interesting socioeconomical and technical challenges
which are getting created because of this.
So, we will go through some of
the issues in energy vertical and
how the computer systems like IoT and AI,
can help solve some of those issues.
So, the first issue is of power shortage.
So, on the left-hand side,
this is the power consumption,
basically the power demand from New York over the year.
So, basically, this axis shows,
over the year, how there are
variations in the power demand.
This particular axis shows during a particular day,
how the demand is varying.
So, during a day, again,
there are one or two peaks depending upon
the power consumption needs
and depending upon the season,
whether it's the summer versus winter,
there's a variability in the demand.
So now, if we look at these plots,
these are again from the New York City.
So, there was a highest peak load demand here,
which was 2006 versus the 1980's.
If you see that, there is no long tails
in this kind of distribution,
so these are basically
hourly loads in terms of megawatt hours,
and as we are moving to 2006 and again to 2018,
we see very long tails of demand which are there.
So, the demand is increasing,
there are short periods of high demand,
and definitely to cater to these kind of demands,
there is no point in adding
the infrastructure just for
catering to that peak demand load,
so definitely there is some other methods or
alternative methods are required to curtail
toward address to these kinds of demands,
which are coming on all the networks.
The second problem is of energy scarcity.
So, in general, there is a problem
of access to clean and cheap energy,
so this is something which we
have very much experienced in India
where there are regular power cuts or there are places
where there is no access to energy as of now.
So, this figure is from 2018,
which shows that across the world,
how much population is without electricity right now.
As we speak right now there is still
approximately 300 million people in
India who do not have access to energy,
and that's because of various reasons
of the cost which is required for
the transmission and distribution of
the energy which is generated at the central level,
as well as the other problem
of the availability of energy itself.
So current energy generation
which is happening through thermal
and coal power plants in India,
which is not sufficient to
increasing demand of energy at various places.
The third issue that we'll talk about is,
mainly the good part which is
happening is there is a lot of
renewable and alternate energy sources
which are coming in picture right now,
in the form of solar,
wind, as well as battery storage et cetera.
Good parts out there are quite in local,
so there are no transmission and
distribution losses involved here.
They're are sustainable and scalable.
So you can scale these plans as required.
Most of these solar and,
solar especially, is reaching grid parity,
which means that the cost of generation
of energy from solar which is
becoming almost same as cost of accessing
the energy from the thermal
or the existing conventional plants.
But there are some disadvantages
in using these renewable energy resources.
The first problem is generation variability.
So, this plot shows the generation
across the year from one, single wind turbine.
So, as we see that there is a huge variability
across the year as well as during one given day,
there is a huge variability in
the generation which is
happening from single wind turbine.
So, when we are taking into account
the large adoption of renewable energy resources,
if we are not able to predict how much amount of
energy that will be generated in next few hours,
then we cannot do the unit commitment and
we cannot rely especially on these kinds of sources.
So, it's very important that how reliably we can
predict how much amount of energy that would be
generated from these renewable energy resources,
and it includes the
accurate weather prediction and on top of that,
accurate prediction of the energy from these sources.
The second problem is that
if you take into account, let's say,
solar, then the solar
is available only during few hours of day.
So, it's available only from, let's say,
6.00 AM to 6.00 PM of the day,
and again there is a variability in amount
of generation which is happening from a solar plant.
So, we cannot rely on
this renewable energy resources completely,
we need to have additional systems in place to
backup whenever these resources are not available,
and that's where basically
we are looking into, storage technologies,
and how storage and other
technologies can be integrated with renewables,
so that we can reliably provide power supply
through these renewable energy resources.
So, that's where now we're
going to talk about the Participatory Grid.
So now, the consumer is not just the endpoint,
consumer is not the dumb or static consumer,
but consumer is participating in the entire process
of energy generation and consumption in various manners.
So, the four main points that I'm going to talk about is
first is how we can address
the power shortage problem by reducing the peak demand?
The second is how we can address the energy scarcity?
On one side there is energy scarcity,
which means that amount of energy which is available,
that is not sufficient to cater to most
of the energy resources,
but on the other hand we see lot of
energy wastage which happens in the residential,
commercial and industrial places.
So, can we identify
the energy wastage and
curtail those energy wastage points?
The third one is how we can
basically maximize the distributed generation,
or basically adoption of
distributed energy resources with
the additional resources like
solar and other technologies,
and the final one is looking at
optimal utilization of resources,
which means that basically including
all the points of preventive maintenance
and predictive analytics.
So, just going back very quickly to
IoT technologies and why
IoT analytics is very suitable
to solve some of these problems.
So, we are looking at basically the things which are
equipment setup at most of
these renewable or the conventional energy resources.
Most of these are capable of
communicating data with standard protocols.
So, we do not have to, most of the cases,
we do not have to add any new sensors as such,
the existing batteries inverters,
wind turbines or even
the smart meters which
are installed at the home premises,
they are capable of sending
the data and the context information as well.
There are some existing processes
in place which need to be
adopted to basically look into the new technologies.
So, I like to basically look at
the IoT stack similar to brain architecture.
So, as we look into the brain,
there are different systems which are
responsible for different types
of actions which are performed in our body.
So, if we look at the Reptilian Brain,
it's basically governing
all the important or survival instincts
of human body including breathing, heart rate,
and balance, which are like
low-latency response events which
are required by the body.
So, those events are basically being
handled very close to the sensor systems.
The second type of brain is Limbic System which
is looking at the feeling and memory formation.
So, looking at more data and generally forming
the patterns or memories through
that and the last one is basically,
Neocortex which is doing
the reasoning forward planning language and other things.
So, this is basically looking at more holistically,
the data from multiple sensors
and making their decisions on top of that.
So, can we look at
the IoT hierarchy of systems in similar view.
So, we have sensors and actuators at the level which are,
many times provided by the equipment.
Then, they have the data loggers or
edge devices which are sensing
and collecting that data
and capable of taking certain actions.
So, the actions which are required as
a low-latency response actions or decisions which
are crucial to be taken in the very near real time,
it's good to have those decisions in
the system which are as close to edge as possible.
Then, there are certain decisions
which require information from multiple resources.
So, let's say, if we are looking at the load across
the community and some
decisions that need to be taken on top of that.
Obviously, we would not be able to take
those decisions at edge devices but it
might be in the Fog or the Local Data
Center or the small computing which is at,
let say, community level where we
can access that and obviously, in the Cloud.
So, there was some initial work we did while I was at
IBM on looking at hierarchical decision system.
So, can we do the analytics and the hierarchical manner,
where one node is doing
the computation which is possible and
essential by that particular node and
passing on information to the next level of nodes?
So, depending upon the problem,
can we decide where the analytics should reside for
that particular problem and can
there be a connection between those nodes?
So, we have not completed this work
after I left IBM research but
we are interested in taking it up.
So, if there are any interest
in the similar kind of systems,
we'll be happy to collaborate on this.
So obviously, the Neocortex is doing
various types of analysis on imaging time-series data.
So, getting back to the Participatory Grid,
I'll start with the load shifting problem which is
basically looking at Peak Power Demand.
So, why reducing the peak is important.
So as we see,
the X-axis is the time of day and
Y-axis is the power production or their demand.
So, there are multiple peaks during the day and
the power is submitted as in the power is
generated by various different types of load systems.
One is basically Baseload which could be a hydrothermal.
So, which is a constant load
which we are getting through out the day.
On top of that there is an intermediate load,
which is again provided by
some conventional energy resources
and then, there are Peak Loads,
which are typically provided by
diesel generators which have low response times,
quickly can be started and run through.
But if we look on the right side plot,
it shows the amount of basically,
the price versus the power generation
and the price does not increase linearly.
As we go to the peak demand,
the price increases exponentially.
So, if we can reduce this particular demand
from a very small tiny amount at least,
it significantly helps in reducing the price in
this particular band when it
goes exponentially increasing.
So, there are historical approaches which have been
taken into account previously to shift these loads.
Some of these approaches have been
Daylight Saving Time or
the Tokyo Brownouts and CA Rolling Blackouts.
Also in India, we see that there
are these kind of load shedding which
happens like a few hours a day
at different localities to cater to the demand.
So, some part of the community would be completely
shut down during a particular point of time.
So, this is more like a top-down approach where
the utility is deciding that I'm
not able to get it to the peak demand.
So, I would shut down part of
the community and that would be
basically a ruling load shaving that we
see often during the summer times.
There are lot of other approaches
which are being tried out in
countries including US and Europe
with variable degree of success,
which are called Demand Response.
So basically, if you look at the left-hand side,
it shows the typical residential demand and what
part of the residential demand
is the deferrable or time shiftable loads.
So, if we're looking at electrical vehicle charging,
that could possibly be
a time shiftable load or if we're looking
at water heater or washing machines and dryers.
Those are time shiftable loads because
it's not like you need it immediately,
but it can be done over a period of three hours of time.
So, there is a large proportion
of loads in the residential,
commercial and industrial loads which
are kind of time shiftable and can be taken
advantage of those loads to
shift from pick-time to off pick-time so that
we can reduce the demand
and basically do the peak shaving during those times.
So, when we started looking into basically,
the problem of peak demand and demand response for
India or the developing countries.
The first challenges in the way it's being tried
out in US and Europe is,
utility collects the data
from the smart meters or
individual appliances, let's say,
washing machine is in-demand response,
then there is IoT device which is
sitting on the washing machine.
It's sending the data to the Cloud and utility
sends the demand response signal either for
the manual intervention or
automatic dispatching the load to the some other times.
So, it requires basically
complete infrastructure for communication,
centralize optimization techniques where
you can send the signals to
the consumers and the consumers
would participate in demand response.
There are few success stories in certain utilities in
US who have successfully done
the demand response process for shifting the load,
but we started looking at the same problem for let's say,
India and Africa, we do not even have
the sufficient infrastructure
for the network communication.
So, we cannot rely on the network communications to
send the signal from the utility to individual equipment.
Also, there are lot of equipments
which are already existing.
So, we cannot change those equipments,
which means that there has to be something which is
sitting outside of the appliance,
but still taking into account the appliance preferences.
The third thing that we wanted to look into is,
as to be as cheap as possible.
So, it has to reside on
a small device and the third thing is,
the same appliance should be
applicable from a range of appliances.
So, it should not be that
one type of appliance is only running
for water heaters versus washing machines.
So, it should be able to handle
multiple different types of time shiftable appliances.
So, taking these multiple things into account,
we built something nPlug.
Again, this is work which
I did while I was at IBM Research.
So, what this nPlug does is,
it's like a smart socket or it's like
over the power strip that we generally use.
So, it sits between the wall socket and their appliance.
The appliance goes into the nPlug and
this nPlug is built using a 16-bit micro-controller,
with very small,
it could be of RAM and four MB of flash memory.
So, what it does is,
it basically does local sensing and decision-making.
So, it senses the grid conditions
especially of voltage and frequency,
to identify whether there
is a high load on the grid or not.
It takes the preferences from the consumer.
So, let me take an example of water heater.
So, let's say most of their time solar water heaters,
the water can be kept heated because there is
insulation available for three to
four hours in the morning.
So, most of the times what happens is we all get
up at six o'clock and start the water heaters.
Water heater takes from
two kilowatt to four kilowatt of energy.
Multiple such appliances come in
the activation and we see the morning peak.
So, can we instead of starting,
everyone starting and manually
switching the appliance at 6:00 AM,
can we switch the appliance
anytime between 3:00 to 6:00 AM,
because water can be any way kept
hot and ready for three hours.
So, the user can set the preference that,
"I want the water to be ready at 6:00 AM",
and it needs to be run
for 30 minutes let's say for water to get heated.
So, then we can actually shift
this demand from 6:00 to 6:30 or
whatever time the user gets up and six to
7:00 AM let's say it's the peak demand time,
we can actually do it anytime from 3:00 to 6:00 AM,
because that's the off-peak time,
there is no much high load on the grid.
The user can set the preferences
that 30 minutes is what I want to run the appliance
and 6:00 AM is
the time by which I want the water to be ready.
Obviously, there are "Override" buttons,
so some days you don't want to set that schedule,
so you can always override and say that I
want to switch it on at whatever time I want.
So, without having any external infrastructure
and just having the sensing
of voltage and frequency inside the appliance itself,
we wanted to see how we can
identify the peak and off-peak demand times.
So, the first part is an autonomous operation or
decision making of identifying
when there is high demand on the grid.
So, this plot shows the time of day versus the voltage,
and each color is a different day of the week.
So, this is the data
which was collected from my house in [inaudible].
So, in the morning you would see that
the when night the voltage
is high because these off-peak times,
the voltage is about 230 isn't
nominal voltage we are expected to have like 230 volts.
But typically, at night we will see
it around 245 or 250 volts.
Whereas in the evening peak times around six to nine,
you would see significant drop which is like 220 volts.
This was in Bangalore where
this local transformer was very close to my house.
We did similar experiments at
various different parts in India,
and we see huge variability.
So, if you go at some of the rural places in [inaudible] ,
you have seen the voltage dropping up to
180 volts or 175 volts.
So, we cannot just fix a particular range because a range
varies across as you go
from one location to other location.
So, we cannot just say that anything above
235 is the off-peak time and anything
below to 220 let say is
the peak time because depending upon the location,
you have to adaptively identify that what
is the peak and off-peak time range.
So, we basically collect the data at
every 30 seconds or one minute time interval and put
the compression on top of it like
this is aggregate approximation,
for aggregating the signals
at every five minutes time interval.
Then we look at basically doing the clustering
of these particular time series
into three clusters of peak,
off-peak, and the median time.
Then we look at the patterns,
so let's say the algorithm looks at
the last 10 days pattern
to identify what are the peak and off-peak times.
So, based on that, it decides that okay,
it identifies first of all the thresholds,
which is the low threshold and
upper threshold for voltage.
Then identifies that based on
last 10 days pattern that this is
a time which is a peak time and this is a time which
off-peak when they need to take the control.
This wasn't previously there were a lot of
techniques which were looking into
the frequency data to identify
these gaps between the supply and demand,
there was lot of work on this,
but no one has looked into
the voltage signals and buttons,
so that was one of the contributions.
Doing these on these small microcontrollers
in streaming manner, is what we did.
The second thing is looking at frequency.
So, frequency would be ideally it
should be 50 hertz and close to,
if we look at the data from yours,
we would not see this much of variability,
it's pretty close to let say 60 Hertz over there,
but here we see
a huge variability from let's say 49.2-50.2.
But still there are events when
it goes beyond the normal threshold.
So, we identify what is the normal behavior for
frequency and which other times where there is imbalance.
So, if the demand is high then generation,
then the frequency will drop.
So, we identify those events and
again capture for shifting the load.
So, now we're able to identify
the off-peak and peak times
and we know that when we
should not be running the appliance.
But let's say, now there is no central controllers,
as such every device is going
to make the decision independently.
So, we should not create
a problem of just shifting the load
from one time to another time
by all of them starting and shifting.
So, let's say all of them sense,
there are thousand influx and all of them sense that it's
a off-peak time and all of them shift to
another time and create another peak demand.
>> So, what's the disadvantage of what the [inaudible].
>> Basically, you are ready to the peak demand.
So, if you are running the water heater
and since utility is not able to supply during
that time because there is a cap on
how much energy utility can
supply at any given point of time.
So, you would have experience let's say
morning and evening power cuts, right?
So, they typically do the power cuts during peak time,
because they are not able to cater
to the demand at that time.
So, that's like a top-down approach,
they force everyone to shut it off.
But what this one is trying
to do is like a bottom-up approach
that the device itself says that I need
not run only during the peak time,
so I can shift my demand.
There are other loads rate,
let say there are lights and
air-conditioner and other things
which are not time shiftable.
So, we have to run those appliances during peak time,
so giving chance to
the appliances which definitely need to run during
that time and shifting the loads
which are kind of voluntarily,
rather than utility forcing
saying that entire community would
go down from 6:00 through 7:00 AM.
>> [inaudible] ?
>> So basically, we then discussed with the utilities
on how they can include as part of the business model.
So, for industrial consumers,
there is time of use pricing, right?
So, even in Karnataka,
there is a time of use pricing where industries are
charged higher rate if they are running it on pick-time.
So, we'll discuss various business models if they
can give a rebate if someone is running the load.
So, let say in this particular devices
and able to send the audit log at the end of the month,
which shows that it was not running during the peak time.
Then, there could be discounts and rebates,
as well as if there is a time used pricing,
there is a straight use case for
the consumer to ship the load during follow-up.
>> [inaudible].
>> So, that's part of the user preferences.
So, there are certain loads which
are like a long-running loads.
So, if you take an example of EV charging,
that would be running over four hours
and you can basically
chop that four hours into let's say six hours with-.
You can say that my window size need to be 30 minutes,
so you should not be switching too often.
So, if I give the chance
for a plants were
uninsured and for at least minimum of 30 minutes,
so that's part of the user preferences.
User can say, "Do I want to run it at one go,
which is a case for let say,
washing machine that I want to run it at one go?"
versus if it's the inverters that we charge
at home or the
electrical vehicles or even the water heaters,
where it can be chopped that
the entire usage can be chopped at multiple time windows.
>> [inaudible].
>> No. So, the way they had design it,
user has to provide the preferences.
We wanted to keep minimal inclusions in the device,
taking into account the small size of that.
But that was a requirement
from a few discussions that we had that if
the device is capable
of providing the usage pattern let's say.
Let say, we can just collect
the data of all consumption of an appliance,
can be categorize what type of device it is?
So, can we categorize,
classify basically the water heaters
from washing machines and
then automatically put the preferences
of what category of device it is?
>> Is it different in plant for each device?
>> It's the same model,
but if you are running it
for multiple different appliances,
you would be putting one for each.
>> [inaudible].
>> In this model, it was not.
But then there was some conversation on,
let's say there are 10 templates within one industry.
Can they talk to each other and do
some even better optimization in terms of distribution?
>> [inaudible]
>> Yeah. So, they did
these experiments on the data collected from,
let's say, US houses as well.
As I said, the range was small.
So, even in case of frequency and voltage,
the range is very tight,
but still, you would see the pattern.
So, the pattern is available over there tight range.
>> [inaudible] ?
>> Yes. So, you would see a small variability still.
So, the variability might be in, let's say,
then this two minus two but still,
the variability is available over there.
Basically, to control this lot of utilities,
put additional devices for doing
multi-stablization and frequency stabilization.
So, a lot of rather in India,
while the utilities, or
I think in few countries in Europe as well,
the pricing in the real-time energy market,
the amount of price that you get
paid for providing energy
at a particular point of time is called ABT,
availability-based tariff,
that's dependent on the frequency at that time.
So, if the frequency is low and
if you're supplying energy at that time,
you get paid more versus
when if you are supplying in this particular region.
>> [inaudible].
>> Yes. So, again,
it's not widely available as of now.
It was initially done as part of the pilots.
Now, as they're putting smart meters,
the demand response is generally done if
there is smart meters available.
But a smart meters giving you timestamp data
already and being pushed to
the Cloud almost in real-time,
so that's where the analysis is happening in the Cloud.
>> [inaudible] ?
>> So, I'll go to the other problem where
we do look at the power consumption of the plants.
But in this case, it was mainly
for identifying what is the condition of the grid,
whether the grid is overloaded or not.
So, if you look at the power consumption part,
you will just get to know that how
your appliances working but you would not
get to know that your transformer
is overloaded right now,
and you need to basically
curtail your demand at this particular point of time.
So, basically, once the problem of
identifying the good condition was solved,
then we wanted to look at how we can basically
distribute spread that demand over time.
So, we looked at
various decentralized scheduling techniques.
We basically looked at computer networks,
CSMA kind of protocols,
which tried to back off
when there is a high demand on the network.
So, we call this algorithm as
a Grid Sense Multiple Access similar to
Carrier Sense Multiple Access.
There are few differences in this one.
So, if we look at this
similar computer network algorithms,
there is no deadline for completing the packet transport.
So, you can basically attempt and send the packets,
whereas in this problem, there is a deadline.
So, if I'm saying that I want
the water to be ready at 6:00 AM,
then 6:00 AM is my deadline.
I need to, no matter what,
we want to give preference to
the user's deadlines which means that the water
has to be ready by the deadline which is given.
So, we had to basically do some modifications for
CSMA algorithms to suit to the deadline-based algorithms.
So, the first thing that we did was contention window,
depending upon the available time slots.
So, let's say if I am at 3:00 AM
and I have a 30-minute slot.
So, I have three hours which means
six slots which are available or five minutes slots,
whatever number of slots which are available,
the current time slot during that
particular 3:00-6:00 AM window
and how much time I need to run that.
So, depending upon that,
every device basically tries
to identify the contention window.
As the time elapses,
as you start going towards their deadline,
your contention window size
starts becoming smaller and smaller.
So, what the device does is,
let's say if I have seven as the contention window size,
device chooses randomly to sense at any
of these initial seven-minute slots,
and identify sense the grid and
identifies whether it's the loaded or not.
So, by doing this, we are
first of all distributing the amount
of endplates with just sensing and
making their decision at the same time.
The second thing is the probabilistic connection.
So, depending upon the voltage levels,
so let's say the voltage is below peak load,
then it says that you
should not be joining because it's a peak time.
If it's above the upper threshold,
which means that it's a peak time and
device can definitely join at any point of time.
If it's in-between, then basically,
it computes probabilistically that with
how much probability that you should be joining the grid,
whether you should be starting.
So, by doing these probabilistic arrangements,
not all endplates would again join
at the same time but they distribute
probabilistically and would not
increase the load beyond a critical point.
>> [inaudible] ?
>> So, we did some experiments on let's say if
there are two controllers,
two endplate controllers in one house.
Obviously, if they can communicate between each other,
there would be better chance
of optimization than this one.
If you are able to do it centrally,
that would be a complete optimal solution.
So, as we go up in the hierarchy or from the H242 Cloud,
it might be giving you a bit.
Because you have higher information available with you,
you would be able to do better optimization.
By taking into account the constraints of,
we didn't want to write any communication infrastructure,
not even the connection with Wi-Fi or the other points.
So, because of that, we decided that we would make
their decisions locally and we would
see how much the difference it comes to.
If there are not many sensor,
if there are not hundreds or thousands of
sensors being controlled at one location,
this one still gives the same performance as if it's two,
10, or 100 still gives the same performance.
>> For households, this is still okay to that?
>> Well, even households
and industries which have less than 100 or 200 devices,
this still works equally as if
you had taken the centralized decision.
>> [inaudible] ?
>> So, that's another as in I can point you to the paper.
Where looking at users load.
We tried to identify what are the preferences
of running their devices and
based on the context of how they
would like to run their devices over weekends
versus let's say weekdays and other things.
So, these are the results of,
we tried it in simulation and
some controlled experiments in real world.
So, basically here it's
creating two peaks during the day.
There are different types of inputs
like water pumps and washing machines,
water heaters and inverters,
which have variable preferences
in terms of their time window,
how much they want to run?
The starting times and durations and other things.
So, we are considering let say close to I think
400 or 4,000 of such appliances.
Even when they are taking
these independent decision in a decentralized manner,
because now they can shift this particular peak
spent three hours or
four hours before the actual time duration.
It's able to actually
significantly spread the demand over time.
>> is this simulation or is it actual measurement?
>> This was in simulation.
So, we did various simulation and
then we tried it in
the control experiments of 100 devices.
So, 100 devices we ran on
multiple different appliances and
users said there are preferences as their need.
So, we didn't have control over
how they said the preferences.
So, now we're moving from the load
shifting to energy efficiency problem.
So, there is one thing is
there is high demand which utility
is not able to get into,
but on the other side, we see that there is
lot of wastage of electricity which happens.
The wastage could be because of the aging appliances like
Aging refrigerators or the cooling systems.
In US itself 30 to 45 percentage of
each HVAC system had to
operate at below efficiency levels.
So, this could be again because of two reasons;
one is that the malfunctioning or aging of their
appliance or inefficient use of appliance.
So, it could be that
the air conditioner is running
by individuals and doors are open.
Which means that it's not running at the optimum levels.
To address this problem,
what we did was something similar where
it's called as a SocketWatch.
Which again sits between
the appliance and the wall socket,
and now it looks at instead of looking at the grid,
it looks at the power consumption
of that particular appliance.
So, we collect data of active or reactive power and
various other parameters which can be sensed
from outside of their plants.
Then we try to model
the benchmark model for that particular plant.
So, here is an example
of washing machine which goes through different modes
depending upon the settings that you
choose on the the washing machine.
It has different signatures of active or reactive power.
So, we identified these states of their plants,
whatever appliances running in during the benchmark time.
So, this is a learning phase when we collect
the data and build a benchmark model.
So, here it's basically
showing that there are different states of
power consumption of their plants and how
the states are transitioned from
one state to another as a Macro model.
So, with the probabilities of going from one state to
another state depending upon various factors.
So, there are two phases,
one is the learning phase when we identify the modes,
depending upon the active or reactive power consumption.
The states depending upon
the time spent in each of the mode.
So, sometimes the same
power consumption but if it's running
for different time intervals means
differently or it's a different state.
Then we look at the periodicity between those modes.
So, for example if we are taking example of refrigerator,
the defrost heating might be running once every day.
So, that's the periodicity of that.
Or, if we see the cycling of compressor,
it might be running at specific time intervals,
which is a periodicity and then the state transition.
So, how the device
transitions from one state to another state.
In the model, once you have the benchmark model,
in the monitoring phase we do two things;
one is the standby mode correction.
So, if we identify that
the appliances is running on standby,
then it's switched off and if
we identify that there is
a change from the benchmark states,
then we basically alert as
a malfunction or inefficient use of the appliance.
So, every appliance while it consumes the power,
it consumes the power in two forms.
So, one is the,
so basically the breakage of your power consumption.
Which is termed as it has
a phase angle and it basically gives you two parameters.
That how much is active or reactive power,
depending upon if your load is
inductive or was capacitive.
The amount of active or reactive power would change.
So, the combination of active or reactive power would
define in some sense what kind of
operation is happening in your plants.
>> [inaudible]
>> So, basically if you
measure any appliance's power consumption,
you get these two factors of active or reactive power.
So, it's just sitting outside
and measuring power consumption, that's all it's doing.
So, there is a lag between,
as in it's power factor between
the active or reactive power
which is a phase angle of those.
So, here are some of
the examples of identifying the issues.
So, this particular one of the refrigerators,
as in this is the benchmark model.
When you see that there is typically,
there are states like compressor cycling,
compressor going on and off.
Which if you look into the clusters,
it basically identifies different clusters
which are states in terms of power consumption.
There is a periodicity,
so how much amount it stays in
the compressor on mode versus compressor off mode.
This refrigerator after building
benchmark model, we identified it.
We basically changing the states and
then they look back and found
out that there was a gasket leakage.
So, what's happening is,
since there is a gasket leakage,
it's not cooling sufficiently.
So, it's not able to take its cooling,
so it was not identified by
the user because it seems to be
working fine but it
wasn't ever going into the compressor off mode,
because thermostat is not able
to activate the compressor off mode.
So, if you see here, it's always on and
there are no cycles of on and off kind of a thing.
If we look into the clustering or the states,
here was the compressor off state,
which was coming periodically,
which was completely missing
from the actual operational model.
So, just moving next
to the distributed generation part.
So, basically, these are called Microgrids,
so the examples which I talked about earlier.
Because they are doing the generation
and consumption at the same location,
so they are called Microgrids.
There are various forms of Microgrids like Offline,
which are completely working independently,
or there are Backup or Online,
which are typically the solar storage
in the house where you are
getting grid power as well as you have
your own system of the grid.
It has been historically, we can say,
the consumption and generation
happening at the same place,
so they're historically different Microgrids,
but obviously, those are not efficient.
So, the problem that we discussed earlier,
previously, the user were just doing consumption.
So, it was simple on and off of a switch.
Whereas now, when the consumer is called the Prosumer,
where there is a local generation and consumption,
the complexity of the problem is
increasing on the consumer side.
So, this is like a time of
use pricing which is provided by the utility.
So, depending upon whether it's a summer
or whether it's a weekday versus weekend,
there are different prices for per unit prices.
You have the solar generation,
which is like we don't have much control over it.
This generation is happening depending upon your weather,
you can just predict
degeneration how much it can be there.
You have a battery. So the control that you have
is how much battery that you can charge and discharge.
Again, the problem at
the small scale is
being handled at multiple different stages as well.
So, there are something which is
called as Virtual power plants,
where you have multiple different types
of power generation equipment.
It could be wind, solar, diesel
generators, solar batteries.
So, how can you activate?
So, basically, virtually, they create a power plant,
and let's say if you have the requirement that
this Virtual power plant need to be
supplying a particular amount of demand during time,
how can you choose the right set of
generation at right time so that
the overall cost of generation is minimized?
So, starting from the individual household level
to the large-scale generation,
the same problem is being
tackled at multiple different levels.
>> [inaudible].
>> So, in earlier two problems,
we're still running with the conventional sources
but tackling to the power shortage
and electricity wastage.
Here, it's mainly we are taking
into account other non-conventional integration,
and that's happening at multiple different levels.
So, you can choose any subside or
of power generators which are available.
Each of them would have
different pricing for per unit generation,
and they have different availability times
and different constraints.
So, taking all these constraints into account,
how can you fulfill your demand in
real time while minimizing the total cost of generation?
So, that's happening at individual
household industry level or the utility level.
The interesting part here is you
need to make the decisions of
which generations to switch on or off in real time.
So, this requires very much computations
which are as close to the generators as possible.
So, if you are let's say collecting
data from wind turbines and solar together,
obviously, you need to collect both data and make
the decisions on the cloud.
But let's say if you're doing
these decisions for the house,
then it's good to make the decision at
individual house level rather than
collecting data at the cloud and making the decision
because of the latency and network issues.
>> [inaudible] what's the cost of each liters
that you need documents of it?
>> So, basically, it's called levelized cost of energy,
which takes into account additional costs of
maintenance and all other things
over the time of whatever period.
So, now, if let say we are taking into
account batteries levelized cost of energy,
you need to take into account how much battery life is.
So, whatever garbages that you are putting in,
you would be expensing it over four year's time.
Depending upon how many
charge and discharge cycles you make,
the battery life is going to go down.
So, that is another constraint on
when you decide how much to
charge and discharge the batteries,
that how many charging-discharging cycles
that you are taking into account.
Each of these resources
have similar kinds of constraints.
So, there are some times
when wind turbine if the wind is too low,
it would still generate energy but it
would run very sub-optimally during those times.
So, your levelized cost of
energy during those times would be
higher so it might be better of to run
some other generator rather than running it on low.
So, here, basically the problem is
our forecasting generation from each of these assets.
So, generally, now even utilities in India have started
posing every plant which
is more than five megawatt capacity.
They need to provide one day in
advance that how much generation
would be happening from the plant
at every 15 minutes time slot.
Then there would be revisions during their day.
Similarly for small houses,
if you are running it for
the individual small house or a commercial entity,
we need to know how much generation is
expected from the sources in real time,
a day in advance so that you can do
the optimization and if there are changes,
you need to be able to adjust to those changes.
The second is similar to the generation forecasting,
you need to do the load forecasting.
So, how much demand is coming from
your residential or commercial place
and how would you be able to cater to that demand.
Next is optimization problem to
comply with the policy. So, yeah.
>> [inaudible] ?
>> So, it depends upon how dividing is done inside house,
but that's something which is possible.
So, because of cost constraints on something,
generally, some people do that,
let's say solar power is supplied to
only few appliances and
grid is supplied to all the houses kind of a thing,
but the combination and mix is possible depending
upon the wiring that you do of your house.
Again, depending upon three-phase
versus one-phase loads and
other things there are few constraints.
>> [inaudible]
depend on the time of device?
>> Yes. So, there are two types of that.
So, there is one is
the consumption tariff where
the consumer is charged depending
upon how much amount of
energy user is consuming at specific time of day.
Then there is other type of
tariff which is called feed-in tariff,
so depending upon how much amount of
energy you are injecting into the grid.
Sometimes it's constant throughout the day,
sometimes there is a time of use as well.
So, this is the other problem
where depending upon the pricing
and the policies which take into account
the additional constraints of
let say levelized cost of charging,
discharging batteries, and battery life,
the optimization needs to be done on the grid.
So, this is a problem which we are working on right now.
So, right now, we do the head forecasting
and push the results to the box or Edge.
Then they do small optimizational,
simple optimization if there are any changes.
So, we are basically working on
how we can do the revisions on the box itself,
because if there is network disconnectivity
or other issues and how we can
basically provide the optimization
and mode of this matter on the Edge.
So, there are penalties associated
if your generation is not correct
beyond 15 percent for each of the 15 minutes point.
So, these are the policies in India right now,
and we need to be able to cater to those.
So, another problem which is not on the Edge but which we
are trying right now is this is
the Microsoft Research Building.
So, if we want to have
larger adoption of renewable energy resources,
we need to be able to identify the user,
need to be able to identify how
much is the potential of solar on
my particular roof and
how much amount of energy that can be generated.
So, basically, this identifies
the area of the roof and classification of the roof.
So, depending upon the dates,
the south-north roof or it's
the east-west roof or it's a flat roof,
the capacity of solar that you can put
on in then and its all our potential word body.
So, we are doing some work on
the image optimization on there,
and there is already some open source work on
that core loop and solar map if somebody is interested.
That was mainly for the French datasets.
We are looking at Indian datasets for this.
Finally, the efficient operations.
So, this is mainly about
the predictive maintenance and root-cause.
So, since we collect data from
various appliances in solar,
there are different types of
failures which are possible across various sensors,
including weather sensors, inverters,
MPPTs, and other things.
So, we tried to identify proactively and
reactively those faults and alert the users.
This is the cycle for data
collected from the IoT platform right now,
and this is how the Real-Time Portfolio View looks like.
So, there is something that we are doing
like online offline module,
where the central user wants to look at the entire data.
So, if I'm the independent power producer
who has 25 plants across India,
I would want to see in real time how
it's performing over my entire set of plants.
But many are times this large plants
are at remote locations.
So, there are always
problems of network connectivity and other things.
So, if the entire model is sitting only on the cloud,
the site engineers who are siting at the ground,
they are not able to access this system as well.
So, we have something like offline sync model,
where it's a subset of functionalities
available on the fog at the plant itself,
and opportunistically, it swings
with the cloud datacenters.
Then identifying various under-performing assets.
So, we have shared some data with [inaudible] team as well,
so we would like to explore
some of the techniques that they have used here.
So, this is the cleaning schedules.
So, cleaning is one of the important issues in solar,
which drives the inefficiencies.
So, without adding any additional sensors,
just like looking at the power consumption traits,
we are able to identify when is
the cleaning required and how much is assaulting losses.
Again, this one we were doing on
the cloud right now and we're trying to push
it in the Edge as part of
the workshop experiment that we're doing.
The last one is combining time-series and image data.
So, one is we are getting this time-series data of
power consumption and other things from
individual equipments.
But we're also trying to get some data from
either tomographic imaging or the normal cameras,
so how we can combine these two to
identify the right fault and decide what type
of dust it is or what type of soiling it is and
activate the auto cleaning mechanism appropriately.
So, there are a lot of interesting, other startups,
and other corporates which are working on
energy problems using
AI and IoT technologies across the world.
We'll be happy if there are any problems which are
of your interest and we'll be
happy to collaborate with you. Thank you.
No comments:
Post a Comment