Monday, January 28, 2019

USA news on Youtube Jan 29 2019

SIMONA ONORI: Good afternoon to everyone, and thank you

for coming here today.

So I would like to thank the organizing committee

for inviting me here and giving me this opportunity to talk

about the research I'm conducting with my students

at Stanford.

And in particular today, we're going

to look at how we can design, now

we can use batteries better through physics

based modeling and control.

And in particular, we're going to look

at automotive applications.

So there are so many things here that we

are going to need to know to understand what we do.

So we are going to talk about energy storage systems, physics

based modeling, control theory, and automotive batteries.

So I did some homework, and I realized-- noticed

that there hasn't been much control covered in this seminar

series.

So there's going to be some equations,

some mathematical modeling.

So I'll try not to go too deep, but it's important for you

to know what type of tools we're using

to understand what are also the challenges

to develop those algorithms.

So let me start off with describing what

is our model-based estimator.

So we have a system, and this system

can be, in this case, a battery, battery cell.

And out of this system, we want to recognize

what are the inputs, outputs, and states.

So the inputs are the what we call actuators.

Those are the signals that we use

to drive the dynamics of a system to our reference points.

Sensors that output are the signals

that we measure through some sensors, some hardware.

And then inside this box, we have

dynamics that evolve over time, which we model through states.

Now an inverse problem is really to try

to reconstruct and know the state by information

from the outputs.

What we want to do is something a little bit more sophisticated

because what we want to try to realize in

the build is an observer, an algorithm, that

from measurements, from the output measurements

and from the input measurements of the real system,

was able to reconstruct to the state of the system itself.

And we also want to guarantee that this estimate converges

to the real volume.

And we do this by, first of all, proving

that the system is observable.

So we want a structural property that we

need to require for the system to have in order

to reconstruct the state.

And we're also going to use feedback.

The feedback allows us to really guarantee

the convergence of the estimates to the real states.

And so there are some important choices that we have to make.

And the first one is what is the model that we can use

or we should be using and what is the algorithm we want

to use for the observer design.

And so many of these are the two main choices.

So we're going to look at in this talk modeling,

and we're going to also show some new observer algorithm.

So energy storage technologies-- today

when we talk about electrochemical energy storage

technologies, we think of lithium ion batteries, right?

But there is more than that.

And one way to summarize the available energy storage

technology current, you know, from the past, also the ones

that we're going to have in the future,

is by using what is called Ragone plot.

Ragone was a professor at Carnegie Mellon at the time he

published this paper in 1968.

What he did was to propose a way to describe energy storages

in terms of two parameters--

specific energy and specific power,

specific energy on the y-axis and specific power

on the x-axis.

And those two metrics were chosen because in his work,

he was looking at selecting systems for electric vehicles

at the time.

He was ahead of his time.

And so specific power is ready to acceleration.

So how quick we can extract the energy out of our devices

basically is a means to describe acceleration properties.

Specific energy gives you an idea or a proxy

of the range you can drive.

And so this is a log log scale.

And what you also can notice here,

those are those diagonal lines.

And those are-- basically, they tell you how fast

you can discharge your device.

So they tell you the time scale of your energy storage.

So super capacitor, capacitor here,

there are devices that you can discharge and charge very fast.

Legacy batteries, lithium ion batteries,

batteries [INAUDIBLE] higher specific energy.

So they go through some electrochemical reactions.

So they are a little bit slower than capacitor.

You can see here that a single device is not

represented by a dot.

It's not located in one particular position.

But there is a curve.

What that says is that the energy decreases.

[INAUDIBLE] more power, you want [INAUDIBLE] out of it.

And so this is an important consideration for the design.

You can also use this Ragone plot to-- and this is what

the Department of Energy has done--

to tell you what are the targets of your energy storage

system in terms of application.

So for EV, all electric vehicles,

plug-in electric vehicles, and self-sustaining hybrid,

those are the requirements for the energy storage.

At the end of the day, we want those devices

to be as close as possible in terms of performance

to the internal combustion engine.

So internal combustion engine is not an energy storage device.

It is an energy conversion.

But you can extract very--

not efficiently, but it's very effective in extracting

and converting the chemical energy in the fuel

into mechanical energy.

So at the end, so it's not just about--

it's a matter of cost for electric vehicles

to be suitable solution, but it's

a matter of having the right energy storage component that

is performing as close as possible

to the internal combustion engine that is used today.

Keep this in mind because you're going to see this later on.

The P to E ratio is very important.

It's [INAUDIBLE] back to those diagonal lines, OK?

It's also called, in the battery world, C rate.

OK, so energy storage--

vehicle technologies.

So today, we have lots of options in terms

of vehicle power trains.

And sometimes, we don't really know what those are, right?

And so one way to plot and to see--

look at them is by describing them

in terms of size of internal combustion engine with respect

to size of electron models.

What that allows you to do is to travel

along this route, this line.

And as you go from here to here, the degree of electrification

increases.

So the electric motor becomes bigger.

So the energy storage becomes bigger.

And at some point, you're going to get rid

of the internal combustion engine.

So the common denominator of those powertrain technologies

is the energy storage.

And I'm not saying that it's battery.

I want to say it's energy storage

because one of the flaws in the design, powertrain design

today, is that we go straight blindly to think

that the lithium ion is the system to be used,

which is probably just not the right approach.

OK, so lithium ion batteries are what

has allowed this wide range of options to exist, OK?

Without this system, we won't be talking today

about electric vehicles.

And so good enough in the late 70s and 80s

showed that we can realize high energy density lithium ion

batteries by introducing a transition metal

oxide [INAUDIBLE] the cars.

And so you could get very high voltage out

of these lithium ion batteries.

And this has revolutionized the portable electronics industry

at the time.

So the first lithium ion batteries

was commercialized by Sony in 1991, 1991.

Today, lithium ion batteries are revolutionizing

the automotive industry.

And then what's happening is that the electric utility

sector is being impacted and as we have seen from PG News

last week by these technologies.

So it's a big deal.

So we need to understand.

We need to know how it works.

And eventually, we also want to learn how to model it.

So lithium ion batteries, they work

based on the redox reactions.

So there is a reduction of oxidation reaction happening.

And so we have materials at the anode that have been selected.

And so lithium ion batteries, let me tell you

that it's a very nice marriage because it's a donor acceptor

couple.

So anode donates electrons, and the cathode

is made of a material that accepts electrons.

So it works very well.

And as the electrons basically travel along

the external [INAUDIBLE],, the ions basically

get intercalated from the anode, and they

pass through the separator into the electrolyte

and go to the cathode.

Those are also called rocket charge battery.

That's just because the lithium basically

go back and forth between the two [INAUDIBLE] electrons.

And the separator is important because it

allows ions to go through it, so but it prevents electrons, OK?

It prevents the two electrodes to touch.

It prevents short circuiting [INAUDIBLE] away and so forth.

OK, so that's lithium ion battery.

So have you ever asked yourself what is a battery?

So sometimes it's not clear.

Some batteries is-- when we make lithium ion batteries

for commercial use, what we do, we

make those cylindrical system.

You know, they're very small.

I should've brought one with me.

But those little guys, they have a 65 millimeter length

and a 10 millimeter diameter.

They also can come-- lithium ion battery cell can

come in the form of patch cells of these type.

They all have a nominal voltage of four volts.

But we don't make huge batteries.

What we do, we create a battery pack

by connecting lithium ion cells in series and parallel

to create a pack that has the voltage and the power that

is needed in your vehicle.

So those are traction battery pack.

Tesla uses the cylindrical cell.

Nissan LEAF uses those pouch cells and so forth.

So the reason why I'm talking about these cells versus pack

is because we are going to look at later on some

of the characterization of those two different domains in which

lithium ion batteries work.

OK, so because of this structure and because

we have many of those simple systems--

relatively simple system in a battery pack,

we need to make sure that everything is working properly.

And that is done by the battery management system, BMS.

And that is basically the equivalent of ECU, the Engine

Control Unit.

What it does-- gets information through sensors,

some measurement.

In particular, we sense measure current, voltage,

and temperature.

And this information is used to prevent the battery

to be overcharged, over discharged,

to go through some short circuit, and to overheat.

This is very important.

And there are many tasks, many functions,

that we ask the BMS to perform and by only using

those measurements.

And so we want them to get the right measurements.

We want to control with the temperature.

We want to balance the cell within the pack

and make sure that they're all equally--

they're all the same amount of energy.

We don't want any imbalance within the pack

for safety reasons.

Monitoring the state of charge, which

is a proxy of-- which is basically information that gets

translated into the driver language

in terms of how many miles I can still drive with my car

with my battery.

We want to make sure if something goes wrong,

we have a means to detect that and to warn the driver.

And why not, you also want to do prognoses.

We want to be able to not only say how old my battery is,

so that's the state of health, but I also

want to know how long in the future

I can still use my battery.

So that is prognosis.

And all those functions are today implemented in BMS.

Prognosis, not yet, but it will very soon.

So what are doing our lab is to design advanced BMS.

And so what that means is that, yes, we

can rely on those measurements.

But we can do more.

We can create virtual measurement, so virtual sensor

by means of estimators.

So estimates are used by us as a means

to look at what's inside--

so to track and monitor the concentration variations,

the overput engine, and some other important and critical

parameters of battery that's also related to age.

And so physical inside are important,

but also the control theory is important to develop estimator

that allows you to rigorously understand

[INAUDIBLE] convention to the right value in the parameters.

And so what industry uses is this safe operating area.

So what you can see-- this is a temperature

versus open secret voltage diagram.

And what you can see is that the battery is only

using this this green area.

So it's overly designed for--

and the design is very conservative.

And that's because we don't want to over-discharge and--

overcharge and over-discharge.

We want to stay away from those edges

because the degradation is going to be more pronounced there

and also because it's safer to be within this.

But the other reason is because we don't really

know when we are here and where we are here.

So that's-- we're being even more conservative.

So if you think about the available power

that you can extract out of the battery in terms of voltage,

that's the only window that you can use, OK,

if you want to stay within the normal function

area within the battery.

So this is the open circuit voltage response.

And so we all operate between usually 20%

and 80% state of charge.

So the rest is not being used at this point in time.

So as I said initially, one of the critical questions

that we have to ask ourselves when we go and develop

an estimator is what is the model we can use

or we should be using.

And so if you look at battery cells, battery,

battery is a multi-scaled system in time and length.

So things that are happening, dynamics

that involve the particle scale, atomic scale,

don't really matter at the system

scale for on-board real time execution.

They do matter for design but not for real time application.

And so it's a tough choice, OK?

We don't really know where to start and what to do.

But that's a very important question.

So if material scientist and electrochemist and chemical

engineers, they work very hard to design better batteries,

it's our job as control engineers

to really be able to extract the maximum performance.

We don't want to waste anything.

And so if we are given a battery with a nominal capacity

of certain values with nominal potential,

we want to be able to really get the most out of it.

And now it can be done through some [INAUDIBLE]

model-based estimation theory.

So that's what we do in our lab.

And there our purpose is to really link science and control

engineering-- so to really understand it

and bring what happens at the small scale up to the system

level scale.

And what you see here is a schematic of a three way

catalyst and the gasoline particulate filters.

It's not a mistake.

Those are the two types of systems

I do research on with other students when

I don't do battery work.

But those systems, they have lots in common,

believe it or not.

And so the main important job is to derive and construct

physically meaningful control-oriented model

because then, we can go ahead and do our advance

model control development and also vehicle-base optimization

and so forth.

So modeling, battery remodeling--

there are mainly two groups of modeling.

One is physics-based, the other one empirical

or control-oriented.

And different people have used to the two groups of modeling

for different purposes.

We do use physics-based modeling for design.

So if you want to design batteries,

you want to change some parameters like porosity

or [INAUDIBLE] and you want to see the performance

at the system level.

For a runtime stage, so for on-board application,

we use control-oriented models.

And these two types of models have been very, very separated

until recently.

And we all know also the model fly, right?

They are wrong, but what we are after here

is to try to understand how trustworthy a battery model is

because the way I see it is that all models are

good if you understand their limitation

and what is that you want to do with it.

So those empirical models, also called equivalent circuit

models, are very good and are today

being very used for BMS implementation

if you want to extract information like power

and predict the voltage response.

The problem with those models-- they

are very easy to run in real hardware,

but they rely on a very heavy calibration campaign.

So those parameters here, they are chemistry-dependent,

and they need to be calibrated across different temperature

ranges and C rate instead of charge and so forth.

But they are yet very, very, very useful.

On the other hand, there is a new-- a different class

of models called physics-based models.

I told you that what they do, they

try to look at the concentration of [INAUDIBLE]

within the particles within the electrodes

and also predict the electrolyte dynamics and the [INAUDIBLE]

potential and so forth.

And so going this way we find in the single particle

model and the [INAUDIBLE] single particle models and also the

[INAUDIBLE] electric models are called the DFN or P2D model.

Now these models increase in complexity as we go this way.

But they also increase in predictability.

So we get better and better response out of this model

as we add dynamics.

And so in this particular case, the single particle model

looks at the electrons as if they

were two spherical particles.

And we basically track the concentration of lithium

through diffusion mechanism inside those particles.

And the enhanced particle models look

at the dynamics of the electrolyte

and the P2D [INAUDIBLE].

And still we make the assumption that they

have two spherical electrons, two spherical particles

in the electrodes.

And then here, basically this assumption, one electrode

is basically removed.

So these are PDE, Partial Differential Equation

model, base model.

And they are based on mass charge consideration

in the solid and electrolyte phase.

And so the question is, can we use

this model for control-oriented purposes,

and also, is this the best model to use?

So the DFN model is the model that's

been commonly used for design purposes, but also for control.

And one of the criticisms that I need to move to my control

community is that people have taken this model for granted,

and they have applied for traction boundaries

in automotive applications.

And this business has created some [INAUDIBLE]

in terms of accuracy.

So if you--

I want to go we started looking at this modeling approach

with a colleague from my department Ilenia

Battiato and my PhD Harikesh, who is now with Faraday Future.

We developed a micro scale modeling framework

and a new micro scale modeling framework

for lithium ion batteries.

And so I want to talk about the limitation of this model.

Again, this is the one that is used today by everyone--

control people and designer.

So a new one developed this model, this micro scale

continuous model, in 1993 and a couple of years

after the first lithium ion batteries was introduced.

And it's based on very strong assumption.

And assumption is that the particles within the electorate

are spherical.

It's very strong.

And also there's no way to accommodate this model

for different shape of those particles.

Another limitation is that when [INAUDIBLE] develop

this model with this student, lithium batteries were being

used for portable electronics.

And for those applications, the power rate is very low.

Plus, another issue is that you wouldn't

mind to replace your battery every two years in your laptop,

right?

You do mind though replacing your battery every two years

in your electric car.

So that is not even an option.

And so this model does not capture any aging dynamics,

and it does not work well, properly,

at a high [INAUDIBLE] rate of operation.

And it also doesn't capture what happens

at the low state of charge.

And so those are very critical, important elements

when you want to design a traction

battery for electric vehicle.

And the literature shows some of the inaccuracy and divergence

of simulation from Neumann model at IC rate and, you know,

towards the end of discharge.

So lower state of charge, capturing dynamics

at lower state of charge is very important

because it gives you the ability to really understand

what's your effective range.

IC rate is important for fast charging and even more.

So if you look at the scanning electron microscope images,

you see also that for different chemistry,

a nickel manganese cobalt and lithium cobalt

oxide and graphite and so forth, those particles

are not spherical.

And if you're trying to use the Neumann

model for those different chemistry,

you don't necessarily-- you won't get an accurate answer.

And so we develop our model, and I

don't think I have the time to go into the detail.

But the main difference is that Neumann model

is based on the volume averaging approach

while my colleague [INAUDIBLE] is an expert in homogenization.

And with her, we developed this new model.

And what we did was to take a unit cell it

within the porous electrode and define this [INAUDIBLE]

separation [INAUDIBLE] separation parameters.

And then we basically--

we did asymptotic expansion technique.

And rigorously and theoretically,

we have obtained a continuous time scale

model which is a 3D model.

The Neumann model is, on the other hand, is P2D.

And P stands for Pseudo 2D model.

And pseudo is because we have spherical coordinates

within the diffusion in the electrodes and then

the x-coordinate for the concentration gradients

within the electrolyte.

And so one of the main differences also between those

two models is that in the electrolyte mass transport

equation, we do have electro-migration mechanism--

so the potential gradient that affects basically

the electrolyte mass transport which

we don't have in the DFM model.

And there are more differences within these two approaches.

An important aspect that we're investigating right now

is the calculation of those effective parameters,

average parameters-- so diffusion and the conductivity.

So the [INAUDIBLE] model, what it does,

it calculates those effective parameters

by using an empirical law, which is the Bruggeman

law that everybody is using.

That is based on a very rough approximation

of some of these parameters.

And it's based on the knowledge of the porosity

and uses this coefficient equal to 1.5

to calculate those effective parameters.

Now if you look at different chemistry and Bruggeman

coefficient that's been mapped over the three

different dimensions, and z is our basically perpendicular

direction, the one across which we do the modeling.

And you can see that for graphite, this coefficient,

this exponent is almost three.

And so if you have two different chemistry which

have the same porosity value, eta, and you use this value,

you come up with the wrong answer

because effectively, the real one is higher than 1.5.

And so those are some of the flaws within the DFM model

that haven't been really addressed by the community.

What we do, on the other hand, is to solve a closure problem.

And that allows you to bring the pore scale information up

to the micro scale or continuous scale.

OK, now in this particular case, we

solved the crucial problem using the spherical particle

assumption just so that we could compare our results.

But the model allows you to really look and change

the particles' size.

And [INAUDIBLE] is looking at this particular work right now.

So this simulation was performed by a research scientist Slava.

And what he got is that there is a 20% difference

between the effective electrolyte diffusion.

This particular case is obtained from the closure problem

with the Bruggeman.

And so we keep working on that route.

At the same time, we have implemented

our full homogenized model.

And we have identified parameters,

and we have set up a cross simulation environment

using MATLAB and [INAUDIBLE] with the physics software.

And the results are that we obtain--

so we do experiments at different temperatures--

same C rate of discharge, 1C, but for four different

temperature--

23, 40, 45, and 52 degrees C. And what we see

is something that we already kind of saw

in the literature is that Neumann model really

starts being less accurate towards the end of discharge

her and even more here.

And also we plot the voltage error

as a function of [INAUDIBLE] and temperature.

So this [INAUDIBLE],, we are going

to present with [INAUDIBLE] these results

in two weeks at the CDC.

So we are doing additional experiments

to identify the models over different C rate of operation.

And the way we see is that this model can really

be used to better designed batteries,

especially for this application, and even more for green storage

application where you really want

to go down to a lower state of charge of charge values.

You want to be able to really have meaningful and accurate

information.

We can use this model also to generate

proxy data or synthetic data for machine learning algorithms.

And even more importantly, we want

to use a-- we are looking at this right

now how to reduce this model and make it

a control-oriented model and for real time estimation.

So as we do this work, we are also working on--

the meantime, we work with my patient student

and you do it on model ID and estimation.

And so the problem here is that we

can't measure state of charge.

[INAUDIBLE] have to, but we really

want to know those two parameters very well.

They're very critical.

And so we go ahead and after the design of [INAUDIBLE]

chemical model based estimator.

Now this has been done by a few researchers.

In the world.

What we found out is that there is a problem

of weak observability there.

And so the reason because this is due to this open circuit

potential at the anode.

Basically, it's pretty flat.

Now when you look at the voltage across the battery

terminal, what you do is do the difference between those two.

So this weak observability problem

has been addressed in different ways in the literature.

But all these ways have some disadvantages.

So some of the approaches would not

allow the observer to be extended

and to predict the aging because they

are based on an assumption that leads

to most conservative which is not true

when the aging happens.

In this particular case that you're going to look at after,

we assume to know very well the concentration of lithium

at the anode through an open loop observer

that we assume to initialize well.

And then the third approach requires

the additional extra temperature sensor

to impose some relationship that will enhance observability.

So we wanted to remove all those limitations

because what we want to do, we wanted to build and define

an observer that we could use to estimate concentration

of lithium at both electrodes that we

could use to estimate aging without adding

an additional sensor.

So we look at the-- we're starting

from the enhanced single particle models

because we wanted to be able to capture also

the electrolyte dynamics for fast charging eventually,

eventually for fast charging application.

And so the key here is to use finite difference approximation

and rewrite the partial differential equation

and to rewrite the system in terms of states

[INAUDIBLE] concentration lithium at the two

electrodes and the electrolyte and also

in terms of output equation, which is a non-linear output

equation as a function of the concentration of temperature

in the electrolyte resistance.

So that's the system we are working with

to develop our estimator.

So before we go ahead, we can go ahead develop the observer,

we have to identify parameters.

There's some parameters which are unknown,

and that is done through some object optimization

by setting up a single object in this case,

but we also look at multi-object optimization problems.

And then we went ahead and proposed our scheme.

So this scheme was the one that we developed with Alex in 2016.

Alex is now [INAUDIBLE] now a [INAUDIBLE]..

So here's the deal with that, OK?

We have one electrode.

We assume that one electrode is perfectly known,

which is the anode.

And therefore, the cathode, we can develop a close loop

observer.

Now we want to get rid of this assumption.

And so what we did was to propose

a concurrent interconnected adaptive observer.

And what it does is that basically, we

estimate in closed loop the concentration of each electrode

separately.

And we assume that the other electrode

is given an open loop.

But the reality-- meaning in practice

is given from the other observer in closed loop.

And so this mechanism works, and it

allows you prove the convergence of the estimates

to the real concentration values.

And so could prove asymptotic stability.

So this is just an idea, very graphic idea, of what

asymptotic stability is.

So if we define the error as being

the difference between the real concentration

values and the estimate, and so this

is the initial error-- initial values of the error.

So after some times we can show that this error goes to zero,

OK?

And that is possible through some design.

So if you set up the cathode server, the anode server,

as I mentioned earlier, in an interconnected manner,

we can show through some [INAUDIBLE] stability theory

that this conversion happens.

And This is some--

so we simulate UDDS cycle here.

And then, you know, we have the blue ones here.

The blue dots represent the real concentration of the two

electrodes, and the red stars here

represent [INAUDIBLE] initialized concentration.

But then we showed we basically convert to the yellow one.

So that's fine.

We did that.

But what that observer didn't include

was we couldn't use that observer to predict the aging.

And we do really want that because that's

the most important critical parameter that we

want to track in real time.

And so when it comes to the aging,

the degradation happens in many, many ways.

There are many interconnected processes

happening within the battery.

But it manifests itself through symptoms

which are basically related to capacity fade and power fade.

Very simple-- so basically, you need

to charge your car more often, and you

won't be able to extract that energy as quickly as you

could initially.

And so among the--

it's a very difficult problem to model.

And so what the literature proposes and shows

is that the SEI layer formation at the anode

is the main aging mechanism that basically

is behind the capacity fade and power fade of the battery.

So the SEI layer means that to the solid electrolyte interface

of the anode is basically--

is this can be represented as this protective layer,

if you will, that basically creates around the anode

during the first cycles of the battery.

So the reaction between the electrolytes, that's what gives

rise to.

The problem with this-- so there is one good point of SEI layer

formation, SEI layer formation, is

that it prevents corrosion of the anode.

But there are so many negative characteristics,

and the main one is that, first of all, it--

so lithium, cyclable lithium, gets tied up within this layer.

And so it's not available for reaction.

And so you have loss in capacity.

Plus, this layer increases in thickness,

and that prevents the ion basically to travel through it.

So that originates increase of resistance or power fade.

So we convince ourselves that SEI layer

is the more important mechanism to model.

And we connected, we linked, this variable to some

of the parameters within the model,

in particular the diffusion coefficient and the also

porosity in the anode and as well as

the stoichiometric window and also the transport parameter

in the SEI layer.

And so these parameters are modeled in basically our model.

And so by tracking adaptive [INAUDIBLE] those parameters,

we can predict capacity fade.

So that's what we did.

The other things that we did was to also include

uncertainties in our model.

So as I showed you earlier, there

are many different types of models.

So whatever you pick, you know that you're

going to make a mistake with respect to what

the real experiments predict, how other high fidelity

models can give you.

So there's always going to be some uncertainties.

So we include that uncertainty in our observer design.

And what we went after this time it was not asymptotic stability

because now we have uncertainties.

But we could show-- we are showing actually

practical stability.

So what we're seeing is that the error between the two estimates

of the states and the real concentration

and also the parameters will not converge to 0,

but it will stay within a [INAUDIBLE] whose

radius is the function of the uncertainties.

So the better you are in designing your model,

the more predictive you are in designing your model,

the better you can also--

the better the performance of your estimator is.

And so we elaborate the previous estimator scheme

a little bit more, and now we have accommodated

the adaptation of the two parameters, [INAUDIBLE]

one [INAUDIBLE] two-- so that the diffusion

coefficient, the transport coefficient, [INAUDIBLE]

electrolyte.

And we did some math.

We prove our practical stability.

And if I can run the simulation--

no, I can't.

OK.

OK, now?

So the this business of adaptability

of estimating those parameters allows

us to also estimate capacity because as the battery ages,

those parameters also basically change.

And that will affect concentration as well as

capacity.

So now we have an observer that not only estimates accurately

with practical stability property state of charge,

but also capacity.

And to the best of our knowledge,

something of this type hasn't been developed yet.

So our next step is to implement these [INAUDIBLE]..

In order to make sure--

to show the validity of real time implementation

of this algorithm.

So battery pack modeling-- initially,

I show you the difference between battery cells

and battery pack.

And so why is that important?

This is a research topic that I started

working on with two visiting scholars, Stefano and Carlo.

And now Anirudh is also full-time working

on this research.

So it's very simple.

The premise is very simple.

Battery cells are not created equal.

And so when you have 10, you know, hundreds of thousands

of those cells that you put in a pack,

something might go wrong, right?

And so the literature shows that there are manufacturing

variances within the cell.

In this particular case, this is the impedance

within the imaginary and the real axis plane.

Impedance of cells, they're different from cathode

thickness and the cathode practical size.

So we also see there are evidence

that shows the temperature of the cells changes,

and the thickness of the electrodes changes.

And so those are again manufacturing gradients effect.

In this study, what they've done, they have two--

basically, they have taken a bunch of cells, [INAUDIBLE]

from [INAUDIBLE].

And they have assessed the initial capacity

of those cells.

And then they let those cells evolve and degrade

and basically age with the same cycle.

And what happens after, you know, 2,000 cycles,

that the trajectory of the capacity has changed.

It's diverged quite a bit.

And then you know, if you look at the Instagram,

there is quite a big of a spread there.

And so on top of that, this study has shown that

in a module of 10 cells in series exposed

to different temperature, so it's

exposed to a different basically temperature gradient,

the agent of the cells basically evolved.

Basically, there is an induced aging of the overall module.

And these last studies were very recent--

shows that these people from Argonne National Lab

have used a battery pack from Nissan LEAF.

And they have tested those packs as well

as they have tested the cells within the pack independently.

And what they've shown is that backpack aging is more

pronounced than cell aging.

So something happens within the pack

that we don't capture at the cell level.

And so if you think about it the majority-- maybe you

don't know, but the majority of the modeling and estimation

work is at the cell level.

I will say that 90% of the papers

today look at the cell level estimation.

And so what we try to do is to understand

what happens at the pack level.

And the way we look at that is that

by saying that the modularity doesn't hold true

within a pack.

So you cannot infer or basically derive the pack level

properties from single cell, OK?

And so modularity, we want to show--

or we thought that modularity is not

preserved upon interconnection of battery pack.

And there is a very nice theory that

has been developed within the synthetic biology

field that looks at what happens when you interconnect system--

so cell in a big module.

And some application, what people have shown

is that when you do connect system in this fashion,

not only does the system downstream--

not only is system downstream is affected

but the system upstream of it.

But also the system upstream of this interconnection

is affected by the system downstream through

this reactive signal.

So something goes back here that you

don't consider when you basically

apply the modularity principle.

What this is called is retroactivity to the output.

So the dynamics of the system--

in this case, downstream, upstream--

not only are affected by whatever inputs we have

but also whatever is basically used

in interconnection settings.

And so we use this idea, and we do some experiments

at the time.

And then we drill the battery.

We look at the core temperature of the battery,

and we measure it.

So that's-- the blue one is the core temperature of a single

battery.

And then we did some experiments by connecting those cells

in series.

And what we observe is that the battery temperature was higher

in this setting.

And we used the same cycle.

And this was just due to the interconnection.

And so we wanted to prove that we had the retroactivity

happening there.

And so by running the thermal equation of a cell which

is connected in series to two cells, one upstream and one

downstream, we can recognize and see

that there are two terms here that--

what they do, they are acting as the retroactivity term.

So the temperature of cell B is not only

a function of the temperature, the skin temperature,

the core temperature itself, but it's

also a function of the temperature upstream, TA,

and the temperature downstream, OK?

And so the environment affects the performance

of the cell in terms of--

and how that happens is through these retroactivity terms.

And so now if we do, you know, our study on a single cell,

I realized that these stress factors, the aging factors,

are merely set of charge, [INAUDIBLE],, and temperature.

With interconnection happening, we

need to account for these thermal interconnections,

those thermal retroactivity terms that

induce some extra dynamics in the system

and also affect the aging of the system.

And so our hypothesis, the one that we're investigating,

is that thermal retroactivity is responsible for aging

to propagate along the cell.

So aging propagates because of thermal retroactivity.

That's the concept that we are developing.

And when you do that, so one proof that that is basically

the right things to do is that if you

look at the rate of change of the SEI layer, SEI layer--

so that is the parameter that represents the age

that we've discussed earlier.

You see that the SEI layer growth of cell B

is not only a function of the parameters of the cell itself

but also the parameters of the adjacent,

you know, neighboring cells through these thermal dynamics.

Now the modeling framework that we are developing

is based on the singular perturbation problem, singular

perturbation approach.

We are recognizing that the model of battery

is evolving across three different scale,

temporal scale--

temperature, concentration, and aging.

And so what we see is that the aging needs to be--

it's enough to look at the aging dynamics on this slow manifold

because the temperature dynamics converge to this slow manifold.

And so that's how [INAUDIBLE].

So this work has lots of implications

because it allow you to accurately estimate

battery state of health.

And that also allows you to mitigate aging during fast

charging within the pack.

Have you seen the work from [INAUDIBLE] National Lab?

It's shown that the battery pack is aging

much faster than single cells.

So if you can somehow mitigate this retroactivity terms

and this [INAUDIBLE] controller that somehow imposes or brings

back that modularity that you wanted

to apply in the first place, that's the way to go.

And that can be done through some thermal control,

thermal management control.

Last application, going to be quick here,

is hybrid vehicle optimization.

And this is a project that I'm doing now

with the army, so Denis Rizzo.

And Zifan and Abdullah were the two postdocs

that were on this project.

And they're now at Bosch and the FCA.

So going back to this vehicle technology,

powertrain technologies options--

you remember that.

Now if you look at what is the requirement of the energy

storage within these vehicles, what

we can see is that in terms of C rate or P to E ratio--

this is the P to E ratio that we defined earlier.

The P to E ratio of the energy storage is different, OK?

It increases as degree of electrification increases.

So you need more battery as the P to E ratio decreases.

So you need more energy than power, OK?

But for micro hybrids, for example,

you need to be able to really use that energy very quickly.

And so the question is, is the lithium ion battery

derived technology to use across all those vehicle powertrains?

And aren't we probably asking too much to this technology?

We are asking the lithium ion batteries

to be as close as possible to internal combustion engine.

On the other hand, there are other technologies

here that people haven't explored, right?

So what if we share the burden, and what if instead,

we maybe tried to use , for example,

super capacitor and trying to have lithium ion batteries

to achieve this goal?

And so can a hybrid storage solution

be valuable in some applications?

And super capacitors are a very neat technology.

They have very high power density.

They're very fast to discharge.

They' have very long life.

You can get up to 1 million of cycles

out of a super capacitor.

And so in a way, super capacitors

have complementary properties of lithium ion batteries so

what is the best way to electrify those?

Should we diversify the energy storage

and see what is the optimal size?

Those are fundamental questions that we

try to address with our work.

So we were asked to develop some tools for the optimal selection

and size of energy storage.

And so we look at this vehicle, the mine resistant ambush

protected all-terrain vehicle.

It's a pretty heavy vehicle.

And what we did was to--

this is a schematic of the powertrain.

And what we did was to hybridize.

So we still have an internal combustion engine.

It's a diesel engine.

And we want to hybridize it with the lithium ion batteries.

In this particular case, we did choose

the nickel manganese cobalt chemistry.

And we also did hybridized with the hybrid energy storage

system-- so a combination of batteries and super capacitor

connected throughout a DC/DC converter,

so in semi-active configuration.

And we use in this case, for vehicle optimization,

we use empirical models that we validate experimentally.

So those models, again, are simple and very,

very effective with this type of work.

And we did that.

We formulate an optimization problem.

The optimization problem has been formulating [INAUDIBLE]

way that we wanted to have on the same cost

function both the energy management,

the power split variables, as well as the design variables.

The design variables are basically

the number of cells of the supercap in parallel

in this series as well as the lithium ion battery.

So these parameters here tell you how big the two packs are.

And these variables here tell you

how you are performing the split in real time,

OK, from the requested power of the wheel.

And so these are the two configurations.

And what we did was to formulate a Hamiltonian function that

was accounting for, again, the design and the energy

management variables at the same time.

So we could solve this problem altogether.

And we defined static states, and we

developed our algorithms.

And the simulations show some very interesting results.

So these are typical.

The concatenation of different driving cycles

that are used by the Army, they're

very different from the driving cycle

we use for passenger vehicle application.

They're very demanding, very high

charge and discharge cycles.

And so our work, our results, have shown that if you use--

so we compare our results with what the army initially

have done.

So they did hybridize the powertrain

by using an ion phosphate battery not optimally designed.

Just put a battery there without any optimal consideration.

By just using a different chemistry

and optimally design the size of it,

we could get a saving in fuel of 7%.

But if you do use supercaps, on the other hand,

with nickel manganese cobalt batteries,

you can get up to a saving of 13% or so.

And those guys here, this vehicle

here in its conventional configuration,

can give you less than five miles per gallon.

So this saving is very significant.

What we also saw, the statue, basically the size

of the supercap with respect to lithium

ion batteries for this particular cycle.

Let me let me just go to this takeaway.

One of the things that we have discovered

is that supercaps have been overlooked.

There is some good reason maybe.

The volumetric density is not high.

So we haven't accounted that in our optimization work.

And so that would be our next step because from a volume

standpoint, they're really very demanding.

The way we want to extend this work

is by trying to give guidelines as

to what energy storage device to use optimum by interrogating

the Ragone plot properly.

And so as you said, as you've seen before,

there is a wide range of vehicles

that needs to be electrified.

And so there's not a one solution fits all, OK?

So there is different requirements

for the applications for different solutions.

The main challenge also that we're trying to solve

is scalability.

How can you scale up and down this solution

across those different vehicles, you know,

where that makes sense?

And so we're also doing some work with my PhD student

Francis, is also now at Tesla part time.

But we are looking at how to identify, out

to address identifiability ability

issues of electrochemical models by using the EIS test data.

So EIS test data, Electrochemical Impedance,

Spectroscopy is a test that you do in frequency domain.

And we try to understand how much information we

can get to solve and address the lack of identifiability

of some parameters in those models.

And I also have the pleasure to work

with Andrew, who is a freshman at Nueva High School.

And Andrew joined our group last summer, and what he's doing

is very ambitious project here.

He's trying to come up with a prognostic algorithm trying

to predict the remaining useful life

from some experimental data collected

on plug-in hybrid vehicles and is

using some correlation in machine learning [INAUDIBLE]..

So what are the next research challenges?

Indeed, grid storage is the next one.

So we have our research tests that we are conducting

in the automotive batteries, but the very next things

that we're doing is to look at how do those batteries behave.

And you know, in a Tesla, the battery pack

has 85 kilowatt hour, right?

So 85 kilowatt hour battery.

Now we are talking about 1.1 gigawatt hour batteries--

much bigger system.

And we don't know what is the duty cycle that those batteries

go through, what type of modeling tools we need,

and how they age.

And you know, there's so many questions

that are there that, you know, create lots of opportunities

for research.

And we are looking at those, some of that,

with Nora, who just joined [INAUDIBLE] this quarter.

And the other point that I want to give

is that we don't need to invent the wheel.

So those batteries are very similar to the batteries

that are used in automotive.

They have differences.

So it's important to understand what

are the differences in terms of usage.

But we can leverage the knowledge and experience

that we have acquired in the automotive field quite a bit.

And that will accelerate the advancement

in the field of grid storage.

And with that, I'm happy to take any questions you might have

and also would like to thank our sponsors without which

I would not be able to do this research.

And thank you so much.

PRESENTER: Thanks, Simona.

[APPLAUSE]

No comments:

Post a Comment