ECE5720: Battery Management and Control 7–1

Physics-Based Optimal Controls

7.1: Degradation as basis for power limits

■ We now approach the frontier of knowledge in battery management:

! The electronics aspects are important, but routine;

! State-of-charge is well defined, and established methods can be

used to get good SOC estimates;
! Similarly, we have seen good methods to estimate resistance and
total capacity of cells, yielding state-of-health estimates;
! Cell energy calculation is straightforward; and

! Several types of cell balancing—with varying complexities and

speed—can be implemented.
■ Improvements can yet be made to all of the above, but the present
state of the art provides adequate BMS for many applications.

! Some questions regarding long-term efficacy of present BMS on

aged battery packs;
! Other issues regarding power calculation, as described in this
chapter.
■ “Using current electronics and knowledge it takes about two years
and $250K to build a custom BMS” [Davide Andrea].

! Not trivial, but very doable.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–2
More power!

■ The big pink elephant in the room that few people talk about is the
way that power estimates are presently calculated.
■ The premise behind these methods is that voltage limits must never
be violated.
■ But why? The real issue is cell degradation. The assumption is that:

! If voltage limits are violated, then the cell will degrade quickly;

! If limits are properly maintained, the cell will have a long and
productive and happy life.
■ But, in fact, voltage limits may be violated for a short time in some
situations without causing any faster aging.
■ And, “normal” voltages may also cause fast degradation in some
situations—particularly for an aged cell.
■ So, real issue is not cell voltage but rather rate of aging/degradation.
■ Cell power limits should really be calculated to more directly optimize
a tradeoff between performance delivered by the cell and the rate of
incremental degradation experienced by the cell.
■ To be able to do this, we must be able to:

1. Model degradation mathematically, and

2. Devise model-based optimized controls to calculate best tradeoff.
■ Some have suggested that if this is done perfectly, battery-pack sizes
may be reduced by up to 50 %, yet still deliver required performance.

! This is ample incentive to make a strong attempt.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–3
Modeling cell degradation

■ We have seen that much is known about cell degradation qualitatively.

■ But, how about quantitatively? That is, can we make accurate
mathematical models of all the degradation mechanisms?
■ We have seen that interactions between mechanisms are complex.
Further, they are not presently well understood.

! Philosophy = “a blind man searching in a dark room for a black cat

that isn’t there.” Are we on the same kind of futile search?
■ At this point, we don’t know. (That’s the nature of research!)
■ But, working in our favor is that we don’t need to model all
mechanisms perfectly to have a useful result.

! For purposes of control, we don’t need to model any mechanism

that is not influenced by a variable that we have control over;
! If we model the most severe mechanisms reasonably well, then we
have a chance at designing controls that make a difference.

Literature that proposes models of degradation

■ The literature on degradation mechanism modeling is quite sparse.

■ Here, we look at two models: SEI formation/growth; lithium plating.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–4
7.2: Full-order model of SEI formation and growth

■ Ramadass and colleagues have proposed a model that describes the

formation and growth of an SEI layer on negative-electrode solid
particles during charging, that uses solvent reduction as the main
side reaction mechanism for degradation (Ramadass 2004).
■ Here, we build on that work to develop a simple incremental model of
SEI growth and associated capacity loss and resistance rise.1
■ The order-reduction method uses volume averaging to create an
algebraic “0-D” model of the infinite-order PDE model.
■ This reduced-order model (ROM) of the SEI growth mechanism is a
first step toward creating a complete coupled reduced-order model of
all dominant cell degradation mechanisms, which then could be used
in an optimal control scheme.

Original model

■ Changes at the electrode/electrolyte interface due to side reactions at

the negative electrode are considered to be one of the primary
causes of cell aging.
■ There are a large number of reduction reactions that can lead to the
deposition of solid SEI products on the electrode surface, and these
are less well understood, being very dependent upon the composition
of the electrolyte solution.

1
Adapted from, Randall, A.V., Perkins, R.D., Zhou, X., Plett, G.L., “Controls Oriented
Reduced Order Modeling of SEI Layer Growth,” Journal of Power Sources, Vol. 209,
July 2012, pp. 282–288.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–5
■ Ramadass makes general assumptions that the side reaction is
considered to be a consumption of the solvent species and lithium
ions, which will form compounds such as Li2CO3, LiF, Li2O, and so
forth, depending on the nature of the solvent.
■ There is significant porosity in the film, and this makes it reasonable
to assume that the SEI layer continues to grow as the solvent diffuses
through the layer during charge.
■ The assumption of the ongoing formation of the SEI layer is also
supported by the research of Aurbach and colleagues (1999), who
propose that the intercalation of lithium into the graphite negative
electrode leads to increase in the lattice volume, which in turn
stretches the SEI layer, causing it to fracture and to expose more of
the active material to the electrolyte, fueling the side reaction, and
contributing to SEI formation.
■ Ramadass’ model assumes:

1. The main side reaction is due to the reduction of an organic

solvent, expressed as S C 2LiC C 2e # ! P, where “S” refers to the
solvent and “P” to the product formed in the side reaction.
2. The reaction occurs only during charging of the cell.
3. The products formed are a mixture of different species, resulting in
averaged mass and density constants used in the later equation
describing the formation and growth of the SEI film.
4. The side reaction is assumed to be irreversible and Usref is chosen
to be 0:4 V versus Li=LiC.
5. The initial resistance of the SEI layer developed during cell
formation is 0:01 !m2.
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–6

6. There is no overcharge reaction considered (i.e., lithium plating is

not modeled).
■ We have somewhat relaxed assumption 2, allowing side reactions to
occur during rest intervals also (and even during discharge).
■ The SEI growth model is tightly coupled with a Newman-style
physics-based model of ideal-cell dynamics.
■ For the negative electrode, the local molar flux jtotal is given by a sum
of the intercalation flux j and the side reaction flux js ,

jtotal D j C js ,

where j is computed via the Butler–Volmer electrochemical kinetic

expression
! " # " #$
i0 ˛a F ˛c F
j D exp " # exp # " ,
F RT RT
which is driven by the overpotential

" D #s # #e # Unref # FRfilmjtotal ,

where i0 [A m#2] is the exchange current density and Unref is the

equilibrium potential in the negative electrode, evaluated as a function
of the solid-phase concentration at the surface of the particle.
■ The kinetics of the side reaction are described using a Tafel equation,
which assume that the side reaction is considered irreversible,
" #
i0;s ˛s F
js D # exp # "s ,
F RT
and the side reaction overpotential is described as

"s D #s # #e # Usref # FRfilmjtotal.

c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–7
■ Once the side reaction flux, js , has been calculated, the change in
the film thickness ıfilm during charging can be calculated by
@ıfilm MP
D# js ,
@t $P
where MP [kg mol#1] is the average molecular weight of the
constituent compounds of the SEI layer and $P [kg m#3] is the
average density of the constituent compounds.
■ This allows the overall film resistance to be calculated as

Rfilm D RSEI C ıfilm=%P ,

where RSEI is the initial film resistance that is produced during the
formation period of the cell, and %P [S m#1] is the conductivity of the
film.
■ In addition to the resistance change, there is a capacity loss caused
by the side reaction current during charge, leading to capacity
changing via the relationship
Z Ln
@Q
D anAFjs dx.
@t 0

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–8
7.3: Simplifying the model

■ To effect an optimal control strategy, the battery management system

must be able to calculate the side reaction flux js very quickly and
accurately.
■ Solving the coupled PDE equations described above (plus the
physics-based ideal-cell model) is too complicated for such a process.
■ The js model needs to be much faster and simpler. In this section, we
present a simpler incremental model for calculating js , Rfilm, and Q.
■ To create a volume-averaged 0-D reduced-order model, three
additional assumptions are made:

1. The cell is always in a quasi-equilibrium state, allowing the

exchange current density i0 to be calculated from the cell SOC
alone, neglecting local variations in electrolyte and solid surface
concentration. The estimated value of js then corresponds to a
suddenly applied current pulse iapp .t/, which is constant over some
time interval &t.
2. The intercalation and the side-reaction fluxes are uniform over the
negative electrode. This allows us to state that the total reaction
flux jtotal is related to the applied cell current iapp by the following
relationship:
iapp
jtotal D ,
anF Voln
where the volume of the active material is described by
Voln D LnA.
3. The anodic and cathodic charge-transfer coefficients are equal
(˛a D ˛c D 0:5/.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–9
■ From the above assumptions, an incremental degradation model can
be formulated as follows. First, at any point in time, the lithiation state
of the negative electrode is calculated as

'n D 'n;min C SOCcell .'n;max # 'n;min/ ,

where:

! 'n;max and 'n;min are the stoichiometric limits of negative-electrode

lithiation (i.e., the value of ' in Li' C6 when the cell is fully charged
and discharged, respectively).
! SOCcell is a value between zero and one, which indicates the cell
state-of-charge.
■ Then Unref is calculated from 'n for the electrode materials being used.
■ We will ultimately iterate to find js . We can initialize its value to zero
and calculate the intercalation flux:
j D jtotal # js
iapp
D # js .
anF Voln
■ From j and assumption 3,
! " # " #$
i0 F F
j D exp " # exp # "
F 2RT 2RT
" #
2i0 F
D sinh "
F 2RT
" #
2RT Fj
"D asinh .
F 2i0
■ Then, we note the similarity between the expressions for " and "s to
find:
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–10

" D #s # #e # Unref # FRfilmjtotal

"s D #s # #e # Usref # FRfilmjtotal

D " C Unref # Usref

" #
2RT Fj
D asinh C Unref # Usref .
F 2i0
■ The film resistance cancels from the calculation. We can now
calculate an updated estimate of the side-reaction flux as
" #
i0;s #F
js D # exp "s .
F 2RT
■ In total, we have the reduced-order model
" # !!
i0;s #F 2RT Fj
js D # exp asinh C Unref # Usref
F 2RT F 2i0
0 % &1
iapp !!
i0;s #F 2RT F an F Voln
# js
B C ref ref
D # exp asinh @ A C Un # Us
F 2RT F 2i0

' (! #iapp !!
i0;s F Usref
# Unref an Voln
C Fjs
D# exp exp asinh .
F 2RT 2i0
■ We’ll see how to solve this equation for js shortly.
■ Once we have solved for js it can then be incorporated into
incremental equations for film resistance and capacity loss.
■ It is assumed that js is constant over some small time interval &t, and
is denoted as js;k for the kth interval.
■ We can convert the continuous-time film thickness relationship to
discrete time as:
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–11
@ıfilm MP
D# js
@t $P
MP &t
ıfilm;k D ıfilm;k#1 # js;k#1,
$P
noting that the sign of js is negative.
■ This result can be used to calculate the film resistance as

Rfilm D RSEI C ıfilm=%P

MP &t
Rfilm;k D Rfilm;k#1 # js;k#1.
$P % P
■ Similarly, we can discretize the capacity equation
Z Ln
@Q
D anAFjs dx
@t 0
Qk D Qk#1 C .anAFLn&t/ js;k#1.

■ In summary, the proposed reduced-order model (ROM) equations

are:

'n D 'n;min C SOCcell .'n;max # 'n;min/

' ref ref
(! #iapp !!
i0;s F Us # Un an Voln
C Fjs;k
js;k D # exp exp asinh ,
F 2RT 2i0
MP &t
Rfilm;k D Rfilm;k#1 # js;k#1
$P % P
Qk D Qk#1 C .anAFLn&t/ js;k#1.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–12
7.4: Simplifying the calculation
■ As it is written now, js;k is an implicit calculation
' ref ref
(! #iapp !!
i0;s F Us # Un an Voln
C Fjs;k
js;k D # exp exp asinh .
F 2RT 2i0
■ One solution methodology would be to use iteration:

1. Guess a value for js;k (e.g., zero),

2. Plug in to RHS of relationship; calculate LHS as new value for js;k ,
3. Repeat step 2 until no significant change in js;k .
■ This method actually works pretty well, and we can arrive at a
solution in fewer than 10 iterations.
■ However, there is also a closed-form solution for js;k (not obvious).
■ First, let’s simplify notation:
0 0 11
' ref ref
(!
i0;s F Us # Un B B #iapp F CC
js;k D # exp exp B
@asinh B
@ C j CC
s;k AA
F 2RT „ i0Vol…n „ƒ‚…
2anƒ‚ 2i0
„ ƒ‚ … B C
A

D A exp .asinh .B C Cjs;k // .

■ Note that A < 0 and C > 0 always.

■ Also note that the value of A can be stored in a lookup table versus
'n, so is not difficult to calculate in real time.
■ A useful identity for simplifying this further is:
p
exp.asinh.x// D x C x 2 C 1.

■ So, we can write

c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–13
! q $
js;k D A .B C Cjs;k / C .B C Cjs;k /2 C 1
q
js;k
.B C Cjs;k /2 C 1
# B # Cjs;k D
A
q
js;k .1 # CA/ # AB D A .B C Cjs;k /2 C 1

.js;k .1 # CA/ # AB/2 D A2 .B C Cjs;k /2 C A2

0 D A2 .B C Cjs;k /2 C A2 # .js;k .1 # CA/ # AB/2

' 2 (
D A B C 2BCjs;k C C js;k C A2
2 2 2

' 2 (
# js;k .1 # CA/2 # 2AB.1 # CA/js;k C A2B 2 .
■ Collecting like terms
' ( 2 ' (
0 D A2C 2 # .1 # CA/2 js;k C 2A2BC C 2AB.1 # CA/ js;k
' 2 2 (
C A B CA #A B .
2 2 2

■ Note that .1 # CA/2 D 1 # 2CA C A2C 2, so this simplifies,

' (
2
0 D .2CA # 1/ js;k C .2AB/ js;k C A2 .

■ Key point: This is a quadratic, so we can easily solve for the root(s)
using the quadratic formula:
p
#2AB ˙ 4A2B 2 # 4A2.2CA # 1/
js;k D
2.2CA # 1/
p
AB ˙ A B 2 C .1 # 2CA/
D .
.1 # 2CA/
■ But, which root to use? The Routh test gives us some guidance:

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–14
2
js;k 2CA # 1, A2
js;k 2AB
1 A2
■ Since 2CA # 1 < 0 and A2 > 0, we are guaranteed two sign changes,
which means that this equation always has exactly one positive real
root and one negative real root.
■ Physically, we know that js;k < 0 (because A < 0), so we want to take
the smaller root of the quadratic solution.
■ So, because A < 0,
p
AB C A B 2 C .1 # 2CA/
js;k D .
.1 # 2CA/

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–15
7.5: Comparing the models

■ The validity of this reduced-order model depends first on the accuracy

of the underlying partial differential equation model, which we
assume here to be exact.
■ It then depends on how closely the reduced-order approximation of js
matches the exact calculation of js .
■ To compare the PDE and reduced-order models, we conducted a
series of simulations.

! In each simulation, the cell was initially at rest.

! A sudden pulse of current was then applied, and the instantaneous

resulting js from the PDE model was compared to the computed js
from the ROM.
■ To simulate the PDE model, we used COMSOL Multiphysics 3.5a
coupled with a MATLAB script to cycle through the series of
simulations and analyze results.

! Specifically, each simulation comprised a 1 s time interval, where

the cell current iapp was modeled as a Heaviside step function,
which was applied half-way through the interval.
! We found that the initial rest interval facilitated convergence of the
solution by allowing the PDE solver to adjust its initial conditions
before applying the step current.
■ The simulation cell parameters that we used are listed in the
appendix. In particular, the cell had a 1:8 Ah capacity.
■ For the full-order PDE simulations:

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–16

! Applied current was varied from 0 A to 5:4 A in steps of 0:1 A;

! Initial cell SOC was varied from 0 % to 100 % in steps of 2 %; and

! Temperature was varied from #35 ı C to 45 ı C in steps of 20 ı C.

■ For the reduced-order simulations, which run much more quickly:

! Applied current was varied from 0 A to 5:4 A in steps of 0:05 A;

! Initial cell SOC was varied from 0 % to 100 % in steps of 1 %; and

ı ı ı
! Temperature was varied from #35 C to 45 C in steps of 10 C.

■ As one point of comparison, the set of 14 025 full-order PDE

simulations took more than eight days to complete on an Intel i7
processor, while the set of 112 200 ROM simulations took a total of
about 2.6 seconds to complete on the same machine.
■ The speedup, on a per-simulation basis, is more than 2 000 000 W 1.
This is the primary advantage of the ROM over the PDE model.
■ The figure below-left shows room-temperature side-reaction flux js as
computed by the reduced-order model (which we now denote as
js;ROM).
■ The figure below-right shows a compilation of js;ROM over a range of
temperatures. We see two trends that match experience: degradation
is worst at high SOC and high charge rates.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–17

■ The figure to the right shows

results of one PDE simulation. Comparing ROM to PDE solutions
−0.23
PDE
■ This example was conducted at −0.235
ROM

25 ıC, 50 % SOC, and by

Js (mA cm )

−0.24
−3

applying a 1C charge pulse at −0.245

t D 0:5 s. −0.25

−0.255
■ The figure shows the raw output −0.26
0 0.2 0.4 0.6 0.8 1
of the simulation, as compared Time (s)

to the ROM.
■ Both the PDE and ROM solutions have a non-zero negative side

reaction flux js even when the cell is at rest.

■ This is due to the fact that we have relaxed assumption 2 of the SEI
growth model to also allow for the side reaction when current in the
external circuit is zero.
■ The figure shows that the ROM matches both the rest SEI
side-reaction rate and the charge-pulse SEI side-reaction rate.
■ Plotted on the same scale, the full PDE solution results are
indistinguishable from the ROM results. So, for comparison purposes,
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–18

we define a relative error between the results as

js;PDE # js;ROM
js;err% D $ 100,
js;PDE
where js;PDE is chosen to be the value of js from the PDE solution
immediately after the application of the current pulse.
■ The figure plots the relative
Relative error
error between the PDE and
ROM solutions for all 25 ıC
simulations. 1.5
1

Js,err%
■ Between 10 % and 90 % SOC 0.5
(e.g., typical extremum 0 3C
operating conditions for EV 100 2C
75 ent
50 1C c urr
cells), the maximum relative SOC (%25 0 harg
e
)
error was 0:44 %. C

■ To further illustrate the performance of the ROM and to see the

dependence of SEI layer growth rate on SOC and charge rate, the
figures below plot these results in a different format (25 ıC).

Degradation as a function of SOC Degradation as a function of rate

0
PDE 0
0% SOC
ROM
−0.2 0C ra
te −0.2
J (mA cm )
Js (mA cm−3)

3C ra
−3

−0.4 te −0.4

−0.6 −0.6
s

−0.8
−0.8
PDE 100%
−1 ROM SOC
−1
0 20 40 60 80 100 0C 0.5C 1C 1.5C 2C 2.5C 3C
State of Charge (%) Charge rate

■ The left frame shows js as a function of SOC at different charge rates

(lines plotted from 0C to 3C in steps of 0.5C).
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–19
■ The right frame shows js as a function of charge rate at different
SOCs (plotted from 0 % SOC to 100 % SOC in steps of 10 % SOC).
■ In all plots, the corrected PDE result is drawn as a solid line, and the
ROM result is drawn as a dashed line. In most cases, it is impossible
to visually distinguish between the PDE and ROM results.
■ The figures below show additional effects on relative error. First, we
see how error varies with temperature.
■ The ROM predictions are best
Effect of temperature on results
at high temperatures, and less 45°C

good at low temperatures. 1.5

25°C
1
Js,err%

! Worst-case js;err% in the 10 % 0.5

5°C

to 90 % SOC range varies 0 −15°C

100
from 0:41 % at 45 ıC to 75
50 2C
3C
SOC 25 1C e cu rrent −35°C
0:55 % at #35 ıC. (%) 0 0C Cha g
r

■ Next, we investigate the effect of &t on the results. Instead of

selecting the value for js;PDE immediately after the application of the
current pulse, js;PDE is now selected to be the PDE solution 0:5 s
seconds after the application of the current pulse, at the t D 1 s point.

Effect of 0.5s ∆t on results

■ The relative error is once again 45°C

10
worst at low temperatures and 25°C

low values of SOC (where the

s,err%

5 5°C
J

absolute amount of degradation

0 −15°C
is small). 100
75 3C
50 2C
SOC 25 1C rent
(%) 0 0C h a r g e cur −35°C
C

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–20
■ Relative errors over 10 % are observed in some cases, but in the
ranges of SOC most important for control, where SOC is greater than
25 %, the worst-case js;err% is far less, varying from 0:85 % at 45 ıC to
1:04 % at #35 ıC.
PDE and ROM solutions over extended period
The figure to the right
0
■ 100 s
1000 s
investigates the effect of a −0.2

J (mA cm−3)
2000 s
prolonged constant-current −0.4

charge at a 1C rate, as might be

s
−0.6 3000 s
1D PDE computation of J
experienced when a cell is −0.8
s
Avg. over width of PDE computation of Js

being charged.
ROM computation of average Js
−1
0 0.2 0.4 0.6 0.8 1
Dimensionless distance across anode

■ The PDE is simulated for 3 000 s, starting with the cell at rest at 10 %
SOC, and 1D profiles of js .x/ across the negative electrode are
plotted at time steps 100 s; 1 000 s, 2 000 s, and 3 000 s.
■ Overlaid on the plot are the average js values predicted by the ROM
at that SOC level, and the actual averaged js values (averaged over
the 1D electrode) from the PDE solution.
■ In the ROM simulation, the SOC is updated on a second-by-second
basis to achieve the present SOC at every point, which is used to
compute the value of js using the method explained herein.
■ We see that even over prolonged constant-current charge profiles,
the ROM is accurate, indicating that assumption 1 of ROM is a
reasonable assumption to make.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–21
7.6: Lithium deposition on overcharge
■ The success of the preceding ROM to predict the performance of a
Tafel equation leads us to expect it will work well for other similar
aging mechanisms (e.g., Darling and Newman, 1998).
■ The PDE model of lithium deposition on overcharge by Arora et al.,
however, does not use a Tafel equation.
! Instead, it uses a modified Butler–Volmer equation.

■ Here, we address creation of a ROM of lithium deposition on

overcharge.2
! It does not work as well as the SEI ROM, especially for prolonged
charging events.
! It is probably better suited for predicting degradation in an HEV
scenario, with random charges.
! We’re working on another method right now, that we think will work
better, but it isn’t ready for public consumption as yet.
■ Lithium deposition/plating is not usually considered a dominant
degradation mechanism, because the cell terminal voltage limits are
designed to avoid conditions that would be conducive to plating.
! However (especially at cold temperatures), the terminal voltages
are poor indicators of internal cell potentials,
! Plating can still happen, and when it does, there is immediate
severe capacity loss.
2
Adapted from, Perkins, R.D., Randall, A.V., Zhou, X., Plett, G.L., “Controls Oriented
Reduced Order Modeling of Lithium Deposition on Overcharge,” Journal of Power
Sources, Vol. 209, July 2012, pp. 318–325.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–22
■ So, modeling lithium deposition is very important to be able to devise
optimized controls to minimize aging.
■ Overcharge manifests first as a metallic lithium deposit on the surface
of the negative electrode solid particles during charge, predominantly
near the separator.
■ Subsequently, the lithium can and does further combine with
electrolyte material to form other compounds such as Li2O, LiF,
Li2CO3, and polyolefins.
■ The nature of the final product is not our major concern; rather, the
issue is that lithium is irreversibly lost.
■ This phenomenon is an unintended side reaction that can lead to
severe capacity fade, electrolyte degradation, possible safety hazard.

Physics-based model of overcharge

■ This work is based on a physics-based model proposed by Arora et

al. (1999).
■ Our goal is to create a high-fidelity reduced-order model of this PDE
degradation model; therefore, we adopt the same assumptions as
they, which were:

1. The main side reaction is expressed as LiC C e# ! Li.s/, which

occurs at Usref D 0 V versus Li/LiC during an overcharge event. This
lithium metal is expected to form first near the electrode-separator
boundary where the surface overpotential is greatest.
2. Lithium metal deposited on the negative electrode reacts quickly
with solvent or salt molecules in the vicinity, yielding Li2CO3, LiF, or

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–23

other insoluble products. A thin film of these products protects the

solid lithium from reacting with the electrolyte. Solid lithium can still
dissolve during discharge, but once lithium is consumed in a
insoluble product, it is permanently lost.
3. The insoluble products formed are a mixture of different species,
resulting in averaged mass and density constants used in a later
equation describing the formation and growth of a resistive film.
4. Only the overcharge reaction is considered (e.g., SEI film growth
and other degradation mechanisms are not modeled).
■ The overcharge model of Arora is tightly coupled with a “pseudo
two-dimensional” Newman-style porous-electrode style model of
ideal-cell dynamics.
■ In the ideal cell, intercalation flux j.x; t/ is expressed as the
Butler–Volmer equation,
! " # " #$
i0 ˛a;nF ˛c;nF
j.x; t/ D exp ".x; t/ # exp # ".x; t/
F RT RT
which is driven by the overpotential

".x; t/ D #s .x; t/ # #e .x; t/ # Unref # FRfilmj.x; t/,

where i0 is the exchange current density,

' max (˛a;n
i0 D kn cs;n # cs;n .cs;n/˛c;n .ce /˛a;n ,

and Unref is the equilibrium potential which is evaluated as a function

of the solid phase concentration at the surface of the particle.
■ Arora expresses the side-reaction flux js (i.e., the rate of irreversible
lithium loss due to lithium plating) as

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–24
" ! !#!
io;s ˛a;s F ˛c;s F
js .x; t/ D min 0; exp "s .x; t/ # exp # "s .x; t/ ,
F RT RT
where ˛a;s ¤ ˛c;s in general,

"s .x; t/ D #s .x; t/ # #e .x; t/ # Usref # FRfilm js .x; t/,

and where the side-reaction exchange current density

i0;s D kn;s .ce /˛a;s .
■ Side reaction is semi-irreversible in the sense that it includes an
anodic rate term, but doesn’t allow overall positive side-reaction flux.
■ The side reaction occurs only at spatial locations in the negative
electrode where "s .x; t/ < 0.
■ This is enforced in the js equation by the “min” function, which sets
js .x; t/ D 0 for values of x where "s .x; t/ % 0, but to the value
computed by the Butler–Volmer equation when "s .x; t/ < 0.
■ A typical scenario is plotted in the
figure, where "s .x; t/ is sketched
across the electrode width.
■ In this example, plating will occur in the
interval from x D x0 to x D Ln.
■ Note that this illustration shows that the cell can be quite far away
from 100 % state-of-charge and still have plating occur near the
separator if a large enough charge-current pulse is applied to the
cell’s terminals.

! The state-of-charge is only one variable of importance—ultimately,

the local overpotential determines whether plating occurs.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–25
■ So, our first goal will be to solve for js . Then, once we have solved for
js , it can then be incorporated into incremental equations for film
resistance and capacity loss.
■ We assume that js is constant over some small time interval &t, and
denote it as js;k for the kth interval.
■ We can convert the continuous-time film thickness relationship to
discrete time as:
@ıfilm MP
D# js
@t $P
MP &t
ıfilm;k D ıfilm;k#1 # js;k#1,
$P
where MP and $P are the average molecular weight and density of
lithium and products, noting that the sign of js is negative.
■ This result can be used to calculate the film resistance as

Rfilm D RSEI C ıfilm=%P

MP &t
Rfilm;k D Rfilm;k#1 # js;k#1.
$P % P
■ Similarly, we can discretize the capacity equation
Z Ln
@Q
D anAFjs dx
@t 0
Qk D Qk#1 C .anAFLn&t/ js;k#1.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–26
7.7: Simulation and results

■ A reduced-order model is derived in the aforementioned paper.

■ The validity of this reduced-order model depends first on the accuracy
of the underlying partial differential equation model, which we
assume here to be exact.
■ It then depends on how closely the reduced-order approximation of jNs
matches the exact calculation of jNs . In this section, results from both
the full and reduced-order models for jNs are compared.
■ To compare the PDE and reduced-order models, we conducted a
series of simulations.

! In each simulation, the cell was initially at rest.

! A sudden pulse of current was then applied, and the resulting jNs
from the PDE model, averaged over a one-second interval
subsequent to the pulse, was compared to the computed jNs from
the ROM.
■ To simulate the PDE model, we used COMSOL Multiphysics 3.5a
coupled with a MATLAB script to cycle through the series of
simulations and analyze results.
■ Specifically, each simulation comprised a 1:2 s time interval, where
the cell current iapp was modeled as a step function, which was
applied at t D 0:2 s.
■ We found that the initial rest interval facilitated convergence of the
solution by allowing the PDE solver to adjust its initial conditions
before applying the step current.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–27
■ The cell parameters that we used in the simulations match those
used in Arora and are listed in the appendix.

! The applied current was varied between 0C and 3C in increments

of C/33;
! The initial cell SOC was varied between 0 % and 100 % in steps of
1 %, and
! Temperature was held constant at 25 ı C.

■ We found that the adjustable tuning factor ˇ D 1:7 worked well (this
implies the change in electrolyte concentration near the separator
changes nearly twice as quickly as it does near the current collector).
■ A total of 10 100 simulations were run.

! As one point of comparison, the set of full-order PDE simulations

took approximately 12 hours to complete, utilizing an average of
three cores, on an Intel i7 processor, while
! The set ROM simulations took approximately 21 seconds to
complete, utilizing an average of one core on the same machine.
! The speedup, on a per-simulation per-core basis, is more than
5 000 W 1. This is the primary advantage of the ROM over the PDE
model.
■ Lithium plating occurs when the side-reaction overpotential is
negative ."s < 0/.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–28
■ The figure illustrates the
side-reaction overpotential Overpotential predictions at 90% SOC, 2.0C rate
20
PDE
across the negative electrode 10
ROM

Overpotential η (mV)
for this cell model, where x D 0

s
0

is adjacent to the
−10
current-collector and x D 85 (m
−20
is adjacent to the separator,
immediately following the onset
−30
0 10 20 30 40 50 60 70 80
Anode location (µm)
of a charge current pulse.
■ From the PDE result, we expect lithium deposit to occur between

about x D 42 (m and x D 85 (m. From the ROM, we expect lithium

deposit to occur between x0 D 49 (m and x D 85 (m.
■ The figure to the right shows the

resulting rate of lithium Overcharge predictions at 90% SOC, 2.0C rate

Overcharge rate js (mA cm−3)

0
PDE
deposition for the PDE and ROM
−50
ROM solutions.
−100

■ The time-average deposition −150

rate of the ROM is somewhat

−200

higher than the time-average

−250
deposition rate of the PDE over −0.2 0 0.2 0.4
Time (s)
0.6 0.8 1

the 1 s interval.
■ The figures below illustrate the predicted overcharge rates over all

scenarios for the PDE and the ROM solutions.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–29

■ As expected, deposition is worse at high SOC and high charge rates.

The PDE and ROM solutions generally agree very well, with greatest
mismatch at high charge rates.
■ The figures below show a different view of the results.

! Cross sections through both the PDE and ROM solution spaces
are plotted and compared.

Instantaneous degradation rate as function of SOC Instantaneous degradation rate as function of current
0 0
)

Overcharge js (mA cm )
−3

−3

−100 −100
Overcharge j (mA cm

−200 −200 SOC = 60%

−300 Charge at 1C rate −300 SOC = 70%
s

−400 −400 SOC = 80%

Charge at 2C rate
−500 −500 SOC = 90%
−600 PDE Charge at 3C rate −600 PDE SOC = 100%
ROM ROM
−700 −700
0 20 40 60 80 100 0C 1C 2C 3C
SOC (%) Charge current

■ The left figure shows how the two methods compare where each pair
of lines represents a specific charge rate.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–30

! As noted before, but perhaps more clearly seen here, the

difference between the PDE and ROM solutions are greatest at
high charge rates.
■ The right figure shows how the two methods compare where each
pair of lines represents a specific initial SOC.

! The difference is greatest at moderate SOC levels.

■ Finally, the figures below illustrate the error between the PDE and
ROM solutions in two ways.

Regions where models predict overcharge

100
Both PDE and ROM
predict overcharge in
75 this region
SOC (%)

50

25 Neither PDE nor ROM

predict overcharge in
this region
0
0C 1C 2C 3C
Charge current

■ The left frame shows the regions where the two methods agree on
whether lithium deposition will occur, and the region where they
disagree.

! The region of disagreement is the very narrow sliver at around

2.4C and 25% SOC, where the ROM predicts overcharge but the
PDE does not.
! Otherwise, the boundaries are identical.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–31
■ The right frame shows the error between the solutions, calculated as
jNs;ROM # jNs;PDE . The maximum error is approximately 65 mA cm#2
(relative error on the order of 10 %).
■ For the purpose of control system design, the results of the left frame
are the most important.
■ Since lithium deposition is such a severe degradation mechanism, a
charging control scheme should avoid ever commanding a control
action that would cause any lithium deposition to occur.
■ A time-optimal charger, based only on the PDE model of lithium
deposition, would select charge pulse current to follow the upper
contour in the left frame.

! This allows the maximum charge rate at any point in time, while
causing no lithium plating.
■ In comparison, a time-optimal charger, based only on the ROM model
of lithium deposition, would select charge pulse current to follow the
lower contour in the figure.

! This will result in somewhat slower charging.

! But, because the ROM over-predicts the amount of lithium

deposition, it will also result in a charging scheme that is
conservative, which is a beneficial feature.
■ We conducted additional simulations to investigate the effect of pulse
length &t in assumption 1.
■ That is, how long can the charge pulse be before the full-order PDE
model and the reduced order model results are significantly different?

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–32
■ We found that pulse lengths less than 10 s are generally well
matched, but pulse lengths much greater than 10 s can give
significant PDE versus ROM mismatch.
■ For long pulse durations, the quasi-static nature of assumption 1 is
violated, and an a significant offset is noted in actual time-varying #e
versus the at-rest #e , moving the crossover point of "s .x; t/.
■ This causes the ROM to under-predict the value of lithium plating
computed by the PDE.
■ For this reason, we propose that the ROM is of most value for
computing current limits in dynamic applications such as
hybrid-electric vehicles, where a bias in #e cannot develop due to the
random nature of power demand, but is of less value for controlling
full charges, such as for electric vehicle applications.
■ We make one final comment regarding efficiency.

! The speedup of ROM vs. PDE can be much greater than 5 000 W 1 if
ROM solutions are pre-computed and stored in a table.
! Then, “computing” any value of jNs;ROM would be nearly
instantaneous, via table lookup.
■ We note that the ROM solution changes as the film resistance
changes, but the film resistance changes very slowly.
■ The entire table might be updated by the battery management system
once per operational period (e.g., once per day), and then utilized
throughout that operational period for significant performance gains.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–33
7.8: Optimized controls for power estimation
■ We’ve now seen that there are quite a few causes of cell degradation,
and an attempt at modeling two of the more significant mechanisms.
■ Much more work remains to be done in this area, first by materials
scientists, then by controls engineers.
■ But, how to use these models to compute power to slow aging? We
look at a few methods next.
■ We have seen that none of the cell degradation mechanisms are tied
directly to the cell terminal voltage, but rather to internal stress factors.
■ Therefore, assuming that degradation mechanisms can be well
modeled, it makes more sense to compute power limits based on
predicted capacity loss and/or impedance rise than on voltage limits.
■ Clearly, there’s a lot of work to do before this is practical, but the
potential benefits are worth it.
■ The next sections of notes very briefly introduce some optimization
methods that might be used with the physics-based degradation
mechanisms to compute better power limits.

Two problems
■ There are (at least) two controls problems to consider.3
■ For EV/E-REV/PHEV, the battery pack is charged from an external
source:

3
A third, well beyond our scope here: Considering xEV as storage units for the “smart
grid,” when does it make sense to “lend” energy to the grid? What should be the rental
fee charged for allowing energy to be borrowed?

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–34

! What is the optimum charge profile?

! Can we “fast charge”?

! For a fixed charge period, what is the best strategy?

■ For all xEV, while the car is being driven,

! What is the maximum charge power that can be maintained over

the next &T seconds?
! What is the maximum discharge power that can be maintained
over the next &T seconds?
■ Different kinds of optimized controls may be better for these two
problems.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–35
7.9: Plug-in charging

■ The plug-in charging problem lends itself well to being solved by a

nonlinear programming method.

! One example is the sequential quadratic programming algorithm,

implemented in MATLAB as fmincon.m.
■ Nonlinear programming is a generic optimization method that
attempts to find solutions to problems that can be posed in the
framework:
8̂
< c.x/ ' 0 Ax ' b
x & D arg min f .x/; such that ceq .x/ D 0 Aeqx D beq
:̂
lb ' x x ' ub,
where f .x/ is a function that we wish to minimize by choosing
optimum input vector x & such that

! Nonlinear inequality constraint vector function c.x/ ' 0 is satisfied,

! Nonlinear equality constraint vector function ceq .x/ D 0 is satisfied,

! Linear inequality constraint vector function Ax ' b is satisfied,

! Linear equality constraint vector function Aeq x D beq is satisfied,

! Bounds lb ' x ' ub for all entries in vector x are satisfied

for user-specified f .x/, c.x/, ceq .x/, A, b, Aeq , beq , lb, and ub.
■ We will choose x to be a vector of cell applied current versus time,
f .x/ to be some estimate of the cell degradation that would be
caused by that applied current, and the other functions and matrices
to make the problem work.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–36
■ For example, we might want to find
K#1
X
i & D arg min #Js .ik ; ´k ; Tk /
kD0
8̂
< ´min ' ´k ' ´max
such that ´K D ´end
:̂
#Imax ' ik ' Imax
X
and ´k D ´0 # ij &t=Q.
j <k

■ This states that we want to minimize capacity loss that would be

experienced by a cell if we were to start at SOC ´0 and end at SOC
´end over a period of K sampling intervals, where current is limited
between ˙Imax, and SOC is limited between ´min and ´max and the
standard SOC equation holds.
■ It takes a little work to recast this problem in the right framework, but
it’s not too bad.
■ First, consider the SOC equation. We can write it in vector form as:
2 3 2 3 2 32 3
´1 1 1 0 0 0 ((( 0 i0
6 7 6 7 6 76 7
6 ´2 7 6 1 7 &t 6 1 1 0 0 ( ( ( 0 7 6 i1 7
6 7 6 7 6 76 7
6 ::: 7 D 6 ::: 7 ´0 # Q 6 ::: ::: ::: ::: : : : ::: 7 6 ::: 7 .
4 5 4 5 4 54 5
´K 1 1 1 1 1 ((( 1 iK#1
„ƒ‚… „ ƒ‚ … „ ƒ‚ …
CV LT x

! Notice that the matrix LT is lower-triangular.

■ Using this formulation, we can write an equation for the ´K constraint

&t h i
´K D ´0 # 1 1 1 ( ( ( 1 x D ´end ,
Q
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–37

or, in the prescribed format for fmincon.m,

h i Q
1 1 1 ((( 1 x D .´0 # ´end / .
„ ƒ‚ … &t
„ ƒ‚ …
Aeq beq

■ The limit ´min ' ´k can be written as

2 3 2 3 2 32 3
1 1 1 0 0 0 ((( 0 i0
6 7 6 7 6 76 7
617 6 7 6 ((( 07 6 7
6 7 ´min ' 6 1 7 ´0 # &t 6 1 1 0 0 7 6 i1 7
6 ::: 7 6 ::: 7 Q 6 : ::: ::: ::: ::: ::: 7 6 ::: 7
4 5 4 5 4 :: 54 5
1 1 1 1 1 1 ((( 1 iK#1
„ƒ‚… „ ƒ‚ … „ ƒ‚ …
CV LT x

&t
.CV/.´min # ´0 / ' # .LT/ x
Q
Q
.LT/ x ' .CV/.´0 # ´min/.
&t
■ Similarly, ´ ' ´
k max can be written as
Q
#.LT/x ' .CV/.´max # ´0 /.
&t
■ Putting the last two constraints together gives
" # " #
LT Q .CV/.´0 # ´min/
x' .
#LT &t .CV/.´max # ´0 /
„ ƒ‚ … „ ƒ‚ …
A b
■ The constraints on input current can be satisfied by setting
lb D #Imax.CV/; and ub D Imax.CV/.
■ Then, all that’s left is to specify the cost function f .x/. (There are no
nonlinear constraints in this problem.)
■ Given what we’ve seen so far, we might consider js to represent the
SEI growth model, or the overcharge model, or the sum of both.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–38
7.10: Fast-charge example
■ We designed optimized controllers to investigate charging strategies
using the degradation models generated to date.
■ In the first controls scenarios, the cell was initially at a SOC value
between 10 % and 90 %, and the charger was required to optimally
charge the cell to 90 % over a period of two hours.
■ SOC was not allowed outside the range of 10 % to 90 %, but current
was unconstrained.
■ We first looked at using only the SEI-growth degradation model in the
control strategy:

Strategy for SEI cost function Strategy for SEI cost function
100
z0 = 90% z0 = 90%
6C
z0 = 70% z0 = 70%
80
z0 = 50% z0 = 50%
4C
z = 30% z = 30%
SOC (%)

0 60 0
Rate

2C z0 = 10% z0 = 10%
40
0
20
−2C
0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (min) Time (min)

■ The optimum charging strategy for this model is to quickly discharge

the cell to the minimum SOC, wait “as long as possible” and then
quickly charge the cell to the desired SOC.
■ The cost of discharging plus charging turns out to be less than the
cost of maintaining a high SOC for an extended period of time(!)
■ We then looked at using the SEI plus the overcharge cost functions
added together:
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–39
Strategy for SEI/ovchg cost function Strategy for SEI/ovchg cost function
100
z0 = 90% z0 = 90%
6C
z = 70% z = 70%
0 80 0
z = 50% z = 50%
4C 0 0
z0 = 30% z0 = 30%

SOC (%)
60
Rate

2C z = 10% z = 10%
0 0
40
0
20
−2C
0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (min) Time (min)

■ These are qualitatively similar, but different in some details.

! In particular, the final charge event is at a much lower rate, and

! Charge current tapers off at high SOC to avoid lithium plating.

■ For grins, we overlay optimum
Optimum charging trajectory
SEI plus overcharge charging −3
x 10

profile on top of the degradation 0

∆Q [% of new−cell]

function. −2

Optimization method
−4
■

automatically avoids the “cliff” −6

0
25
0C
50 1C
where degradation starts to get SOC 75 100 3C
[%] r
2C
g e cur
rent
Cha
much worse.
■ Third, we looked at fast charging strategies, with results plotted below.

■ The cell was at an initial SOC of 50 %, then was allowed 15, 30, 45,
60, 75, 90, 105, or 120 minutes to charge to 90 %.
■ Strategies using the SEI cost function and the combined SEI plus
overcharge cost function are shown.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–40
Strategy for SEI cost function Strategy for SEI/ovchg cost function
100 100

80 80
SOC (%)

SOC (%)
60 60

40 40

20 20

0 0
0 15 30 45 60 75 90 105 120 0 15 30 45 60 75 90 105 120
Time (min) Time (min)

■ Again, they are similar, but not identical.

! If sufficient time is granted, the charger will discharge the cell to

the minimum allowed SOC, and then charge the cell.
! If less time is granted, the charger will only partially discharge the
cell before charging.
! If even less time is granted, charger charges cell immediately.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–41
7.11: Dynamic power calculation
■ As discussed in Ch. 6, the goals of dynamic power calculation are:
a) Discharge power: Based on present battery-pack conditions,
estimate the maximum discharge power that may be maintained
constant for &T seconds without violating pre-set design limits on
cell voltage, SOC, maximum design power, or current.
b) Charge power: Based on present battery-pack conditions, estimate
the maximum battery charge power that may be maintained
constant for &T seconds without violating pre-set design limits on
cell voltage, SOC, maximum design power or current.
■ As before, we handle this problem by looking for the maximum
dis/charge current the cell can withstand, and then convert that value
to power by multiplying by voltage.
■ Unlike before, we now consider degradation mechanisms to be the
limiting factor, rather than cell terminal voltage limits.
■ The method proposed here is not yet thoroughly tested, but with
some work should give good results.
■ It is closely related to a control-system design paradigm called Model
Predictive Control (MPC). The idea is to:
! Determine an N -length sequence of control signals, using a model
of the system to be controlled to predict future system
performance, that will cause the system’s controlled variables to
converge toward desired values;
! Implement only the first of these N signals;
! Repeat.
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–42
■ This allows us, for example, to predict a constant-current input that
would not violate limits and would optimize a cost function if applied
for the full &T seconds (N sample periods), but only implement the
first of these, then repeat.
■ Standard MPC is a little different from what we will look at here:

! MPC reformulates the system model to use &uk as the input

signal rather than uk ;
! This formulation implicitly adds an integrator to the dynamics,
which is good for control, but is unnecessary for power estimation.
! Also, seems appropriate for set-point control: when &uk D 0, then
u is a constant and y approaches a steady-state constant.
◆ Again, not necessary for power estimation.
! Also, standard MPC does not allow the state-space model to have
a direct feedthrough “D” term, which we need here.
■ We’ll use a similar idea to MPC, leading up to the same form of
quadratic optimization used by MPC.
■ The system model we assume is:

xkC1 D Axk C Buk

yk D C xk C Duk ,

where yk are the performance variables that we would like to control

to some limit, or to maintain within some hard constraints.

! That is, yk may be different from the normal system outputs that
we have called yk in the past.

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–43
■ Define the vectors:
h iT h iT
Y D yk ykC1 ( ( ( ykCN and U D uk ukC1 ( ( ( ukCN .

■ Then, we can write,

2 3 2 3 2 32 3
yk C D 0 uk
6 7 6 7 6 76 7
6 ykC1 7 6 CA 7 6 CB D 7 6 ukC1 7
6 7 6 7 6 76 7
6 ykC2 7 D 6 CA2 7 xk C 6 CAB CB 7 6 ukC2 7
6 7 6 7 6 76 7
6 :: 7 6 :: 7 6 ::: ::: ::: 7 6 :: 7
4 : 5 4 : 5 4 54 : 5
ykCN CAN CAN #1B CAN #2B ( ( ( D ukCN
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
Y F ˆ U
Y D F xk C ˆU .

■ Now, we define a cost function that we wish to optimize:

J D .Rs # Y /T Q.Rs # Y / C U T RU

D .Rs # ŒF xk C ˆU )/T Q.Rs # ŒF xk C ˆU )/ C U T RU

D RsT QRs # RsT QF xk # RsT QˆU

# xkT F T QRs C xkT F T QF xk C xkT F T QˆU

# U T ˆT QRs C U T ˆT QF xk C U T ˆT QˆU C U T RU .

■ To simplify this, note that each term is a scalar, and hence equal to its
own transpose:
J D ŒRsT QRs # 2RsT QF xk C xkT F T QF xk ) (not a function of U )

C 2ŒxkT F T Qˆ # RsT Qˆ)U

C U T ŒˆT Qˆ C R)U .

■ Let,
c 2013, 2015, Gregory L. Plett
Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–44

H D 2ŒˆT Qˆ C R)

f T D 2.xkT F T Qˆ # RsT Qˆ/.

■ Then,
1
J D U T H U C f T U C constant.
2
■ Further, we can put constraints on Y via

Ymin ' F xk C ˆU ' Ymax,

which can be written as
ˆU ' ŒYmax # F xk )

#ˆU ' ŒF xk # Ymin),

which can both be combined in the matrix inequality
" # " #
ˆ Ymax # F xk
U ' ,
#ˆ F xk # Ymin
„ ƒ‚ … „ ƒ‚ …
Aineq bineq

or, Aineq U ' bineq .

■ So, we now have defined vectors/matrices H , f T , Aineq , and bineq that
match a quadratic programming problem, which is:
1
U & D arg min U T H U C f T U
2
such that
Aineq U ' bineq .
■ In MATLAB, the solution is found via quadprog.m.
■ Note, we can use h iT
U D 1 1 1 ((( 1 u

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–45

to make a fast single-variable optimization problem, to give us the

maximum dis/charge current value that would apply to all times.
■ But, what values to use?
! Reference Rs value for model SOC state on charging set to 1.0;
! Reference Rs value for model SOC state on discharging set to 0.0;
! Need to define other reference variables for soft constraints to
minimize degradation mechanisms (how?);
! Hard constraints must be set to prohibit lithium plating in the
negative electrode.

Where from here?

■ We’ve reached the edge of what we presently know how to do for
battery management.
■ There’s plenty of work yet to do:
! How do we efficiently implement the optimized power controls, and
how do we tune to accommodate various aging mechanisms and
cost tradeoffs?
! How do we perform system identification of physics-based model
parameters to give a good enough model to match a real cell well?
! How do we model new degradation mechanisms in efficient ways,
for implementation in embedded systems?
■ And many more we haven’t even thought of yet.
■ I hope some of this material has sparked your imagination, and I
hope you will be able to contribute to making battery management
systems of the future even better!

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–46
Appendix: Parameters used for SEI simulations

Symbol Units Neg. electrode Separator Pos. electrode

L (m 88 20 80
R (m 2 - 2
A m2 0.0596 0.0596 0.0596
* S m#1 100 - 100
"s - 0.49 - 0.59
"e - 0.485 1 0.385
brug - 4 - 4
csmax mol m#3 30 555 - 51 555
ce0 mol m#3 1 000 1 000 1 000
'i;min - 0.03 - 0.95
'i;max - 0.886 - 0.487
Ds m2 s#1 3:9 $ 10-14 - 1:0 $ 10-14
De m2 s#1 7:5 $ 10-10 7:5 $ 10#10 7:5 $ 10-10
tC0 - 0.363 0.363 0.363
k A m5=2 mol#3=2 4:854 $ 10-6 - 2:252 $ 10-6
˛a - 0.5 - 0.5
˛c - 0.5 - 0.5
Usref V 0.4 - -
i0;s A m#2 1:5 $ 10-6 - -

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "
ECE5720, Physics-Based Optimal Controls 7–47
Appendix: Parameters used for overcharge simulations

Symbol Units Neg. electrode Separator Pos. electrode

L (m 85 76.2 179.3
R (m 12.5 - 8.5
A m2 1 1 1
* S m#1 100 - 3.8
"s - 0.59 - 0.534
"e - 0.36 1 0.416
%e S m#1 0.2875 0.2875 0.2875
brug - 1.5 - 1.5
csmax mol m#3 30 540 - 22 860
ce0 mol m#3 1 000 1 000 1 000
'i;min - 0.10 - 0.95
'i;max - 0.90 - 0.175
Ds m2 s#1 2:0 $ 10-14 - 1:0 $ 10-13
De m2 s#1 7:5 $ 10-11 7:5 $ 10#11 7:5 $ 10-11
tC0 - 0.363 0.363 0.363
k A m5=2 mol#3=2 2 $ 10#6 - 2 $ 10#6
˛a - 0.5 - 0.5
˛c - 0.5 - 0.5
˛a;s - 0.3 - -
˛c;s - 0.7 - -
Usref V 0.0 - -
RSEI - m2 0.002 - -
i0;s A m#2 10 - -

c 2013, 2015, Gregory L. Plett

Lecture notes prepared by Dr. Gregory L. Plett. Copyright "